Special Cluster in Conjunction with IEICE General Conference 2020

IEICE Communications Express, Vol.9, No.12, 559–560

FOREWORD to Vol. 9 Special Cluster in Conjunction with IEICE General Conference 2020

The IEICE Communications Express (ComEX) is an online open-access letter journal with a binary peer-review publishing system established in June 2012 by the IEICE covering the entire field of communications. The average period from the submission to the decision was less than one month, and the acceptance rate was 48% in 2019. By the way, the IEICE holds a General Conference in March every year, in which the fruitful and valuable research seeds are presented. Because most of the manuscripts are written in Japanese, however, there are few opportunities for world researchers and developers to know the existence of these excellent works. The editorial committee of ComEX considered that it is our mission to prepare a place for opening these technologies to the world and decided to edit the special cluster related to the IEICE General Conference. In 2019, we planned the first special cluster related to the IEICE General Conference and could publish 34 letters. Fortunately, good reputations were received from the authors and readers. Therefore, we decided to edit the second special cluster this time. Unfortunatelly, all the presentations of the IEICE General Conference 2020 were canceled due to the COVID-19 outbreak. Additionally, most of the university campuses were closed around the submission deadline. Therefore, we worried about the establishment of this special cluster. However, many letters, 40 letters, were submitted for this cluster. After careful reviews, 17 letters were accepted in total. Because most of the rejected papers had sufficient potential to be published on ComEX, we recommended authors re-submitting the revised version to regular issues. I expect that this special cluster encourages researchers and will promote further research activities in the communication research fields. Finally, I would like to express sincere appreciation to all the authors for their excellent contributions and reviewers and editorial committee members for their great effort to make a success in this special cluster.

Guest Editor-in-Chief: Hiroo Sekiya (Chiba Univ.)

Special Cluster Editorial Committee Members Guest Editors: Kazunori Hayashi (Kyoto Univ.) and Noriaki Kamiyama (Fukuoka Univ.) © IEICE 2020 Guest Associate Editors: Yuyuan Chang (Tokyo Inst. of Tech.), Chiao-En Chen DOI: 10.1587/comex.2020COF0001 Published December 1, 2020

559 IEICE Communications Express, Vol.9, No.12, 559–560

(National Chung Cheng University), Young-June Choi (Ajou University), Chun- I Fan (National Sun Yat-sen University), Mariusz Glabowski (Poznan Univ. of Tech.), Bo Gu (Sun Yat-sen University), Guan Gui (Nanjing University of Posts and Telecommunications), Hiroaki Harai (National Inst. of Info. and Commun. Tech.), Ezra Ip (NEC Laboratories America), Hiroshi Kubo (Ritsumeikan Univ.), Ayumu Kubota (KDDI R&D), Zhetao Li (Xiangtan University), Richard T.B. Ma (National University of Singapore), Tzyh-Ghuang Ma (National Taiwan Univ. of Sci. and Tech.), Ryutaroh Matsumoto (Tokyo Inst. of Tech., Aalborg Univ.), Hiroaki Morino (Shibaura Inst. of Tech.), Hoang Nam Nguyen (Vietnam National Univer- sity Hanoi), Wakaha Ogata (Tokyo Inst. of Tech.), Masakatsu Ogawa (Sophia Univ.), Chuwong Phongcharoenpanich (King Mongkut’s Inst. of Tech. Ladkrabang), Nordin Ramli (Malaysian Inst. of Microelectronic Systems), Kentaro Saito (Tokyo Institute of Technology), Yuji Sekiya (The Univ. of Tokyo), Takatoshi Sugiyama (Kogakuin Univ.), Hidenori Takahashi (KDDI R&D), Ryo Yamaguchi (Softbank Mobile), Shuto Yamamoto (NTT), Hui Zhang (Nankai University), Miao Zhang (Xiamen Univ.)

560 IEICE Communications Express, Vol.9, No.12, 561–566 Special Cluster in Conjunction with IEICE General Conference 2020 Channel prediction of wideband OFDM systems in a millimeter-wave band based on multipath delay estimation

Yuta Takano1, Toshihiko Nishimura1, Takeo Ohgane1, Yasutaka Ogawa1, a), and Junichiro Hagiwara1 1 Graduate School/Faculty of Information Science and Technology, Hokkaido University, Kita 14, Nishi 9, Kita-ku, Sapporo, Hokkaido, 060-0814, Japan a) [email protected]

Abstract: Multi-user MIMO systems enable high capacity transmission. A base station, however, needs accurate channel state information (CSI). In time-varying environments, the CSI may be outdated at the actual transmission time. One of the solutions to this issue is channel prediction. The authors have proposed the prediction method using FISTA, a compressive sensing technique, for OFDM systems in a millimeter-wave band. Unfortunately, in realistic multipath environments, the prediction performance of the proposed technique degrades. In this letter, we examine the prediction performance in wider OFDM systems. It will be shown that FISTA reveals excellent performance in a suﬃciently wide band case. Keywords: channel prediction, multipath, delay estimation, compressive sensing, OFDM, IDFT Classiﬁcation: Wireless Communication Technologies

References

[1] A. Duel-Hallen, “Fading channel prediction for mobile radio adaptive transmission systems,” Proc. IEEE, vol. 95, no. 12, pp. 2299–2313, Dec. 2007. DOI: 10.1109/JPROC.2007.904443 [2] A. Adhikary, E.A. Safadi, M.K. Samimi, R. Wang, G. Caire, T.S. Rappaport, and A.F. Molisch, “Joint spatial division and multiplexing for mm-wave channels,” IEEE J. Sel. Areas Cummun., vol. 32, no. 6, pp. 1239–1255, June 2014. DOI: DOI:10.1109/JSAC.2014.2328173 [3] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear inverse problems,” SIAM J. Imag. Sci., vol. 2, no. 1, pp. 183–202, March 2009. DOI: 10.1137/080716542 [4] Y. Takano, Y. Ogawa, T. Nishimura, T. Ohgane, and J. Hagiwara, “Channel prediction of wideband OFDM systems in a millimeter-wave band using delay-

561 IEICE Communications Express, Vol.9, No.12, 561–566

[5] M.K. Samimi and T.S. Rappaport, “3-D millimeter-wave statistical channel model for 5G wireless system design,” IEEE Trans. Microw. Theory Techn., vol. 64, no. 7, pp. 2207–2225, July 2016. DOI: 10.1109/TMTT.2016.2574851 [6] K. Hayashi, M. Nagahara, and T. Tanaka, “A user’s guide to compressed sensing for communications systems,” IEICE Trans. Commun., vol. E96-B, no. 3, pp. 685–712, March 2013. DOI: 10.1587/transcom.E96.B.685

1 Introduction A base station (BS) needs downlink channel state information (CSI) for multi-user MIMO transmission that enables high capacity communication. Radio propagation, however, usually varies due to the motion of user equipment (UE) and/or scatters. In such time-varying multipath environments, the CSI estimated with pilot symbols may be outdated at the actual transmission time, and the precoding performance may degrade. If we predict CSI from observed past one, we can reduce the degradation. Among prediction techniques, the sum-of-sinusoids method [1] predicts channels by resolving an arrival signal into individual multipath components and summing the predicted ones. This corresponds to predicting a future delay proﬁle. If the complex amplitude of each multipath component is estimated accurately, we can predict reliable channels for a long prediction range. At millimeter-wave frequencies, channel responses are sparse [2], and we can apply a compressive sensing technique such as the fast iterative shrinkage-thresholding algorithm (FISTA) [3] to obtain the complex amplitudes of multipath components. The authors have proposed the CSI prediction method in which FISTA resolves the arrival signal in a wideband OFDM system into multipath components in the delay domain, and recently reported the prediction performance [4]. Unfortunately, in realistic multipath environments [5], the prediction performance of the proposed technique degrades and is almost the same as that of the method using the conventional inverse discrete Fourier transform (IDFT). In this letter, we will show the prediction performance in wider OFDM systems.

2 Formulation of the CSI prediction The goal of this letter is to examine the CSI prediction using delay proﬁle estimation in wideband OFDM systems, and we deal with SISO-OFDM for the sake of simplicity. Extending to MIMO systems is a future work. The prediction technique can be applied to both of FDD and TDD.

We express the downlink channel state at subcarrier frequency fs at time t as H( fs; t) (s = 1,2, ··· , Sc), where Sc denotes the number of subcarriers in OFDM. We assume that the BS has the CSI at time t0 and t1, i.e., H( fs; t0) and H( fs; t1) (t0 < t1) with pilot symbols. Representing the pilot symbol transmission interval as Tint, we have t1 = t0 + Tint. Note that Tint is the channel measurement interval. We predict the channel state at time t2, H( fs; t2), from H( fs; t0) and H( fs; t1) (t1 < t2). We assume that all the multipath components arrive in the delay range from © IEICE 2020 τ τ DOI: 10.1587/comex.2020COL0002 min to max. We discretize the delay range into P sampling points. The ith delay Received April 8, 2020 τ′ = τ + ( − ) τ τ = (τ − τ )/( − ) Accepted April 23, 2020 sampling point is given by i min i 1 ∆ , where ∆ max min P 1 Publicized May 14, 2020 Copyedited December 1, 2020

562 IEICE Communications Express, Vol.9, No.12, 561–566

and i = 1,2, ··· , P. Expressing the complex amplitude of multipath component at τ′ ′( ) delay i at time t as Ai t , we approximately have ∑P ′( ) (− π τ′) = ( )( = , , ··· , ). Ai t exp j2 fs i H fs; t s 1 2 Sc (1) i=1

At time t0 and t1, the right-hand side is H( fs; t0) and H( fs; t1), respectively. As stated previously, they have been obtained with pilot symbols. At each time, we have ′( ) Sc linear equations, and the P complex amplitudes Ai t are unknowns.

(a) When P = Sc, (1) is an even-determined case, and should theoretically be solved. However, because the subcarrier frequency interval ∆ fs and/or delay sampling point interval ∆τ are/is small, the equations are usually ill-conditioned, and we cannot solve them.

(b) When P < Sc, (1) is an over-determined case, and the least squares method with the pseudo-inverse matrix should be used. Unfortunately, the method usually does not work because of the ill condition also in this case.

(c) When P > Sc, (1) is an under-determined case. The solution is not determined uniquely, and we should use the pseudo-inverse matrix. Due to the ill-conditioned situation, the matrix inversion cannot be done in most cases.

To overcome the above ill condition, we tried the Tikhonov regularization and the rank reduction technique. Unfortunately, they did not work well in our cases. On the other hand, in millimeter-wave bands, channels are sparse, and most of ′( ) ′( ) ℓ ℓ Ai t are 0. We can expect that Ai t are obtained by the 1- 2 optimization [6], which minimizes the squared error with the ℓ1-norm regularization term. This is a compressive sensing technique. In this study, we use FISTA [3] for the ℓ1-ℓ2 ′( ) ′( ) optimization. The detail procedure for obtaining Ai t0 and Ai t1 using FISTA, and ′( ) ′( ) prediction of Ai t2 are stated in [4]. From Ai t2 , we can predict the future channels H( fs; t2) by the following equation:

∑P ˆ ( ) = ′( ) (− π τ′). H fs; t2 Ai t2 exp j2 fs i (2) i=1

3 Simulations In this section, we show simulation results of the channel prediction method with FISTA. In addition to them, we also consider the performance using the conventional IDFT for multipath detection in the delay domain. As for the channel model, we used the 28-GHz non-line-of-sight model proposed in [5]. Table I shows the simulation parameters. Refer to [4, 5] for symbols and terms

that are not deﬁned in this letter. Strictly speaking, since ∆ fs = 120 kHz, the bandwidth B is 199.92 MHz for Sc = 1666. In this letter as shown in Table I, we write B = 200 MHz for the sake of simplicity. This is the same also for B = 400 MHz, 800 MHz. In a frame conﬁguration in the 5G system employing TDD for 28 GHz,

the downlink duration is 0.25 ms, and the channel measurement interval Tint was set © IEICE 2020 DOI: 10.1587/comex.2020COL0002 0.25 ms as shown in Table I. Furthermore, in the simulations, we ignored noise, that Received April 8, 2020 Accepted April 23, 2020 is, we assumed that H( fs; t0) and H( fs; t1) were obtained accurately. Publicized May 14, 2020 Copyedited December 1, 2020

563 IEICE Communications Express, Vol.9, No.12, 561–566

Table I. Simulation parameters

Lowest subcarrier frequency f1 28 GHz Subcarrie frequency interval ∆ fs 120 kHz Bandwidth B 200 MHz 400 MHz 800 MHz Number of subcarriers Sc 1666 3333 6666 Number of unknowns for 313 626 1251 multipath components P Delay sampling point interval 0.8 ns 0.4 ns 0.2 ns for FISTA ∆τ Delay sampling point interval 5 ns 2.5 ns 1.25 ns for IDFT 1/B δ for FISTA 2.08 × 104 4.17 × 104 8.33 × 104 Distance between BS and UE 60.2 m Minimum delay point τmin 150 ns Maximum deley point τmax 400 ns UE velocity 20 m/s Channel measurement interval Tint 0.25 ms Channel measurement time t0, t1 t0 = 0 s t1 = t0 + Tint = 0.25 ms t = t + T Channel prediction time t 2 1 2 T = 0.025,0.050,0.075, ··· ,0.250 ms µ for FISTA 1.0 ε for FISTA 1 × 10−6 Number of clusters 3 Number of subpaths for each cluster 10 Cluster excess delay interval 25 ns Subpath delay interval 2.5 ns Shadowing Disregarded Number of trials K 100

We determined the subpath phases, mean azimuth and elevation angles of arrival from the lobes, and angle spreads by random numbers [5]. Changing the random numbers, we conducted K = 100 trials, and evaluated the channel prediction performance using the normalized mean square error (MSE) deﬁned by ( ) ∑K ( ) 1 E{|∆H k ( f ; t )|2} Normalized MSE [dB] = 10 log s 2 , (3) 10 K E{|H(k)( f ; t )|2} k=1 s 2

(k) where E{} is the mean for all the subcarrier frequencies fs, and ∆H ( fs; t2) is the channel prediction error at the sth subcarrier for the kth trial. Fig. 1 shows examples of multipath component estimation results in the delay

domain at t0 for diﬀerent frequency bandwidths. Note that Fig. 1 shows only the delay range where all the multipath components arrive. We see that three clusters exist and each cluster has ten subpaths expressed by × marks. IDFT cannot resolve the subpaths for any bandwidth. FISTA cannot detect them correctly for B = 200 MHz, 400 MHz either. However, when B = 800 MHz, it is seen that FISTA can resolve all the subpaths. Estimated amplitudes look rather low, but we see several estimated responses in the neighborhood of each subpath. Combining them, the estimated amplitude will have a higher value. © IEICE 2020 DOI: 10.1587/comex.2020COL0002 Fig. 2 shows the channel prediction performance, Normalized MSE versus T. Received April 8, 2020 Accepted April 23, 2020 As stated in Section 2, t2 = t1 +T holds, that is, we predict the channel at T from the Publicized May 14, 2020 Copyedited December 1, 2020

564 IEICE Communications Express, Vol.9, No.12, 561–566

Fig. 1. Examples of multipath component estimation results at t0.

estimation time t1. The maximum prediction range T = 0.25 ms is the same as the downlink duration in the 5G system employing TDD for 28 GHz as stated previously. The performance of FISTA using 200 MHz or 400 MHz bandwidth data is almost the same as that of IDFT. Considering that we ignored noise, the performance is © IEICE 2020 DOI: 10.1587/comex.2020COL0002 poor. When the bandwidth is 800 MHz, however, the CSI prediction performance Received April 8, 2020 Accepted April 23, 2020 Publicized May 14, 2020 Copyedited December 1, 2020

565 IEICE Communications Express, Vol.9, No.12, 561–566

Fig. 2. Channel prediction performance.

of FISTA is improved greatly and is much better than that of IDFT. This is because each multipath component is resolved by the wider bandwidth data as shown in Fig. 1(c). It is seen that even though we use a compressive sensing technique, wider bandwidth data are required to predict CSI accurately.

4 Conclusions In this letter, we have examined the CSI prediction of OFDM systems based on delay proﬁle estimation. We used FISTA to resolve multipath components in the delay domain, and compared the CSI prediction performance with that of IDFT. It has been shown that in realistic millimeter-wave propagation environments, if the bandwidth of OFDM is suﬃciently wide, FISTA reveals excellent performance, which is much better than that of IDFT.

566 IEICE Communications Express, Vol.9, No.12, 567–572 Special Cluster in Conjunction with IEICE General Conference 2020 Measurement accuracy of Wi-Fi FTM on actual devices

Masakatsu Ogawa1, a) and Hoyeon Choi1 1 Faculty of Science and Technology, Sophia University, 7-1 Kioi-cho, Chiyoda-ku, Tokyo 102-8554, Japan a) [email protected]

Abstract: Demand for positioning services such as object tracking systems and navigation systems is increasing. The current Wi-Fi positioning system uses the algorithm using mainly the signal strength. However, the signal strength is susceptible to multipath propagation, thus the ranging error generally large. The IEEE802.11mc specifies the fine timing measurement (FTM) protocol based on round trip time (RTT). As of now, few commercial products support the FTM function. The objective of this letter is to evaluate the accuracy of the Wi-Fi FTM using the actual devices and to perform one-dimensional positioning by a simple algorithm. In the one-dimensional positioning, average positioning accuracy is 0.10 m in outdoor and 0.23 m in indoor. Keywords: Wi-Fi, fine timing measurement (FTM), positioning Classification: Wireless Communication Technologies

References

[1] IEEE Std 802.11-2016 (revision of IEEE Std 802.11-2012), “Part 11: Wireless LAN medium access control (MAC) and physical layer (PHY) specifications,” IEEE Standard for Information technology — Telecommunications and information exchange between systems. Local and metropolitan area networks — Specific requirements, 2016. [2] M. Ibrahim, H. Liu, M. Jawahar, V. Nguyen, M. Gruteser, R. Howard, B. Yu, and F. Bai, “Verification: accuracy evaluation of WiFi fine time measurements on an open platform,” Proc. 24th Annual International Conference on Mobile Computing and Networking (MobiCom ’18), New Delhi, India, pp. 417–427, Oct. 2018. DOI: 10.1145/3241539.3241555 [3] Google, Documentation for app developers, Wi-Fi location: ranging with RTT, https://developer.android.com/guide/topics/connectivity/wifi-rtt, accessed April 2, 2020.

1 Introduction Location information is expected for several applications, such as object tracking systems and navigation systems. The global navigation satellite system (GNSS) is © IEICE 2020 the most widely used in an outdoor environment, and whose main application is the DOI: 10.1587/comex.2020COL0001 Received April 3, 2020 navigation for such as cars and smartphones. However, the GNSS cannot apply to Accepted April 28, 2020 Publicized May 19, 2020 Copyedited December 1, 2020

567 IEICE Communications Express, Vol.9, No.12, 567–572

the indoor environment because there is no signal from satellites. A solution for an indoor environment is the use of wireless communication systems such as cellular systems, Wi-Fi, Bluetooth, and Zigbee. Most of the positioning algorithms are based on the received signal strength (RSS). However, signal strength is susceptible to multipath propagation, thus the ranging error generally large. There is the round-trip time (RTT) based measurement to overcome the inherent difficulties. The IEEE802.11mc specifies the fine trimming measurement (FTM) based on RTT [1]. As of now, few commercial products support the FTM function, especially in Japan, there are very few products that have certificated the technical standard conformity certificate. As far as we searched, we find two kinds of devices that support the FTM. One is Intel Dual Band Wireless-AC 8260 NIC (Intel 8260) [2]. The other is Google Wi-Fi (product official name is Google Wifi) and Google Pixel 3 [3]. In the case of Intel 8260, RTT is obtained by the debug information; thus, RTT cannot be acquired by the general use. On the other hand, Android 9.0 Pie and above version support the FTM. Because of this situation in Japan, the objective of the letter is to clarify the accuracy of measurement distance using actual commercial devices and to clarify the accuracy of one-dimensional positioning by a simple algorithm.

2 Fine timing measurement (FTM) protocol The IEEE802.11mc speciﬁes the FTM protocol to estimate the distance between two devices without association to a particular access point (AP). The distance is estimated by Time of Flight (ToF) using signal propagation. This protocol is used for the measurement of RTT based distance between two devices. In the FTM process, an initiator ﬁrst transmits an FTM request to a responder corresponding to a device for which the distance is desired to be measured. After the initiator receives the response regarding the FTM request from the responder, the ranging

process is started. The responder sends the FTM frame at time t1 and waits for its acknowledgment (ACK). The initiator receives the FTM frame at time t2 and responds with its ACK at time t3. After that, the responder receives the ACK at time t4 and sends the FTM frame, including t1 and t4. Through this process, the initiator has four types of times, t1, t2, t3, and t4. The diﬀerence time (t4 − t1), includes the processing time between the reception and transmission at the initiator (t3 − t2). The RTT excludes the processing time (t3 − t2); thus, it is (t4 − t1) − (t3 − t2). The RTT based distance is calculated by ((t4 − t1) − (t3 − t2)) · c/2, where c is the velocity of the radio signal.

3 Experimental setup We use the following devices: Intel 8260 operated as an AP and a station (STA), Google Wi-Fi acted as an AP, and Google Pixel 3 (Pixel3) acted as an STA. The STAs act as the initiator, and the APs act as the responder. The limitation of the device is the frequency setting in the AP. The frequency settings of Intel 8260 operated as AP and Google Wi-Fi are 2.4 GHz and 5 GHz, respectively. Besides, © IEICE 2020 Intel 8260 as STA and Pixel3 support both 2.4 GHz and 5 GHz, respectively. Note DOI: 10.1587/comex.2020COL0001 Received April 3, 2020 that the AP and STA are used, but the station need not associate with the AP. Accepted April 28, 2020 Publicized May 19, 2020 Copyedited December 1, 2020

568 IEICE Communications Express, Vol.9, No.12, 567–572

In this letter, two experiments in indoor and outdoor are conducted. The propagation between the AP and STA is a line-of-sight (LOS) environment. The objective of the first experiment is to clarify the accuracy of measurement distance using actual commercial devices. The combination of AP and STA is four types: AP (Intel 8260) and STA (Intel 8260), AP (Intel 8260) and STA (Pixel3), AP (Google Wi-Fi) and STA (Intel 8260), and AP (Google Wi-Fi) and STA (Pixel3). In these combinations, the antenna height of the AP and STA set the same height, and we measure the distance between the AP and STA three times. The height of the antenna is 1, 2, and 3 m for outdoor, and 1 and 2 m for indoor. The ceiling height limits the height of the antenna indoor. The first experiment also clarifies the accuracy difference in these combinations. The second experiment is to clarify the accuracy of one-dimensional positioning by a simple algorithm in the case of the best accuracy combination in the first experiment. The difference from the first experiment, the height of STA is 1 m. This is the height at which people usually use their smartphones. Besides, the AP is fixed at the height of 3 m for outdoor and 2 m for indoor.

4 Experimental evaluation We evaluate the accuracy measurement through two experiments. The ﬁrst experiment is to evaluate the distance accuracy. The second experiment is to evaluate the one-dimensional positioning accuracy. The measurement is conducted three times in each STA’s position. We extract 200 data excluding data that could not be measured correctly, delete the outlier that is more than three scaled median absolute deviations (MAD) away from the median and use the mean of distance and RSS.

4.1 Accuracy of measurement distance We experimented indoor and outdoor. Fig. 1(a) and (b) show the experimental environment outdoor. The antenna at the AP and STA is the same height. The number of measurement points is five; 5 m, 10 m, 15 m, 20 m, and 25 m from the AP. The measurement results of distance and RSS in each devices’ combination show Fig. 1(c) to (f). In all combinations of devices, it is evident that the RSS does not decrease with the actual distance and varies widely for each measurement. Therefore, it is clarified that the distance based on the RSS is low accuracy. In terms of the measured distance, the combination of AP (Intel 8260) and STA (Pixel3) has a great variation for each measurement compared with other combinations, and it is also evident that the measurement error is significant. To further analysis of the measurement distance, we evaluate it by the mean absolute error (MAE). Fig. 2(a) shows the MAE in outdoor. The combination of AP (Google Wi-Fi) and STA (Pixel3) achieves the minimum MAE. In the 3 m height, the MAE is 0.26 m. The other combinations, the MAE is over 1 m. As with outdoor, in terms of indoor shown in Fig. 2(b), the combination of AP (Google Wi-Fi) and STA (Pixel3) achieves the minimum MAE. Unlike the outdoor environment, since the indoor environment is affected by multipath propagation, the MAE in indoor is

© IEICE 2020 larger than outdoor. DOI: 10.1587/comex.2020COL0001 Received April 3, 2020 In the ﬁrst experiment, it was found that the accuracy of measurement distance Accepted April 28, 2020 Publicized May 19, 2020 Copyedited December 1, 2020

569 IEICE Communications Express, Vol.9, No.12, 567–572

Fig. 1. Experimental environment and its results in outdoor.

Fig. 2. Mean absolute error of measurement distance in outdoor and indoor. AP_STA denotes the combination of AP and STA.

depends on the device, and the best combination is AP (Google Wi-Fi) and STA (Pixel3).

570 IEICE Communications Express, Vol.9, No.12, 567–572

4.2 Accuracy of one-dimensional positioning In the second experiment, the combination of AP (Google Wi-Fi) and STA (Pixel3) is used. An algorithm of one-dimensional positioning is straightforward. In Fig. 3(a),

let D, d1, and d2 be the distance between AP1 and AP2, the measurement distance from AP1 and AP2, respectively. Assuming that the measurement distance, d1, and d2, includes the same error, the one-dimensional positioning point, P, is expressed as follows:

  (D − (d1 + d2)  d1 + if D ≥ d1 + d2  2 P = (1)  ( − ( + )  D d1 d2  d1 − if D < d1 + d2  2 We experimented indoor and outdoor. Fig. 3(b) shows the experimental environment. The number of measurement points is four; 5 m, 10 m, 15 m, and 20 m from the AP1. We evaluate the accuracy of one-dimensional positioning by the MAE and compare the positioning point using both AP1 and AP2 with that using only AP1 and only AP2. Fig. 3(c) and (d) show the MEA in outdoor and indoor, respectively. In these ﬁgures, “Total” denotes the MAE in all points. Even if both AP1 and AP2 are used, the MAE may not be decreased at each point. In all points, MAE with the use of both AP1 and AP2 is decreased compared with the use of only AP1 or only AP2. In this case, average positioning accuracy is 0.10 m in outdoor and 0.23 m in indoor.

Fig. 3. One-dimensional positioning algorithm, experimental environment, and mean absolute error of positioning in outdoor and indoor.

5 Conclusion © IEICE 2020 DOI: 10.1587/comex.2020COL0001 We have compared Wi-Fi FTM supported devices by the measurement distance. As Received April 3, 2020 Accepted April 28, 2020 of now, there is few Wi-Fi FTM supported devices, and it is necessary to clarify Publicized May 19, 2020 Copyedited December 1, 2020

571 IEICE Communications Express, Vol.9, No.12, 567–572

the measurement accuracy depends on devices. The experimental results show that the combination of Google Wi-Fi as AP and Google Pixel as STA achieves the best accuracy in indoor and outdoor. Besides, we have experimented with the one-dimensional positioning using two APs. We demonstrated that positioning accuracy is 0.10 m in outdoor and 0.23 m in indoor. Our future work is to consider a two-dimensional positioning algorithm.

572 IEICE Communications Express, Vol.9, No.12, 573–579 Special Cluster in Conjunction with IEICE General Conference 2020 Propagation delay time estimation by neural network using urban environment parameters

Koyo Tategami1, a), Mitoshi Fujimoto1, Koshiro Kitao2, Minoru Inomata2, Satoshi Suyama2, and Yasuhiro Oda2 1 Graduate School of Engineering, University of Fukui, 3–9–1 Bunkyo, Fukui 910–8507, Japan 2 NTT DOCOMO, INC., 3–6 Hikarino-oka, Yokosuka-shi, Kanagawa 239–8536, Japan a) [email protected]

Abstract: In the 5th generation mobile communication system, introduction of small cells using low-height base station is being examined. Small cells are installed in urban street cell environments where mobiles are densely populated. Therefore, actual measurement and complicated modeling of radio wave propagation are necessary to grasp the propagation characteristics. In this paper, a simple propagation delay estimation method is proposed. In the proposed method, neural network using urban area structure parameters is applied. It is shown that the propagation delay time can be easily estimated by the proposed method without considering moving objects and trees. Keywords: 5G mobile communication system, urban street cell, neural network, propagation delay time estimation Classiﬁcation: Antennas and Propagation

References

[1] NTT DOCOMO, “DOCOMO 5G White Paper, 5G Radio Access: Requirements, Concept and Technologies,” July 2014. [2] K. Tategami, M. Fujimoto, K. Kitao, and T. Imai, “Angle of arrival characteristics at 20 GHz band in LOS urban area street cell environment,” Proc. 2019 IEEE International Workshop on Electromagnetics, PS-7, Sep. 2019. DOI: 10.1109/iwem.2019.8887950 [3] M. Inomata, T. Imai, K. Kitao, Y. Okumura, M. Sasaki, and Y. Takatori, “Radio propagation prediction method using point cloud data based on hybrid of ray- tracing and effective roughness model in urban environments,” IEICE Trans. Commun., vol. E102-B, no. 1, pp. 51–62, Jan. 2019. DOI: 10.1587/transcom. 2017EBP3436 [4] M. Sasaki, M. Inomata W. Yamada, N. Kita, T. Onizawa, M. Nakatsugawa, K. Kitao, and T. Imai, “Path loss model considering blockage effects of traffic signs up to 40GHz in urban microcell environments,” IEICE Trans. Com-

© IEICE 2020 mun., vol. E101-B, no. 8, pp. 1891–1902, Aug. 2018. DOI: 10.1587/transcom. DOI: 10.1587/comex.2020COL0008 2017EBP3255 Received May 27, 2020 Accepted June 18, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

573 IEICE Communications Express, Vol.9, No.12, 573–579

[5] S.B. Kotsiantis, “Supervised machine learning: a review of classiﬁcation techniques,” Informatica, vol. 31, pp. 249–268, July 2007. [6] K. Saito, Y. Jin, C. Kang, J. Takada, and J. Leu, “Two-step path loss prediction by artiﬁcial neural network for wireless service area planning,” IEICE Com- mun. Express, vol. 8, no. 12, pp. 611–616, Sep. 2019. DOI: 10.1587/comex. 2019GCL0038 [7] S.N. Livieratos and P.G. Cottis, “Rain attenuation along terrestrial millimeter wave links: a new prediction method based on supervised machine learning,” IEEE Access, vol. 7, pp. 138745–138756, Oct. 2019. DOI: 10.1109/ ACCESS.2019.2939498 [8] R. Rojas, Neural Networks: A Systematic Introduction, Springer-Verlag, Berlin, 1996.

1 Introduction In the 5th generation mobile communication systems, introduction of small cells using low-height base station is being examined. The major installation environments of small cells are urban street cell environments where pedestrians and vehicles are densely populated [1]. In order to understand the propagation characteristics in such environments, it is necessary to analyze the data obtained by field measurements [2] or to perform simulation analysis by complex modeling of moving objects and trees [3, 4]. In this paper, a simple propagation delay estimation method is proposed. In the proposed method, the delay profile measured by the channel sounder is applied to the trainings data of neural network [5, 6, 7]. And also the urban area structure and the parameters calculated from them are applied as the inputs data of the neural network. This paper is organized as follows. The proposed propagation delay time estimation method is explained in Section 2. The measurement environments and specifications are explained in Section 3. Comparison of measured results and estimated results is explained in Section 4. The conclusion is provided in Section 5.

2 Propagation delay time estimation by neural network 2.1 Outline of the proposed estimation method In the proposed method, neural network is used to realize simple propagation delay time estimation. The process of the proposed method is as follows. Step1. Model the simplified street model on the urban area. Step2. Calculate delay time using the simplified street model. Step3. Learn the neural network model. The input data for the learning is dimension of the simplified street model (in step1) and calculated delay times (in step2). The training data for the learning is the measured delay profile. Step4. Estimate the delay profile by inputting unmeasured parameters into the learned model in step3. In order to estimate the propagation delay time without considering moving objects and trees, the actually measured delay profile is used as training data. Since the © IEICE 2020 DOI: 10.1587/comex.2020COL0008 measured delay profile reflects the effects of moving objects and trees, these effects Received May 27, 2020 Accepted June 18, 2020 are also reflected in the model obtained by learning. Also, by using the results Publicized June 30, 2020 Copyedited December 1, 2020

574 IEICE Communications Express, Vol.9, No.12, 573–579

Fig. 1. Proposed propagation delay time estimation method by the neural network

measured at many points for learning, it is possible to estimate the results at points where no measurement is performed.

2.2 Input parameters First, the urban area is modeled to the simplified street model consisting of ground and buildings as shown in Fig. 1 (a). This model is constructed using the dimension parameters shown in Fig. 1 (b) of the actual measurement environment. The dimension parameters can be easily obtained. The delay time shown in Fig. 1 (b) is calculated using the dimension parameter of the simplified street model. The path length of the radio waves is calculated geometrically from the simplified street

model, and the delay time is calculated. The delay time τ1 is the direct wave, the delay time τ2 is the reﬂected wave from the building near the transmitting station, © IEICE 2020 and the delay time τ3 is the reﬂected wave from the building far from the transmit- DOI: 10.1587/comex.2020COL0008 Received May 27, 2020 ting station. Besides the delay time from buildings, the delay time of the ground Accepted June 18, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

575 IEICE Communications Express, Vol.9, No.12, 573–579

reflection wave can be calculated, but the delay time difference between the direct wave and the ground reflection wave is very small. This means that the delay time of the ground reflection wave has a small role as a feature. Therefore, it is not used as an input parameter. The respective delay times are expressed by the following equations. √ d2 + (h − h )2 τ = Rx Tx × 109 (1) 1 c √ (2W + d sin |θ|)2 + (d cos |θ|)2 + (h − h )2 τ = Tx Rx Tx × 109 (2) 2 c √ (2(W − W ) − d sin |θ|)2 + (d cos |θ|)2 + (h − h )2 τ = Tx Rx Tx × 109 (3) 3 c Here, θ is the angle of the transmitting station with respect to the receiving station, and c is the speed of light. These parameters are normalized by the maximum value per each environment and input to the neural network model in Fig. 1 (c).

2.3 Output parameters The output layer of the neural network shown in Fig. 1 (c) corresponds to the delay bin, and outputs the probability value that a wave arrives within each delay from each neuron. Since the delay profile used for learning is normalized by the maximum value per delay profile obtained for each environment, the value of the delay profile ranges from 0 to 1. This relative power can be regarded as a probability value. Therefore, the output probability value corresponds to the relative power.

3 Measurement environment and specifications Figures 2 (a) and 2 (b) show the measurement area of the actual measurement data, Fig. 2 (c) shows a schematic diagram of the measurement system, and Fig. 2 (d) shows the measurement specifications. The measurement environments were urban street cell environments in which buildings lined up on both sides of the road. Only the LOS measurements are discussed in this paper. In addition, there were running and parking vehicles on the road. Also, there were trees and subway entrances, lights, pedestrians, etc. on the sidewalk. The measurement was conducted in an uplink, and the transmitting station (Tx) transmitted an OFDM signal with a bandwidth of about 45 MHz in a 20 GHz band using a sleeve antenna. The transmission power was 30dBm. On the base station (receiving station) [BS (Rx)], signals were received using a 256 (16 × 16) elements planar patch array antenna. The base station antenna height was 5m. The transmitting station antenna heights were 1.5m or 2.5m in Kayaba, and only 2.5m in Hatchobori. When the transmitting station was installed on the sidewalk, its height was 1.5m, and when it was installed on the roadway, its height was 2.5m. The measurements were conducted at fixed-point. At each position of the transmitting station, data was recorded 30 times at 1-second intervals on the receiving station side, © IEICE 2020 DOI: 10.1587/comex.2020COL0008 and five snapshots were recorded at 1-ms intervals for each recording. The delay Received May 27, 2020 Accepted June 18, 2020 profile was obtained by performing IFFT on the data (complex amplitude of each Publicized June 30, 2020 Copyedited December 1, 2020

576 IEICE Communications Express, Vol.9, No.12, 573–579

Fig. 2. Measurement environment and speciﬁcations

subcarrier in the frequency domain) acquired by each element. An average delay profile was calculated by averaging the delay profiles obtained by the respective antenna elements. The largest delay time of the measured delay profile is 10000ns. The largest delay time was set to 2209 ns in the analysis, the median value after that time (between 2232 and 10000 ns) was found, and the value beyond 5 dB from the median value was taken as the noise floor. All values below the noise floor were set to 0. The delay profile obtained by the process was used for learning.

4 Comparison of measured results and estimated results In the neural network model [8], the number of hidden layers is 2, the activation function is sigmoid function, the loss function is binary cross-entropy, and the optimization algorithm is Adam. The number of neurons in the input layer is 9, the number of neurons in the output layer is 100, the number of neurons in the hidden layer is 200 and 150, and the dropout rate is 0.2. Here, the loss function © IEICE 2020 DOI: 10.1587/comex.2020COL0008 is a function that represents the magnitude of the diﬀerence between the estimation Received May 27, 2020 Accepted June 18, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

577 IEICE Communications Express, Vol.9, No.12, 573–579

Fig. 3. Comparison of measurement and estimation results

and the actual value, and evaluates the estimation accuracy of the model. The optimization algorithm is an algorithm for updating the weight so that the value obtained by the loss function is minimized. By repeatedly updating the weights, the delay time estimation model is constructed. Since the constructed delay time estimation model has 100 delay bins with delay resolution of 22.3ns, the largest © IEICE 2020 DOI: 10.1587/comex.2020COL0008 delay time that can be estimated is 2209ns. Received May 27, 2020 Accepted June 18, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

578 IEICE Communications Express, Vol.9, No.12, 573–579

Figure 3 (a) shows an example of the measured and estimated results when the transmitting station (height 1.5m) is installed at K1 and K7 as verification data. In order to show that the estimation model can estimate the delay time of unmeasured points, those used as training data is not used for verification data. Here, the data for verification (the number of data: 30) was set to one of the transmitting stations, and the data for learning was data different from the data for the verification (number of data: 960). The measurement result when a transmitting station with a height of 1.5 m is installed in K1 shows that the peak of the first delayed wave of the estimated result matches the measured result, and the delay time can be estimated with high accuracy. In the case of K7, although the peak positions are partially different, many peaks coincided. Figure 3 (b) shows an example of the measured and estimated results when the transmitting station (height 2.5m) is installed at H1 and H4 in Hatchobori. Except for the difference in the peaks of the first delayed wave between the estimated and measured results for H4, the peaks coincide in many parts. So it can be said that the delay time estimation is possible by the proposed method. Figure 3 (c) shows the cumulative distribution function of the mean delay time and delay spread at all transmitting stations. It is a combination of multiple cases. Different transmitters are used as verification data for each case, and are separated from the training data. In the learning for constructing the delay time estimation model, the noise floor processed delay profile was used. Therefore, the delay profile output by this model is the noise-floor-processed one. When calculating the average delay time and delay spread, the output delay profile was used without the noise floor processing. It can be confirmed that the distribution of the measured results and the estimated results are almost the same for both the mean delay time and the delay spread. The median of the measured delay spread is 125.52 ns, and the median of the estimated delay spread is 131.68 ns. These differences are as small as 6.16 ns. This suggests that the propagation delay time can be estimated not only direct waves but also delayed waves by the proposed method.

5 Conclusion A simple propagation delay estimation method by neural network using urban structure parameters as input was proposed. It was shown that the proposed method can accurately estimate the propagation delay time without considering moving objects and trees.

579 IEICE Communications Express, Vol.9, No.12, 580–585 Special Cluster in Conjunction with IEICE General Conference 2020 A proposal on virtual massive array using fast beam switching and blind algorithm

Issei Watanabe1, a), Kentaro Nishimori1, Ryotaro Taniguchi1, and Tomoki Murakami2 1 Graduate School of Science and Technology, Niigata University, Ikarashi 2-nocho 8050, Nishi-ku Niigata, 950–2181, Japan 2 Access Network Service Systems Laboratories, NTT Corporation, 1–1 Hikari-no-oka, Yokosuka-shi, Kanagawa, 239–0847, Japan a) [email protected]

Abstract: In future wireless communication systems, massive multiple- input multiple-output (MIMO) is attracting attention for improving transmission rate. However, there is an issue that a large scale of hardware is needed due to massive numbers of antennas, transmitters, receivers, A/D and D/A convertors in massive MIMO. In this study, we propose a method to realize a virtual massive array with only one receiver by changing the beam pattern and performing A/D conversion at high speed within one symbol. The basic performance and eﬀectiveness of the proposed method is demonstrated via a computer simulation. Keywords: massive MIMO, virtual massive array, propagation environmental control, Robust ICA Classiﬁcation: Antennas and Propagation

References

[1] C.V. Forecast, “Cisco Visual Networking Index: Global Mobile Data Traffic Forecast Update, 2010-2015,” Feb. 2011. [2] C.V. Forecast, “Cisco Visual Networking Index: Global Mobile Data Traffic Forecast Update, 2016-2021,” Feb. 2017. [3] G.J. Foschini and M.J. Gans, “On limits of wireless communications in a fading environment when using multiple antennas,” Wireless Personal Commun., vol. 6, pp. 311–335, 1998. [4] D. Gesbert, M. Kountouris, R.W. Heath, Jr., C.-B. Chae, and T. Salzer, “Shifting the MIMO Paradigm,” IEEE Signal Process. Mag., vol. 24, no. 5, pp. 36–46, Sep. 2007. DOI: 10.1109/msp.2007.904815 [5] Y. Takatori and K. Nishimori, “Application of downlink multiuser MIMO transmission technology to next generation very high throughput wireless access systems,” IEICE Trans. Commun., vol. J93-B, no. 9, pp. 1127–1139, Sep. 2010. [6] N. Kikuma, K. Nishimori, and T. Hiraguri, “Effect of user antenna selection on block beamforming algorithms,” IEICE Trans. Commun., vol. E101-B, no. 7, pp. 1523–1535, 2018. DOI: 10.1587/transcom.2017cqi0001 © IEICE 2020 DOI: 10.1587/comex.2020COL0011 [7] E.G. Larsson, “Very large MIMO system,” ICASSP 2012 Tutorial, 2012. Received June 4, 2020 Accepted June 18, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

580 IEICE Communications Express, Vol.9, No.12, 580–585

[8] F. Rusek, D. Persson, B.K. Lau, E.G. Larsson, T. L . Marzetta, O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with very large MIMO,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60, Jan. 2013. DOI: 10.1109/msp.2011.2178495 [9] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL of cellular networks: how many antenas do we need?,” IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013. DOI: 10.1109/JSAC.2013.130205 [10] W. Tidd, Y. Huang, and Y. Zhao, “Sequentical beamspace beamforming,” IEEE Aerospace Conference, March 2012. DOI: 10.1109/AERO.2012.6187088 [11] V. Zarzoso and P. Comon, “Robust independent component analysis by iterative maximization of the kurtosis contrast with algebraic optimal step size,” IEEE Trans. Neural Netw., vol. 21, no. 2, pp. 248–261, Feb. 2010. DOI: 10.1109/ TNN.2009.2035920

1 Introduction With the fast spread of wireless communication devices in recent years, mobile communication traffic is increasing rapidly. The total wireless communication traffic volume in 2021 is estimated to be nearly 200 times the traffic volume in 2010 [1, 2]. Concurrently, a high-speed wireless communication technology is needed. Multiple-input multiple-output (MIMO) technology using multiple antennas both at the transmitting and receiving sides has been introduced to improve the communication speed in a limited frequency band [3]. In addition, long-term evolution (LTE)-Advanced and IEEE 802.11ac standards introduce multiuser MIMO (MU-MIMO) [4, 5, 6]. In MU-MIMO, the nulls are created for other users, and high-speed communication is realized through digital signal processing. However, it has been observed that the transmission efficiency per user considerably reduces when the total number of antennas at the user terminals approaches the number of antennas at the base station [6]. To solve this problem, in the fifth-generation (5G) mobile communication systems, the application of massive MIMO using a large number of antennas at the base station has attracted attention [7, 8, 9]. Massive MIMO assumes that the number of antennas at the base station is sufficiently larger than the total number of antennas at the user terminals so that the transmission rate does not significantly decrease even if the number of users increases [9]. However, the issue in massive MIMO is that transmitters, receivers, A/D and D/A convertors are required for each antenna. Hence, the hardware scale is larger and the cost regarding the hardware increases as the number of antennas increases. So far, there has been research on direction of arrival (DOA) estimation using beam-space processing instead of element space one with one receiver [10]. To apply beam-space beamforming for the communication with one receiver, we propose a method to realize a virtual massive array with only one receiver by changing the beam pattern and performing A/D conversion at high speed within one symbol. Hence, our purpose is different from the purpose in [10]. Moreover, the proposed

© IEICE 2020 method realizes the multiple signal separation without even channel state information DOI: 10.1587/comex.2020COL0011 Received June 4, 2020 (CSI) which is essential in massive MIMO by using blind algorithm. In this letter, Accepted June 18, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

581 IEICE Communications Express, Vol.9, No.12, 580–585

the basic performance and effectiveness of the proposed method is presented by computer simulation when considering the different pattern with over sampling within one symbol using an actual modulation scheme. The remainder of this paper is organized as follows. In Sect. 2, the basic configuration and procedure of the proposed method are presented. The principle of the robust independent component analysis (ICA) is denoted. In Sect. 3, the basic performance and effectiveness of the proposed method is shown by using minimum shift keying (MSK) with over sampling.

2 Proposed method 2.1 Basic idea of proposed method Figure 1(a) shows the basic configuration of the proposed method. As the function changing the beam pattern, a total of P beam patterns are generated by changing the weight at high speed while receiving a signal within one symbol. As configuration examples for changing the beam pattern, we consider the rotation of the directional antenna, the rotation of the parasitic element, antenna switching, and antenna combination. In this study, we evaluated the case of antenna combination and changed the weight of a four-element half-wave circular array to make variable beams. Figure 1(b) shows an example of the relationship between sampling timings and beam patterns. By sampling multiple times at different timings while changing the beam pattern within one symbol, it is possible to obtain multiple received signals with different propagation environments even with one receiver. We call this method Propagation Environment Control. Next, Q (≤ P) received signals are selected from the P received signals. Fi- nally, to completely cancel the interference signal, we use the Robust independent component analysis (ICA) [11] with the selected signal as the input. The detail of the Roboust ICA is shown in Sect. 2.2. The Robust ICA is a blind adaptive array algorithm that uses only the received signal and does require neither user-terminal

Fig. 1. Proposed conﬁguration and an example of the relation- © IEICE 2020 DOI: 10.1587/comex.2020COL0011 ship between sampling timings and beam patterns Received June 4, 2020 Accepted June 18, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

582 IEICE Communications Express, Vol.9, No.12, 580–585

transmission timing synchronization nor the CSI estimation. The weight of the Ro- bust ICA is estimated not by using amplitude such as the constant modulus algorithm (CMA) but by using kurtosis, which is a measure of non-Gaussianity.

2.2 Principle of Robust ICA Robust ICA is a method for optimizing the step size and maximizing the kurtosis [11]. If the weight after m iterations is W(m), the equation for updating the weight using Robust ICA is given by

W(m + 1) = W(m) − µopt ∇W κ(m), (1)

µopt = arg max |κ(W(m) − µ∇W κ(m))|, (2) µ

where κ is the kurtosis and µ is the step size. In addition, ∇W κ(m), which is the gradient of κ with respect to W, is expressed as { 4 2 ∗ ∇W κ(m) = E{|y(i)| y(i) X(i)} E2{|y(i)|2} − E{y(i)X(i)}E{y(i)∗2} } (E{|y(i)|4} − |E{y(i)2}|2)E{y(i)∗ X(i)} − , (3) E{|y(i)|2}

where X(i)(i = 1, ··· , n) and y(i)(i = 1, ··· , n) are the input vector and output signal, respectively, at the i-th sample. W is updated to maximize the absolute value

of the kurtosis using the optimized step size µopt for each iteration. Robust ICA realizes a rapid and stable convergence performance. Moreover, it is reported that the calculation load per iteration is small because the root of the fourth

degree polynomial of µ is a candidate for µopt [11]. In this study, µ is calculated from 0.1 to 50, and µ with the maximum kurtosis is derived as µopt .

3 Effectiveness of proposed method The results after evaluating the eﬀect of the proposed method via computer simulation is demonstrated. Table I shows the simulation conditions. An environment with one desired wave and one interference wave is assumed and the direction of arrival is set to random. The SNR was 30 dB and the modulation scheme MSK, whose amplitude is constant. The smoothing size is the number of samples used for weight update. In the simulation, P over sampling is applied within one symbol

Table I. Simulation conditions Parameter Value SNR 30 dB Modulation scheme MSK Number of symbols 10000 smoothing size 1000

583 IEICE Communications Express, Vol.9, No.12, 580–585

and the received signals are obtained by each sampling signal. Note that fading is not considered, and suﬃcient symbols are given, because the basic performance is evaluated in this study. Note that we conﬁrmed that the weight of Robust ICA can be converged with a few iterations. We should evaluate the performance using a realistic number of symbols as future work. Figure 2 shows the cumulative distribution function (CDF) characteristics versus Bit Error Rate (BER) when the pattern variation, when P is changed. Here, the results when the pattern variation P is 2, 4, 8, and 16 and when Q is 2 are shown. From Fig. 2, it can be seen that the BER is reduced by increasing P. The dotted line shows the results when the MMSE with the CSI estimation is used in a two-element half-wave linear array. When P is 16, the BER is lower than that of the two-element linear array. The higher performance can be achieved by further increasing the value of P. From the results, there is a possibility that the proposed method can realize a virtual massive array even with one receiver without the CSI estimation.

Fig. 2. CDF characteristics versus BER

4 Conclusion This paper proposed a method to realize a virtual massive array with only one receiver by changing the beam pattern and performing A/D conversion at high speed within one symbol. We have evaluated the basic performance of the proposed method and we can confirm that the BER is reduced by increasing the pattern variation P. From this result, we revealed that even if one receiver was used, the same effect as using multiple receivers could be expected by utilizing multiple antenna patterns at high speed in one symbol and acquiring multiple signals with different propagation environments.

584 IEICE Communications Express, Vol.9, No.12, 580–585

Acknowledgments This part of the study was supported by SCOPE #185004002, KAKENHI, Grant- in-Aid for Scientiﬁc Research (B) (17H03262).

585 IEICE Communications Express, Vol.9, No.12, 586–592 Special Cluster in Conjunction with IEICE General Conference 2020 Limitation of parallel assumption in repeat-pass InSAR using nonparallel orbits

Masanori Gocho1, 2, a), b), Shoichiro Kojima1, and Hiroyoshi Yamada2 1 National Institute of Information and Communications Technology, 4-2-1 Nukui-Kitamachi, Koganei, Tokyo 184–8795 Japan 2 Faculty of Engineering, Niigata University, 8050 Ikarashi 2-no-cho, Nishi-ku, Niigata 950–2181, Japan a) [email protected] b) [email protected]

Abstract: In this report, we show the critical condition of the slope angle between orbits in InSAR (interferometric synthetic aperture radar) analysis. InSAR analysis, where the phase difference between two SAR observation data is measured, has been used for terrain height monitoring, and ground deformation detection, for example. However, InSAR data observed from nonparallel orbits degrades the performance of analysis. In the case of repeat- pass InSAR using airborne SAR, it is difficult to realize sufficiently parallel orbits. The purpose of this report is to formulate the nonparallel component of repeat-pass InSAR using an airborne SAR system and clarify the limitation of its parallel assumption. Then, we show the critical slope angle for use in InSAR analysis. Keywords: InSAR, SAR interferometriy, airborne SAR, orbital error, nonparallel orbit, critical baseline Classification: Antennas and Propagation

References

[1] P.A. Rosen, S. Hensley, I.R. Joughin, F.K. Li, S.N. Madsen, E. Rodriguez, and R.M. Goldstein, “Synthetic aperture radar interferometry,” Proc. IEEE, vol. 88, no. 3, pp. 333–382, March 2000. DOI: 10.1109/5.838084 [2] R.F. Hanssen, Radar Interferometry: Data Interpretation and Error Analysis, Kluwer Academic, 2001. DOI: 10.1007/0-306-47633-9 [3] C.V. Jakowatz, D.E. Wahl, P.H. Eichel, D.C. Ghiglia, and P.A. Thompson, Spotlight-Mode Synthetic Aperture Radar: A Signal Processing Approach, Springer Science & Business Media, 1996. DOI: 10.1007/978-1-4613-1333-5 [4] M. Soumekh, Synthetic Aperture Radar Signal Processing with MATLAB Al- gorithms, Wiley-Interscience, April 1999. [5] I.G. Cumming and F.H. Wong, Digital Processing of Synthetic Aperture Radar Data Algorithms and Implementation, Artech House remote sensing library, © IEICE 2020 Artech House, 2005. DOI: 10.1587/comex.2020COL0013 Received June 10, 2020 Accepted June 17, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

586 IEICE Communications Express, Vol.9, No.12, 586–592

[6] M. Gocho, H. Yamada, Y. Yamaguchi, and R. Sato, “Study on phase calibration for TomoSAR by using at reference surfaces with known altitude,” Progress in Electromagnetics Research Symposium (PIERS), Aug. 2017. [7] M. Gocho, H. Yamada, M. Arii, S. Kojima, R. Sato, and Y. Yamaguchi, “3- D imaging using SAR tomography with Pi-SAR2-X dataset,” IEICE Trans. Commun., vol. E101-B, no. 2, pp. 409–417, 2018. DOI: 10.1587/transcom. 2017isp0020 [8] J.J. Mohr and J.P. Merryman Boncori, “An error prediction framework for interferometric SAR data,” IEEE Trans. Geosci. Remote Sens., vol. 46, no. 6, pp. 1600–1613, June 2008. DOI: 10.1109/TGRS.2008.916213 [9] G. Liu, R.F. Hanssen, H. Guo, H. Yue, and Z. Perski, “Nonlinear model for InSAR baseline error,” IEEE Trans. Geosci. Remote Sens., vol. 54, no. 9, pp. 5341–5351, 2016. DOI: 10.1109/TGRS.2016.2561305 [10] X. Chen, J. Peng, and H. Yang, “Orbital error modeling and analysis of spaceborne InSAR,” IEEE 2017 SAR in Big Data Era: Models, Methods and Applications (BIGSARDATA), pp. 1–6, 2017. DOI: 10.1109/ BIGSARDATA.2017.8124930

1 Introduction InSAR (interferometric synthetic aperture radar) [1, 2] analysis, where the phase difference between two SAR [3, 4, 5] observation data is measured, has been used for terrain height monitoring, and ground deformation detection, for example. Since an interferogram consists of some phases corresponding to the desired geometric information and the flat-earth phase generated by the observation baseline, preprocessing to remove the flat-earth phase is necessary in InSAR analysis. The flat-earth phase is a function of the baseline parameters and a reference surface, so it can be removed by calculating the baseline using flight path information. Therefore, accurate orbit information is required for the removal process. In this report, we show the critical condition of the slope angle between orbits in InSAR analysis. To simplify the problem in InSAR, we often assume that two orbits are perfectly parallel. Since the baseline is independent of the flight direction under the parallel assumption, the flat-earth phase is a function of only the slant range and it can simplify the analysis. This assumption is not always valid for repeat-pass InSAR using airborne SAR, because guaranteeing parallelism between the orbits is more difficult than in single-pass InSAR or spaceborne SAR. In fact, we confirmed the fluctuation of flat-earth phase residuals fluctuated along the azimuth direction [6, 7], but it cannot be explained under the parallel assumption. Clarifying the mathematical model of the phase residuals is important for orbital error estimation, and phase residuals have been studied in satellite InSAR [8, 9, 10]. However, the slope of the orbit has not been focused on yet, and the limitation of this assumption has not been clarified. The purpose of this report is to formulate the nonparallel component of repeat-pass InSAR using an airborne SAR system and clarify the limitation of its parallel assumption.

587 IEICE Communications Express, Vol.9, No.12, 586–592

2 Repeat-pass InSAR Let us consider a SAR dataset observed by a linear orbit whose altitude from the ground surface is H, and define a cylindrical coordinate system whose cylindrical axis is the observation orbit of the SAR data, as shown in Fig. 1(a). Under this definition, any point in the coordinate system can be represented as (u,r, θ) with azimuth u, slant range r, and elevation angle θ. In this letter, this SAR dataset and the corresponding orbit/coordinate system are called the master dataset and the master orbit/coordinate system, respectively. We assume that we have another SAR data whose observation orbit is approximately parallel to the master orbit, where the data and orbit are called the slave data/orbit, respectively. Let us define four parameters to represent the slave orbit as follows (also see Fig. 1):

1. B: The distance between master/slave orbits at the slow-time center t = 0.

588 IEICE Communications Express, Vol.9, No.12, 586–592

2. α: The elevation angle of the slave orbit at t = 0.

3. b: The velocity of the slave orbit projected to the polar plane. If b = 0, then the two observation orbits are perfectly parallel.

4. β: The movement direction on the polar plane. If β = 0, then the movement is vertical; if β = 90◦, then the movement is horizontal.

These parameters represent two straight orbits gradually approaching (or leaving) each other. Note that what these parameters represent is the relative movement of the slave orbit normalized by that of the master orbit. In addition, we deﬁne the angle δ between the master and slave orbits; it can be expressed as δ = b, tan v (1) where v is the platform velocity. Also, azimuth u and slow-time t satisfy the linear = v ∈ [−T , T ] relationship u t, t 2 2 , where T is the time width of the slow-time. In this letter, we do not discuss dependence of the nonlinear error on the slow- time t, or other factors such as acceleration, vibration, or rotation. The propagation distance of the slave orbit ρ(u,r, θ) observing a scatterer (u,r, θ) can be expressed as

ρ(u,r, θ|t) = √ (r cos θ − B cos α − bt cos β)2 + (r sin θ − B sin α − bt sin β)2 ≈ r − B cos(θ − α) − bt cos(θ − β) ≈ r − B cos(θ − α) − uδ cos(θ − β), (2)

u where B, bt ≪ r, and bt = b v = u tan δ ≈ uδ. In this situation, the interferometric phase ϕ(u,r, θ) can be expressed as

ϕ(u,r, θ) = 2k{ρ(u,r, θ) − r} ≈ −2kB cos(θ − α) − 2kuδ cos(θ − β)

= ϕ∥(θ) + ϕ∦(u, θ) (3)

ϕ∥(θ) = −2kB cos(θ − α) (4)

ϕ∦(u, θ) = −2kuδ cos(θ − β), (5)

where 2π 2π f k = = c (6) λ c is the spatial frequency corresponding to the wavelength λ and center frequency fc. Equations (4) and (5) are the parallel and nonparallel components, respectively, and the interferometric phase ϕ(u,r, θ) is their sum. Obviously, if δ → 0, then tan δ → 0 and ϕ∦(u, θ) → 0.

3 Discussion 3.1 Computer simulation © IEICE 2020 DOI: 10.1587/comex.2020COL0013 We show some examples of ϕ∦(u, θ) in Fig. 2. From these results, we can see that Received June 10, 2020 Accepted June 17, 2020 a small δ generates a phase shift over one cycle. If we use an InSAR dataset that is Publicized June 30, 2020 Copyedited December 1, 2020

589 IEICE Communications Express, Vol.9, No.12, 586–592

Fig. 2. Examples of Eq. (5), where azimuth u ∈ [−1000 m, 1000 m], elevation angle θ ∈ [30°,65°], and β = 0. The title of each subimage (δ, fc) denotes the orbit slope angle (δ) and the center frequency ( fc) of SAR systems.

not suﬃciently parallel for interferometric analysis, such a phase residual may occur in the analysis results. Therefore, we should not use the parallel assumption unless the dataset satisﬁes the condition described below.

3.2 Parallelism condition (a): perfect parallel Obviously, ϕ∦(u, θ) can be ignored under the condition

max|ϕ∦(u, θ)| = kD|δ cos(θ − β)| ≪ 2π, λ © IEICE 2020 → |δ cos(θ − β)| = |δ∥ | ≪ , (7) DOI: 10.1587/comex.2020COL0013 D Received June 10, 2020 Accepted June 17, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

590 IEICE Communications Express, Vol.9, No.12, 586–592

where D = umax − umin is the azimuth image width and δ∥ = δ cos(θ − β) is the component of δ parallel to the line of sight. This condition indicates that even a small angle of the wavelength λ relative to the azimuth image width D cannot be ignored. Since the azimuth width D of airborne SAR is several kilometers whereas the wavelength λ is about a few centimeters to 1 m, there is a diﬀerence of at least three orders of magnitude.

3.3 Parallelism condition (b): azimuth direction The second condition is derived by the sampling theorem along the azimuth direction as follows:

∂ϕ∦ 2π max ≤ , (8) ∂u ∆u

where ∆u is the sampling interval along the azimuth direction, which is equal to the azimuth pixel size. As long as the condition Eq. (8) is satisﬁed, the vibration of ϕ∦(u, θ) along the azimuth direction is observable and can be corrected. By substituting Eq. (5) into Eq. (8), we can obtain the second condition

∂ϕ∦ = 2kδ cos(θ − β), ∂u λ → |δ∥ | ≤ . (9) 2∆u λ The ratio ∆u is greater than 0.1 in most SAR systems, as shown in Fig. 3(a), and 0.1 rad ≈ 6°, so the limit of δ∥ is about 3°. Therefore, satisfying the condition Eq. (9) is much easier than satisfying the condition Eq. (7).

3.4 Parallelism condition (c): slant range direction The third condition is derived by the sampling theorem for the ﬂat-earth surface as follows:

∂ϕ∦ 2π max ≤ , (10) ∂rref ∆r where H r = (11) ref cos θ is a dependent slant range r constrained on the reference surface and ∆r is the slant range sampling interval of the SAR system. By solving Eq. (10) for δ, Eq. (12) can be obtained: ∂ϕ δ (θ − β) ∦ = 2ku sin , ∂rref rref tan θ λ rref λ H tan θ → |δ⊥| ≤ tan θ = , (12) ∆r D ∆r D cos θ δ = δ (θ − β) δ λ where ⊥ sin is the perpendicular component of . In Eq. (12), ∆r is . tan θ . generally about 0 1, as shown in Fig. 3(b), and the function cos θ is greater than 1 0 for almost all θ, as shown in Fig. 3(c). Therefore, the condition Eq. (12) is not severe © IEICE 2020 DOI: 10.1587/comex.2020COL0013 as long as we choose a suitable D less than or equal to H. Received June 10, 2020 Accepted June 17, 2020 Publicized June 30, 2020 Copyedited December 1, 2020

591 IEICE Communications Express, Vol.9, No.12, 586–592

λ Fig. 3. The markers in (a) and (b) denotes the ratio ∆u and λ ∆r of some SAR systems respectively, and the dashed λ = . λ = . lines are ∆u 0 1 and ∆r 0 1.

4 Conclusion In this report, we clarified the limitation of the parallel condition of repeat-pass InSAR using nonparallel orbits. The first condition Eq. (7) indicates that even a small slope of one wavelength cannot be approximated as parallel. Therefore, we should consider the nonparallel component of the interferometric phase for repeat-pass InSAR analysis unless the orbits of the data are perfectly parallel. The second and third conditions, Eq. (9) and Eq. (12), show the limitation of the critical slope angle to observe and correct the nonparallel component of the flat-earth phase. The nonparallel component ϕ∦ will be observable and correctable only if the conditions, Eq. (9) and Eq. (12), are satisfied.

592 IEICE Communications Express, Vol.9, No.12, 593–598 Special Cluster in Conjunction with IEICE General Conference 2020 Experimental investigation of communication quality degradation of 1000BASE-T1 by pulse disturbance

Yusuke Yano1, a), Tohlu Matsushima2, b), and Osami Wada1, c) 1 Graduate School of Engineering, Kyoto University, Kyoto Daigaku Katsura, Nishikyo-ku, Kyoto 615–8510, Japan 2 Graduate School of Engineering, Kyushu Institute of Technology, 1–1 Sensuicho, Tobata-ku, Kitakyushu-shi Fukuoka 804–8550, Japan a) [email protected] b) [email protected] c) [email protected]

Abstract: Adverse effect of pulse disturbances on the communication quality is investigated experimentally on in-vehicle Ethernet (1000BASE-T1, IEEE 802.3bp). Common- and differential-mode pulse disturbances were injected into communication cables by a method referring to the coupling network described in IEC 62228-5. As a result, it was revealed that steep change of differential signal due to pulse disturbance degrades the communication quality at a specific pulse width. Keywords: in-vehicle ethernet, 1000BASE-T1, immunity, pulse disturbance Classification: Electromagnetic Compatibility (EMC)

References

[1] ISO 7637-2 ed. 3 (2011), Road vehicles – Electrical disturbances from conduction and coupling – Part 2: Electrical transient conduction along supply lines only, March 2011. DOI: 10.3403/30174646u [2] ISO 7637-3 ed. 3 (2016), Road vehicles – Electrical disturbances from conduction and coupling – Part 3: Electrical transient transmission by capaci- tive and inductive coupling via lines other than supply lines, July 2016. DOI: 10.3403/30139628u [3] S. Matsushima, T. Matsushima, T. Hisakado, and O. Wada, “Experimental evaluation of communication quality of ethernet in relation to parameters of pulse disturbances,” IEICE Commun. Express, vol. 6, no. 12, pp. 639–644, Sep. 2017. DOI: 10.1587/comex.2017XBL0116 [4] IEEE Std. 802.3bp-2016, Amendment 4: Physical Layer Speciﬁcations and Management Parameters for 1 Gb/s Operation over a Single Twisted-Pair Copper Cable, Sep. 2016. DOI: 10.1109/IEEESTD.2016.7564011 [5] IEC 62228-5 ED1, 47A/1093/CD, Integrated circuits – EMC evaluation of transceivers – Part 5: Ethernet transceivers, Feb. 2020. © IEICE 2020 DOI: 10.1587/comex.2020COL0019 Received June 23, 2020 Accepted June 26, 2020 Publicized July 9, 2020 Copyedited December 1, 2020

593 IEICE Communications Express, Vol.9, No.12, 593–598

1 Introduction Recently, in-vehicle Ethernet is attracting attention as a high-speed and large-capacity communication method, and communication systems from 10 Mbps to multi-gigabit are being discussed for practical use. Since automotive networks require high- reliability and safety, immunity testing is essential. However, conventional automotive electromagnetic compatibility (EMC) test standards do not support high-speed communication systems. The transient disturbance and the system-level test methods defined by ISO 7637- 2 [1] and ISO 7637-3 [2] do not enough cover possible disturbances that should be considered in recent automotive environment. The purpose of this research is to clarify the dominant characteristics of disturbances which degrade communication quality of in-vehicle Ethernet, and to propose an appropriate EMC evaluation method for it. In the previous study [3], as a first step, we focused on 100BASE-TX which is commonly used as a consumer Ethernet and investigated what degrades its communication quality. As a result, experiments demonstrated that pulse disturbances with durations of about 1-symbol transmission time of the signal affected the communication quality the most. In this paper, as the second step, we focus on an in-vehicle Ethernet (1000BASE- T1 [4]) and investigate the adverse effect of pulse disturbances on the communication quality and the appropriate test method.

2 Pulse disturbance injection For in-vehicle Ethernet, a single unshielded twisted pair (UTP) cable is used for a communication line to reduce its weight. Disturbances couple to the cable as common mode, and it is converted to differential mode due to the imbalance of cables, connectors, electronic components, and board wiring. The differential-mode disturbance causes communication quality degradation. In this investigation, we adopt an injection method using the coupling network (CN) described in IEC 62228-5 [5] to simulate both the common- and differential- mode disturbances. As a parameter of the pulse disturbance, we change the pulse width at which the peculiarity result has obtained in [3]. The degradation of the communication quality is evaluated using frame error rates (FERs).

2.1 Immunity test system Figure 1 shows an overview of the immunity test system used in this study. Two communication modules are connected by two UTP cables through the CN. Each of the modules is connected to the ports of an Ethernet tester (MT1000A, Anritsu) by using optical ﬁber cables (i.e., 1000BASE-SX). Ethernet frames are transmitted from one of the tester port and are received by the other port through the 1000BASE- T1 communication system. Pulse disturbances are generated by a pulse generator (M8195A, Keysight) and are injected into the communication system via the CN. The modules, the cables, and the CN are placed at 50 mm in height over the system ground. © IEICE 2020 To investigate the dependence on the cable length l, two lengths of the cable, DOI: 10.1587/comex.2020COL0019 Received June 23, 2020 3.0 m and 6.4 m for each, are used. Accepted June 26, 2020 Publicized July 9, 2020 Copyedited December 1, 2020

594 IEICE Communications Express, Vol.9, No.12, 593–598

Normally, in [5], communication errors due to differential-mode disturbance, which is caused by slight mode conversion of −40 dB or less, are evaluated; however, in this study, the amount of mode conversion is set to larger to build a test system without an expensive RF amplifier. The resistance and capacitance values of the CN are changed from [5]. The values of resistance are 0 and 300 Ω, and the values of capacitance are 4.7 nF. In this case, the amount of mode conversion at the CN is −16 dB. The impedance inserted into the differential line increases from 240 Ω [5] to 300 Ω; as a result, the insertion loss in the frequency band of 1000BASE-T1 communication (375 MHz) decreases from 1.62 dB to 1.31 dB.

Fig. 1. Overview of immunity test system.

2.2 Measurement of FER Each of the communication modules is set to master- and slave-mode, respectively. The FERs on the master and slave side are individually measured by the Ethernet tester. The preamble length is 8 bytes, the Ethernet frame length is 100 bytes, the interframe gap is 12 bytes, and the test period for each FER measurement is 30 s. Since the communication speed of both 1000BASE-SX and 1000BASE-T1 is 1 Gbps (125 MBps), a total of 3.125 × 107 frames are transmitted and received in 30 s; therefore, the PER of about 10−7 can be evaluated. The transmission period of the frame is 960 ns, and the repetition rate of the pulse disturbance (as stated later) is 1000 ns, the disturbance thus collides about once per frame.

2.3 Pulse width and communication quality The parameters of the pulse generator are set as follows: the pulse repetition period

T = 1000 ns, the peak-to-peak amplitude of the source voltage Vs = 250 mV, and the © IEICE 2020 DOI: 10.1587/comex.2020COL0019 rise and fall times tr = tf = 0.7 ns. The pulse width tp is a variable. Received June 23, 2020 Accepted June 26, 2020 Publicized July 9, 2020 Copyedited December 1, 2020

595 IEICE Communications Express, Vol.9, No.12, 593–598

Figure 2 shows the relationships between tp and the FER. In this result on 1000BASE-T1, the FER changes depending on the pulse width as a similar change in 100BASE-TX [3]; moreover, the dependency on the cable length is also seen. As seen in the region “A”, the FER increases by more than two orders of mag-

nitude at a speciﬁc tp, and in the region “B”, the FER becomes almost constant regardless of tp. However, the FER increases with tp longer than 1-symbol time of 1000BASE-T1 (1.33 ns). This is diﬀerent from the experimental result of 100BASE- TX [3] (i.e., the result that the pulse disturbance with the duration of about 1-symbol time of the signal degraded the communication quality most).

Lengthening l shifts “A” in the direction of increasing tp. When l are 3.0 m and 6.4 m, tp of 25 and 50 ns degrade the FER most, respectively. Besides, the overall increase in the FER depending on l is due to the decrease in diﬀerential signal level with the long cables. The causes of the pulse width dependence and the cable length dependence are clariﬁed in the next subsection.

Fig. 2. Pulse width and communication quality.

2.4 Causes of pulse width and cable length dependence

The common-mode voltage Vcom and the differential-mode voltage Vdiff shown in Fig. 3 are observed at different positions shown in Fig. 1. Vcom and Vdiff are measured at the input connector of the CN and the on-board connector of the module, respectively. We conclude that the drastic increase in the region “A” is causes by the steep

change of Vdiff as seen in Fig. 3(b), and the convergence to constant values in the FER in the region “B” is due to the cancellation of Vdiff caused by common-mode reflection and mode conversion as seen in Fig. 3(f). The causes are described below. © IEICE 2020 DOI: 10.1587/comex.2020COL0019 According to the pulse width, a peculiar change is seen in Vdiff due to the Received June 23, 2020 Accepted June 26, 2020 common-mode reflected wave and the mode conversion. Figures 3(a)–(c) show the Publicized July 9, 2020 Copyedited December 1, 2020

596 IEICE Communications Express, Vol.9, No.12, 593–598

waveforms when l is 3.0 m.

Figure 3(a) shows the waveforms when tp is 10 ns, and a reﬂected wave in Vcom is observed from the time of 23 ns, which corresponds to the round-trip time of the

common-mode disturbance propagating along the 3.0 m cable. For Vdiff, the first rising edge is observed at 16 ns after Vcom is injected. The time 16 ns corresponds to the one-way propagation time of the differential mode from the injection point of disturbance to the end of the cable at 3.0 m. From 39 ns, the second rising edge of

Vdiff is observed in the negative direction, which corresponds to the reflected Vcom reaching the injection point and mode-converted to differential mode.

Figure 3(b) shows the waveforms when tp is set to 25 ns. The falling edge of the first Vdiff and negative rising edge of the second Vdiff almost coincide; consequently, the variation of disturbance waveform on Vdiff becomes steep. When tp is 25 ns, the FER increases drastically as shown in the region “A”, so we speculate that pulse disturbances with a steep change degrades the communication quality.

Figure 3(c) is for tp = 50 ns. Vdiff keeps nearly zero from 39 ns until 66 ns. This is because the first and second Vdiff cancel each other. When tp is longer than the pulse width which generates the steep change, only the cancellation period is

changed even if tp is changed; in other words, the degradation of the communication quality occurs equally in terms of the pulse collision probability. We thus infer that the convergence to constant values seen in the region “B” is due to the cancellation. Figures 3(d)–(f) show the voltages when l is 6.4 m. The longer l increases the

propagation delay. Consequently, when the tp is around 50 ns, the steep change of

597 IEICE Communications Express, Vol.9, No.12, 593–598

Vdiﬀ occurs as seen in Fig. 2(e). When the tp is longer than 50 ns, the cancellation occurs as seen in Fig. 2(f).

3 Conclusion We investigated experimentally the relationship between the communication quality of 1000BASE-T1 and the pulse width of the disturbance. As a result, FER increased by more than two orders of magnitude with a specific pulse width. It was clarified that the differential-mode pulse disturbance changes steeply at the specific pulse width due to the common-mode reflection and the mode conversion. Therefore, we concluded that the pulse disturbance with the steep change significantly degrades the communication quality.

Acknowledgments This work is supported by Japan Automotive Software Platform and Architecture (JASPAR).

598 IEICE Communications Express, Vol.9, No.12, 599–604 Special Cluster in Conjunction with IEICE General Conference 2020 Initial veriﬁcation of a bandwidth tunable Ku-band power ampliﬁer designed by the HySIC concept

Satoshi Yoshida1, a), Kenjiro Nishikawa1, and Shigeo Kawasaki2 1 Electrical and Electronics Engineering Course, Graduate School of Science and Engineering, Kagoshima University, 1–21–40 Korimoto Kagoshima 890–0065 Japan 2 Institute of Space and Astronautical Science (ISAS), Japan Aerospace Exploration Agency (JAXA), 3–1–1 Yoshinodai, Chu-oh, Sagamihara 252–5210, Japan a) [email protected]

Abstract: In this letter, an initial verification results of a bandwidth tunable power amplifier in Ku-band designed by the hybrid semiconductor integrated circuit (HySIC) concept is reported. A GaAs monolithic microwave integrated circuit (MMIC) and a Si radio frequency integrated circuit (RFIC) were utilized as the HySIC configuration in the circuit design. For the purpose of initial confirmation of this design validity, the GaAs and Si chips were fabricated and packaged onto the copper tungsten plate with gold plating. As measured results, the lower cut-off frequency was changed from 9.8 GHz to 11.9 GHz, resulting in the bandwidth tunability in Ku-band. Keywords: bandwidth tuning, amplifier, varactor, Ku-band, integrated circuit Classification: Wireless Communication Technologies

References

[1] K. Sakamoto and Y. Itoh, “L-band SiGe HBT frequency-tunable dual-bandpass or dual-bandstop differential amplifiers using varactor-loaded series and parallel LC resonators,” IEICE Trans. Electron., vol. E95-C, no. 12, pp. 1839–1845, Dec. 2012. DOI: 10.1587/transele.E95.C.1839 [2] R. Malmqvist, A. Gustafsson, M. Alfredsson, and A. Ouacha, “A tunable active MMIC filter for on-chip X-band radar receiver front-ends,” 2002 IEEE MTT-S International Microwave Symposium Digest (Cat. No.02CH37278), Seattle, WA, USA, vol. 3, pp. 1907–1910, 2002. DOI: 10.1109/MWSYM.2002.1012236 [3] B. Park, S. Choi, and S. Hong, “A low-noise amplifier with tunable interference rejection for 3.1- to 10.6-GHz UWB systems,” IEEE Microw. Wireless Componen. Lett., vol. 20, no. 1, pp. 40–42, Jan. 2010. DOI: 10.1109/LMWC. 2009.2035963 [4] K. Yamanaka, K. Sugaya, Y. Horiie, T. Yamaguchi, N. Tanahashi, and Y. Itoh, © IEICE 2020 “A Ku-band frequency-tunable active matched feedback MMIC amplifier us- DOI: 10.1587/comex.2020COL0007 Received May 15, 2020 ing variable-capacitance elements,” 2000 IEEE MTT-S International Microwave Accepted June 22, 2020 Publicized July 9, 2020 Copyedited December 1, 2020

599 IEICE Communications Express, Vol.9, No.12, 599–604

Symposium Digest (Cat. No.00CH37017), Boston, MA, USA, vol. 3, pp. 1895– 1898, 2000. DOI: 10.1109/MWSYM.2000.862352 [5] H. Uchida, K. Ogura, Y. Konishi, and S. Makino, “A frequency-tunable amplifier with a simple tunable admittance inverter,” 2005 European Microwave Conference, Paris, pp. 1–4, Oct. 2005. DOI: 10.1109/EUMC.2005.1608854 [6] S. Kawasaki and A. Miyachi, “The dawn of the new RF-HySIC semiconductor integrated circuits: an initiative for hybrid ICs consisting of Si and compound semiconductors,” IEICE Trans. Electron., vol. E99–C, no. 10, pp. 1085–1093, Oct. 2016. DOI: 10.1587/transele.E99.C.1085 [7] T. Shimura, S. Maki, S. Fujiwara, K. Fujii, Y.Takahashi, S. Suzuki, M. Miyashita, K. Yamamoto, H. Seki, M. Hieda, and Y. Hirano, “A multiband power amplifier using combination of CMOS and GaAs technologies for WCDMA hand- sets,” 2014 IEEE Radio Frequency Integrated Circuits Symposium, Tampa, FL, pp. 141–144, 2014. DOI: 10.1109/RFIC.2014.6851680 [8] D. Leipold, W. Allen, P. Sheehy, and G. Hau, “A WCDMA 41% power efficiency direct DC coupled hybrid CMOS/GaAs power amplifier with pre-distortion linearization,” 2012 IEEE Radio Frequency Integrated Circuits Symposium, Montreal, QC, pp. 279–282, 2012. DOI: 10.1109/RFIC.2012.6242281 [9] M. Horberg, T. Emanuelsson, P. Liganderi, H. Zirath, and D. Kuylenstiema, “An X-band varactor-tuned cavity oscillator,” 2017 IEEE MTT-S International Microwave Symposium (IMS), Honololu, HI, pp. 1938–1941, 2017. DOI: 10.1109/MWSYM.2017.8059041

1 Introduction Nowadays, wireless communication system has become one of the indispensable infrastructures. Analog circuit technology contributes in the hardware realization such as base stations and terminals. Focusing on a power amplifier circuit that operates in high frequency band, frequency tunable one has been developed for several decades. L-band frequency tunable amplifier using the varactor [1], X- band frequency tunable low noise amplifier (LNA) using tunable filter [2], tunable LNA for 3.1 to 10 GHz ultra wide band system [3] are examples for the frequency tunable amplifiers. Frequency tunable amplifiers in more higher frequency band at Ku-band [4, 5] have been proposed. On the other hand, packaging technique is also important for fabrication and manufacturing. The hybrid semiconductor integrated circuit (HySIC) concept [6] is one of the attractive technique for integration. We focus on the HySIC concept which utilizes GaAs monolithic microwave integrated circuit (MMIC) and a Si radio frequency integrated circuit (RFIC) for a design of a bandwidth tunable Ku-band power amplifier. Conventional proposals of the combination of GaAs MMIC and Si RFIC [7, 8] were utilized for different power range circuit, for example, low power driver amplifier for Si RFIC and medium or high power amplifier for GaAs MMIC. However, we utilize Si RFIC for frequency tuning function and GaAs MMIC for main circuit of the power amplifier. This paper presents a validity of bandwidth tunability of a Ku-band single stage power amplifier designed by the HySIC concept. For the tuning capability, © IEICE 2020 DOI: 10.1587/comex.2020COL0007 a varactor [1, 9] is used in the output matching circuit. Section 2 presents a basic Received May 15, 2020 Accepted June 22, 2020 circuit design of the proposed amplifier, Section 3 introduces and discusses the Publicized July 9, 2020 Copyedited December 1, 2020

600 IEICE Communications Express, Vol.9, No.12, 599–604

measurement results, and Section 4 presents our conclusions.

2 Circuit design and analysis Figure 1(a) shows a basic circuit diagram of the proposed bandwidth tunable Ku- band power amplifier. A GaAs MMIC by UMS PH25 process and a Si RFIC by TSMC 0.18 µm CMOS process were utilized as the HySIC integration in the tunable amplifier design. A simple circuit of a single stage common source amplifier was designed in the GaAs MMIC. For the bandwidth tuning operation, a part of the output matching circuit was designed and placed in the Si RFIC. For the tuning operation, a MOS varactor was utilized in the Si RFIC. In this report, only the varactor, a capacitor, and a RF choke inductor were placed on the Si RFIC as an initial trial, but some digital-assisted control circuit for optimum control function (control circuit in Fig. 1(a)) should be integrated at the next phase in near future. This is one of the reasons why Si RFIC was utilized in the initial design. Circuit

parameters of C1, C2, C3, C4, C5, C6, L1, L2, L3, L4, L5, L6, L7, L8, R1, R2 were 1.6 pF, 1.6 pF, 1.4 pF, 1.6 pF, 3.6 pF, 1.6 pF, 6.0 nH, 0.3 nH, 0.3 nH, 0.3 nH, 0.4 nH, 0.4 nH, 0.1 nH, 9.0 nH, 150 ohm, 5 ohm, respectively. Figure 1(b) shows photos of the proposed ampliﬁer under measurement using on-wafer probe. Measurement results will be discussed in the next section. At ﬁrst, for the initial design, bandwidth tuning capability was evaluated by S-parameter and input-output characteristic using an ideal capacitor instead of the

MOS varactor model Cv1. Figure 2(a) shows a simulation result of |S21| using the proposed ampliﬁer circuit. The frequency characteristics were dynamically changed

around 10 GHz by the capacitance of Cv1 tuning from 0.85 pF to 1.83 pF, while the characteristics change in upper frequency bands (over 13 GHz) were negligibly small. From this simulation result, we expect the tuning range of lower cut-oﬀ frequency from 10.1 GHz to 12.1 GHz, resulting in the bandwidth tunability. In this

© IEICE 2020 Fig. 1. A basic circuit diagram of the proposed bandwidth DOI: 10.1587/comex.2020COL0007 Received May 15, 2020 tunable Ku-band power ampliﬁer. Accepted June 22, 2020 Publicized July 9, 2020 Copyedited December 1, 2020

601 IEICE Communications Express, Vol.9, No.12, 599–604

Fig. 2. Summary of the simulation results.

report, the cut-off frequency was defined as the frequency when the |S21| becomes lower than 10 dB. Based on the previous initial simulation results, a varactor was designed. We selected a MOS varactor in the Si RFIC instead of a varactor diode in GaAs MMIC, because capacitance variable range of the PH25 process was limited from 80 fF to 300 fF, resulting in out of the desired capacitance tuning range. This was another reason why the part of the output matching circuit was designed in the Si RFIC. Structural parameters of the MOS varactor were determined based on the simulation results shown in Fig. 2(a). Finalized values of Finger number, group number, width per finger, and finger length were 46, 6, 1 µm, and 1 µm, respectively. Simulation results of the relationship between bias voltage and varactor capacitance were summarized in the Fig. 2(b). The capacitance was able to be changed from 0.77 pF to 1.84 pF while the bias voltage was changed from −3 V to 3 V. The capacitance range from 0.85 pF to 1.83 pF is realized by changing the bias voltage from −1 V to 1 V.

Figure 2(c) shows a |S21| simulation result of the proposed power amplifier circuit using the varactor model. Obtained results have same trend as the initial simulation result, shown in Fig. 2(a). The frequency tuning range of the lower cut-off frequency was from 10.2 GHz to 12.2 GHz, as for the Fig. 2(c). Figure 2(d) shows the simulation results of the input-output characteristics of the proposed power amplifier circuit using the varactor model. We simulated at © IEICE 2020 DOI: 10.1587/comex.2020COL0007 10.0 GHz, 12.0 GHz, 14.0 GHz, and 19.5 GHz, however, only to be shown one result Received May 15, 2020 Accepted June 22, 2020 Publicized July 9, 2020 Copyedited December 1, 2020

602 IEICE Communications Express, Vol.9, No.12, 599–604

at 12.0 GHz because the variation of the characteristic is simply recognized from the

|S21| simulation results. The simulation results at Vdctr = 0 V and 1 V have same trend, but little bit high gain and power added eﬃciency (PAE) as for the result of

Vdctr = 1 V. And the result at Vdctr = −1 V have degraded characteristic compared with other simulation results. This discussion was easily expected from the |S21| simulation results at 12.0 GHz, shown in the Fig. 2(c). Other simulation results at 14.0 GHz and 19.5 GHz have little variation with around 30 % maximum PAE

when the Vdctr was changed. On the other hand, at 10.0 GHz, the simulation results greatly moved by the Vdctr change. Summarizing the Fig. 2(d), maximum value of the PAE was 28.5 % when Vdctr = 1 V under the condition of 6.0 dBm input power. 1 dB compression point P1dB was 13.2 dBm. These S-parameters and harmonic balance simulations were performed by Advanced Design System (ADS) software of Keysight Technologies.

3 Fabrication and measurement The proposed circuits of GaAs MMIC and Si RFIC were fabricated using foundry service provided by UMS PH25 and TSMC 0.18 µm processes, respectively. Both two chips of the MMIC and the RFIC were mounted onto CuW (copper tungsten) plate with gold plating shown in Fig. 1(b). Silver paste was used to ﬁx these two chips onto the CuW plate. Ground-signal-ground pads with pitch of 150 µm were used for input and output interconnection terminals of the two chips. Wedge type gold wire bonding technique was used for the interconnection between two chips. Also, this wire bonding technique was utilized for another interconnection. Ideal thickness of the GaAs MMIC and the Si RFIC were 100 µm and 300 µm, respectively. Therefore, 200-µm-thickness CuW subcarrier with 1 mm square was used under the GaAs MMIC for height adjustment. This was because level distance should be minimized for the short gold wire bonding between the two chips. The Air Coplanar Probe (ACP) by FormFactor was used for the RF input and output. S-parameters and input-output characteristics were measured to evaluate the bandwidth tunability function in Ku-band. Bias conditions of the GaAs MMIC were Vg = −0.3 V and

Vd = 3.0 V as same as the simulation. Bias voltage Vdctr dependence on small signal S-parameters was measured using a vector network analyzer. Figure 3(a) shows the measured result of |S21|. As

603 IEICE Communications Express, Vol.9, No.12, 599–604

expected in the simulation, frequency characteristics changes depending on Vdctr , and same trend that the cut-off frequency gradually raises while Vdctr reduces, was observed. The tuning range of the lower cut-off frequency was from 9.8 GHz to 11.9 GHz. Also, we have confirmed same trend as expected as the initial simulation result shown in Fig. 2(c).

Bias voltage Vdctr dependence on input-output characteristics was measured using a signal generator and a spectrum analyzer. Figure 3(b) shows the measured

results. At ﬁrst, we have conﬁrmed the gain under the condition of Vdctr = 1 V at small signal region was higher than other bias conditions, as same as the simulation.

Maximum value of the PAE was 30.3 % when Vdctr = 1 V under the condition of 10.8 dBm input power. 1 dB compression point P1dB was 13.2 dBm. Also, both simulation and measurement results have same trend, therefore, we have confirmed the validity of the simulation. From these simulation and measurement results, we have confirmed the bandwidth tunable capability at lower frequency region in Ku-band by using the proposed power amplifier circuit.

4 Conclusion Applicability of the HySIC concept to a Ku-band bandwidth tunable power amplifier was tested in this letter. A GaAs MMIC and a Si RFIC were utilized for the amplifier circuit. As measured results, frequency tunable range of the lower cut-off frequency from 9.8 GHz to 11.9 GHz was measured. Maximum value of the PAE was 30.3 %

when Vdctr = 1 V under the condition of 10.8 dBm input power. 1 dB compression point P1dB was 13.2 dBm. From the simulation and measurement results, we have conﬁrmed the validity of the proposed concept as an initial proposal for future digital-assisted bandwidth tunability of a Ku-band power ampliﬁer.

Acknowledgment This work was supported in part by JSPS KAKENHI Grant Number JP16K18106 and I–O data foundation. Part of this work was carried out under the Cooperative Research Project Program of the Research Institute of Electrical Communication, Tohoku University.

604 IEICE Communications Express, Vol.9, No.12, 605–609 Special Cluster in Conjunction with IEICE General Conference 2020 Estimating music listener’s emotion from bio-signals by using CNN

Nanami Tanizawa1, a), Mutsumi Suganuma2, and Wataru Kameyama2 1 Department of Computer Science and Communications Engineering, Graduate School of Science and Engineering, Waseda University, 3–4–1 Okubo, Shinjuku-ku, Tokyo, 169–8555 Japan 2 Faculty of Science and Engineering, Waseda University, 3–4–1 Okubo, Shinjuku-ku, Tokyo, 169–8555 Japan a) [email protected]

Abstract: The purpose of this paper is to estimate emotions for music pieces with lyrics. We investigate whether four emotions (happy, sad, angry and relaxed) can be accurately estimated by obtained bio-signals of subjects during music listening by analyzing them with convolutional neural network. Questionnaire responses by subjects are used as the correct labels, and the four emotions are classified by each and combined. The results of the analysis show that the accuracies of the both classification methods highly exceed the chance level. It suggests the possibility of emotion estimation for music listeners by bio-signals. Keywords: bio-signal, emotion estimation, machine learning, CNN Classification: Multimedia Systems for Communication

References

[1] Y. Pan, C. Guan, J. Yu, K.K. Ang, and T.E. Chan, “Common frequency pattern for music preference identiﬁcation using frontal EEG,” 2013 6th International IEEE/EMBS Conference on Neural Engineering (NER), San Diego, CA, USA, pp. 505–508, Nov. 2013. DOI: 10.1109/NER.2013.6695982. [2] B. Gingras, M.M. Marin, E. Puig-Waldmüller, and W.T. Fitch, “The eye is listening: music-induced arousal and individual diﬀerences predict pupillary responses,” Frontiers Human Neuroscience, vol. 9, 619, 2015. DOI: 10.3389/ fnhum.2015.00619, 2015.

1 Introduction In recent years, streaming services like Apple Music are commonly used for music listening. Because a huge number of music pieces are distributed in these services, © IEICE 2020 it is diﬃcult for users to ﬁnd the music pieces that they want to listen to. Therefore, DOI: 10.1587/comex.2020COL0026 Received June 29, 2020 music recommendation system is used to solve this problem. The recommendation Accepted July 15, 2020 Publicized July 28, 2020 Copyedited December 1, 2020

605 IEICE Communications Express, Vol.9, No.12, 605–609

systems in current use recommend music pieces based on user’s listening history with associated metadata. However, we presume that recommendation in this way cannot always reflect user’s preference because the systems don’t know exactly that how the user feel about the recommended one. Therefore, we assume that if the emotion and the impression to music pieces can be estimated from bio-signals, we can recommend music pieces more accurately reflecting user’s preference by using the results as additional metadata. In some previous studies, likes and dislikes of music pieces and emotions of music pieces by using bio-signals are well classified. For example, [1] reports, by analyzing the electroencephalographs (EEG) of two channels with SVM, likes and dislikes of music pieces are able to be estimated with the average accuracy of 74.77%. Therefore, it can be inferred that it is possible to estimate the music listener’s preferences by using bio-signals. However, such the rough classifications, i.e. only like and dislike, is not sufficient to generate rich metadata for music pieces. Motivated by above, the purpose of this paper is to investigate whether it is possible to estimate emotions to music pieces in detail by bio-signals.

2 Experiment Bio-signals are collected from subjects while listening to music through three devices: an EEG of EMOTIV EPOC+ with 14 channels of AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8 and AF4 in the international 10-20 system with 128 [Hz] sampling, a heartbeat sensor of Polar H10 with V800, and an eye-tracker of Tobii X60 with 60 [Hz] sampling. The eye-tracker is used even for music listening in the experiment because [2] reports that pupil diameter changes even in listening to music reﬂecting human emotion change such as arousal, tension, pleasantness and familiarity. There were 10 subjects in the experiment including three males and seven fe- males, where the average age is 21.3 and the standard deviation is 0.64. In the experiment, 40 music pieces were used, where 20 of them were brought in by the subjects, and the other 20 were prepared by the experimenters. All the music pieces were with lyrics in Japanese, and the listening time for each music piece was 60 seconds. The 60-second parts of the music pieces were selected by the experimenters to include the most characteristic parts of them. After listening to each, the subjects were asked to rate it by an 11-point scale on each of four emotions: happy, sad, angry, and relaxed. The subjects wore the EEG, the heartbeat sensor and the noise-canceling ear- phone during the experiment. To measure the pupil diameter, the subjects were asked to look at the center of the gray screen on the PC monitor while listening to the music pieces. First, the subjects listened to a silent track for 60 seconds in order to obtain bio-signals under the normal condition. Then, they listened to each of the 40 music pieces one by one in random order. In between the playback of each music piece, they were given time to answer the questionnaire. And they were also given break

606 IEICE Communications Express, Vol.9, No.12, 605–609

3 Data used for analysis Of the 60-second data measured, the data for 58 seconds excluding the first and last 1 second are used in the analysis, and the data with loss of more than 1 second are excluded from the analysis data. Therefore, the length of the data is different from music piece to music piece. For the EEG, the data are transformed by Fast Fourier Transform with a 1- second window and a 0.25-second slide. Then, alpha wave (8–13 [Hz]), beta wave (14–30 [Hz]) and gamma wave (31–45 [Hz]) are calculated for each channel. Then beta/alpha and gamma/alpha are calculated for each channel. In addition, using the symmetry of the electrode positions, the left-right ratio of the corresponding channels is calculated about alpha wave and beta wave. From the measured heartbeats, first, the RRI (R-wave and R-wave Interval) is calculated, then change rate of RRI is calculated by dividing the RRI per second by the average RRI per music piece. After that, linear interpolation is performed every 0.25 seconds. For the pupil diameter, first, the data in 0.5 seconds before and after blinks are excluded as they are still in blinks. Then, linear interpolation is performed, and the mean of the left and right pupil diameters is calculated. Finally, the average values with a 1-second window and a 0.25-second slide are calculated. The pupillary light reflex compensation is not applied because subjects watch the gray screen of no brightness change. The questionnaire responses are used as the correct labels. They are labeled as positive (6 to 11) or negative (1 to 5) for each emotion in the 11-point scale. In addition to this labeling, the questionnaire responses are also categorized into the combinations of the four emotions (like “happy and sad” or “happy and sad and relaxed”). In this case, the maximum number of the classifications is 16 including the case of no-raised emotions, but the number of classifications differs from subject to subject because there are no answered emotions in the questionnaire responded by some subjects. Therefore, the chance level for estimation differs from subject to subject, as well. The preprocessed EEG, RRI and pupil diameter data are used for the analysis as the explanatory variables, while the questionnaire responses are used as the explained variables. Table I summarizes them.

4 Analysis method Convolutional Neural Network (CNN) is used to analyze the data. We implement the model using Keras and Tensorflow as the backend. The network configuration of the CNN used for the analysis is shown in Fig. 1. First, 1D-Conv with kernel = 2, stride = 1 and filter = 64 is applied to the data of 58 dimensions by the step of 32. Then, 1D-Conv with kernel = 2, stride = 2 and filter = 128 is applied to the data of 64 dimensions by the step of 31. The output is flattened, and three fully connected (FC) layers are appled. For the CNN parameters, the optimization function is Adam, the loss function is categorical_crossentropy, the activation function is ReLU, the batch size is 28, the number of epochs is 50, and the learning rate is 10−3. The input dataset is normalized by appyling z-score for each dimension. 80 percent of the total © IEICE 2020 data in random selection are used for the training, and the remaining 20 percent are DOI: 10.1587/comex.2020COL0026 Received June 29, 2020 used for the test data. Accepted July 15, 2020 Publicized July 28, 2020 Copyedited December 1, 2020

607 IEICE Communications Express, Vol.9, No.12, 605–609

Table I. The explanatory variables and the explained variable in analysis

608 IEICE Communications Express, Vol.9, No.12, 605–609

5 Result and discussion Table II shows the results of the classiﬁcation accuracy for all the subjects.

Table II. Results of the classiﬁcation accuracy

The accuracies are above the chance levels by 0.5–0.8. The training losses of the combination classifications converged well, while those of the binary classifications do not so well. This may be caused by the same network structure applied for both classifications (labeled as positive or negative for each emotion and labeled as combinations of the four emotions), which means it may be suitable only for the combination classifications. We assume that a simpler network structure would be better for the binary classification.

6 Conclusion In this paper, we estimate user’s emotions while listening to music pieces with lyrics from bio-signals in order to realize a music recommendation system that reflects user’s preference as a final goal. The results suggest that the four emotions of “happy”, “sad”, “angry” and “relaxed” can be accurately estimated from bio-signals of EEG, RRI and pupil diameter by using CNN. However, because the numbers of music pieces and subjects are limited, it is not confirmed whether the proposed model is general enough for other music pieces and subjects. Therefore, as for the future study, we will collect more data from more music pieces and subjects to ensure the results. In addition, we will reconsider the labeling method and optimize the network parameters.

609 IEICE Communications Express, Vol.9, No.12, 610–615 Special Cluster in Conjunction with IEICE General Conference 2020 Effects of miners’ location on blocks selection in blockchain

Kosuke Toda1, a), Naomi Kuze1, b), and Toshimitsu Ushio1, c) 1 Graduate School of Engineering Science, Osaka University, 1–3 Machikaneyama-cho, Toyonaka, Osaka 560–8531, Japan a) [email protected] b) [email protected] c) [email protected]

Abstract: Blockchain is a distributed ledger technology for recording transactions. To guarantee the immutability of the blocks, miners need a lot of computational resources for generating blocks (mining). Since mining is conducted distributedly by many miners, chain forks can occur in the blockchain. In this paper, we investigate the eﬀects of miners’ locations, the size of the mining pool (MP), and the network structure on the selection rate of blocks generated by the MP when chain forks occur through simulations. Keywords: blockchain, fork, mining pool, regular graph, scale-free network Classiﬁcation: Network

References

[1] S. Nakamoto, “Bitcoin: a peer-to-peer electronic cash system,” http://bitcoin. org/bitcoin.pdf, 2008. [2] Y. Xiao, N. Zhang, W. Lou, and Y.T. Hou, “Modeling the impact of network connectivity on consensus security of Proof-of-Work blockchain,” arXiv preprint, arXiv:2002.08912, 2020. [3] D.J. Watts and S.H. Strogatz, “Collective dynamics of ‘small-world’ networks,” Nature, vol. 393, no. 6684, pp. 440–442, 1998. DOI: 10.1038/30918 [4] A.L. Barabáshi and R. Albert, “Emergence of scaling in random networks,” Sci- ence, vol. 286, no. 5439, pp. 509–512, 1999. DOI: 10.1126/science.286.5439.509

1 Introduction Blockchain is a distributed ledger technology for recording transactions and underlies the digital currency such as Bitcoin [1]. In a blockchain-based service, users called miners distributedly generate blocks consisting of several transactions. These blocks are stored and managed as a chain of blocks (a blockchain). Miners are connected as a network, and the generated blocks are propagated hop by hop. To guarantee the immutability of the blocks, Proof-of-Work (PoW) is used as a consensus algorithm. In this algorithm, miners need a lot of computational resources for generating blocks (mining), which results in the resistance to block tampers. A lot of miners try to © IEICE 2020 DOI: 10.1587/comex.2020COL0012 generate blocks, and a miner that succeeds in generating a block get a reward. The Received June 9, 2020 Accepted July 7, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

610 IEICE Communications Express, Vol.9, No.12, 610–615

more resources the miner has, the easier the block generation is to succeed. This results in an incentive for miners to have many resources. In order to get rewards efficiently, several miners form a mining pool (hereinafter referred to as an MP) for cooperating in mining. Since mining is conducted distributedly by many miners, chain forks can occur in the blockchain. When a chain fork occurs, each miner needs to select a highest block as a parent block for a new block, which complicates the consensus of the blockchain. However, properties of chain forks have not been clarified. In this paper, we investigate the effects of miners’ locations, the size of the MP, and the network structure on the selection rate of blocks generated by the MP when chain forks occur through simulations.

2 Blockchain forks 2.1 Problem statement We focus on chain forks. Since new blocks are propagated hop by hop on the network, we clarify how the occurrence of chain forks and the blocks selection are related to the network structure and the location of miners belonging to the MP. Xiao, et al. [2] studied the eﬀects of network connectivity on consensus security such as chain fork rate and mining reward. We investigate the eﬀects of the location of miners on the occurrence of chain forks and the blocks selection.

2.2 A blockchain model We model the network of miners (miners’ network) by the undirected graph G = (V, E). V is a ﬁnite set of nodes and each node corresponds to a miner. E is a set of edges, indicating that there is a direct data communication between miners. Let N be the number of miners, i.e. |V | = N. Moreover, we assume that there is only one MP and P ⊆ V be the set of miners belonging to the MP. Miner m < P performs mining independently. Miners belonging to the MP cooperate with each other by sharing the information of blocks generation of other miners belonging to the MP.

We define the information βk about the kth block Bk generated in the miners’ network as βk = (nk, pk, dk, wk), where nk is the identifier of Bk, pk is the identifier of the parent block of Bk, dk is the block height of the block Bk, and wk ∈ V is the miner that generated block Bk. Block B0 represents a genesis block. A chain consisting of a group of blocks held by miner m is called a local chain in miner m. m m = Let tk be the time when miner m received block Bk. In this paper, t0 0 for all m. β m The block Bk held by the miner m is represented by the pair of k and tk , that is, (β , m) ℓ k tk . If the miner m holds blocks in addition to the genesis block at time t, the t,m m set of blocks held by miner m is represented as C = {(β0,0),...,(βℓ,tℓ )}. When there are multiple blocks with the same hight, a chain fork occurs. That ′ is, for different blocks Bk and Bk′ (k , k ) held by miner m, a chain fork occurs t,m t,m when dk = dk′ . Let H be the set of (β, τ) ∈ C such that β = (n, p, d, w) satisfies d = max{dk | 0 ≤ k ≤ ℓ}. We consider a discrete-time model of the block generation. Miners share their local chains per unit time through local interactions. We assume that each miner © IEICE 2020 DOI: 10.1587/comex.2020COL0012 m ∈ V performs the following behavior at time t ∈ T = {0,1,...}: Received June 9, 2020 Accepted July 7, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

611 IEICE Communications Express, Vol.9, No.12, 610–615

1. If miner m receives a block that it does not hold from adjacent nodes, add that block to its local chain.

2. Miner m selects one block contained in Ht,m as a parent block and conducts mining for generating a new block. The block selection is described in detail below.

< P (β , m) ∈ t,m m = { m | If m , m selects a block p tp H that satisfies tp min tk (β , m) ∈ t,m} ∈ P (β , m) ∈ t,m w ∈ P k tk H . If m , m selects a block p tp H that satisfies p (β , m) w ∈ P if it exists. If there is no block p tp that satisfies p , m selects a block in the same way as m < P. In this paper, we assume that it takes 1 unit time for a miner m to transmit a block to its adjacent miners. Let f be the probability that a block is generated within a unit time. Assuming that all miners have the same computational resources, the probability that each miner generates a block within a unit time is given by f /N. We consider the following two locations of miners belonging to the MP. One is the location in which a subgraph whose nodes belong to P is connected (Location 1). The other is the location in which no edge in V belongs to P × P (Location 2). Here, it is assumed that both Location 1 and 2 include a node with the maximum degree (hereinafter referred to as the hub node).

3 Simulation and evaluation In this section, we conduct simulation evaluations for investigating the eﬀects of miners’ location in the MP on the blocks selection.

3.1 Simulation settings 3.1.1 Settings We consider a network modeled by both a regular graph where every vertex has the same degree and a graph whose degree distribution follows a power-law (scale-free network). The former graph is generated by the Watts-Strogatz model (WS model) WS(N, k, p) with p = 0 [3]. The latter graph is generated by the Barabási-Albert

model (BA model) BA(N, m0) [4]. Note that N is the number of nodes, k is the mean degree, p is randomness, and m0 is the initial number of nodes. We consider the regular graph N = 100 and k = 6,8,10 and the BA model N = 500 and

m0 = 1,2,3,4,5. |P|/N = 0.1, f = 0.02, and T = 10,000 in this simulation.

3.1.2 Metrics Among the blocks included in the longest chain of a certain miner m obtained at the end of each simulation, the percentage of the blocks generated by the miners belonging to the MP is used as the index (I). Let X and Y be the samples of I obtained in Location 1 and 2, respectively. 2 2 Assuming that X and Y follow the Gaussian distributions N(µ1, σ ), N(µ2, σ ) with the same variance, conduct the following t-test. Here, when the variance σ2 is unknown, we test the following two-sided hypoth- α = . © IEICE 2020 esis at the signiﬁcance level 0 05. DOI: 10.1587/comex.2020COL0012 Received June 9, 2020 Accepted July 7, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

612 IEICE Communications Express, Vol.9, No.12, 610–615

  H0 : µ1 = µ2 (Null hypothesis), (1)  µ , µ ( ). H1 : 1 2 Alternative hypothesis

The null hypothesis H0 indicates that there is no significant difference in the block selection by the location, and the alternative hypothesis H1 indicates that there is a significant difference in the blocks selection by the location. The simulation is performed n times for each of the Location 1 and 2. Let the sample mean of X and Y be X¯ and Y¯, and the estimator of variance σ2 be σˆ 2. Using two-sided α points with 2n − 2 degrees of freedom, the rejection range W for the significance level α in the t-test |X¯ − Y¯ | T = √ (2) σ 2 ˆ n is W = {T ≥ 1.96} for sufficiently large n. We set n = 100. In other words, the t-value obtained by Eq. (2) satisfies T ≥ 1.96, it can be said that there is a significant difference in the blocks selection by the location.

3.2 Results and discussions The results of the t-test are shown in Table I. With the regular graphs, there is no signiﬁcant diﬀerence by the location regardless of the degree. With the BA models,

there is a significant difference by the location when m0 = 1, but the others do not. First, we focus on the reason why there is no significant difference in the regular graph. In the regular graphs, all nodes have the same degree. The results show that the location of miners belonging to the MP has no effect on the selected block when chain forks occur if all nodes in the network have the same degree. Next, we focus on the relationship between the average length in the BA model and the ratio of the occurrence of chain forks. Let ⟨l⟩ be the average length of

the network and ⟨lh⟩ be the average length from the hub node to other nodes. For m0 = 1,2,3,4,5, we obtain a pair (⟨l⟩, ⟨lh⟩) = (6.45,4.49),(3.73.2.21),(3.27,2.17), (2.93,1.89),(2.77,1.95), respectively. We evaluate the ratio of the occurrence of chain forks over all generated blocks, denoted by F [%]. Shown in Fig. 1 is the relationship between the average length ⟨l⟩ and the ratio F. The horizontal axis

Table I. The results of the simulation and t-test. Regular graph k X¯ Y¯ σˆ 2 T 6 9.81 10.09 6.26 7.91 × 10−1 8 10.12 9.98 5.01 4.56 × 10−1 10 10.36 9.98 4.88 1.24 BA model 2 m0 X¯ Y¯ σˆ T 1 10.30 9.47 5.61 2.47 2 10.34 9.97 4.56 1.22 3 9.71 9.62 4.47 3.04 × 10−1 − © IEICE 2020 4 10.11 9.96 4.97 4.67 × 10 1 DOI: 10.1587/comex.2020COL0012 − Received June 9, 2020 5 9.74 9.53 4.55 6.71 × 10 1 Accepted July 7, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

613 IEICE Communications Express, Vol.9, No.12, 610–615

Fig. 1. The relationship between the average length ⟨l⟩ and the ratio of chain forks F.

indicates the average length and the vertical axis indicates the ratio of the occurrence of chain forks. The plotted points are the average value of F in 100 simulations, and the error bar indicates the 2σ conﬁdence bound. The results show that ⟨l⟩ and F are in a linear relationship. Finally, we focus on the relationship among the average length in the BA model, the location of miners belonging to the MP, and the ratio at which blocks generated by miners belonging to the MP are selected when chain forks occur. We classify the chain forks into four types. The ﬁrst one is chain forks “PP” such that the highest block of each branch of the fork is generated by miners belonging to the MP. The second one is chain forks “II” such that the highest block of each branch of the fork is generated by miners who does not belong to the MP. The third one is chain forks

“PIP” where there exists the highest block of a branch generated by a miner who does not belong to the MP while the selected highest block is generated by a miner

who belongs to the MP. The fourth one is chain forks “PII” where there exists the highest block of a branch generated by a miner who belongs to the MP while the selected highest block is generated by a miner who does not belong to the MP. Let n(A) be the number of the occurrence A. We then deﬁne the following index I and perform the t-test as in Section 3.1.2.

© IEICE 2020 n(PP) + n(PI ) DOI: 10.1587/comex.2020COL0012 I = P . (3) Received June 9, 2020 ( ) + ( ) + ( ) Accepted July 7, 2020 n PP n PIP n PII Publicized August 4, 2020 Copyedited December 1, 2020

614 IEICE Communications Express, Vol.9, No.12, 610–615

Table II. t-test about the fork. 2 (N, m0) X¯ Y¯ σˆ T (500,1) 0.570 0.406 0.090 3.87 (500,2) 0.535 0.550 0.144 2.82 × 10−1 (500,3) 0.453 0.428 0.157 4.40 × 10−1 (500,4) 0.498 0.400 0.170 1.68 (500,5) 0.403 0.401 0.175 5.07 × 10−2

The results of the t-test are shown in Table II. There is a signiﬁcant diﬀerence

by the location when m0 = 1, but the others do not. When m0 = 1, ⟨l⟩ and ⟨lh⟩ are large, and the ratio of selected blocks generated by miners belonging to MP in

Location 2 is signiﬁcantly lower than that in Location 1. When m0 = 2,3,4,5, ⟨l⟩ and ⟨lh⟩ are small, and the distance between the miners belonging to the MP in Location 2 is also small. Furthermore, Location 1 and 2 make little diﬀerence in the ratio of selection of the blocks generated by the miners belonging to the MP when the chain forks occur.

We summarize that in networks with large ⟨l⟩ and ⟨lh⟩, the ratio of the occurrence of chain forks is large, the ratio of selected blocks depends on the location of miners belonging to MP, and a block generated by a miner with a small distance from the hub node belonging to MP is likely to be selected.

4 Conclusion We have clariﬁed the relationship among the average length of the network, the location of miners belonging to the MP, the ratio of the occurrence of chain forks, and the ratio of selected blocks when chain forks occur. Our future work is to investigate the relationship between the miners’ locations and attacks such as selﬁsh mining.

Acknowledgments This research was supported by Grant-in-Aid for Early-Career Scientists No. 18034267 from the Japan Society for the Promotion of Science (JSPS).

615 IEICE Communications Express, Vol.9, No.12, 616–621 Special Cluster in Conjunction with IEICE General Conference 2020 Distributed topic management in publish-process-subscribe systems on edge-servers for real-time notiﬁcation service

Tomoya Tanaka1, a), Tomio Kamada1, and Chikara Ohta2 1 Graduate School of System Informatics, Kobe University, Rokkodai-cho, Nada-ku, Kobe, Hyogo 657–8501, Japan 2 Graduate School of Science, Technology and Innovation, Rokkodai-cho, Nada-ku, Kobe, Hyogo 657–8501, Japan a) tomoya.tanaka@ﬁne.cs.kobe-u.ac.jp

Abstract: The importance of real-time and data-driven notification has been growing for social services and Intelligent Transporting System (ITS). As an advanced version of Pub/Sub systems, publish-process-subscribe systems with MEC (Multi-access Edge Computing), where published messages are spooled and processed on edge servers, have been proposed. In this paper, we present a topic-based publish-process-subscribe system that allows a topic to be managed on multiple edge servers so that messages are processed near publishers and transferred to subscribers immediately. However, managing each topic on numerous edge servers can cause exhaustion of storage resources on edge servers. We introduce a simple topic allocation method on edge servers to discuss the problem. Experiments show the feasibility of our proposed system. Keywords: real-time notification, data-driven notification, multi-access edge computing, publish-subscribe model Classification: Network

References

[1] A. Renznik, L.M.C. Murillo, Y. Fang, W. Featherstone, M. Filippou, F. Fontes, F. Giust, Q, Huang, A. Li, C. Turyagyenda, D. Wehner, and Z. Zheng, “Cloud RAN and MEC: a perfect paring,” ETSI White Paper, https://www.etsi.org/images/ files/ETSIWhitePapers/etsi_wp23_MEC_and_CRAN_ed1_FINAL.pdf, accessed May 29, 2020. [2] S.A. Shaheen and R. Finson, “Intelligent transportation systems,” in Reference Module in Earth Systems and Environmental Sciences, Elsevier, 2013. DOI: 10.1016/B978-0-12-409548-9.01108-8 [3] V. Setty, G. Kreitz, G. Urdaneta, R. Vitenberg, and M. van Steen, “Maximizing the number of satisfied subscribers in pub/sub systems under capacity constraints,” Proc. IEEE INFOCOM 2014, Toronto, ON, Canada, pp. 2580–2588, April/May 2014. DOI: 10.1109/INFOCOM.2014.6848205 © IEICE 2020 [4] F. Zhang, B. Jin, W. Zhuo, Z. Wang, and L. Zhang, “A content-based pub- DOI: 10.1587/comex.2020COL0040 Received June 30, 2020 lish/subscribe system for efficient event notification over vehicular ad hoc net- Accepted July 15, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

616 IEICE Communications Express, Vol.9, No.12, 616–621

works,” Proc. 9th International Conference on Ubiquitous Intelligence and Com- puting and 9th International Conference on Autonomic and Trusted Computing, Fukuoka, Japan, pp. 64–71, Sept. 2012. DOI: 10.1109/UIC-ATC.2012.21 [5] B. Krishnamachari and K. Wright, “The publish-process-subscribe paradigm for the internet of things,” USC ANRG Technical Report, http://anrg.usc.edu/www/ wp-content/uploads/2017/07/ANRG_TechReport_ 201704_PublishProcess SubscribeForIoT.pdf, accessed May 29, 2020. [6] S. Khare, H. Sun, K. Zhang, J. Gascon-Samson, A. Gokhale, X. Koutsoukos, and H. Abdelaziz, “Scalable edge computing for low latency data dissemina- tion in topic-based publish/subscribe,” Proc. 2018 IEEE/ACM Symposium on Edge Computing (SEC), Seattle, WA, USA, pp. 214–227, Oct. 2018. DOI: 10.1109/SEC.2018.00023 [7] R. Kawaguchi and M. Bandai, “Edge based MQTT broker architecture for geographical IoT applications,” Proc. 2020 International Conference on Informa- tion Networking (ICOIN), Barcelona, Spain, pp. 232–235, March 2020. DOI: 10.1109/ICOIN48656.2020.9016528 [8] J. Kreps, N. Narkhede, and J. Rao, “Kafka: A distributed messaging system for log processing,” Proc. 6th International Workshop on Network- ing Meets Databases (NetDB 2011), Athens, Greece, pp. 1–7, June 2011, http://notes.stephenholiday.com/Kafka.pdf, accessed May June 26, 2020.

1 Introduction In recent years, the importance of real-time and data-driven notification has been growing dramatically for applications such as social services, IoT applications, and Intelligent Transporting System (ITS) [1]. For example, real-time decision-making services proposed in ITS are expected to react immediately to changes in traffic conditions, analyze the current conditions, and provide optimal vehicle behavior [2]. Multi-access Edge Computing (MEC) and Pub/Sub messaging models have been exploited to achieve instant reactions to changing conditions and immediate data- driven notifications to clients [3, 4]. In a more sophisticated version, the publish- process-subscribe model with MEC, where edge servers process messages published from clients followed by disseminating the generated notification to subscribers, have great potential to achieve real-time and data-driven notification [5]. We present a topic-based publish-process-subscribe system which allows a topic to be managed on multiple edge servers, though many researchers assume one server to manage a topic [6, 7]. In our system, publishers/subscribers can exchange messages immediately using nearby edge servers. Besides, our system duplicates messages for a topic on multiple servers so that notification can be generated locally using all messages published to the topic, which can include messages that first received by remote edge servers. Kafka also stores messages duplicated to enable fault-tolerance, but pay little attention to the geographical locality [8]. However, duplicating messages on numerous edge servers causes exhaustion of the storage capacity because the available storage capacity on edge servers is limited. In this paper, we discuss the feasibility of our publish-process-subscribe system, investigating the trade-offs between the network path length needed for notifications and the © IEICE 2020 DOI: 10.1587/comex.2020COL0040 consumption of storage capacity. Received June 30, 2020 Accepted July 15, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

617 IEICE Communications Express, Vol.9, No.12, 616–621

2 System overview In this section, we describe our publish-process-subscribe system, where a topic is managed on multiple edge servers. Figure 1(a) shows the overview of our system. Each publisher is assigned to an edge server that manages all topics the client publishes to. For example, Client B publishes messages to Topic P and Q, so the assigned server of Client B manages both of the topics. Besides, each subscriber is assigned to an edge server, from which they receive messages. We call the server assigned to each client the “home server”. Note that messages are spooled and processed only on publishers’ home servers, and subscribers’ home servers do not manage the subscribed topic such as Edge C in Fig 1.

Fig. 1. Overview of our publish-process-subscribe system

Each edge server functions as a broker, which receives messages from publishers and delivers notifications to subscribers. A message processor, which functions as a message spooler and an analyzer of the spooled messages, is prepared for each managed topic. In Fig. 1(a), Edge B manages Topic P and Q, so Edge B prepares the message processor for each topic. Our system generates notifications and delivers them to subscribers in the steps shown in Fig. 1(b). When an edge server receives a message from a publisher, it spools messages in the message processor prepared for the topic. Next, the message processor analyzes the spooled messages and generates a notification. Finally, the generated notification is disseminated to subscribers’ © IEICE 2020 DOI: 10.1587/comex.2020COL0040 home servers and delivered to subscribers. Received June 30, 2020 Accepted July 15, 2020 Managing a topic on multiple edge servers can enable publishers/subscribers Publicized August 4, 2020 Copyedited December 1, 2020

618 IEICE Communications Express, Vol.9, No.12, 616–621

to access nearby edge servers. However, it increases the storage consumption and may cause exhaustion of the storage capacity. We introduce a simple home server assignment method that divides the considered field into clusters and prepares only one server which manages a topic in a cluster. In each cluster, publishers to the same topic share spooled messages for the topic, so that the consumption of storage capacity is reduced. The proximity between clients and their home server, as well as the usage of storage capacity on edge servers, are strongly influenced by the number of clusters. Let us consider the proximity and the storage capacity usage, changing the number of clusters under the condition that the number of servers is 16. There are 11 clients with a weak locality in the 20 km square field. Six of the clients publish to Topic P, while the others publish to Topic Q. Figure 2(a) shows a home server assignment when the number of clusters K is 1, which means that each topic is managed only one server in the field. On the other hand, Fig. 2(b) shows a home server assignment when K = 16, which means that numerous edge servers manage one topic. In Fig. 2(a), the distance between clients and their home servers can be long, whereas the usage of storage capacity is minimized. In Fig. 2(b), the storage consumption is higher than the case of Fig. 2(a), whereas the distance between clients and their home server is minimized. Our objective is to discuss the feasibility of topic allocations shown in Fig. 2(b) considering limited storage capacity on edge servers. In the experiment in Section 4, we increase the number of clusters and measures the increase of storage consumption when topics are managed by numerous edge servers. We first formulate storage consumption in the next section.

Fig. 2. Geographical view of home server assignment to clients

3 Problem formulation

We consider a ﬁeld with L edge servers denoted S = {s1, s2, ··· , sL }. There is M active clients that include both publishers and subscribers, denoted C =

{c1, c2, ··· , cM }, in the ﬁeld. Each client cm ∈ C has a home server sl ∈ S. © IEICE 2020 DOI: 10.1587/comex.2020COL0040 Let Cl denote the set of clients whose home server is sl. The home server assign- Received June 30, 2020 Accepted July 15, 2020 ment, which determined the topic allocation on edge servers, is denoted by a binary Publicized August 4, 2020 Copyedited December 1, 2020

619 IEICE Communications Express, Vol.9, No.12, 616–621

matrix X = {xm,l |cm ∈ C, sl ∈ S}. If cm’s home server is sl (i.e. cm ∈ Cl), xm,l = 1, whereas xm,l = 0 in the other cases. We formulate storage consumption on edge servers as well as the proximity

between publishers/subscribers and their home server. Let ul denote the total size of spooled messages managed by the edge server sl. As for the proximity between home servers and clients, let gm denote the distance between a client cm and its home server. The value of gm and ul is determined by the assignment X. We increase the number of clusters in the assignment method described in Sec-

tion 2 and measures the influence on the proximity gm and the storage consumption ul. The assignment method is elaborated by the following. We divide a field into K clusters based on edge servers’ position using the K-means clustering algorithm. A client is assigned to a server belonging to the closest cluster. Dividing the field into

numerous clusters reduces the value of gm, but increase the value of ul.

4 Experiments We conduct simulations to investigate the increase of the edge server resource usage

ul as the field is divided into clusters finely, which reduces the value of gm. The assignment X, and the value of ul, gm would be different depending on the number of clusters K.

We adopt a metric Y1, and Y2, which denote the average of gm and the average of ul, respectively. We observe the value ofY1 andY2 changing the number of clusters K. In the simulation, we set the number of edge servers L as 100. As preparation for the simulations, we distribute 1865 clients who can take the publisher’s role following a Gaussian distribution with a standard deviation d centered on a randomly chosen point in 10 km square ﬁeld. We conduct simulations in the case of d = 0.1 km, d = 1 km and d = 10 km. We assign a home server to each client after dividing the ﬁeld into K clusters. We give 1 MB spooled messages in advance in all message processors. Therefore,

ul is calculated as the number of topics managed by edge server sl, multiplied by 1 MB. Figure 3(a) shows the distance between publishers and their home server against the number of clusters, while Fig. 3(b) shows the average storage capacity usage. We can observe that dividing the field into small pieces (i.e. K is close to L = 100) reduces the average distance between clients and their home server smaller than 1 km in the 10 km square field. As the value of K increases, the average storage capacity usage increases. But, the difference of used storage capacity between K = 1 to K = 100 is smaller than a certain value, even when the locality of clients are weak (i.e. d = 10). In the case of these simulations, the difference is less than 50 MB. If the extra preparation for the storage capacity is practicable, we can manage topics on numerous edge servers and achieve proximity between clients and their home servers.

5 Conclusion

© IEICE 2020 In this paper, we presented a publish-process-subscribe system that allows a topic DOI: 10.1587/comex.2020COL0040 Received June 30, 2020 to be managed on multiple edge servers so that information can be exchanged Accepted July 15, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

620 IEICE Communications Express, Vol.9, No.12, 616–621

Fig. 3. Trade-oﬀ measurements

immediately between nearby publishers/subscribers. We modeled the presented system. Numerical experiments demonstrated the feasiblility of our system, where each topic are managed on numerous edge servers. Future work involves constructing cooperative management method of the spooled messages between edge servers, so that edge servers which is not heavily used can spool messages instead of servers which is heavily used. Moreover, we will consider two-hierachical publish-process-subscribe system with edge servers and cloud servers.

621 IEICE Communications Express, Vol.9, No.12, 622–626 Special Cluster in Conjunction with IEICE General Conference 2020 Applying hard positive mining and its evaluation for person re-identiﬁcation

Yuki Hiroi1, a) and Wataru Kameyama2 1 Department of Computer Science and Communications Engineering, Graduate School of Fundamental Science and Engineering, Waseda University, 3–4–1 Okubo, Shinjuku-ku, Tokyo 169-8555, Japan 2 Faculty of Science and Engineering, Waseda University, 3–4–1 Okubo, Shinjuku-ku, Tokyo 169-8555, Japan a) [email protected]

Abstract: “Person Re-identification” technology has attracted more attention because of the proliferation of surveillance cameras and the increased awareness of crime prevention. However, there are various problems. One of them is that the same person is likely to be recognized erroneously as a different person due to the differences in persons’ appearance. Therefore, in this paper, we propose a method of hard positive mining to achieve better performance in identifying the same person images that are difficult to identify as the same person. The evaluation results using MARS dataset show that the CMS by applying the proposed hard positive mining is higher than the existing method of CNN with temporal pooling. It suggests that it is effective to apply hard positive mining in MARS dataset. Keywords: person re-identification, hard positive mining, hard sample mining, convolutional neural network Classification: Multimedia Systems for Communication

References

[1] K. Chen, Y. Chen, C. Han, N. Sang, C. Gao, and R. Wan, “Improving person re-identification by adaptive hard sample mining,” 2018 IEEE International Con- ference on Image Processing (ICIP), Athens, Greece, pp. 1638–1642, Oct. 2018. DOI: 10.1109/icip.2018.8451129 [2] H. Sheng, Y. Zheng, W. Ke, D. Yu, X. Cheng, W. Lyu, and Z. Xiong, “Mining hard samples globally and efficiently for person re-identification,” IEEE Internet Things J., March 2020. DOI: 10.1109/JIOT.2020.2980549 [3] Q. Qian, L. Shang, B. Sun, J. Hu, T. Tacoma, H. Li, and R. Jin, “SoftTriple loss: deep metric learning without triplet sampling,” 2019 IEEE International Con- ference on Computer Vision (ICCV), Seoul, Korea, pp. 6449–6457, Oct./Nov. 2019. DOI: 10.1109/iccv.2019.00655 [4] L. Zheng, Z. Bie, Y. Sun, J. Wang, C. Su, S. Wang, and Q. Tian, “Mars: a video benchmark for large-scale person re-identification,” European Conference on Computer Vision, vol. 9910, pp. 868–884, 2016. DOI: 10.1007/978-3-319- © IEICE 2020 46466-4_52 DOI: 10.1587/comex.2020COL0021 Received June 26, 2020 Accepted July 21, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

622 IEICE Communications Express, Vol.9, No.12, 622–626

[5] J. Gao and R. Nevatia, “Revisiting temporal modeling for video-based person ReID,” arXiv preprint, arXiv:1805.02104v2, 2018. [6] Z. Zhong, L. Zheng, D. Cao, and, S. Li, “Re-ranking person re-identiﬁcation with k-reciprocal encoding,” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, pp. 3652–3661, July 2017. DOI: 10.1109/cvpr.2017.389

1 Introduction In recent years, with the proliferation of surveillance cameras and the increased awareness of crime prevention, “Person Re-identification (ReID)” becomes one of the hot research issues, which is the task to identify the same person between non- overlapped cameras. It is useful for many different applications including crime prevention and marketing. A lot of studies have been done on ReID, but it’s still difficult to identify the same person because of persons having some item or not, different postures, and different view angles taken by different cameras. A sample image in which the same person is recognized erroneously as a different person is called a hard positive sample. On the other hand, a sample image that is likely to identify another person erroneously as the same person is called a hard negative sample. The hard positive samples and the hard negative samples are together called to hard samples in ReID. Some methods of hard sample mining which improve to recognize hard samples are proposed, such as [1] and [2]. [3] focuses on hard negative mining which improve to recognize hard negative samples by using the proposed triplet loss. However, the ever-proposed methods need to modify network models and loss functions, which make it difficult to include them into the existing network models. And the more the data is required for hard sample mining to be processed, the more the computational cost is required. Moreover, nowadays, since a huge size of persons’ images can be collected, it makes much larger diversity in images for a single person. Therefore, unless selecting good hard samples for training without considering the diversity, a non-essential model may be created to produce poor performance. Therefore, in this paper, we focus on hard positive samples to consider the diversity in images for one person, and propose a hard positive mining (hereafter referred to as HPM) method to achieve better performance in recognition of hard positive samples. The proposed method trains the networks with optimized learning- image orders and sets considering the hard positive samples, and it is independent from any network architectures and loss functions. It utilizes CMS (Cummulative Match Score), which is mainly used as the evaluation metrics in ReID, as the index for detecting hard positive samples. Using the CMS, time-series of images (hereafter referred to a clip) of the same person are ranked, and the clips of lower ranks that are taken as the detected hard positive samples are trained at the end of the epochs where the loss function is converging in order to reduce the computational cost and to prevent the overfitting. We compare the results using the MARS dataset [4] with

© IEICE 2020 [5] which achieves the highest performance in the dataset where CNN with temporal DOI: 10.1587/comex.2020COL0021 Received June 26, 2020 pooling are used without HPM. Accepted July 21, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

623 IEICE Communications Express, Vol.9, No.12, 622–626

2 Proposed method In [5], clips of the same class and different classes are trained per batch, where the batch size is B and the number of clips per class in each batch is N. Each batch contains B/N different classes that are randomly selected, and each class contains N clips. In addition, the number of clips per a class in dataset is M whose value varies depending on the class in the dataset, but the value of N is determined as a fixed one depending on, for example, memory and computational costs. And N clips of the same class to be included in a batch are randomly selected from the M clips. Due to this random selection, there is an issue to increase the possibility of recognizing the same class as a different class. In order to improve this, we propose an HPM method that positively trains the same person clips that are difficult to identify as the same class. In order to detect the hard positive samples, the CMS values are used to rank M clips of each class and to select N lower rank clips that are difficult to identify among M clips. In the proposed method, first, N and M are compared. If N is greater than or equal to M, classes to be in a batch are randomly selected, then clips in a selected class are also randomly selected, allowing duplications, as same in [5]. However, if N is less than M, first, classes are randomly selected as same in [5], but after that, M clips of each class are ranked by CMS. In general, CMS is used for the evaluation index of ReID, which is calculated from samples (images or clips) in various different classes, and those of all classes are ranked according to the distance of features from small to large for each sample in a class by CMS. However, in this proposal, CMS is used to detect clips including hard positive samples by calculating it from clips in the same class, where the clips in the same class are ranked according to the distance of features from small to large for each clip in a class by the calculated CMS. Then, the HPM index is calculated by Eq. (1).

N∑−1 = M−i( − )α HPK xK N i (1) i=0 ∈ Z| ∈ [ , − ] R Where HPK is the HPM index of clip K (K K 0 M 1 ), xK is the number of clip K in all the Rank R of the clips calculated by CMS, and α is a weight. After

calculating HPK , N clips are selected in the descending order of HPK for each class. As mentioned above, each batch contains B/N diﬀerent classes. Therefore, CMS and Eq. (1) are calculated B/N times per batch in the proposed method. Finally, all the selected B clips are put in a batch for training. Table I shows an example of the proposed HPM method, where a class is randomly selected, M = 5 (Clip 0 to Clip 4), N = 3 and α = 10. Because N < M,

each HPK has to be calculated. As shown in Table I, Clip 0 appears once in Rank 5, 5 = 4 = 3 = none in Rank 4 and twice in Rank 3. Therefore, x0 1, x0 0 and x0 2. Then, HP0 is calculated to 50 by Eq. (1). Similarly, HP1, HP2, HP3 and HP4 are 40, 60, 90 and 80, respectively. Because N= 3, i.e. three clips are to be selected, Clip 3,

Clip 4 and Clip 2 are selected in the descending order of HPK as they are to contain hard positive samples for training. © IEICE 2020 DOI: 10.1587/comex.2020COL0021 Received June 26, 2020 Accepted July 21, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

624 IEICE Communications Express, Vol.9, No.12, 622–626

Table I. An example of the proposed hard positive mining

3 Experiment In the experiment, we use MARS which is a large person image dataset taken by diﬀerent surveillance cameras. In MARS, there are about 1,260 persons and about 1 million images in total. The ratio of the training data and the test data used in this experiment is 5:5 of random selection. This experiment is to compare the baseline [5] applying the proposed HPM with the baseline of no HPM. CMS is used for the evaluation metrics. In addition, we apply Re-rank [6] to both which improves CMS [5]. Table II shows the parameters of this experiment. As shown in Table II, applying HPM is started from the 700th epoch in order to positively train the hard positive samples only at the end of epochs where the loss function is converging. In addition, the value of weight α is set to 10 from the results of our preliminary experiment.

Table II. Parameters of experiment

4 Results Table III shows the results in mAP (mean Average Precision) and CMS of ranks. As shown in Table III, the accuracies with the proposed HPM are higher than those © IEICE 2020 of the baseline except for Rank 10. In particular, in Rank 1, the accuracy of the DOI: 10.1587/comex.2020COL0021 Received June 26, 2020 proposed HPM is about 2.3 points higher than that of the baseline. Therefore, it is Accepted July 21, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

625 IEICE Communications Express, Vol.9, No.12, 622–626

Table III. Result of experiment

considered that the proposed HPM is eﬀective in MARS.

5 Conclusion In this paper, we propose an HPM method by using the CMS values in order to achieve better performance in recognizing the hard positive samples. The results suggest that it is effective to use the proposed HPM method in MARS. Since our experiment up to now has been done only with MARS, we will carry out a further experiment on another dataset to investigate whether the proposed HPM method is stably effective even in a different dataset. In addition, we will study to reduce the computational cost of the HPM indices much more.

626 IEICE Communications Express, Vol.9, No.12, 627–631 Special Cluster in Conjunction with IEICE General Conference 2020 Rate adaptation for robust data transmissions utilizing multi-AP reception and packet-level FEC

Ryota Bingo1 and Hiroyuki Yomo1, a) 1 Graduate School of Science and Engineering, Kansai University, 3–3–35, Yamate-cho, Suita, Osaka 564–8680, Japan a) [email protected]

Abstract: This letter proposes a rate control for robust data transmissions utilizing multiple access points (APs) combined with packet-level forward error correction (FEC). Beacon frames received from multiple APs are used to select an appropriate set of PHY and FEC rates to be employed for uplink transmissions, which is based on a machine-learning technique. With computer simulations, we show that the proposed rate control achieves the required reliability while signiﬁcantly reducing the occupancy period of the shared channel. Keywords: wireless LAN, broadcast, packet-level FEC, rate control Classiﬁcation: Wireless Communication Technologies

References

[1] K. Ikeda, K. Yamamoto, and H. Yomo, “Robust data transmissions utilizing multiple access points reception and packet-level FEC,” 2019 2nd World Sym- posium on Communication Engineering (WSCE), Dec. 2019. DOI: 10.1109/ WSCE49000.2019.9041039 [2] H.J. Zhu and D. Kidston, “The impact of link adaptation on Wiﬁ 802.11n,” 2016 IEEE International Conference on Networking, Architecture and Storage (NAS), Long Beach, CA, pp. 1–5, 2016. DOI: 10.1109/NAS.2016.7549424 [3] T. Matsuda, T. Noguchi, and T. Takine, “Broadcasting with randomized network coding in dense wireless ad hoc networks,” IEICE Trans. Commun., vol. E91-B, no. 10, pp. 3216–3225, Oct. 2008. DOI: 10.1093/ietcom/e91-b.10.3216 [4] A. Goldsmith, Wireless Communications, Cambridge Univ. Press, 2005. DOI: 10.1017/CBO9780511841224 [5] G. Pei and T.R. Henderson, “Validation of OFDM error rate model in ns-3,” Boeing Research Technology, Technical Report, pp. 1–15, 2010. [6] G.E. Hinton and R.R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,” Science, vol. 313, no. 5786, pp. 504–507, 2006. DOI: 10.1126/ science.1127647

1 Introduction © IEICE 2020 Improving robustness of wireless data transmissions is desired in many industrial DOI: 10.1587/comex.2020COL0038 Received June 30, 2020 ﬁelds, especially in factories with the advent of the concept of factory automation. A Accepted July 20, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

627 IEICE Communications Express, Vol.9, No.12, 627–631

common use-case is data transmissions (e.g., video data) from a station (STA) to an access point (AP) through wireless LAN (WLAN) interface, which are further forwarded to a central server. In general, unicast transmissions are employed between each STA and its connected AP. However, unicast transmissions have the problem of single point of failure: once a connected link suffers from communication failures for a long period of time due to shadowing, fading, and interference, quality of services (QoSs) of users become unacceptable. In order to solve this problem, we have proposed robust data transmissions exploiting broadcast (Multi-AP) transmissions combined with packet-level forward error correction (FEC) [1]. With our Multi-AP transmissions, each STA transmits data to multiple APs located inside its communication range with a single transmission by exploiting broadcast nature of wireless medium. Some APs can receive data transmitted by a STA even if the other APs failed. We also apply packet-level FEC in order to further enhance the reliability of data transmissions: a data packet lost at all surround APs can be successfully recovered thanks to redundant packets. In [1], we have confirmed with simulations and experiments that the proposed Multi-AP transmissions improve robustness in comparison to uni-cast transmissions. The important parameters to be controlled for Multi-AP transmissions are the physical layer (PHY) rate and FEC rates, which affect not only packet delivery rate (PDR) but also airtime (channel occupancy period) of the shared channel. In conventional unicast transmissions, these rates can be controlled with the approach of trial-and-error by using the link-level acknowledgement (ACK) or received signal strength indicator (RSSI) of a single target receiver [2]. However, the link-level ACK is not implemented for the broadcast mode of WLAN used in our Multi- AP transmissions. Furthermore, unlike unicast transmissions, there can be multiple points of receptions for broadcast transmissions, which requires us to consider RSSIs for multiple receivers. In this letter, we propose an adaptive rate control for Multi-AP broadcast transmissions. The novelty lies in the design of rate selector based on a machine-learning technique, which solely counts on the information obtained from beacon frames periodically transmitted by surrounding multiple APs.

2 System model We consider a scenario where a STA transmits data to a central server through multiple APs and gateway (GW) (see Fig. 1). The data transmissions between STA and APs are supported by IEEE 802.11 wireless LAN while APs, GW, and server are connected by wired connections, e.g., Ethernet. The STA employs a broadcast mode of IEEE 802.11 and transmits data to multiple APs within its communication range. The packets received at each AP are forwarded to GW. In order to enhance the reliability, we employ packet-level FEC (hereafter, we call it simply as FEC). With packet transmissions using FEC, when STA has K data packets to transmit, it transmits N packets (N ≥ K) at an application level. Here, N − K packets are redundant packets, and its FEC rate is calculated as K/N. At GW, if any set of K

© IEICE 2020 packets out of N packets are successfully received, it can recover the original K data DOI: 10.1587/comex.2020COL0038 Received June 30, 2020 packets [3]. Accepted July 20, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

628 IEICE Communications Express, Vol.9, No.12, 627–631

Fig. 1. Proposed rate control and model to construct training data.

3 Proposed adaptive rate control In order to adaptively control PHY and FEC rates in our Multi-AP transmissions described in Sec. 2, we propose a rate control that exploits information obtained from beacon frames periodically transmitted by surrounding APs. Specifically, we exploit RSSIs and beacon delivery ratio (BDR). The RSSIs reflect the channel conditions for STA-AP links while BDR accounts for interference level at the communication area. Here, interference is considered to be caused by radio devices sharing the same unlicensed band, which do not necessarily follow carrier sense multiple access (CSMA) protocols. Based on these information, STA attempts to select the best rate for uplink transmissions to multiple APs. In this letter, in order for each STA to select an appropriate rate, we resort to a machine-learning approach. A rate selector is designed, which outputs the optimal set of PHY/FEC rates as shown in the bottom of Fig. 1. The training data is constructed by computer simulations. In this letter, as shown in Fig. 1, we generate RSSIs and BDR at different locations within the considered 130 m × 130 m area where 9 APs are deployed. Then, at these locations, the best PHY/FEC rate, which satisfies the target application-level (i.e., after FEC decoding) PDR (APDR) with minimum fractional airtime, is obtained. PDR is defined as the ratio of the number of successfully decoded data to that of data transmitted by STA at application level. On the other hand, fractional airtime is the fraction of time during which signals transmitted by STA occupy the channel. The best rate is recorded as labelled data for the considered input of RSSIs and BDR. We consider path loss with its coefficient of 3, correlated shadowing with a first-order auto regression (AR) model

with σΨdB = 4 dB and Xc = 75 m [4], and block Rayleigh fading. The transmission power of STA and APs are set to be 20 mW, and application data with 1496 bytes are © IEICE 2020 DOI: 10.1587/comex.2020COL0038 assumed to be transmitted by STA with the period of 1/30 s. The packet/beacon error Received June 30, 2020 Accepted July 20, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

629 IEICE Communications Express, Vol.9, No.12, 627–631

rates are calculated based on NIST error model [5]. We also generate packet/beacon loss caused by interference randomly with a parameter called interference error rate, ranging from 0% to 30%. The considered PHY rates are 6, 12, 18, 24, 36, 48, and 54 Mbps while FEC rates of 1, 2/3, 1/2, 2/5, and 1/3 are employed. The PHY rate of beacon frame is ﬁxed at 6 Mbps. The training data is constructed every 1 s. As a learning algorithm, we employed stacked autoencoder with two hidden layers [6].

4 Simulation results We test the performance of the proposed rate control by using a STA moving around the considered area shown in Fig. 1, where the same simulation model as training phase is employed. The velocity of STA is 1 m/s with direction uniformly selected from [0, π/2]. The initial position is set to be (0,0) in Fig. 1. Figs. 2 and 3 show APDR and fractional airtime against interference error rate, respectively. Since there is no existing scheme to control PHY/FEC rate for Multi-AP transmissions considered in our work, we compare the performance of Multi-AP transmissions employing the proposed rate control with that employing the ﬁxed PHY/FEC rate.

Fig. 2. APDR against interference error rate of transmissions with the proposed rate control and those with ﬁxed PHY/FEC rates.

Fig. 3. Fractional airtime against interference error rate of © IEICE 2020 transmissions with the proposed rate control and those DOI: 10.1587/comex.2020COL0038 Received June 30, 2020 with ﬁxed PHY/FEC rates. Accepted July 20, 2020 Publicized August 4, 2020 Copyedited December 1, 2020

630 IEICE Communications Express, Vol.9, No.12, 627–631

Here, we consider two extreme cases: one is the most optimistic selection where PHY and FEC rates are respectively set to be 54 Mbps and 1, and the other is the most pessimistic one with 6 Mbps and 1/3. The target APDR for the proposed rate control is set to be 99%. From these figures, we can first see that the case with PHY rate of 6 Mbps and FEC rate of 1/3 achieves high APDR close to 100% while it exhibits large fractional airtime. This is because this set of PHY and FEC rates generates a large number of redundant packets with large packet size. On the other hand, the case with PHY rate of 54 Mbps and FEC rate of 1 achieves the smallest fractional airtime thanks to the reduced redundancy while it shows the worst APDR, which degrades as the interference error rate increases. On the other hand, the trade-off between APDR and fractional airtime is well-controlled by the proposed rate control as seen in Figs. 2 and 3: the proposed rate control achieves high APDR with small fractional airtime. The fractional airtime of the proposed rate control gradually increases as the interference error rate increases as shown in Fig. 3. This is because the proposed rate control tends to select more conservative rates as learned by the training data for larger interference error rate, which is required to achieve the target APDR. In fact, we have confirmed that the proposed rate control achieves the target APDR of 99% for any value of interference error rate as shown in Fig. 2. From these results, we can confirm that the proposed rate control adaptively selects the appropriate set of PHY and FEC rates according to the link and interference conditions over the communication area.

5 Conclusions In this letter, we have proposed an adaptive rate control for robust data transmissions utilizing Multi-AP receptions combined with packet-level FEC. The proposed scheme solely counts on the beacon frames received from surrounding APs for STA to decide the appropriate set of PHY/FEC rates to be employed for its uplink transmissions. The best set of PHY/FEC rates is selected by a rate selector, which learns the best rate for a given input by training data. Our numerical results show that the proposed rate control can well-control the trade-oﬀ between reliability and airtime.

Acknowledgments This work includes results of the project entitled “R&D on Technologies to Densely and Eﬃciently Utilize Radio Resources of Unlicensed Bands in Dedicated Areas,” which is supported by the Ministry of Internal Aﬀairs and Communications as part of the research program “R&D for Expansion of Radio Wave Resources (JPJ000254)”.

631 IEICE Communications Express, Vol.9, No.12, 632–635 Special Cluster in Conjunction with IEICE General Conference 2020 Estimating subjective video quality while using smartphone

Shotaro Mishina1, a), Mutsumi Suganuma2, and Wataru Kameyama2 1 Department of Computer Science and Communications Engineering, Graduate School of Fundamental Science and Engineering, Waseda University, 3–4–1, Okubo, Shinjuku-ku, Tokyo 169–8555 Japan 2 Faculty of Science and Enginnering, Waseda University, 3–4–1, Okubo, Shinjuku-ku, Tokyo 169–8555 Japan a) [email protected]

Abstract: In this paper, we investigate to estimate the subjective video quality of mobile users in a real situation of daily life. We conduct an experiment to collect bio-signals and sensor data from three subjects for a week as well as their contexts by questionnaire, and machine learning algorithms including Random Forest and k-Nearest Neighbor are applied for the estimation. The experimental results show that the subjective video quality is estimated accurately by Random Forest, while the contribution of users’ context is not conﬁrmed. It suggests that the subjective video quality can be well estimated by bio-signals and sensor data of mobile users in a real situation. Keywords: QoE, subjective video quality, mobile user, bio-signals, random forest Classiﬁcation: Multimedia Systems for Communication

References

[1] ITU-T Rec. P.10/G.100, “Vocabulary for performance, quality of service and quality of experience,” 2017. [2] A. Oishi, M. Suganuma, and W. Kameyama, “Time-wise QoE estimation of video quality using multi-modal bio-signals,” IEICE Tech. Rep., vol. 118, no. 503, CQ2018-94, pp. 13–18, March 2019 (in Japanese). [3] E. Ponder and W.P. Kennedy, “On the act of blinking,” Quarterly Journal of Experimental Physiology, vol. 18, no. 2, pp. 89–110, 1927. DOI: 10.1113/ expphysiol.1927.sp000433 [4] H. Shimizu, W. Kameyama, and M. Suganuma, “Context analysis and estimation of mobile users by using bio-signals and sensor data,” 2019 IEEE 8th Global Conference on Consumer Electronics (GCCE), pp. 263–266, Osaka, Japan, Oct. 2019. DOI: 10.1109/gcce46687.2019.9015475 [5] ITU-R Rec. BT.500-13, “Methodology for the subjective assessment of the quality of television pictures,” 2012. © IEICE 2020 DOI: 10.1587/comex.2020COL0035 Received June 30, 2020 Accepted July 27, 2020 Publicized August 6, 2020 Copyedited December 1, 2020

632 IEICE Communications Express, Vol.9, No.12, 632–635

1 Introduction Estimating QoE (Quality of Experiment) rather than QoS (Quality of Signal) has become an important issue for network management, for example. The conventional method to estimate QoE relies on MOS (Mean Opinion Score) obtained by questionnaire responses from subjects. However, not only does it take much time to collect results, but also cannot be done while viewing videos. Furthermore, it doesn’t take into account the fact that QoE may vary depending on many inﬂuencing factors [1]. [2] reports that subjective video quality can be estimated from subjects’ bio-signals and their interests on the videos they watch. However, as the experiment has been done in a controlled environment in laboratory, it is not clariﬁed yet that such a methodology can be applied to a real situation of mobile users watching video in a variety of contexts. Therefore, in this paper, we focus on QoE estimation of mobile users under real situation of daily life, utilizing bio-signals, sensor data and questionnaire responses.

2 Experiment 2.1 Devices and procedure JINS MEME ES_R (hereafter JINS MEME) is used to collect data of eye-ball movement, head movement and blinks in 20 [Hz]. [3] reports that blink has a profound relationship with emotion, for example, blinking ramps-up when people get angry. Polar H10 with V800 are adopted as a heartrate monitor to collect RRI (R-wave and R-wave Interval) [4]. Subjects use Galaxy S9 (Android 9.0 smartphone) to watch YouTube videos and answer a questionnaire. The questionnaire has five questions: the subjective video quality, interest in video, willingness to recommend the video to others, willingness to watch the video again, and their context at the time of watching the video. The first four questions have five multiple choices of 1–5, and the context has three options: “home”, “outside home”, “in public transportation”. All the questions and options are prepared in referencing to [5]. Three subjects of two men and one woman joined the experiment where the average age was 22.6 and the standard deviation was 0.47. The subjects were asked to wear JINS MEME and Polar H10 with V800, and to use the smartphone to watch YouTube videos freely. The experiment was conducted for one week. The subjects were instructed to answer the questionnaires after watching each video and use the provided smartphone freely if they use it only for the experiment.

2.2 Data collection Four applications are used to collect data installed on the smartphone: YouTube app, AZ screen recorder, JINS MEME app and MEME demo (3rd). YouTube app enables the subjects to watch videos using their own Google accounts in order to make it easier to ﬁnd videos each subject want to watch. AZ screen records the smartphone screen to monitor the subjects’ usage of the smartphone. JINS MEME app establishes the connection between the smartphone and JINS MEME. And © IEICE 2020 MEME demo (3rd) records data from JINS MEME in CSV ﬁle. DOI: 10.1587/comex.2020COL0035 Received June 30, 2020 Accepted July 27, 2020 Publicized August 6, 2020 Copyedited December 1, 2020

633 IEICE Communications Express, Vol.9, No.12, 632–635

3 Data The total amount of data collected from the three subjects is about 19 hours or 68,079 seconds. Subject 1 watched 58 videos for 15,384 seconds, Subject 2 watched 82 videos for 23,911 seconds, and Subject 3 watched 38 videos for 28,783 seconds. Table I summarizes all the responses to the questionnaire, where context 1 is “home”, 2 is “outside home”, and 3 is “in public transportation”. The subjects reiterate watching videos for this experiment during the experiment period. All the following data preprocessing is performed for each video viewing and for each subject. Data of JINS MEME in 20 [Hz] are converted to 10 [Hz] by calculating the mean and the variance of every 0.1 [s] by six consecutive data from t [s] to t + 0.25 [s] in referencing to [4], which suggests that this conversion method is eﬀective on behavior classiﬁcation. Polar H10 with V800 records the times when it detects R waves, then linear interpolation on the data is applied to transform them into 10 [Hz]. RRI change rate is calculated by dividing the transformed RRI data by the average RRI of each video viewing. Regardless of each video duration, the questionnaire responses are labeled to the data of the bio-signals and the sensor data from the start time to the end time of each corresponding video.

Table I. Summary of questionnaire responses

4 Analysis results and discussion Table II shows all the explanatory variables. We prepare two datasets: one includes all the variables shown in Table II, and the other excludes the questionnaire responses from the ﬁrst, in order to evaluate the contribution of the questionnaire responses. The subjective video quality is taken as the objective variable in the analysis. Random

Table II. Explanatory variables

634 IEICE Communications Express, Vol.9, No.12, 632–635

Forest classifier (RF) and k-Nearest Neighbor classifier (KNN) of scikit-learn are used for the analysis, where all the parameters are set as the defaults except for the random seed. Using the two data sets, 5-fold cross validation on each subject are conducted with “StratifiedKfold” from scikit-learn to level off the variation of the datasets. Table III shows the average accuracies of RF and KNN for the two datasets. Given the accuracies, it can be said that the subjective video quality can be accurately estimated by RF from the data including bio-signals and sensor data even under the mobile environment of daily life. However, comparing the results of the two datasets, the contribution of the context data is not clear. The reason could be that the three contexts defined in this experiment are too simple and too rough to reflect the actual mobile user’s context.

Table III. Results of 5-fold cross validation

RF: Random Forest, KNN: k-Nearest Neighbor

By the analysis on the important features of the RF results, the questionnaire responses, the head angular acceleration in yaw direction and the eye movement on X axis are found to be the signiﬁcant explanatory variables.

5 Conclusion We have investigated to estimate subjective video quality from bio-signals and sensor data of mobile users in daily life. According to the experimental results, RF is found to classify and estimate the subjective video quality with very high accuracy. However, on the context data contribution to the classification and the estimation, its effectiveness is not confirmed. As for the further study, we will conduct the experiment with more subjects, refine the questionnaire, investigate to introduce much detailed contexts to reflect the actual mobile user’s context, and optimize the parameters of the classifiers.

Acknowledgments This work is supported by JSPS KAKENHI Grant Number JP19K11932.

635 IEICE Communications Express, Vol.9, No.12, 636–641 Special Cluster in Conjunction with IEICE General Conference 2020 Increasing symbol rate by symbol decision based on spatial luminance distribution for rolling-shutter optical camera communication

Taiki Tanemura1 and Wataru Chujo1, a) 1 Department of Electrical and Electronic Engineering, Meijo University, 1–501 Shiogamaguchi, Tempaku-ku, Nagoya 468–8502, Japan a) [email protected]

Abstract: In rolling-shutter (RS) optical camera communication (OCC), as the ratio of the exposure time to the symbol length increases, it becomes difficult to achieve error-free transmission due to inter-symbol interference. In this study, to increase the symbol rate, symbol decision based on spatial luminance distribution is adapted for RS OCC. When the symbol rate increases and the ratio is close to 1, it is difficult to achieve error-free transmission by conventional symbol decision based on pixel value. However, even when the ratio increases up to 1, error-free transmission is achieved by the symbol decision based on spatial luminance distribution. Keywords: visible light communication, optical camera communication, image sensor, rolling shutter, exposure time, spatial luminance distribution Classification: Wireless Communication Technologies

References

[1] C. Danakis, M. Afgani, G. Povey, I. Underwood, and H. Haas, “Using a CMOS camera sensor for visible light communication,” IEEE Globecom Work- shops, Anaheim, USA, pp. 1244–1248, Dec. 2012. DOI: 10.1109/GLOCOMW. 2012.6477759 [2] H. Aoyama and M. Oshima, “Line scan sampling for visible light communication: theory and practice,” IEEE International Conference on Communications, London, UK, pp. 5060–5065, June 2015. DOI: 10.1109/ICC.2015.7249126 [3] K. Ohshima, T. Naramoto, K. Yamaguchi, and W. Chujo, “Rolling-shutter-based asynchronous optical camera communication by a cycle pattern of received symbols using smartphones,” IEICE Commun. Express, vol. 8, no. 3, pp. 49–54, March 2019. DOI: 10.1587/comex.2018XBL0141 [4] W. Chujo and M. Kinoshita, “Rolling-shutter-based 16QAM optical camera communication by spatial luminance distribution,” IEICE Commun. Express, vol. 8, no. 12, pp. 566–571, Dec. 2019. DOI: 10.1587/comex.2019GCL0055

636 IEICE Communications Express, Vol.9, No.12, 636–641

1 Introduction In rolling-shutter (RS) optical camera communication (OCC) between LED trans-

mitter and RS camera receiver, exposure time, Tex, of oﬀ-the-shelf camera is longer than exposure time interval, Tt [1, 2]. As the symbol length, Ts, decreases and comes close to Tt , it becomes diﬃcult to achieve error-free transmission owing to long exposure time. However, in some smartphones’ built-in cameras, error-free transmission has

been achieved even when Ts/Tt = 1 at IEICE Communications Express [3]. Al- though Tex of the built-in image sensor is unknown, error-free transmission has been achieved by adjusting the threshold pixel value, pth. In this study, to increase the symbol rate, symbol error rate (SER) is measured

while changing the ratio, Tex/Ts. In general, 8-bit pixel value of the captured LED image is used for symbol decision. When the ratio is much less than 1, error-free transmission is expected with conventional symbol decision based on the pixel value. However, error-free transmission based on the pixel value is diﬃcult when the ratio is close to 1. In order to achieve error-free transmission when the ratio increases up to 1, symbol decision based on spatial luminance distribution [4] is adapted for RS OCC.

2 Relationship between exposure time and symbol length

SER is most likely determined by the ratio, Tex/Ts. Figure 1 shows the eﬀect of exposure time on received symbol pattern when repetition symbol of “1” and “0”

is transmitted with on-oﬀ keying (OOK). In Figs. 1(a), (b), and (c), Ts/Tt = 1 is assumed and the sensitivity of the image sensor is assumed to be high.

Figure 1(a) shows a schematic diagram illustrating the relationship between Tex and received pixel value when Tex/Ts ≈ 0, where ∆t is oﬀset time between shutter timing of the camera and the symbol timing, and p1 and p0 is the received pixel value from 0 to 255 when symbol “1” and “0” is transmitted, respectively. When

Tex/Ts ≈ 0, error-free transmission is expected at any shutter timing because Tex does not overlap between symbols “1” and “0.” Therefore, the pixel value, p1 and p0, is expected to be 255 and 0, respectively.

On the other hand, Fig. 1(b) shows the diagram when 0 ≪ Tex/Ts ≪ 1. The exposure time, Tex, overlaps between symbols “1” and “0.” The pixel value, p1, is always expected to be 255. On the other hand, the pixel value, p0, is given by

po = a(∆t + Tex − Ts) ≥ 0, (1)

where a is an arbitrary constant determined with the sensitivity of the image sensor,

light intensity of the transmitter, and so on. As ∆t andTex increase, p0 increases more. However, p1 and p0 have diﬀerent pixel values when 0 ≪ Tex/Ts ≪ 1. Error-free transmission is expected by choosing pth appropriately for symbol decision. Moreover, Fig. 1(c) shows the diagram when Tex/Ts ≈ 1. The pixel value, p0, increases further. Both p1 and p0 are expected to be 255. It seems to be impossible to make a symbol decision based on the pixel value. © IEICE 2020 DOI: 10.1587/comex.2020COL0023 To clarify the eﬀect of Tex on received symbol pattern, Fig. 1(d) shows examples Received June 29, 2020 Accepted July 27, 2020 of measured pixel values while changing the ratio, Tex/Ts, where Ts/Tt = 3.A Publicized August 6, 2020 Copyedited December 1, 2020

637 IEICE Communications Express, Vol.9, No.12, 636–641

Fig. 1. Eﬀect of exposure time, Tex, on received symbol pattern while changing the ratio, Tex/Ts, where Ts is the symbol length, Tex is the exposure time, Tt is the exposure time interval, ∆t is oﬀset time between shutter timing of the camera and the symbol timing, and p1 and p0 is the received pixel value from 0 to 255 when symbol “1” and “0” is transmitted, respectively.

symbol consists of three pixels. When Tex/Ts = 0.3817, three pixels values, p1, are always 255. Although three pixels values, p0, gradually increase, average of the pixel values, p0, is less than 128. If pth ≥ 128, error-free transmission is achieved. When Tex/Ts = 0.7150, two pixels values, p0 = 255. However, the value of the remaining one pixel, p0 < 255. Although average of three pixels values, p0, increases further, error-free transmission is possible by choosing pth < 255. In contrast, when Tex/Ts = 1.0483, both the pixel values, p1 = p0 = 255 owing to long exposure time. There is no diﬀerence between p1 and p0. It is impossible to make a symbol decision based on the pixel value.

638 IEICE Communications Express, Vol.9, No.12, 636–641

3 SER measured with symbol decision based on pixel value SER was measured with symbol decision based on the received pixel value. Fig. 2

shows relationship between measured SER and the ratio, Tex/Ts. Fig. 2(a) shows technical parameters of LED transmitter and RS camera receiver. Repetition symbol

of “1” and “0” is transmitted with OOK. The symbol length, Ts, is chosen to integer multiple of the exposure time interval, Tt . Relationship between SER and Tex/Ts is investigated at various symbol rate by changing the ratio, Ts/Tt .

Fig. 2. SER measured with symbol decision based on pixel value.

Figure 2(b) shows relationship between SER and Tex/Ts when pth = 160, where −4 SER = 10 indicates error-free transmission. Since pth is not high, errors may occur as Tex/Ts increases. Actually, errors occur when Tex/Ts is more than 0.5363. This is due to increase of the pixel value, p0. When 0 ≪ Tex/Ts ≪ 1, it is obvious that long exposure time degrades SER.

In contrast, Fig. 2(c) shows SER and Tex/Ts when pth = 254. Since pth is high, error-free transmission is achieved up to Tex/Ts = 0.858. The pixel value, p0, increases as Tex/Ts increases. However, if p0 < 255, it is possible to make a symbol decision by increasing pth up to the maximum value, 255. Possibility of the symbol decision depends on the sensitivity of the image sensor. We considered the reason © IEICE 2020 DOI: 10.1587/comex.2020COL0023 why error-free transmission was achieved using smartphones’ built-in RS cameras Received June 29, 2020 Accepted July 27, 2020 even when Ts/Tt = 1 at IEICE Communications Express [3]. The reason may be that Publicized August 6, 2020 Copyedited December 1, 2020

639 IEICE Communications Express, Vol.9, No.12, 636–641

the sensitivity of the image sensor was not high. On the other hand, in this study, the sensitivity of the image sensor is too high to achieve error-free transmission when

Tex/Ts = 1 with symbol decision based on the pixel value.

4 SER measured with symbol decision based on spatial luminance distribution

When Tex/Ts increases up to 1, p0 is saturated to 255 and whiteout occurs in the captured image. It is impossible to make a symbol decision based on the pixel value.

However, even when p0 is saturated on the LED surface image, spatial luminance distribution of the surface gradually decreases towards end part of the surface. Therefore, spatial luminance distribution is eﬀective for symbol decision [4]. RS camera captures the spatial luminance distribution of LED surface at each time interval. In the captured image, white pixels indicate symbol “1” after adjusting

pth. Although p0 is saturated to 255 on the surface, number of white pixels varies on each interval in accordance with OOK modulation. Figure 3(a) shows an example of captured LED image when repetition symbol

of “1” and “0” is transmitted with OOK, where Ts/Tt = 6 and Tex/Ts = 1.0242.

© IEICE 2020 Fig. 3. SER measured with symbol decision based on spatial DOI: 10.1587/comex.2020COL0023 luminance distribution. Received June 29, 2020 Accepted July 27, 2020 Publicized August 6, 2020 Copyedited December 1, 2020

640 IEICE Communications Express, Vol.9, No.12, 636–641

Each pixel line of the camera is sequentially exposed from top to bottom. Since p0 is saturated on the LED surface image, whiteout occurs at the middle part of the image. It is impossible to make a symbol decision with the pixel value. However, the spatial luminance distribution of LED surface gradually decreases towards end part of the

surface. Pixel values, p1 and p0, on right- and left-edge parts of the image vary in accordance with OOK. Therefore, after adjusting pth, number of white pixels varies in accordance with OOK. Figure 3(b) shows measured pixel values on the middle part of the captured

image, where Ts/Tt = 3 and Tex/Ts = 1.0483. Symbol length is three times as long as the exposure time interval. Since all the pixel values are saturated at the maximum value, it is impossible to make a symbol decision. On the other hand,

Fig. 3(c) shows number of white pixels on each interval after adjusting pth = 254. It is possible to make a symbol decision by averaging the number of white pixels on symbol “1” and “0” every 3 pixels.

Figure 3(d) shows relationship between measured SER and the ratio, Tex/Ts, with symbol decision based on the spatial luminance distribution, where pth = 254. Error-free transmission is achieved up to Tex/Ts = 1.0725. When the ratio of the symbol length to the exposure time interval, Ts/Tt = 2, symbol rate increases up to 40.388 kilo-symbols per second (symbols/s). Even when Tex/Ts ≈ 1, error-free transmission is achieved.

5 Conclusion In this study, to increase the symbol rate for RS OCC, the symbol decision based on spatial luminance distribution was adapted for RS OCC with OOK. When the

ratio of the exposure time to the symbol length, Tex/Ts ≈ 0, error-free transmission was achieved regardless of the threshold pixel value, pth, with conventional symbol decision based on pixel value.

On the other hand, when 0 ≪ Tex/Ts ≪ 1, error-free transmission was achieved by carefully choosing pth with the symbol decision based on pixel value. However, error-free transmission depends on the sensitivity of the image sensor.

To increase the symbol rate further, when Tex/Ts ≈ 1, error-free transmission was achieved with the symbol decision based on spatial luminance distribution. Even

though pixel values, p1 and p0, on the middle part of LED image are saturated and whiteout occurs, p1 and p0 on the edge part of the image help to make a symbol decision. Error-free transmission was achieved up to Tex/Ts = 1.0725 with the symbol decision based on spatial luminance distribution. Symbol rate increased up to 40.388 kilo-symbols/s.

641 IEICE Communications Express, Vol.9, No.12, 642–645 Special Cluster in Conjunction with IEICE General Conference 2020 Study on path loss model for macrocell environment using percentage of area occupied by buildings for 5G

Koshiro Kitao1, a), Tetsuro Imai1, Minoru Inomata1, Satoshi Suyama1, and Yusuhiro Oda1 1 NTT DOCOMO, INC., 3–6 Hikarino-oka, Yokosuka-shi, Kanagawa 239–8536, Japan a) [email protected]

Abstract: We propose a correction formula using the percentage of area occupied by buildings, α, for a path model in the urban macrocell environment defined in ITU-R M.2135 based on measurements performed in Japan for below 6 GHz for 5G. Based on analysis, the regression results of the correction formula in which S, the median of the difference between the measured path loss and that of the reference path model is proportional to α, is better than when S is proportional to the logarithm of α. Keywords: mobile radio, 5G, path loss model, percentage of area occupied buildings Classification: Antennas and Propagation

References

[1] NTT DOCOMO, INC., “DOCOMO 5G white paper, 5G radio access: Require- ments, concept and technologies,” July 2014. [2] ITU-R Report M. 2412-0, “Guidelines for evaluation of radio interface technologies for IMT-2020,” Oct. 2017. [3] S. Kozono and K. Watanabe, “Inﬂuence of environment building on UHF band mobile radio propagation,” IEEE Trans. Commun., vol. 25, no. 10, pp. 1133– 1143, Oct. 1977. DOI: 10.1109/tcom.1977.1093736 [4] K. Kitao and S. Ichitsubo, “Path loss prediction formula for urban and suburban areas for 4G systems,” Proc. VTC 2006, vol. 6, pp. 2911–2915, May 2006. DOI: 10.1109/vetecs.2006.1683401 [5] ITU-R Report M. 2135-1, “Guidelines for evaluation of radio interface technologies for IMT-Advanced,” Dec. 2009.

1 Introduction Research and development on the ﬁfth generation mobile communication system © IEICE 2020 (5G) were pursued to actualize high-speed and high-capacity communications, and DOI: 10.1587/comex.2020COL0033 Received June 30, 2020 commercial services have started. Use of the microwave band under 6 GHz is Accepted July 2, 2020 Publicized August 20, 2020 Copyedited December 1, 2020

642 IEICE Communications Express, Vol.9, No.12, 642–645

considered for 5G [1, 2]. In general, a path loss model that can be applied to various environments such as urban and suburban areas is required in order to design service areas for mobile communication systems. In previous studies, applying an empirical correction formula that employs the percentage of area occupied by buildings based on measurement to a reference path loss model for urban environments were proposed in order to predict path loss with high accuracy in a macrocell environment, which is a typical cell deployment for the microwave band [3, 41]. Regarding the path loss model for macrocell environment for 5G, the ITU-R M. 2412 model [2] is widely recognized; however, a correction formula using the percentage of area occupied buildings for this model has not yet been proposed. In ITU-R M. 2412, there are two path loss models for a macrocell environment which is UMa_A and UMa_B. UMa_A is based on the ITU-R M.2135 model that was originally developed for 4G systems and is widely used for system design and evaluation. We investigate a correction formula for the model applied to an urban macrocell environment in ITU-R M.2135 based on measurements [5]. We ﬁrst outline the measurements and then derive the correction formula for the ITU-R M.2135 model.

2 Measurements Path loss measurements were performed to derive the path loss correction formula in ﬁve areas in Japan. Table I gives the measurement parameters. In this experiment, the base station (BS) antenna transmits a continuous wave (CW) and the mobile station (MS) antenna receives the wave and records the received level. The maximum distance between the BS and MS is 3 km. The BS antenna is mounted on a tower or building roof at the height of 10 m to 117 m. The MS antenna is mounted on a measuring car at the antenna height of 2.1 m or 2.7 m. Both antennas are sleeve antennas. The measurement frequency is 0.81 GHz to 8.45 GHz. Path loss is obtained based on the average received level over the interval of 50 m. Table I gives the percentage of area occupied by buildings, α, and is expressed as

α = Ab/Ap (1)

2 2 where Ab (m ) is the area occupied by buildings and Ap (m ) is the area of the prediction target. In the Samukawa area, there are areas that have two diﬀerent α values and values are obtained for each area. Moreover, only data from non-line-of sight (NLOS) areas are analyzed.

Table I. Measurement parameters.

© IEICE 2020 DOI: 10.1587/comex.2020COL0033 Received June 30, 2020 Accepted July 2, 2020 1In this letter, empirical correction formula for ITU-R M.2135 is newly studied. Publicized August 20, 2020 Copyedited December 1, 2020

643 IEICE Communications Express, Vol.9, No.12, 642–645

3 Path loss correction formula using percentage of area occupied by buildings Similar to traditional methods to derive the path loss correction formula using α, the formula is derived by applying regression analysis to the diﬀerence between the measured path loss in various environments and the value from the reference path loss model for an urban area [3]. In this study, the reference path loss model is the model for an urban macro (UMa) NLOS environment from ITU-R M. 2135, which is expressed as

= . − . ( ) + . ( ) − ( . − . ( / )2) ( ) PLUMa 161 04 7 1 log10 W 7 5 log10 h 24 37 3 7 h hBS log10 hBS + ( . − . ( ))( ( ) − ) + ( ) − ( . ( ( . ))2 43 42 3 1 log10 hBS log10 d 3 20 log10 fc 3 2 log10 11 75hUT − 4.97) (2)

where PLUMa is the path loss (dB), W is the road width (m), h is the average building height (m) in the target area, hBS is the BS antenna height (m), d is the distance between the BS and the MS (m), fc is the frequency (GHz), and hUT is the MS antenna height [5]. hBS and hUT are the same for the measurements and typical values W = 20 (m) and h = 20 (m) described in [5] are used for analysis.

The path loss model using the correction formula, S(α), and the above PLUMa of ITU-R M. 2135 are expressed as

PL = PLUMa − S(α) (3)

Similar to a previous study [3], correction value S is the median of the diﬀerence between the measured path loss and that of the reference model within a 500 m by 500 m area. Figure 1 shows the distribution of α per 500 m by 500 m area in the measurement area. Moreover, Fig. 2 shows the relationship between α and S. Similar to [4], results of S regressed by the logarithm of alpha is expressed as

S(α) = 35.1 − 21 log α (4)

The root mean square (RMS) of the residual error, which is the diﬀerence between the measured path loss and calculated value of (4), is 9.4 (dB). Figure 2 shows that

© IEICE 2020 Fig. 1. Distribution of percentage of area occupied by build- DOI: 10.1587/comex.2020COL0033 Received June 30, 2020 ings, α. Accepted July 2, 2020 Publicized August 20, 2020 Copyedited December 1, 2020

644 IEICE Communications Express, Vol.9, No.12, 642–645

Fig. 2. Correction value S vs. percentage of area occupied by buildings, α.

S tends to be proportional to α. The results of S regressed by α is express as

S(α) = 22.5 − 0.6α (5)

The RMS of the residual error, which is the diﬀerence between the measured path loss and calculated RMS value of the residual error of (5), is 8.8 (dB). This value is better than that obtained in (4). From these results, (5) can be used for the correction formula for (2).

4 Conclusion We investigated a path loss correction formula for a path loss model in a UMa NLOS environment in ITU-R M. 2135 based on measurements in order to predict the path loss in a macrocell environment for frequency bands below 6 GHz for 5G. The regression results of the correction formula in which S is proportional to α are better than that in which S is proportional to the logarithm of α. A correction formula for higher frequencies such as the millimeter wave band is left for future study.

Acknowledgments The authors thank Ms. Shihoko Takahashi of DOCOMO Technology, Inc. for her helpful suggestions and encouragement.

645 IEICE Communications Express, Vol.9, No.12, 646–649 Special Cluster in Conjunction with IEICE General Conference 2020 Development of W-band waveguide based on plastic additive manufacturing with Ni electroless plating

Kohei Takizawa1, a), Kohei Fujiwara1, Yuta Watanabe1, Ryuichi Kobayashi1, Satoshi Kuwahara1, and Shota Takemura1 1 Tokyo Metropolitan Industrial Technology Research Institute, 2–4–10 Aomi, Koto-ku, Tokyo 135–0064, Japan a) [email protected]

Abstract: In general, a millimeter-wave component, such as a waveguide, is made of metal, therefore, it is heavy weight and expensive. In this development, a W-band waveguide based on an additive manufacturing with an electroless plating was demonstrated. We conﬁrmed that the performance of the transmission is approximately −2 dB in 75–110 GHz on a 50 mm long WR-10 waveguide. Keywords: millimeter-wave, waveguide, additive manufacturing, electroless plating Classiﬁcation: Antennas and Propagation

References

[1] P.Rousseau, Y.Avelino, and R. Lacoste, “How can 3D-printed plastic waveguides enable V-band applications?,” Microwaves & RF, Jan. 2019. [2] M. D’Auria, W.J. Otter, J. Hazell, B.T.W. Gillatt, C. Long-Collins, N.M. Ridler, and S. Lucyszyn, “3-D printed metal-pipe rectangular waveguides,” IEEE Trans. Compon. Packag. Manuf. Technol., vol. 5, no. 9, pp. 1339–1349, Sep. 2015. DOI: 10.1109/tcpmt.2015.2462130 [3] C. Tomassoni, O.A. Peverini, G. Venanzoni, G. Addamo, F. Paonessa, and G. Virone, “3D printing of microwave and millimeter-wave ﬁlters,” IEEE Microw. Mag., vol. 21, no. 6, pp. 24–45, June 2020. DOI: 10.1109/mmm.2020.2979153 [4] I. Kojima, “Process contol technology of stereolithography apparatus and its application,” J. Soc. Instrum. Contr. Eng., vol. 54, no. 6, pp. 416–420, June 2015. DOI: 10.11499/sicejl.54.416 [5] T. Hagiwara, “Current status of 3D printing material and its future aspect,” J. Imag. Soc. Jpn, vol. 54, no. 4, pp. 293–300, Aug. 2015. DOI: 10.11370/isj.54.293 [6] K. Takizawa, K. Fujiwara, Y. Watanabe, R. Kobayashi, S. Kuwahara, and S. Takemura, “Fabrication of millimeter-wave band waveguides by using additive manufacturing and plating technique,” IEICE Technical Report, C2-73, Oct. 2019.

646 IEICE Communications Express, Vol.9, No.12, 646–649

1 Introduction The usage of millimeter-wave in anti-collision radar and ultra-high-speed telecommunications has led to the development of the millimeter-wave industry over recent years. The millimeter-wave circuit is heavy because its components, such as waveguides, are made of metal. Furthermore, since the metal parts are manufactured by precision milling, the manufacturing cost tends to being higher. The use of millimeter-wave technology requires lightweight and inexpensive millimeter-wave components. In order to solve these issues, an additive manufacturing (AM) technology is suitable method. In paticular, AM is a more inexpensive method than milling or die casting. Several researches have been previously conducted on manufacturing millimeter-wave components by the metal AM and stereolithography [1, 2, 3]. However, the metal AM is not lightweight, and stereolithography has a problem of mechanical strength [4, 5]. In this development, we demonstrated a WR-10 waveguide in the length of 50 mm based on an AM technology using polyamide 11 and Ni electroless plating.

2 AM based waveguide fabrication The plated AM based waveguide and metal waveguide are shown in Fig. 1. The AM based waveguide is compliant with the WR-10 standard with the UG-387/U ﬂange, and the length is 50 mm. The cross section is 2.54 mm × 1.27 mm. In the built parts with polyamide 11, we employed a RaFaEl 300F (Aspect, Japan), which has an ability to form a 100-µm-thickness-layer with a 10 W ﬁber laser (λ = 1064 nm) output. A Ni electroless plating process on the waveguide surface involves the following two steps. First, repeat a sensitizer process for 1 minute at 30 ℃ and an activator process for 1 minute at 30 ℃ twice. Next, a Ni electroless plating process for 100 minutes at 70 ℃.

Fig. 1. Photograph of AM based waveguide and metal waveguide.

647 IEICE Communications Express, Vol.9, No.12, 646–649

3 Experimental setup and result S-parameters were evaluated using a vector network analyzer (Keysight N5247A) with a waveguide extender (OML V10VBA2-T/R-A). The measurement frequency range was 75–110 GHz. Figure 2 (a) shows transmission losses (S21). The transmission loss in the current trial is improved by more than 5 dB comparison with the

Fig. 2. (a) is comparison of measured transmission losses of metal, current, and previous trial. (b) is comparison © IEICE 2020 of measured reﬂection of metal, current, and previous DOI: 10.1587/comex.2020COL0024 Received June 29, 2020 trial. Accepted July 10, 2020 Publicized August 20, 2020 Copyedited December 1, 2020

648 IEICE Communications Express, Vol.9, No.12, 646–649

previous trial [6]. The difference between the current trial and the metal waveguide is approximately 2 dB. Figure 2 (b) shows reflections (S11). The difference between the maximum reflection of the current trial and the metal waveguide is approximately 13 dB. Figure 3 shows group delays. It is confirmed that group delay of the current trial waveguide has stable phase characteristic.

Fig. 3. Comparison of group delay of metal, current, and previous trial.

In the current trial, the plating thickness was increase in the waveguide in order to improve the conductivity. The plating temperature is increased from 45 ℃ to 70 ℃. The diﬀerence of the metal and the current trial in the S-parameter characteristics might be caused a defect of Ni electroless plating inside the AM based waveguide. Therefore, it is necessary to observe the inside of the waveguide with a non-destructive inspection, such as an X-ray computed tomography and improve a plating quality.

4 Conclusion In this development, a 50 mm long WR-10 waveguide was developed based on the AM technology and Ni electroless plating. The transmission loss of approximately −2 dB and the reﬂection of approximately −17 dB was obtained in the frequency range of 75–110 GHz. It is improved than the previous trial by higher plating temperature of 25 ℃. The group delay of the AM based waveguide is comparable with a metal waveguide. The performance diﬀerence might be caused that the non-plated area is existing inside the AM based waveguide. In the future, we will improve the plating quality and analyze the inside plating layer using a non- destructive inspection instrument. © IEICE 2020 DOI: 10.1587/comex.2020COL0024 Received June 29, 2020 Accepted July 10, 2020 Publicized August 20, 2020 Copyedited December 1, 2020

649 IEICE Communications Express, Vol.9, No.12, 650–655 Special Cluster in Conjunction with IEICE General Conference 2020 Electromagnetic wave pattern detection using cepstral features in the manufacturing ﬁeld

Ayano Ohnishi1, a), Michio Miyamoto1, Yoshio Takeuchi1, Toshiyuki Maeyama1, 2, Akio Hasegawa1, and Hiroyuki Yokoyama1 1 Advanced Telecommunications Research Institute International, 2–2–2 Hikaridai, Seika-cho, Sorakugun, Kyoto 619–0288, Japan 2 Takushoku University, 815–1 Tatemachi, Hachioji-shi, Tokyo 193–0985, Japan a) [email protected]

Abstract: In manufacturing fields such as factories, multiple wireless communication systems often operate in the same area simultaneously. More- over, it is known that industrial equipment emits electromagnetic noise over channels for wireless communications [1]. In order to ensure reliable communications under such an environment, monitoring radio wave environments specific to each manufacturing field and finding channels and timing which enable stable communications are required. The authors have studied technologies to efficiently analyze a large amount of monitoring data including signals which show unknown spectrum such as wide band electromagnetic noise [2, 3]. This paper proposes performing machine learning using cepstrum vectors as features to grasp types of noise and signals from data measured under environments in which electromagnetic noise and communication signals coexist. By using this features, the authors demonstrate that reduction of computational loads and improvement of detection accuracy can be expected. Keywords: wireless communication, factory, electromagnetic noise, machine learning, clustering, cepstrum Classification: Sensing

References

[1] IEEE 802 Nendica Report, “Flexible factory IoT: Use cases and communication requirements for wired and wireless bridged networks,” IEEE SA Industry Connections White Paper, April 2020. [2] A. Ohnishi, M. Miyamoto, Y. Takeuchi, A. Hasegawa, and H. Yokoyama, “A study of noise pattern detection for frequency sharing among various communications under high electromagnetic noise in the manufacturing ﬁeld,” Proc. 2020 IEICE General Conference, B-17-28, 2020 (in Japanese). © IEICE 2020 [3] M. Miyamoto, A. Ohnishi, Y. Takeuchi, A. Hasegawa, and H. Yokoyama, DOI: 10.1587/comex.2020COL0039 Received June 30, 2020 “Development of measurement technology for frequency sharing among various Accepted July 15, 2020 Publicized October 23, 2020 Copyedited December 1, 2020

650 IEICE Communications Express, Vol.9, No.12, 650–655

communications under high electromagnetic noise in the manufacturing ﬁeld,” Proc. 2020 IEICE General Conference, B-17-27, 2020 (in Japanese). [4] D.G. Childers, D.P. Skinner, and R.C. Kemerait, “The cepstrum: A guide to processing,” Proc. IEEE, vol. 65, no. 10, pp. 1428–1443, 1977. DOI: 10.1109/ PROC.1977.10747 [5] A.V. Oppenheim and R.W. Schafer, “From frequency to quefrency: A history of the cepstrum,” IEEE Signal Process. Mag., vol. 21, no. 5, pp. 95–106, 2004. DOI: 10.1109/MSP.2004.1328092 [6] Y. Kishino, Y. Shirai, K. Takeuchi, T. Suyama, F. Naya, and N. Ueda, “Regional garbage amount estimation and analysis using car-mounted motion sensors,” Proc. 2018 ACM Int. Joint Conf. and 2018 Int. Symp. Pervasive Ubiquitous Comput. Wearable Comput., 2018. DOI: 10.1145/3267305.3267610 [7] H. Cheng, B.L. Mark, and Y. Ephraim, “Wideband temporal spectrum sensing using cepstral features,” 2019 IEEE 20th Int. Symp. A World of Wire- less, Mobile and Multimedia Networks (WoWMoM), 2019. DOI: 10.1109/ WoWMoM.2019.8793002 [8] R.M. Al-Makhlasawy, M.M. Abd Elnaby, and H.A. El-Khobby. “Automatic modulation recognition in wireless systems using cepstral analysis and neural networks,” 2012 IEEE 29th National Radio Science Conference (NRSC), C24, 2012. DOI: 10.1109/NRSC.2012.6208542 [9] F. Mo, Y.-H. Lu, J.-L. Zhang, Q. Cui, and S. Qiu, “A support vector machine for identiﬁcation of monitors based on their unintended electromagnetic em- anation,” Progress in Electromagnetics Research, vol. 30, pp.211–224, 2013. DOI: 10.2528/PIERM12122406 [10] M. Riera-Guasp, J.A. Antonino-Daviu, and G.-A. Capolino, “Advances in electrical machine, power electronic, and drive condition monitoring and fault detection: state of the art,” IEEE Trans. Ind. Electron., vol. 62, no. 3, pp.1746– 1759, 2014. DOI: 10.1109/TIE.2014.2375853 [11] D. Arthur and S. Vassilvitskii, “k-means++: The advantages of careful seeding,” Proc. Annu. ACM-SIAM Symp. Discrete Algorithms, pp. 1027–1035, 2006. [12] S.J. Phillips, “Acceleration of k-means and related clustering algorithms,” Work- shop on Algorithm Engineering and Experimentation, 2002. DOI: 10.1007/ 3-540-45643-0_13

1 Introduction In recent years, sampling and recording data for a long time have become possible when monitoring and analyzing radio wave environments because software defined radio receivers and storage media of a large capacity and high speed performance have become widely available. Therefore, methods to extract the necessary information by efficiently analyzing types of signals and their generation timing from a large amount of data are needed. Electromagnetic noise emitted from industrial equipment especially does not have known spectrum and time wave forms such as communication signals. In addition to this, the spectrum and time wave forms are not fixed because kinds of equipment or their layouts in each factory are different. Therefore, algorithms to detect known spectrum and known wave forms are not applicable. Machine learning is useful when generally grasping types of signals from data © IEICE 2020 DOI: 10.1587/comex.2020COL0039 measured in environments in which electromagnetic noise and communication sig- Received June 30, 2020 Accepted July 15, 2020 Publicized October 23, 2020 Copyedited December 1, 2020

651 IEICE Communications Express, Vol.9, No.12, 650–655

nals coexist. An assumed classification process is as follows: a classifier is built by learning prepared supervised data. After that, examination data is input to the classifier and classified by types of signals. In order to do this, firstly, it needs to be prepared supervised data efficiently by performing clustering of unclassified data. In this paper, cepstrum vectors which are calculated and extracted from IQ samples are used as features when performing clustering process of unclassified data including electromagnetic noise emitted by industrial equipment. Cepstral analysis is a widely known technique in the field of speech recognition and vibration analysis. It is used to extract envelopes of spectrum and detect peri- odicity of spectrum and so on [4, 5]. Cepstral features have been used to classify data in their field. Cepstral features calculated from vibration data of on-vehicle motion sensors are used to detect some events in [6]. In the wireless communication fields, cepstral features are used to analyze the channel utilization state of primary users in [7] and identify modulation technique [8] but there are few examples using cepstral features for detection of electromagnetic noise. Focusing on features for detection of noise and signal, if data is investigated previously and information about characteristics of frequency is not needed, only using average and deviation values of power spectrum will work [9]. However, features including information about spectrum pattern should be used when analyzing data containing signals which have obvious characteristics related to frequency such as narrow band communications. Alternatively, it is necessary to examine computation speed and reduce the amount of data when using spectrum as features to keep information about frequency [10]. We propose using cepstral features aiming to reducing the amount of data and improving the accuracy of clustering. We describe advantages of using cepstral features in Section 2. Then, we show the effectiveness of using them to detect electromagnetic noise and communication signal by showing clustering results with the actual data recorded in a factory in Section 3. Finally, we conclude in Section 4.

2 Calculation of cepstral features Clustering process is performed to grasp various signal patterns such as communication signals and electromagnetic noise. Cepstral vectors are used as the features in this paper. The cepstrum is deﬁned by performing inverse Discrete Fourier Trans- form (DFT) of the sequence after calculating logarithm of magnitude of the DFT of a signal. The cepstrum is given by

−1 cepstrum = DFT (log |DFT(xn)|) (1)

where log(·) denotes the logarithm with any base and xn is a signal sequence obtained in a certain period and DFT(·) is DFT and DFT−1(·) is inverse DFT.

2.1 Dimension of cepstral features Extracting low-dimensional elements of a cepstral vector calculated from monitoring data (IQ samples) of radio wave environments corresponds to applying low-pass ﬁlter © IEICE 2020 to a spectrum pattern. This facilitates discriminating of spectrum patterns. Fig. 1 DOI: 10.1587/comex.2020COL0039 Received June 30, 2020 shows a part of cepstrum vectors extracted from 15000 vectors calculated using Accepted July 15, 2020 Publicized October 23, 2020 Copyedited December 1, 2020

652 IEICE Communications Express, Vol.9, No.12, 650–655

Fig. 1. Magnitude cepstrum

actual data. Elements of cepstrum vectors are close to zero in higher dimensions. 16-dimensional cepstrum of signal and noise have diﬀerent characteristics which enable to classify them. Therefore, we decided to extract 16-dimensional cepstrum and use as features.

2.2 Advantage of cepstral features in clustering process (1) Reducing the number of dimensions When extracting low-dimensional elements from cepstral vectors and using as features, it is possible to reduce computational loads of clustering process and memory consumption by reducing the number of dimensions of features drastically while keeping characteristics of spectrum. It is also possible to reduce the amount of data recording in storage media keeping their characteristics.

(2) Suppression of variations in clustering results When using spectrum as features, some results are not appropriate depending on the random seed because of ambiguity of spectrum patterns including white noise and randomness found in initial settings of clustering algorithms (k-means++ [11], which can improve initial settings in k-means algorithm, is used in this paper). On the other hand, when using cepstrum as features, generating inappropriate results are suppressed by reducing white noise. (An example of clustering results is shown in Chapter 3.)

(3) Estimation of the number of clusters As explained in (2), because variations in clustering results are suppressed, the same results are generated regardless of the values of random seed as long as values below a value of the proper number of clusters are set to the parameter for clustering analysis. Here the proper number of clusters refers to the situation in which each classified pattern has different characteristics and multiple signal patterns are classified properly. When clustering process is performed setting values exceeding the proper number of clusters, the results are different depending on the random seed because each cluster is classified in an inappropriate boundary. It is possible to simply estimate the proper number of clusters using this.

653 IEICE Communications Express, Vol.9, No.12, 650–655

3 Clustering of electromagnetic wave pattern As shown in Fig. 2 [2], a pattern of narrow band communication signals and two patterns of electromagnetic noise are contained in the actual data (IQ samples) analyzed in this paper.

Fig. 2. (a) Spectrogram of analyzed data. (b) Communication signal. (c) Type1 of electromagnetic noise. (d) Type2 of electromagnetic noise (modiﬁed based on [2] Figs. 1, 2)

We investigated whether generation timing of electromagnetic noise and communication signals are detected properly by performing clustering process. We extracted and used 16-dimensional cepstral features from a 15000-dimensional spectrum vector at every millisecond and analyzed them for one second in total. A 15000-by-1000 array is input to clustering process at every second when using spectrum. In contrast, a 16-by-1000 array is input when using cepstral features. Therefore, it is performed the reduction of data of approximately one-thousandth when running clustering process at every one second. The k-means algorithm have a computation complexity of O(ndk) [12], where n is the number of data points and d is the number of dimensions, and k is the number of clusters. The number of dimensions is reduced by using cepstral features. The clustering results when using cepstral features and setting the number of clusters within the range of 3 to 5 are shown in Fig. 3. The ordinate of the graph is cluster Id and it is shown the cluster Id into which signals have been classified at every millisecond. When setting the number of clusters to 4, communication signals and two types of noise patterns are classified properly in both random seeds. The clustering process was performed ten times changing the value of random seed to investigate suppression of variations in clustering results and possibility of estimating the number of clusters. As an example, Fig. 3 shows the clustering results in two different seeds while using cepstral features. When setting the number of © IEICE 2020 DOI: 10.1587/comex.2020COL0039 clusters to 3 or 4 which are less than the appropriate number, no matter what random Received June 30, 2020 Accepted July 15, 2020 seed was used, the results were the same and appropriate. However, when using Publicized October 23, 2020 Copyedited December 1, 2020

654 IEICE Communications Express, Vol.9, No.12, 650–655

Fig. 3. Clustering results when setting the number of clusters within the range of 3 to 5 ((a) random seed 35 (b) random seed 171)

spectrum as features, even if the number of clusters less than the appropriate number was set, these results were not the same and inappropriate in some cases. On the other hand, when setting the number of clusters to values exceeding the appropriate number while using cepstral features, the diﬀerent results are shown depending on the values of random seed. From this, it is possible to estimate the proper number of clusters by seeing if the results are variable while changing the value of a random seed.

4 Conclusion This paper demonstrated that the reduction of data of approximately one-thousandth and detecting of multiple type of signals with high accuracy were performed by using cepstral features when analyzing data measured under environments in which electromagnetic noise and communication signals coexist. We also conﬁrmed that estimation of the proper number of clusters was possible by comparing clustering results while changing the value of random seeds.

Acknowledgments This work includes results of the project entitled “R&D on Adaptive Media Access Control for Increasing the Capacity of Wireless IoT Devices in Factory Sites,” which is supported by the Ministry of Internal Aﬀairs and Communications as part of the research program “R&D for Expansion of Radio Wave Resources (JPJ000254)”. The authors would like to thank to Murata Machinery, Ltd. for supporting experiments in their factory.

655 IEICE Communications Express, Vol.9, No.12, 656–661 Video communication for teleconferencing using edge computing

Kouichi Genda1, a), Mitsuru Abe2, and Shohei Kamamura3 1 Department of Computer Science, College of Engineering, Nihon University, 1 Nakagawara, Tokusada, Tamura, Koriyama-shi, Fukushima 963–8642, Japan 2 NTT Communications Corporation, 2–3–1 Otemachi, Chiyoda-ku, Tokyo 100–8019, Japan 3 NTT Network Service Systems Laboratories, NTT Corporation, 3–9–11 Midori-cho, Musashino-shi, Tokyo 466–8555, Japan a) [email protected]

Abstract: This paper proposes a backbone network resource optimization algorithm for video communications of teleconferencing that use edge computing. In the current video communication architecture, the key component, the multi-point control unit (MCU), is deployed in the central cloud server, and its bandwidth consumption in the backbone network becomes huge as the video resolution and the frequency of use increase. By using edge computing, MCU can be deployed at the entrance node of the backbone network. This can reduce the bandwidth consumption of the backbone network. However, the edge deployment and routing (EDR) problem, which is classified as NP-hard, should be solved to achieve sufficient bandwidth reduction. To solve the EDR problem within a feasible time, we propose a divide and merge algorithm using the linear programming approach. We demonstrate that, with our algorithm, bandwidth consumption using edge computing is reduced by approximately 30% compared to the current architecture on a world-wide network. Keywords: video communication, edge computing, linear programming, network optimization, multi-access edge computing Classification: Network

References

[1] “Multi-access edge computing (MEC),” https://www.etsi.org/technologies/ multi-access-edge-computing, accessed March 2020. [2] P. Mach and Z. Becvar, “Mobile edge computing: a survey on architecture and computation offloading,” IEEE Commun. Surveys Tut., vol. 19, no. 3, pp. 1628– 1656, 2017. DOI: 10.1109/comst.2017.2682318 [3] “Bandwidth optimization within Pexip Infinity,” https://docs.pexip.com/admin/ bandwidth_management.htm, accessed March 2020. [4] J. Chu and C.-T. Lea, “Optimal link weights for maximizing QoS traffic,” Proc. IEEE International Conference on Communications (ICC), pp. 610–615, 2007. DOI: 10.1109/icc.2007.105 © IEICE 2020 [5] C.-Y. Hong, S. Mandal, M. Al-Fares, M. Zhu, R. Alimi, K. Naidu B., C. Bhagat, DOI: 10.1587/comex.2020XBL0123 Received September 3, 2020 S. Jain, J. Kaimal, S. Liang, K. Mendelev, S. Padgett, F. Rabe, S. Ray, M. Accepted September 10, 2020 Publicized September 24, 2020 Copyedited December 1, 2020

656 IEICE Communications Express, Vol.9, No.12, 656–661

Tewari, M. Tierney, M. Zahn, J. Zolla, J. Ong, and A. Vahdat, “B4 and after: managing hierarchy, partitioning, and asymmetry for availability and scale in Google’s software-deﬁned WAN,” Proc. ACM SIGCOMM, pp. 74–87, 2018. DOI: 10.1145/3230543.3230545

1 Introduction Video communication for teleconferencing has become widespread because of the evolution of high-efficiency compression and mobile communication technologies. Especially, video communication is indispensable for smooth remote work and education during pandemics such as COVID-19. In the current video communication architecture, the key component, called the multi-point control unit (MCU), is deployed in the central cloud server. MCU combines received multiple video streams into a single video stream. As the video resolution and the frequency of use increase, a big concern is that the bandwidth consumption in the backbone network is too large to provide stable video communication. Multi-access edge computing (MEC) architecture in the 5G backbone network is standardized by the ETSI [1]. Various functions that are originally deployed on the central cloud server are deployed on the edge nodes. Although many researchers are tackling a variety of topics that are related to MEC architecture [2], the existing literature focuses on mobile communication, not on video communication for teleconferencing. We focus on applying the MEC architecture to video communication because of its large potential for reducing the bandwidth consumption of the backbone network. However, it is necessary to solve the edge deployment and routing (EDR) problem to achieve sufficient bandwidth reduction, where the EDR problem is defined to optimize the placement of the edge nodes with the MCUs and the routing between the users. As the EDR problem is generally formulated as a mixed integer linear programming (MILP) problem, and is classified into NP-hard problems, it is difficult to solve the problem within a feasible computation time. Here, we propose a novel but simple method to solve the EDR problem of the MEC architecture within a feasible time frame. This study has two main contributions. First, we provided a model for applying the MEC architecture to video communication, and defined the problem called the EDR problem. Second, we provided the simple algorithm, and its effectiveness was demonstrated on a world-wide network.

2 Video streams for teleconferencing in MEC architecture Fig. 1 illustrates the video streams over the backbone network in the MEC architecture. The video streams are routed from the users to the edge nodes (ENs) with the MCUs, via subscriber nodes (SNs) where the users’ video streams are accommodated. © IEICE 2020 DOI: 10.1587/comex.2020XBL0123 Received September 3, 2020 Accepted September 10, 2020 Publicized September 24, 2020 Copyedited December 1, 2020

657 IEICE Communications Express, Vol.9, No.12, 656–661

Fig. 1. Video streams in the MEC architecture.

When the video streams from all users who join the video conference are accommodated in the MCU, such as users a and b, they are combined into a single video stream at the MCU, and then the video stream is sent back to the users. This is called local loopback. In contrast, when the video streams from the users who join the video conference are accommodated in the different MCUs, they are exchanged between the MCUs, where the video streams are compressed to thumbnailed video streams with a reduced bit rate, about 1/10 [3]. This is called traffic compression. In Fig. 1, when user B is the current main speaker, the video streams of users A and C are sent as thumbnailed videos between MCU1 and MCU2. Video streams have specific characteristics of being terminated at the MCUs on their routes. This is different from the typical characteristics of end-to-end mobile communication traffic.

3 Divide and merge algorithm 3.1 Overview We propose the divide and merge algorithm to solve the EDR problem within a feasible computation time. The idea is based on the characteristics that the video traﬃc is terminated at the ENs. The EDR problem is divided into two sub-problems under a practical assumption that the number and the location of the ENs are designed by the business policy. The ﬁrst sub-problem is to determine the appropriate network resources and routings for the video streams between the ENs and SNs. The second sub-problem is to determine the resource allocation and routing for the video streams between the ENs. We solved these sub-problems not by using the MILP but by using the LP formulation to reduce the computation time. Finally, the results of the sub-problems are merged to obtain the total bandwidth consumption in the backbone network. The LP formulation in the sub-problems is discussed in the following Sections. The proposed algorithm provides a sub-optimal solution that is the upper bound of the optimal solution that provides the minimum bandwidth consumption of the backbone network. This is because the bandwidth is sequentially determined in a stepwise manner. © IEICE 2020 DOI: 10.1587/comex.2020XBL0123 Received September 3, 2020 Accepted September 10, 2020 Publicized September 24, 2020 Copyedited December 1, 2020

658 IEICE Communications Express, Vol.9, No.12, 656–661

3.2 Video streams between ENs and SNs We assumed that the total ingress/egress traffic demand for each SN is estimated in advance, but the traffic matrix between the SNs is inexplicit because the video communication between the arbitrary users would be frequently used under the mature video communication environment. In addition, we assumed that the video streams at the SN may be adequately routed between the SN and multiple ENs. This assumption is realistic because an appropriate route per video stream can be designed using the segment routing over MPLS technique. The limitations of the MCU, such as the maximum capacity, are ignored here because we can reflect them if necessary. The LP is formulated as follows. ∑ minimize ρij · Cij (1a) (i, j)∈E

subject to ∑ ∑ st − st = st, = , xij xji f i s (1b) j:(∑i, j)∈E j:(∑j,i)∈E st − st = , , , , xij xji 0 i s t (1c) j:(i, j)∈E ∑ j:(j,i)∈E st ≤ ρ · , xij ij Cij (1d) (s∑,t)∈P f st · E s = Ds, (1e) t:(s,t)∈P

Eq. (1a) is an objective function minimizing the bandwidth consumption between SNs and ENs in the network G(V, E), where V is a set of nodes consist of SNs

and ENs, and E is a set of links. Cij is the capacity of (i, j) ∈ E, where ρij, 0 <= ρij <= 1, is the utilization rate of (i, j). Eqs. (1b) and (1c) are the constraints for the flow conservation. f st , a variable, represents the video traffic of (s,t) ∈ P, st where P is a set of logical paths between the SN and EN. xij , a variable, is the portion of the video stream from s to t that is routed through (i, j). Eq. (1d) indicates the capacity constraint of (i, j). Eq. (1e) states the constraint of the video stream bandwidth at the SN. E s indicates that s belongs to SN, where E s is 1 if s is SN. Ds, a given parameter, represents the total traffic demand at the SN, where Ds is 0 if s is EN.

3.3 Video streams between ENs We treat the video streams between the ENs as a hose model. This is because we design appropriate resources by using the total ingress/egress video streams at each EN under the assumption that the traffic matrix between the SNs is inexplicit. The total video stream at each EN is the sum of the user traffic that the EN directly accommodates and the traffic between the EN and SNs. All the video streams between the ENs are assumed to be thumbnailed to simplify the problem. ∑ r © IEICE 2020 minimize ρij · C , (2a) DOI: 10.1587/comex.2020XBL0123 ij Received September 3, 2020 (i, j)∈E Accepted September 10, 2020 Publicized September 24, 2020 Copyedited December 1, 2020

659 IEICE Communications Express, Vol.9, No.12, 656–661

subject to ∑ ∑ yuv − yuv = , = , ij ji 1 i u (2b) j∑:(i, j)∈E j∑:(j,i)∈E yuv − yuv = , , , v, ij ji 0 i u (2c) ∑j:(i, j)∈E j:(j,i∑)∈E πu · u + λv · v ≤ ρ · r , ij R ij C ij Cij (2d) u:(u,v)∈P v:(u,v)∈P yuv ≤ πu + λv , ij ij ij (2e)

Eq. (2a) is an objective function minimizing the bandwidth consumption between r ( , ) ∈ the ENs. Cij is the residual capacity of i j E after the bandwidth assignment between ENs and SNs. Eqs. (2b) and (2c) are the constraints for the ﬂow conservation, yuv ( , v) ∈ where ij , a variable, is the portion of the video traﬃc of a logical path u P from the EN u to the EN v that is routed through (i, j). Eqs. (2d) and (2e) state the capacity constraint of (i, j), where the hose model formulation is converted into a regular LP problem by using the duality theorem [4]. Ru is the total ingress video v v πu λv stream to u. C is the total egress video stream from . ij and ij are non-negative variables.

4 Performance evaluation We evaluated the effectiveness of the MEC architecture for the video communication in terms of the bandwidth consumption in the backbone network. The bandwidth consumption in the MEC architecture is computed using the proposed algorithm. We compared our results to the bandwidth consumption of a central cloud architecture, which is set as the benchmark. The performance on a world-wide network in Fig. 2 [5] is simulated. The traffic demand of video communication at each location is in proportion to the population density. In the MEC architecture, the traffic streams between the ENs are reduced to 1/8 by traffic compression. The bandwidth consumption between the ENs is evaluated under the worst condition where there are no local loopback. GLPK was used as the LP solver. Fig. 3 shows the bandwidth consumption in the backbone network. The shaded area indicates the bandwidth consumption between the SNs and ENs while the white

660 IEICE Communications Express, Vol.9, No.12, 656–661

area indicates that between the ENs. The bandwidth consumption in the MEC architecture with three MCUs is reduced by about 30% in comparison to that in the benchmark. Because the computed bandwidth in the MEC architecture is the upper bound of the optimal bandwidth, the results indicate that the MEC architecture is eﬀective in reducing the bandwidth consumption.

Fig. 3. Bandwidth consumption in the backbone network using the MEC and the central cloud architectures.

Although we omitted the detailed discussion of the computation time, we con- ﬁrmed that the time for calculating the bandwidth between the SNs and ENs and the bandwidth between the ENs was less than 1 minute for every evaluated model by using a general Intel CORE i7 CPU and 32 GB of memory.

5 Conclusion We proposed a sub-optimal algorithm to solve the EDR problem for video communication of teleconferencing in MEC architecture. With our algorithm, the bandwidth consumption of the MEC architecture is reduced by about 30% on a world-wide network in comparison to the current architecture. As the next works, we will tackle the optimal EDR problem and other sub-optimal solutions.

661 IEICE Communications Express, Vol.9, No.12, 662–667 Communication environment analysis of textile antenna using ray tracing method

Daisuke Yamanaka1, a) and Masaharu Takahashi1, b) 1 Graduate School of Engineering, Chiba University, 1–33 Yayoi-cho, Inage-ku, Chiba-shi, Chiba 263–8522, Japan a) [email protected] b) [email protected]

Abstract: Biological information monitoring systems are widely used to obtain and monitor the vital signs of patients in hospitals. However, the current system mainly uses the 420–450 MHz band, making it diﬃcult to manage channel switching, adding and removing equipment and so on, unless you are an expert in wireless communications. There is also the problem of having to carry a transmitter that is the same size as a small terminal. Therefore, we proposed a textile antenna using the 5.2 GHz band as a replacement for the system’s transmitter antenna. In this paper, we report the antenna characteristics of the transmitter antenna designed by the authors when it is attached to the human body, the received power results of the transmitter power by analyzing the textile antenna in a simulated real-world environment model. Keywords: ray-tracing method, textile antenna, dual polarized, patch antenna Classiﬁcation: Antennas and Propagation

References

[1] Nihon Kohden Corporation, “Patient monitoring,” https://eu.nihonkohden.com/ en/products/patientmonitoring, accessed on: Dec. 18, 2017. [2] Y. Nakatani and M. Takahashi, “Textile antenna for biological information monitoring,” International Symposium on Antennas and Propagation 2016, Okinawa, Japan, pp. 986–987, Oct. 2016. [3] H.S. Berger and H.M. Gibson, “Managing your hospital RF spectrum,” Biomed- ical Instrumentation & Technology, vol. 47, no. 3, pp. 193–197, May/June 2013. DOI: 10.2345/0899-8205-47.3.193 [4] D. Yamanaka and M. Takahashi, “5.2 GHz band textile antenna for biological information monitoring system,” 2018 IEEE International Symposium on An- tennas and Propagation and USNC-URSI Radio Science Meeting, Boston, U.S., pp. 1295–1296, July 2018. DOI: 10.1109/APUSNCURSINRSM.2018.8609143 [5] D. Yamanaka and M. Takahashi, “5.2 GHz band textile antenna for biological information monitoring,” IEICE Trans. Commun. (Japanese Edition), vol. J101- B, no. 7, pp. 584–591, July 2018. DOI: 10.14923/transcomj.2017JBP3057 [6] D. Yamanaka and M. Takahashi, “Investigation of the characteristics of a 5.2 © IEICE 2020 DOI: 10.1587/comex.2020XBL0116 GHz textile antenna on a human body,” 2019 Wireless Days (WD), Manchester, Received August 23, 2020 United Kingdom, pp. 1–4, 2019. DOI: 10.1109/WD.2019.8734212 Accepted September 15, 2020 Publicized September 30, 2020 Copyedited December 1, 2020

662 IEICE Communications Express, Vol.9, No.12, 662–667

[7] T. Nagaoka, S. Watanabe, K. Sakurai, E. Kunieda, S. Watanabe, M. Taki, and Y. Yamanaka, “Development of realistic high-resolution whole-body voxel models of Japanese adult male and female of average height and weight, and application of models to radio-frequency electromagnetic-ﬁeld dosimetry,” Phys. Med. Biol., vol. 49, no. 1, pp. 1–15, Jan. 2004. DOI: 10.1088/0031-9155/49/1/001

1 Introduction It is important to grasp a patient’s biological information in a hospital because of the sudden situation change may occur accidentally [1]. For here the vital signs are heart rate, blood pressure, respiratory rate, body temperature and the like. This information has been obtained from various sensors attached to the human body surface, and a transmitter which installed outside the body transmits that information to a bedside monitor or a central monitor at a nurse station [2]. When the transmitter falling beneath the bed would cause an unstable communication problem. The transmitter must always be carried around to satisfies the stable quality of radio communication. Besides, cables connecting the terminal and various sensors interfere with movement. Also, frequencies used in the current system are in the 420 to 450 MHz band, and operations are being administered to avoid interference by assigning and managing channels for each ward [2]. In addition, it is clinical engineering technicians, not radio professionals, who manage the wireless environment in actual medical practice. For them, the specialized tasks of managing channels on a ward-by-ward basis and adding or removing equipment are not easy. That is why we propose the use of wireless LANs as a solution to this problem. Meanwhile, in recent years, congestion and interference of 2.4 GHz band wireless LAN in hospitals have been reported. It is because of the popularization of devices equipped with wireless LAN for medical staff and patients and an increase in medical devices using wireless LAN [3]. However, there are almost no investigations as a single package on the design of an antenna for the off-body network, analysis of characteristics when installed close to the human body, and 5.2 GHz indoor propagation analysis by ray-tracing method when using its antenna for off-body communication. In our previous study [4, 5] and [6], a textile antenna was proposed as a transmitting antenna for a biological monitoring system using the 5.2 GHz band which is not congested [3] show in Fig. 1. In this paper, we report the results of antenna characteristic analysis when the proposed textile antenna is placed on a high-definition human body model [7], and the results of received power analysis using the ray-tracing method in real environment models.

2 Analysis of a received power using the ray-tracing method Fig. 2 is ray-tracing simulation models that simulated a hospital room, and Fig. 3 (a) shows the result of calculating the received power. The textile antenna in Fig. 1 are used for the transmitting textile antennas (Tx), and a half-wavelength dipole antenna is used for the receiver antenna (Rx). Tx are ﬁxed on the chest and the back of the © IEICE 2020 DOI: 10.1587/comex.2020XBL0116 human body, which height is 1.2 m. Received August 23, 2020 Accepted September 15, 2020 Publicized September 30, 2020 Copyedited December 1, 2020

663 IEICE Communications Express, Vol.9, No.12, 662–667

Fig. 1. The textile antenna and the realized gain

The designed transmitting textile antenna is shown in Fig. 1. In this antenna, using a conductive fabric (conductive cloth/electromagnetic shielding cloth MK- KTN 260 manufactured by Tanimura Co., Ltd.) for both a radiating element and a 6 ground plane. A conductivity σe set to 8.29 × 10 S/m for the conductive fabric in numerical simulation. We used a felt textile for a dielectric substrate; we set its relative permittivity to 1.36. In the process of designing this antenna, the dimensions of the radiating element at 5.25 GHz, the desired frequency, were calculated in the initial stage of the antenna design. We did not use a coaxial power supply from the grounding side, which requires a power supply structure on the back of the antenna. This is because it is assumed that the grounding side of the antenna is used to face the human body. Therefore, we have adopted a method to supply power from the same plane as the radiation elements using MSL (Micro Strip Line). Thus, the resonance frequency was adjusted by adjusting the path length of the current flowing through the radiation element in a right-angled upward direction. The end of the MSL is set near the endpoint of the radiating element. For impedance matching, slits 1 mm wide and 10 mm long were made along the MSL. The characteristics of the antenna were evaluated by the FDTD simulation and measurement. As a result, it was confirmed that the present antenna emits two orthogonal polarizations through the single feed point. In addition, it was confirmed that the double polarization is effective in RSSI measurement to respond to the patient’s posture change [5]. For the receiver antenna, this is because most wireless LAN access points use dipole or monopole antennas. Therefore, in this numerical analysis, we use a

© IEICE 2020 numerical antenna model with characteristics almost similar to those of the actual DOI: 10.1587/comex.2020XBL0116 Received August 23, 2020 receiver. The imaging method was used for the ray-tracing simulation. The number Accepted September 15, 2020 Publicized September 30, 2020 Copyedited December 1, 2020

664 IEICE Communications Express, Vol.9, No.12, 662–667

Fig. 2. Ray tracing model

of reflections for each wall of x, y, and z was set to 3 in all, and the maximum number of diffraction in each direction was set to 1. In these simulated conditions, the sum of the generated paths results in 340. The number of reflections and the number of passes are the minimum number of times that the analysis results converge in this model. The FDTD program has calculated both Tx and Rx antennas radiation patterns. The values from these radiation patterns have used to calculate the ray- tracing according to the ray’s angle. This received power is defined as the average received power through the receiving antenna port. The received power included all direct, reflected, and diffracted wave paths, and the sum of the ray power has calculated. The total received power from the chest and back antennas when the receiving antenna was fixed at y = 0 m and the human body with transmitting antennas were moved from 0.0 to 7.0 m in the y direction is shown in Fig. 3 (a). The total received power was greater than −70 dBm at all points, and the variation of received power was very small. Furthermore, it was also observed that the received power decayed as it approached the window side. The reason for this is that the reflective and transmission characteristics of concrete and glass are different. The © IEICE 2020 results indicate that diversity can be transmitted using these chest and back antennas DOI: 10.1587/comex.2020XBL0116 Received August 23, 2020 for biological monitoring in a real-world environment and that stable communication Accepted September 15, 2020 Publicized September 30, 2020 Copyedited December 1, 2020

665 IEICE Communications Express, Vol.9, No.12, 662–667

Fig. 3. Calculated received power

is possible. The results of the received power analysis when the human body is ﬁxed at d = 5.0 m, and the receiving antenna was displaced horizontally y direction from 0 m to 7.2 m are shown in Fig. 3 (b) and (c). At the chest antenna, it has been shown © IEICE 2020 DOI: 10.1587/comex.2020XBL0116 that the power can be transmitted at an almost constant rate regardless of the position Received August 23, 2020 Accepted September 15, 2020 Publicized September 30, 2020 Copyedited December 1, 2020

666 IEICE Communications Express, Vol.9, No.12, 662–667

of the receiving antenna. Again, it can be observed that the received power level decays as it approaches the window side. On the dorsal side, the standing position of the human body is y = 5 m, which indicates that almost all points are in the NLOS environment and the received power intensity is low. Further evaluated the received power when the patient was lying down with the model in Fig. 2 (b). In the model, the bed is made up of PEC (Perfect Electric Conductor) and felt ﬁbers, and these electrical characteristics are set to be the same as in the antenna model. The textile antenna placed on the chest and back of the human model is described in Fig. 1. This is an approach to ensure a reliable wireless connection, regardless of the patient’s posture. The results of the received power analysis of the supine model in the supine position are the thoracic antenna is −59.4 dBm, and the dorsal antenna is −100.6 dBm. You can’t communicate with the back antenna alone, but it can be seen that it is possible to communicate well enough with the chest antenna. These results shows the importance of using multiple textile antennas in a distributed manner because it is diﬃcult to achieve stable communication with a single textile antenna alone.

3 Conclusion From the results, we can conclude that in a hospital room, when the receiving antenna is a half-wavelength dipole, the textile antenna attached to the chest or back of the human body can provide suﬃcient received power strength of more than −70 dB, regardless of the human standing position. By operating these two antennas as a diversity of antennas, it is possible to ensure stable communication. Furthermore, we have found that even when lying down, received power can be provided by at least one textile antenna in the LOS environment. In the future, we plan to analyze the received power in lying down and seated conditions.

667 IEICE Communications Express, Vol.9, No.12, 668–673 Good reversible quasi-cyclic codes via unfolding cyclic codes

Ramy Taki ElDin1, a) and Hajime Matsui1 1 Toyota Technological Institute, 2–12–1 Hisakata, Tempaku, Nagoya, Aichi 468–8511, Japan a) [email protected]

Abstract: In this paper, we consider the reversibility problem in the class

of quasi-cyclic (QC) codes Q over Fq of length nℓ and index ℓ generated C F by unfolding cyclic codes over qℓ of length n. We prove a necessary and sufficient condition on C that ensures the reversibility of Q. Using computer search, we offer some good reversible QC codes that are generated by unfolding cyclic codes. Keywords: reciprocal polynomial, binary code, computer search, best known parameters, minimum distance Classification: Fundamental Theories for Communications

References

[1] R. Taki ElDin and H. Matsui, “Run-length constraint of cyclic reverse- complement and constant GC-content DNA codes,” IEICE Trans. on Funda- mental of Electronics, Communications and Computer Sciences, vol.E103-A, no.1, pp.325–333, Jan. 2020. DOI: 10.1587/transfun.2019eap1053 [2] R. Taki ElDin and H. Matsui, “Quasi-cyclic codes via unfolded cyclic codes and their reversibility,” IEEE Access, vol.7, pp.184500–184508, Dec. 2019. DOI: 10.1109/access.2019.2960569 [3] M. Grassl, “Bounds on the minimum distance of linear codes and quantum codes,” 2020. [Online] Available: http://www.codetables.de/ [4] J.L. Massey, “Reversible codes,” Information and Control, vol.7, no.3, pp.369– 380, Sep. 1964. DOI: 10.1016/s0019-9958(64)90438-3

1 Introduction

Let Fq denote a ﬁnite ﬁeld of q elements. A cyclic code over Fq of length n is a linear Fn subspace of q invariant under the cyclic shifts of its codewords. A quasi-cyclic F ℓ ℓ Fnℓ (QC) code over q of length n and index is a linear subspace of q invariant under cyclic shifts of its codewords by ℓ coordinates. Let Γ denote the class of QC F ℓ ℓ F codes over q of length n and index generated by unfolding cyclic codes over qℓ of length n. By unfolding we mean a one-to-one map φθ that represents elements F Fℓ { , θ, . . . , θℓ−1} θ of qℓ by vectors in q using the basis 1 , where is a root of an © ( ) ∈ F [ ] ℓ F = F (θ) IEICE 2020 irreducible polynomial p x q x of degree , i.e., qℓ q . A code is called DOI: 10.1587/comex.2020XBL0117 Received August 24, 2020 reversible if it is invariant under reversing the coordinates of codewords. Reversible Accepted September 17, 2020 Publicized October 1, 2020 Copyedited December 3, 2020

668 IEICE Communications Express, Vol.9, No.12, 668–673

codes are important in some applications, e.g., DNA codes [1]. In this paper, we extend our work [2] by presenting some good reversible QC codes in Γ. Here, by a good code, we mean an optimal or suboptimal code, where we say a code is optimal (resp. suboptimal) if its minimum distance is the optimal value (resp. the optimal value minus one or two) provided by [3]. In Theorem 1, we prove a necessary and sufficient condition for the reversibility of QC codes generated F by unfolding cyclic codes over qℓ . Theorem 1 in this paper is more useful than Theorem 3 in [2] because it does not assume a predetermined θ, does not require to build the defining set of the cyclic code, and shows that the self-reciprocity of the minimal polynomial of θ is a sufficient condition for the reversibility but not necessary. As an application of Theorem 1, we present in Table I some good reversible QC codes in Γ of even ℓ obtained by computer search. The rest of this paper is organized as follows. Preliminaries are summarized in Section 2. We show the main contribution in Section 3. Our work is concluded in Section 4.

2 Preliminaries C F We denote by a cyclic code over qℓ of length n and dimension k, where n is C F [ ]/⟨ n − ⟩ coprime to q. That is, is an ideal in qℓ x x 1 generated by a generator polynomial g(x) that divides xn − 1. Massey [4] showed that a cyclic code is reversible if and only if g(x) is self-reciprocal, i.e., g∗(x)|g(x), where g∗(x) = deg(g(x))g( / ) g( ) θ ∈ F x 1 x is the reciprocal polynomial of x . Let qℓ be a zero to an p(x) ∈ F [x] ℓ α ∈ F ℓ irreducible∑ polynomial q of degree . Any element q is written α = ℓ−1 θ j ∈ F ≤ ≤ ℓ − as j=0 aj for uniquely determined aj q (0 j 1). Hence, the = (α , α , . . . , α ) ∈ C α ∈ F ≤ ≤ − codeword c 0 1 n−1 , where i qℓ (0 i n 1), is represented by the polynomial ∑n−1 ∑n−1 ∑ℓ−1 i i © jª c(x) = αi x = x ai, j θ ®, (1) i=0 i=0 « j=0 ¬ ∑ ∈ F α = ℓ−1 θ j ≤ ≤ − ≤ ≤ ℓ − where ai,j q and i j=0 ai,j for 0 i n 1 and 0 j 1. φ F → Fℓ Deﬁnition 1. Deﬁne the map θ : qℓ q such that ∑ℓ−1 j φθ : α = aj θ 7→ (a0, a1,..., aℓ−1). j=0 −1 We refer to φθ as the unfolding map, while φθ : (a0, a1,..., aℓ−1) 7→ α is the folding map. The unfolding map φθ can be applied to codewords of C as follows

−1 φθ c = (α0, α1, . . . , αn−1) ∈ C ⇆ (a0,0,..., a0,ℓ−1,...... , an−1,0,..., an−1,ℓ−1). (2) φθ

From (2), the reverse of φθ (c) is the unfolding of the word r ∈ Fn which is given in qℓ the polynomial form by

∑n−1 ∑ℓ−1 n−1 ℓ−1 −i © −jª r(x) = x θ x ai, j θ ® . (3) i=0 « j=0 ¬ © IEICE 2020 C F DOI: 10.1587/comex.2020XBL0117 Unfolding a cyclic code over qℓ of length n and dimension k generates a QC Received August 24, 2020 Accepted September 17, 2020 code, which we refer to as Q = φθ (C), over Fq of length nℓ, index ℓ, and dimension Publicized October 1, 2020 Copyedited December 3, 2020

669 IEICE Communications Express, Vol.9, No.12, 668–673

kℓ. In fact, Q is an Fq-linear code invariant under cyclic shifts by ℓ coordinates. That is,

(a0,0,..., a0,ℓ−1, a1,0,..., a1,ℓ−1,..., an−1,0,..., an−1,ℓ−1) ∈ Q

=⇒ (an−1,0,..., an−1,ℓ−1, a0,0,..., a0,ℓ−1,..., an−2,0,..., an−2,ℓ−1) ∈ Q.

Let Γ be the class of QC codes over Fq of length nℓ and index ℓ generated by F unfolding cyclic codes over qℓ of length n.

3 Good Reversible QC Codes in Γ C F Hereinafter, denotes a cyclic code over qℓ of length n, dimension k, and generator polynomial g(x). Moreover, Q = φθ (C) refers to a QC code in the class Γ over Fq of length nℓ and dimension kℓ. The following theorem gives a necessary and suﬃcient condition for reversibility of QC codes in Γ. Contrary to [2, Theorem 3], Theorem 1 does not require constructing the deﬁning set of the cyclic code C.

Theorem 1. Let C be a cyclic code over F ℓ = F (θ) of length n and dimension ∑ q q n−k i k. Let g(x) = α x ∈ F ℓ [x] be the monic generator polynomial of C, where ∑ i=0 i q α = ℓ−1 θ j ∈ F Q = φ (C) F ℓ i j=0 aij for some aij q. The QC code θ over q of index , length nℓ, and dimension kℓ is reversible if and only if any of the following conditions is true:

A) g(x) is a self-reciprocal polynomial over Fq. B) p(x) is a self-reciprocal polynomial and g(x)|bg∗(x), where

σ F F [ ]/⟨ n − ⟩ σ(θ j i) = Proof. Let be the q-linear transformation on qℓ x x 1 such that x θ−j x−i for 0 ≤ j ≤ ℓ − 1 and 0 ≤ i ≤ n − 1. Informally, σ corresponds to the reverse of the codewords in Q, and in particular ( ) ∑n−k ∑n−k ∑ℓ−1 ∑n−k ∑ℓ−1 i © j iª j i σ(g(x)) = σ αi x = σ aij θ x ® = σ(aij θ x ) i=0 « i=0 j=0 ¬ i=0 j=0 ∑n−k ∑ℓ−1 ∑n−k ∑ℓ−1 j i −j −i = aij σ(θ x ) = aij θ x . i=0 j=0 i=0 j=0

∗ n−k i −i bg (x) = x σ(g(x)) σ(αx ) = σ(α)x α ∈ F ℓ Thus . Similarly, for∑ any q and ≤ ≤ − ( ) ∈ C ( ) = k−1 β ι 0 i n 1. For any c x , there exists b x ι=0 ι x such that ( ) = g( ) ( ) β ∈ F ≤ ι ≤ − c x x b x , where ι qℓ for 0 k 1. In addition, (1) and (3) show that r(x) = xn−1θℓ−1σ (c(x)). Hence ( ) ∑n−k ∑k−1 n−1 ℓ−1 n−1 ℓ−1 i ι r(x) = x θ σ(g(x)b(x)) = x θ σ αi x βι x i=0 ι=0 ∑n−k ∑k−1 ∑n−k ∑k−1 n−1 ℓ−1 i+ι n−1 ℓ−1 −i −ι © IEICE 2020 = x θ σ(α βι x ) = x θ x σ(α βι)x . DOI: 10.1587/comex.2020XBL0117 i i Received August 24, 2020 i=0 ι=0 i=0 ι=0 Accepted September 17, 2020 Publicized October 1, 2020 Copyedited December 3, 2020

670 IEICE Communications Express, Vol.9, No.12, 668–673

Firstly, we show that condition A implies the reversibility of Q. Let g(x) α ∈ F ≤ i ≤ n − k σ be self-reciprocal with coefficients i q for 0 ∑ ∑ . Since is − ℓ− n−k − k−1 −ι F -linear, σ(α βι) = α σ(βι) and r(x) = xn 1θ 1 α x i σ(βι)x = q i ∑ i i=0 i ι=0 n−1θℓ−1g( −1) k−1 σ(β ) −ι g∗( )| ( ) g( ) x x ι=0 ι x . Consequently, x r x . But x is self- reciprocal, then g(x)|r(x), r(x) ∈ C, and Q is reversible. Secondly, we show that condition B implies the reversibility of Q. If p(x) is ℓ θ−1 ( ) σ F self-reciprocal, then is even and is a root of p x . The restriction of to qℓ is −1 the field automorphism θ 7→ θ that fixes Fq. Thus σ(αi βι) = σ(αi)σ(βι). Using σ(α)x−i = σ(αxi) and xn−k σ(g(x)) = bg∗(x), we get ( ) ∑n−k ∑k−1 ∑n−k ∑k−1 n−1 ℓ−1 −i −ι n−1 ℓ−1 i −ι r(x) = x θ σ(αi)x σ(βι)x = x θ σ αi x σ(βι)x i=0 ι=0 i=0 ι=0 ∑k−1 ∑k−1 n−1 ℓ−1 −ι k−1 ℓ−1 ∗ −ι = x θ σ(g(x)) σ(βι)x = x θ bg (x) σ(βι)x . ι=0 ι=0

Since g(x)|bg∗(x), we have g(x)|r(x), r(x) ∈ C, and Q is reversible. Finally, we show that the reversibility of Q implies A or B. Since Q is reversible, g(x)|r(x) b(x) b(x) = r(x) = then ∑ for every choice of ∑. By choosing 1,( we∑ get ) n−1θℓ−1 n−k σ (α ) −i = n−1θℓ−1 n−k σ(α i) = n−1θℓ−1σ n−k α i = x i=0 i x x i=0 i x x i=0 i x xk−1θℓ−1 xn−k σ(g(x)) = xk−1θℓ−1bg∗(x). This proves that g(x)|bg∗(x). In addition, we distinguish the following two cases:

a) If αi ∈ Fq for all 0 ≤ i ≤ n − k, then bg(x) = g(x). The self-reciprocity of g(x) is a consequence of g(x)|bg∗(x). Condition A follows. b) α < F ≤ h ≤ n − k a , < t ≤ If h q for some∑ 0 , there exists ht 0 with 0 ℓ − α = t θ j ( ) = θℓ−t ( ) = 1 such that h j=0 ahj . By choosing b x , we have r x xk−1θℓ−1 xn−k σ(θℓ−t g(x)) is a codeword of C. Thus, g(x)|xn−k σ(θℓ−t g(x)). In fact, deg(bg∗(x)) = deg(xn−k σ(θℓ−t g(x))) = n − k which is the least polynomial degree in C. Moreover, the constant terms of bg∗(x) and xn−k σ(θℓ−t g(x)) are 1 and σ(θℓ−t ), respectively. Therefore, xn−k σ(θℓ−t g(x)) = σ(θℓ−t )bg∗(x). In particular, equating n−k−h ℓ−t ℓ−t the coeﬃcients of x yields σ(θ αh) = σ(θ )σ(αh). Consequently,

∑t−1 ∑t−1 ∑t −ℓ+t−j ℓ © ℓ−t+jª ℓ © ℓ−t+jª ahj θ + aht σ(θ ) = σ ahj θ ® + σ(aht θ ) = σ ahj θ ® j=0 « j=0 ¬ « j=0 ¬ ∑t ∑t−1 ℓ−t ℓ−t −ℓ+t −j −ℓ+t−j −ℓ = σ(θ αh) = σ(θ )σ(αh) = θ ahj θ = ahj θ + aht θ . j=0 j=0

σ(θℓ) = θ−ℓ From the outermost∑ sides of the last equation, we conclude . Let ℓ ℓ−1 v p(x) = x + pv x be the monic minimal polynomial of θ over F . Then, 0 = ( v=∑0 ) ∑ ∑ q σ( ) = σ θℓ + ℓ−1 θv = σ(θℓ) + ℓ−1 θ−v = θ−ℓ + ℓ−1 θ−v = (θ−1) 0 v=0 pv v=0 pv v=0 pv p . Thus, θ−1 is a root of p(x), p(x) is self-reciprocal, and condition B follows. □

Example 1. Let C be the cyclic code over F4 of length n = 43, dimension k = 35, 2 3 5 6 8 © IEICE 2020 and generator polynomial g(x) = 1 + θx + (1 + θ)x + θx + (1 + θ)x + x ∈ F4[x], DOI: 10.1587/comex.2020XBL0117 θ ∈ F ( ) = + + 2 ∈ F [ ] φ F → F2 Received August 24, 2020 where 4 is a zero to p x 1 x x 2 x . The map θ : 4 2 Accepted September 17, 2020 Publicized October 1, 2020 Copyedited December 3, 2020

671 IEICE Communications Express, Vol.9, No.12, 668–673

unfolds C to a binary QC code Q in Γ of index ℓ = 2, length nℓ = 86, and dimension kℓ = 70. The minimum distance of Q is dQ = 6. According to [3], a linear binary code of length 86, dimension 70, and minimum distance 6 is optimal. Moreover, Q is reversible as it meets condition B of Theorem 1. Hence, Q is an optimal reversible code in Γ.

Other examples of good binary reversible QC codes in Γ of even index are shown in Table I, where all suboptimal codes have their minimum distances which

are equal to the optimal values minus two provided by [3]. In this table, p(x) ∈ F2[x] θ ∈ F φ is the monic minimal polynomial of 2ℓ that deﬁnes θ . Using condition B of Theorem 1, one can verify the reversibility of the QC codes listed in Table I.

Remark 1. It is derived from Theorem 1 that, for odd index, Q = φθ (C) is reversible

code if and only if g(x) is a self-reciprocal polynomial over Fq.

Remark 2. As far as we have searched, all good codes satisfy condition B and only one of them also satisfies condition A in Theorem 1. It seems that good reversible QC codes generated by unfolding are likely to satisfy condition B instead of condition A. One reason for this observation is that, under condition A, dQ ≤ wt(g(x)) holds, where dQ denotes the minimum distance of Q and wt(g(x)) denotes the number of nonzero coefficients of g(x). On the other hand, under condition B, this upper bound for dQ does not hold, e.g., the first code in Table I satisfies condition B but dQ = 8 > wt(g(x)) = 6. On a lower bound for dQ, in general we have dQ ≥ dC, which suggests that larger dC tends to increase dQ.

4 Conclusion Theorem 1 proposed an equivalent condition for the reversibility of QC codes generated by unfolding cyclic codes over an extension ﬁeld. Applying this theorem to computer search, we found various good reversible QC codes in Γ as shown in Table I.

Acknowledgments Ramy Taki ElDin is on leave of absence from Faculty of Engineering, Ain Shams University, Cairo, Egypt. This work was supported in part by the JSPS KAKENHI under Grant JP19K22850.

672 IEICE Communications Express, Vol.9, No.12, 668–673 Opt./Subopt. Optimal Suboptimal Optimal Optimal Suboptimal Optimal Suboptimal Suboptimal Suboptimal Suboptimal Suboptimal Suboptimal ] ] ] ] 6 10 10 ] ] ] ] ] ] ] ] , , , 8 6 6 4 10 6 6 6 5 , , , , , , , , , 90 96 70 Q , , , 10 44 70 12 24 36 24 36 24 , , , , , , , , , 22 66 86 20 52 52 42 54 114 40 136 110 Parameters [ [ [ [ [ [ [ [ [ [ [ [ 7 + x 2 + θ 6 x + ) θ 4 3 ( x θ + + + 2 2 3 θ x ( x ) ) 7 + 3 θ 5 θ x + ) + Γ 4 3 5 x 2 θ θ θ in + + + 2 ℓ + 4 3 θ θ x θ + ) + 4 1 + ( θ 2 1 + θ ( 4 + 4 3 x + x 3 + x ) θ θ 2 + 3 ( + x 11 θ + 3 ) 2 x + x + 5 θ x ) 2 x θ ( θ ) + x 8 ) θ 5 5 ( ( 6 7 x + + + θ x g 10 x θ + 2 2 4 x + 3 + + x x θ + + x ) ) 6 4 + θ 4 ) 4 3 5 x + x 3 θ 8 ) θ θ x ) + 2 θ x θ θ 4 + θ + + 3 1 + θ ( 3 + + x + 2 2 2 + θ θ + 7 θ θ 4 1 + ( + θ ( x x 3 + ( x + + + 2 2 θ θ ) 2 + + 2 x x + θ 1 5 θ ) x 5 ( + + 6 ( θ x 5 ) + x + x 2 2 ) 3 θ + + θ + 6 θ x x θ 1 ) + ) x ) ( θ + 3 7 + + 5 ) + θ 5 θ 4 2 3 θ + θ + 3 x 2 θ θ θ θ + x + x 5 x ( + + ) Good binary reversible QC codes of even + + ) + θ 1 + θ 6 θ 3 4 ( + + 1 4 θ 1 2 ( θ θ + x + ( x x + + θ + + 4 ) ) + + 1 1 + 5 3 x x + θ 5 ( + ( ) ) 2 2 θ θ 3 θ 3 3 θ θ θ θ + + x + x θ + + + ) 3 ) Table I. + + + + 2 7 + 4 3 + θ 4 1 x 2 θ ( ( θ θ 1 1 1 1 θ x θ ( ( ( ( ( θ ( + + + + + + + + + + + + + 3 6 6 θ 1 θ θ θ 1 1 1 1 1 1 ( 1 1 ( 8 8 x x + + 7 7 x x + + 6 6 x x ) + + x 4 4 ( 6 6 6 x x p 2 2 2 x x x + + x x x 2 2 + + + i i i i x x + + + x x x x 3 3 3 + + x x x x x x 0 0 0 0 x x = = = = 4 i 4 i 4 i 10 i + + + + + + + + 1 1 1 ∑ ∑ ∑ 1 1 1 1 1 ∑ k 5 3 6 9 4 6 3 7 22 35 15 12 n 5 7 9 5 11 33 43 13 13 19 17 11 © IEICE 2020 DOI: 10.1587/comex.2020XBL0117 ℓ 2 2 2 4 4 4 6 6 6 8 8

Received August 24, 2020 10 Accepted September 17, 2020 Publicized October 1, 2020 Copyedited December 3, 2020

673 IEICE Communications Express, Vol.9, No.12, 674–678 The attenuation characteristics of millimeter-wave by snow accretion

Kazuki Nakamura1, 2, a), Nagateru Iwasawa1, Kunihiro Kawasaki1, Shotaro Yoshida3, and Masaharu Takahashi4 1 Signalling and Transport Information Technology Division, Railway Technical Re- search Institute 2–8–38, Hikari-cho, Kokubunji-shi, Tokyo 185–8540, Japan 2 Graduate School of Science and Engineering, Chiba University 1–33 Yayoi-cho, Inage-ku, Chiba-shi, Chiba 263–8522, Japan 3 Signalling and Communications Division, JR Soken Electric Consulting 2–8–38, Hikari-cho, Kokubunji-shi, Tokyo 185–0034, Japan 4 Center for Frontier Medical Engineering, Chiba University 1–33 Yayoi-cho, Inage-ku, Chiba-shi, Chiba 263–8522, Japan a) [email protected]

Abstract: The millimeter-wave band is being focused as a frequency band in which high-speed and large-capacity communication can be realized, and various approaches for practical use have been made. In this paper, we studied the attenuation characteristics of the 45 GHz with the moisture contents and thickness of snow, which had never been studied and is expecting to be used in next-generation train communication systems and 5G. The characteristics are also compared with ones of 60 GHz which can be used for new Wi- Fi standard (IEEE 802.11ad). As the result, it is found that when snow accretes to the surface of the antenna radome, the attenuation increases as the moisture content increases and we can not neglect the inﬂuence on wireless communication using millimeter-waves. Keywords: snow, millimeter-wave, attenuation, thickness, moisture content Classiﬁcation: Antennas and Propagation

References

[1] G. Yue, D. Yu, L. Cheng, Q. Lv, Z. Luo, Q. Li, J. Luo, and X. He, “Millimeter- wave system for high-speed train communications between train and trackside: system design and channel measurements,” IEEE Trans. Veh. Technol., vol. 68, no. 12, pp. 11746–11761, Dec. 2019. DOI: 10.1109/TVT.2019.2919625 [2] K. Nakamura, N. Iwasawa, K. Kawasaki, N. Shibagaki, Y. Sato, and K. Kashima, “Study of the new application using the millimeter-wave in the railway,” Proc. Int. Conf. 2017 IEEE CAMA, MB1-2, pp. 20–23, Dec. 2017. DOI: 10.1109/CAMA.2017.8273400 [3] A. Kanno, P.T. Dat, N. Yamamoto, T. Kawanishi, N. Iwasawa, N. Iwaki, K. Nakamura, K. Kawasaki, N. Kanada, N. Yonemoto, Y. Sato, M. Fujii, K. © IEICE 2020 Yanatori, N. Shibagaki, and K. Kashima, “High-speed railway communication DOI: 10.1587/comex.2020XBL0128 system using linear-cell-based radio-over-ﬁber network and its ﬁeld trial in 90- Received September 8, 2020 Accepted September 28, 2020 Publicized October 8, 2020 Copyedited December 1, 2020

674 IEICE Communications Express, Vol.9, No.12, 674–678

GHz bands,” J. Lightw. Tech., vol. 38, no. 1, pp. 112–122, Jan. 2020. DOI: 10.1109/JLT.2019.2946691 [4] ITU-R Recommendation P.838-3, “Speciﬁc attenuation model for rain for use in prediction methods,” 2015 [5] Y. Takimoto and M. Kotaki, “Millimeter-wave characteristics of fallen snow,” Electronics Communications in Japan, Part 1, vol. 82, no. 4, pp. 33–44, April 1999. DOI: 10.1002/(sici)1520-6424(199904)82:4<33::aid-ecja4>3.0.co;2-3 [6] ITU-R Report M.2442-0, “Current and future usage of railway radiocommuni- cation systems between train and trackside,” 2018 [7] M. Shimba, T. Sato, H. Koike, and K. Sato, “Signal level degradation due to snow accretion on a radome,” Electronics Letters, vol. 23, no. 14, pp. 739–741, July 1987. DOI: 10.1049/el:19870524 [8] G. Wakahama and Y. Mizuno, “Studies on tensile strength of wet snow,” Trans- portation Research Board Special Report, vol. 185, pp. 18–22, June 1979. [9] K. Kawashima, T. Endo, and Y. Takeuchi, “A portable calorimeter for measuring liquid-water content of wet snow,” Annals of Glaciology, vol. 26, pp. 103–106, 1998. DOI: 10.3189/1998AoG26-1-103-106

1 Introduction In recent years, the various systems utilizing millimeter-waves are being studied for railways [1, 2, 3]. It is known from previous studies that the propagation characteristics of the wireless communication systems using millimeter-waves are aﬀected by rain [4] and snow [5]. In particular, on the assumption of using millimeter- wave radar [5], the characteristics of snow had been studied for 50–75 GHz band. However it was not a study focusing on the moisture content. Furthermore, no studies have been done on the eﬀect of snow against 40 GHz band millimeter-wave, which is expected to be used in next-generation train communication systems and 5G. In this paper, we report the results of measurement and comparison of the attenuation, assuming the case that snow accrete to the surface of the antenna radome at the 45 GHz, which is being considered for next generation ground-to-train communication system [6], and the results are compared with 60 GHz, which can be used without a license.

2 The measurement method of attenuation of snow accretion The measurement tests of attenuation of snow accretion were carried out in a cold room in which the temperature was kept at −5 to 0 degrees Celsius. The test conﬁguration is shown in Fig. 1. The antennas used for the transmitter and the receiver are both standard horn antennas with the gain of 23.5 dBi and the half- power beam width of E-plane and H-plane are approximately 10 degrees. The measurement was done using CW waves of linear polarization of 45 GHz and 60 GHz. The snow plates with various moisture content and thickness were set on the acrylic plate between the antennas, assuming snow accrete to the surface of the antenna radome. The acrylic plate was intended to be a radome, not just a snow support. The size of the snow plate was 400 mm in length and 400 mm in © IEICE 2020 DOI: 10.1587/comex.2020XBL0128 width, and the height (thickness) was prepared from 10 to 80 mm, referencing to the Received September 8, 2020 Accepted September 28, 2020 results of the previous study [7]. The size of the exposure surface of this snow plate Publicized October 8, 2020 Copyedited December 1, 2020

675 IEICE Communications Express, Vol.9, No.12, 674–678

is sufficiently large compared to the exposure area calculated from the half-power beam width of the antenna, so it is considered that the wraparound can be ignored. In addition, the plastic racks, which have almost no effect on radio propagation, are used to set the antennas and a snow plate. Furthermore, the setting distance is 225 mm, which is more than twice the Fresnel radius (44.4 mm @ 45 GHz). The moisture content of snow was set to 5.5%, 10%, and 14.6% by adding water to the natural snow stored in winter, referring to the previous study, where 10–15% is said to have the highest adhesion strength [8]. In the test, the receiving power was measured through the snow plate with varying thickness and moisture content. The amount of attenuation was defined as the difference between the receiving power when the snow plate was placed and not placed.

Fig. 1. Measurement conﬁguration

3 Measurement and analysis results The test results and approximation lines in terms of attenuation versus thickness of snow are shown in Fig. 2, and in terms of attenuation versus thickness of snow are shown in Fig. 3. The moisture contents shown in these results were measured by the melting calorimetry method used by an Endo-type snow water content meter [9]. In the range of moisture contents and thicknesses of snow where the tests were done, when the moisture contents of snow are same, the attenuation increases as the thickness increases. And the slope of attenuation with increasing thickness of snow is almost the same at 45 GHz and 60 GHz, but the attenuation at 60 GHz is larger than at 45 GHz, in Fig. 2. In Fig. 3, the difference in the amount of attenuation due to the differences in moisture contents of snow are that the thinner the thickness of snow, the larger the slope of the approximation line, and more significant at 45 GHz than at 60 GHz. On the other hand, for the same snow thickness exceed 40 mm, the amount of attenuation at 60 GHz almost does not change even if the moisture content © IEICE 2020 DOI: 10.1587/comex.2020XBL0128 increases. Further, it can be seen that as the moisture contents of snow becomes Received September 8, 2020 Accepted September 28, 2020 higher, the difference in the attenuation amount, depending on the thickness and Publicized October 8, 2020 Copyedited December 1, 2020

676 IEICE Communications Express, Vol.9, No.12, 674–678

frequency of the snow, becomes smaller and tends to converge. Considering the above results and the thickness of snow accretion to the antenna surface was about 70 mm in the previous research (in the case of a Cassegrain antenna of 1 m diameter with an elevation angle of 27.5 degree) [7], it is necessary to assume that the attenuation is about 45–50 dB due to the eﬀect of snow accretion.

Fig. 2. Measurement result of attenuation vs thickness of snow

Fig. 3. Measurement result of attenuation vs moisture content of snow

677 IEICE Communications Express, Vol.9, No.12, 674–678

4 Conclusion In this paper, to contribute to designing a ground-to-train communication system using millimeter-wave, we studied the attenuation due to adhesion of snow with varying water content and thickness at 45 GHz, which had not been studied. This result is also compared with the measurement results of 60 GHz. As a result, it is found that the received power is significantly attenuated by the containing moisture in the snow on the surface of the antenna. For example, the maximum attenuation at 45 GHz is 45 dB at a thickness of 80 mm and a moisture content of about 15%. It is also found that the effect tends to more significant at 60 GHz than at 45 GHz even with same moisture content. From these results, it is considered that the snow accretion on the surface of the antenna radome can not be neglected regarding the influence on the wireless communication using millimeter-waves. Therefore, it turned out that when designing the radio circuit, it is necessary to have an enough margin or to design an antenna structure that snow does not accrete to the surface. In the future, we plan to deeply study the propagation characteristics of millimeter- waves in the railway environment, which are necessary for designing the wireless circuits of millimeter-wave communication systems for trains, through simulations and experiments.

678