arXiv:1712.01154v4 [eess.SP] 28 Dec 2018 etri rtflyakolde.Tefis w uhr con authors two first The acknowledged. gratefully is Center suc techniques high-resolution classical the and studied, single system. a surveillance using the transm controller at dedicated required its no is Therefore, to receiver. can wireless drone d system channel the the processing The from by controller. bearing) transmitted drone ground (or its a direction drone’s and presents estimate drone communicatio paper the a this on between end, eavesdropping by this system, To surveillance present. cont its always and drone is the between level), battery and altitude, iaefeunl ihtercnrles n h downlink all the at and and controllers, night their commu- and i.e. drones with day commercial frequently the work nicate of can Most it conditions. weather as (RF) Frequency for importance. system self paramount surveillance of ample is a even give drones developing and events suspect privacy that These the examples [8]. violate evident activities to have used criminal drones being and carry of [7]; drone a accused Airport with International been colliding of Angeles feet Los 200 near within 20 came March jet in concerns Lufthansa building; raising government a [6], the House to January risks White in security the about example, at crashed For drone problematic popularity concerns. a a 2015 public gaining in stirred use fast has drone However, manner are [1–5]. world drones the around Likewise, years. recent tracking sn igecanlR eevr h ae aiae the validates paper The well. receiver. as ar RF experimentally implement pattern method be single-channel proposed radiation can method a antenna DF proposed the using phase the of a Also, essential. model that not analytical sense mechan the the calibration DF and antenna in proposed an low-complex The mechanism, is and ESPRIT. clas synchronization practical and It with MUSIC (UAVs). is associated as scheme vehicles direction challenges such aerial algorithms practical the DF unmanned the for small by (DNN) motivated of network (DF) neural finding deep (SDAE)-based FbsdDrcinFnigo AsUigDNN Using UAVs of Finding Direction RF-based h upr fNF 15102 n UDMTInternational SUTD-MIT and 61750110529 NSFC of support The Fbsddrcinfidn D)tcnqe aebe well been have techniques (DF) finding direction based RF h uhr f[]suh odtc rn sn Radio using drone a detect to sought [9] of authors The in dramatically increased has drones civilian of use The ne Terms Index Abstract ie inladtlmtysgas(ih pe,position speed, (flight signals telemetry and signal video , Ti ae rsnsasas eosn autoencoder denoising sparse a presents paper —This aihAbeywickrama Samith Doesrelac,drcinfidn,UAV finding, direction surveillance, —Drone Email: .I I. NTRODUCTION [email protected] † igpr nvriyo ehooyadDsg,Singapore Design, and Technology of University Singapore ∗† ∗ aiuJayasinghe Lahiru , ainlUiest fSnaoe Singapore Singapore, of University National , rbtdequally. tributed { aruna Design roller as h sical itter ism, ata jayasinghe,hua 16 ed † n u Fu Hua , e - , , tsol entdta hs ycrnzto mechanism, synchronization sparse signa phase this drone a the utilizes that of noted network be direction the should the It of classify pow rest to received representation the the network of Then, the representation of sparse values. sparse layer robust proposed hidden network a first the extracts neural the to precisely, deep More fed (DNN). (SDAE)-based Then, are mechanism. autoencoder values switching denoising RF power power a obtained the signal using controller, that the antenna received ground each the signals the its measures the at to receiver drone processing com- channel the single which By from receiver. system transmitted having a channel are array on single the antenna focus in a directional We network a DF. neural prises drone deep involves of that context method other any low- a fallows. contributions and as our summarized and practical paper, be this practic can a in some method DF propose with drone bounds complex we it still Therefore, methods, challenges. ant these the of for of direction vital pattern radiation the is actual estimate the to As transmitter. utilized switched WiFi are [16 a array from In antenna obtained [13–16]. is are beam signals that DF measurements such power performing with signal utilized, even feasible are t practically measurements only po if power However, characteristics techniques. signal DF signal classical and Such with wideband, challenges burst-mode. unknown, in usually transmitted are They communication. hardw consumption. the coherent power a increases the and realize this complexity to Clearly, necessary receiver. becomes multi-channel channel RF the each phas throughout of start-up of constant this channels calibrating remains Therefore, the [12]. but when operation initialized, time each are phase receiver therefore random CORDIC The a a up. creates power has on which position (CORDIC), start-up random computer uses digital chain (DDC) rotation converter formulated. coordinate down be digital However, the can coherent. receivers, be matrix most simulta- should correlation channels extracted data multiple be base-band a Therefore, should that the elements so means, antenna neously This all the from observation. be data snapshot to algori a those are considered require because ESPRIT techniques and are multi-channel MUSIC [11] inherently However, algorithms. ESPRIT popular most and [10] MUSIC otebs fatos nweg hr a o been not has there knowledge authors’ of best the To their for OFDM WiFi-like use drones civilian the of Most † uahn Nissanka Subashini , fu,yuenchau } @sutd.edu.sg † n huYuen Chau and , N nens and antennas, † values e thms enna the are he se in er al ], l. a a where Gn(θ) is the real numbered antenna gain for the azimuth

N-element angle θ, λ is the signal wavelength, and Directional Antenna Array 2π(n − 1) β (θ)= d cos − θ , (3) n N Single-pole-N-throw " # er Antenna Switch ¡ where d is the radius of the circular antenna array [18]. Since

Single Channel we consider a practical DF method in this paper, s(k) and SDR . . . an(θ) are assumed to be unknown. Therefore, our objective ...... is to recover the azimuth angle θ, while the parameters s(k) . . .

. . . Softmax Classi Data Preprocessing . . . and an(θ) are unknown...... We focus on a power measurements based approach. To this end, the ensemble averaged received signal power at the n−th

Encoder antenna element can be given as

SDAE-DNN 2 Pn =E |rn(k)| Fig. 1. The System Model. h i ∗ =E an(θ)s(k)+ nn(k) an(θ)s(k)+ nn(k)

h 2 2  2  i an antenna gain calibration mechanism, and the analytical = |an(θ)| E |s(k)| +E |nn(k)| model of the antenna radiation pattern are not essential for 2 h 2 i h i = Gn(θ)Ps + σ , (4) this single channel implementation. The paper validates the proposed method experimentally through a software defined where 2 2 radio (SDR) implementation in conjunction with TensorFlow Gn(θ)= |an(θ)| ,

[17]. Furthermore, such an experimental validation for drone 2 DF is not common in the literature, and can be highlighted as Ps =E |s(k)| , another contribution of this paper. h i 2 2 The paper organization is as follows. The system model is σ =E |nn(k)| , presented in Section II. Section III discusses the proposed deep h i and E[·] denotes the expectation operator. Here, (4) follows architecture. Then, in Section IV, we validate the proposed from the fact that s(k) and n (k) are independent and method using experimental results. Section V concludes the n uncorrelated, i.e., paper. ∗ ∗ To promote reproducible research, the codes for generating E[s(k)nn(n)] = E[s (k)nn(n)]=0. most of the results in the paper are made available on the website: https://github.com/LahiruJayasinghe/DeepDOA. It can be observed that the received power values at the antenna elements ni and nj , where i, j ∈ {1, ··· ,N} and i 6= j, are 2 2 II. SYSTEM MODEL not identical, since Gi (θ) 6= Gj (θ). We have this property thanks to the gain variation of the directional antenna array We consider a system which consists of a single channel in [0 2π). Therefore, it is desirable to have an underlying receiver, and a circular antenna array equipped with N direc- N relationship (or a pattern) between {Pn}n=1 and θ. tion antennas, see Fig. 1. The antenna array is connected to the The proposed method is as follows. The receiver sequen- receiver using a non-reflective Single-Pole-N-Throw (SPNT) tially activates one antenna element at a time using the RF switch. The switching period is Ts. Suppose that a far- SPNT RF switch, and measures the corresponding received field drone signal impinges on the antenna array with azimuth power value. During the activation of n−th antenna, Pn is angle θ ∈ [0 2π). The received signal at the n−th antenna measured, where n ∈{1, ··· ,N}. A single switching cycle is element can be given as equivalent to N activations, starting from the first antenna to the N−th antenna. Let p = [P , ··· , P ]⊤ denote the power r (k)= a (θ)s(k)+ n (k), (1) 1 N n n n measurements corresponding to a single switching cycle. As ⊤ it is depicted in Fig. 1, x = [x , ··· , xN ] is obtained during where k is the sample index, an(θ) is the n−th antenna 1 response vector for the azimuth angle θ, s(k) is the drone the preprocessing stage, where transmitted signal as it arrives at the antenna array, nn(k) is P x n . (5) circularly symmetric, independent and identically distributed, n = N Pi complex additive white Gaussian noise (AWGN) with zero i=1 2 This means, x is the ratio between P and the summation mean and variance σ , and n ∈ {1, ··· ,N}. Here, an(θ) n P n follows the form of of all power values within the same switching cycle. In the next section, we discuss how the proposed network recovers 2π j βn(θ) an(θ)= Gn(θ)e λ , (2) θ from x. eevdpwrvle tteotu ae fteSDAE. the of layer output the at values power received W arx hc nue htteotu ae eosrcsth reconstructs ( layer possible output as precisely the as that input ensures which matrix, nt.Teeoe h otfnto a egvnas sparsity given be This can function constraint. l cost hidden the sparsity the Therefore, of a units. activation sparse to the encourages minimiz subjecting constraint is error it reconstruction the while that such optimized are { where lmn-ieo t argument, its on element-wise (6), eevdpwrvalues power received noe egtmatrix, weight encoder [ i orpe hc dsnieacrigt oeniemodel noise input, some its to to according noise adds which corrupter tic n uptlyrvle r acltdas calculated are values layer output and DE(i.2(),tepercse eevdpwrvalues the power of received phase preprocessed training the the softmax { During 2-(a)), fully-connected 1. (Fig. a Fig. SDAE by see followed layer, DNN, classifier trained a and idnlyruisaecluae as calculated are units element layer of number hidden the to equal is layer N input the in neurons of b x x d oti n,teprmtr fSA ( SDAE of parameters the end, this To h rpsdde rhtcuecmrssatandSDAE trained a comprises architecture deep proposed The n n 1 .Hr,w atclrytre nrcntutn h input the reconstructing on target particularly we Here, ). L ntedrcinlatnaary hn h auso the of values the Then, array. antenna directional the in . . . . (a) SDAE Training Phase b , . . . , } } f ( n N n N W c =1 =1 f snndtriitc ic tcrut h aestof set same the corrupts it since non-deterministic, is ,

( Data Corruption r sindt h nu nt.Teeoe h number the Therefore, units. input the to assigned are b · spse hog it. through passed is d ) e N , i.e. snnlna ciainfnto htoperates that function activation non-linear is ] I.SA-N A SDAE-DNN III. b . . . . ⊤ Encoder d , = ) eoeteba etr,and vectors, bias the denote x ′ ...... h Decoder X i = x ˆ =1 i.2 riigPhases. Training 2. Fig. T = = f { ( c f x x . . . . ˆ ( f ( b n i x W ( e } − ) W n N where , . . . . =1 f [ = x Encoder ⊤ c W i ( h ) x W ndfeetwy vr time every ways different in W 2 ⊤ RCHITECTURE (b) DNN Training(b) DNN Phase b + ) + ...... + e ⊤ stemti rnps of transpose matrix the is 1 b x b , . . . , β ∈ d b ′ stedcdrweight decoder the is m e ) X = , M ...... ) =1 R , h M e KL( f x W M × c Cross-Entropy 1 ′ ( ...... N x , . . . , ] , · ⊤ ) ρ b sastochas- a is || eoe the denotes e and ρ and , . . . . m N ′ ) , i b ⊤ d ayer

In . b Softmax Classi er ed, (8) (7) (6) d = e ) s parameter where x process. steaeaeatvto ee fthe of level activation average the is oanof domain ihdirection with KL( results. experimental using steKlbc-ebe K)dvrec 1] rm() tc it (9), that From observed [19]. be divergence (KL) Kullback-Leibler the is nrae oooial as monotonically increases h aus hn h eto h ewr tlzsti sparse this utilizes network power the input estimate) of (or the classify rest to of representation the representation Then, sparse values. (or robust noisy layer a a in hidden extracts operates first system constra the the sparsity when a environment, even to th Therefore, subject to well. is fed as function are cost values This power power network. reconstructed noise-corrupted betwe the the calculated and while is values, values In error power follows. squared as non-corrupted the summarized the (8), be function can cost it the and property, sparse this n hssast a eepotdt estimate to exploited be can sparsity this and savr ml au ls ozr.Teeoe hntecost the when Therefore, parameter zero. the to minimized, close is (8) value function small very a is rm hrfr,telann taeyi uevsdi thi in coming supervised signal is drone Since, strategy the phase. learning of training direction the the Therefore, is from. label the where elble into labelled be noe,i sstesm ciainfunction activation same L the uses it Since encoder, classification. of task ciainfnto.Aan os orpe eevdpowe received corrupted values noise Again, function. activation hidden connected fully three comprises layers, DNN fully depicts, a as (b) DNN the to connected is layer. encoder connected discard trained is represen decoder the that the neurons and Now, dominant non-zero. the stay while features zero, specific to close be to arx twl o eotmzdaan hrfr,ol the only Therefore, again. optimized be matrices not weight will it matrix, hstann phase. training this m 3 i 1 ntenx eto,w ilvldt u rpsdmethod proposed our validate will we section, next the In et h N riigpaei omne.A i.2- Fig. As commenced. is phase training DNN the Next, eak1: Remark and , ( and KL( ρ β x || i ρ prtsa h rd-f aaee ewe h qae err squared the between parameter trade-off the as operates ) m T n i.e. ρ L eoe h ciaino the of activation the denotes ) x || n t au a eeprclydcdddrn h trainin the during decided empirically be can value its and , 4 1 stesz ftetann aaset, data training the of size the is n ′ , 02 [0 ρ , o s etfidlnrUi RL)[12]a their as [21–23] (ReLU) Unit liner Rectified use L m ρ n N 2 tsol entdta h noigdoesignal drone incoming the that noted be should It = ) stesast parameter, sparsity the is =1 , θ π L Q ) W cuisacranioae on nteangle the in point isolated certain a occupies KL( Therefore, . 3 ρ r h riigipt.Nw aane to need data Now, inputs. training the are and , lse u oteueo ota classifier, softmax of use the to due classes D1 log ρ ρ m , || W  W = ρ L ρ m ρ 4 m D2 T n ota ae 2]frthe for [20] layer softmax a and , 0 = ) stepetandecdrweight encoder pre-trained the is 1 ρ  and , L m X θ i (1 + =1 T 2 ssas ntesaildomain, spatial the in sparse is iegsfrom diverges sa lmn ftetrained the of element an is if , h W m − ρ ( D3 m m x m ρ i θ log ) t idnui where unit hidden -th − ) = . ρ r piie during optimized are W hui o h input the for unit th enforces ρ n tews it otherwise and , ftenetwork the of )  f θ 1 idnlayers Hidden . 1 ρ ee euse we Here, . − Typically, . β − ρ sahyper a is ρ m { ρ  m } rand or m M ed, (9) =1 int an en ρ g e s r t TABLE I TABLE II CONFUSIONMATRIXWHENTHEPROPOSEDNETWORKISUSED. CONFUSION MATRIX WHEN ONLY THE DNN IS USED.

00 450 900 1350 1800 2250 2700 3150 00 450 900 1350 1800 2250 2700 3150

00 95 1 1 0 3 0 0 0 00 94 5 1 0 0 0 0 0

450 1 97 0 0 1 1 0 0 450 10 88 1 0 0 1 0 0

900 1 1 98 0 0 0 0 0 900 4 1 88 1 3 0 0 3

1350 0 0 0 100 0 0 0 0 1350 0 6 1 93 0 0 0 0

1800 3 3 0 0 92 2 0 0 1800 30 4 3 0 60 2 0 1

2250 0 0 4 0 1 95 0 0 2250 1 0 4 0 2 91 0 2

2700 0 0 3 0 0 1 95 1 2700 0 0 1 0 2 1 94 2

3150 0 0 0 0 0 0 1 99 3150 0 0 0 0 3 0 0 97

IV. EXPERIMENTAL VALIDATION Our experimental setup comprises a SDR (USRP B210), and a four element sector antenna, which is a variant of the antenna implemented in [24]. We use only a single RF receiving channel of the SDR. Therefore, the SDR is connected to the antenna using a non-reflective Single-Pole-4- Throw (SP4T) RF switch. DJI Phantom 3 is considered as the target drone throughout the experiment. The drone downlink channels occupy the bandwidth from 2.401 GHz to 2.481 GHz, each has 10 MHz bandwidth OFDM signal. This OFDM signal (a) Training Field (b) Direction Configuration transmitted by the drone provides the main source to perform the DF task. Fig. 3-(a) represents the environment that we used for Fig. 3. The Training Field and The Direction Configuration. the training data collection. This is a large ground with an open area. Also, there was negligible RF interference on the robust, and it outperforms the baseline method. 2.401 GHz - 2.481 GHz range. To simplify the experiment, Since our implementation does not use multiple RF channels we virtually divided the area into eight octants, see Fig. 3- and any information about the antenna radiation pattern, it (b). Each octant is considered as one direction during the is not desirable to compare our results with conventional experiment. For example, the first octant is considered as 0 techniques. Therefore, we omit such simulation/experimental degrees direction, while the second octant is considered as 45 results. Further interesting experimental evaluations and in- degrees direction, and so on. Therefore, when the drone is sights will be presented in future extensions of this work. flying, its direction is indicated by its corresponding octant. ONCLUSION In the trained network, L5−th layer has eight neurons (we V. C have eight classes for the direction classification, or, Q = 8) This paper has proposed a novel DF method to be used and L1−th layer has four neurons (the antenna array has four in a drone surveillance system. The system comprises a elements, or, N =4). The hidden layers L2, L3, and L4 have single channel receiver and a directional antenna array. The 200, 12, and 12 neurons, respectively. These values have been receiver sequentially activates each antenna in the array, and empirically decided during the training process. measures the received power values. The power measurements After the training phase, the evaluation is done in a different corresponding to each switching cycle are fed to the pro- environment. Now, the frequency spectrum (2.401 GHz - posed deep network. Then, it performs DF by exploiting the 2.481 GHz) suffers from WiFi and bluetooth interferences. sparsity property of the incoming drone signal, and the gain To this end, two experiments have been carried out. First, we variation property of the directional antenna array. The paper evaluated the proposed deep architecture, and its confusion has validated the proposed method experimentally. Also, it matrix is given in the Table I. Next, we considered a baseline has been proven that a phase synchronization mechanism, an method, where only a conventional DNN is implemented antenna gain calibration mechanism, and the analytical model without the L2 layer (other layers have same number of nodes). of the antenna radiation pattern are not essential for this single Its confusion matrix is given in the Table II. Note that the channel implementation. In future the scheme will be applied confusion matries represent the percentage (%) values. It can to portable, SDR-based prototype design [9] and field test and be observed that the proposed deep architecture is certainly experiment. REFERENCES Proc. Instituto Tecnologico de Aeronautica (ITA), Sep. 2014. [1] Y. Sun, H. Fu, S. Abeywickrama, L. Jayasinghe, C. Yuen, and J. Chen, [13] H. Fu, S. Abeywickrama, and C. Yuen, “A robust phase-ambiguity- “Drone classification and localization using micro-doppler signature immune doa estimation scheme for antenna array,” To appear in IEEE with low-frequency signal,” To appear in IEEE International Conference International Conference on Communication Systems, 2018. on Communication Systems, 2018. [14] S. Maddio, M. Passafiume, A. Cidronali, and G. Manes, “A closed-form [2] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication formula for RSSI-based doa estimation with switched beam antennas,” design for multi-UAV enabled wireless networks,” IEEE Trans. Wireless in Proc. 2015 European Radar Conference, Sep. 2015. Commun., vol. 17, no. 3, pp. 2109–2121, 2018. [15] R. P¨ohlmann, S. Zhang, T. Jost, and A. Dammann, “Power-Based [3] Q. Wu and R. Zhang, “Common throughput maximization in UAV- Direction-of-Arrival Estimation Using a Single Multi-Mode Antenna,” enabled OFDMA systems with delay consideration,” To appear in IEEE ArXiv e-prints, Jun. 2017. Trans. Commun., 2018. [16] S. Maddio, M. Passafiume, A. Cidronali, and G. Manes, “A scalable dis- [4] Q. Wu, J. Xu, and R. Zhang, “Capacity characterization of UAV-enabled tributed positioning system augmenting wifi technology,” in Proc. IEEE two-user broadcast channel,” To appear in IEEE J. Sel. Areas Commun., International Conference on Indoor Positioning and Indoor Navigation, 2018. Oct. 2013. [5] Q. Wu, L. Liu, and R. Zhang, “Fundamental tradeoffs in communication [17] M. Abadi and A. Agarwal, “Tensorflow: Large-scale machine learning and trajectory design for UAV-enabled wireless network,” Submitted to on heterogeneous distributed systems,” ArXiv e-prints, Mar. 2016. IEEE Trans. Wireless Commun., 2018. [18] B. R. Jackson, S. Rajan, B. Liao, and S. Wang, “Direction of arrival [6] B. Jansen, “Drone crash at white house reveals security risks.” [online] estimation using directive antennas in uniform circular arrays,” IEEE https://www.usatoday.com/story/news/2015/01/26/, 2015. [Accessed on Trans. Antennas and Propagation, vol. 63, pp. 736–747, Feb. 2015. 26 November 2017]. [19] S. Kullback and R. Leibler, “On information and sufficiency,” The [7] J. Serna, “Lufthansa jet and drone nearly collide near lax.” [online] Annals of Mathematical Statistics, vol. 22, pp. 79–86, Mar. 1951. www.latimes.com/local/lanow/la-me-ln-drone-near-miss-lax-20160318, [20] C. Bishop, Pattern Recognition and Machine Learning. Springer-Verlag, 2016. [Accessed on 26 November 2017]. 2006. [8] A. Morrow, “Couple accuses neighbor of stalking with drone.” [online] [21] K. Hara, D. Saito, and H. Shouno, “Analysis of function of rectified https://www.suasnews.com/2014/12/, 2014. [Accessed on 26 November linear unit used in deep learning,” in Proc. IEEE International Joint 2017]. Conference on Neural Networks, Jul. 2015. [9] H. Fu, S. Abeywickrama, L. Zhang, and C. Yuen, “Low complexity [22] L. Jayasinghe, N. Wijerathne, and C. Yuen, “A deep learning approach portable passive drone surveillance via SDR-based signal processing,” for classification of cleanliness in restrooms,” To appear in Proc. IEEE IEEE Commun. Mag., vol. 56, pp. 112–118, Apr. 2018. International Conference on Intelligent and Advanced Systems, Aug. [10] R. Schmidt, “Multiple emitter location and signal parameter estimation,” 2018. IEEE Trans. Antennas and Propagation, vol. 34, pp. 276–280, Mar. [23] L. Jayasinghe, T. Samarasinghe, C. Yuen, and S. S. Ge, “Temporal 1986. convolutional memory networks for remaining useful life estimation of [11] R. Ray and T. Kailath, “ESPRIT-estimation of signal parameters via industrial machinery,” To appear in IEEE International Conference on rotational invariance techniques,” IEEE Trans. Acoustics, Speech, Signal Industrial Technology, 2019. Processing, vol. 37, pp. 984–995, Jul. 1989. [24] L. Yang, Z. Shen, W. Wu, and J. D. Zhang, “A four-element array of [12] J. van der Merwe, J. Malan, F. Maasdorp, and W. D. Plessis, “Multi- wideband low-profile H-Plane horns,” IEEE Trans. Vehicular Technol- channel software defined radio experimental evaluation and analysis,” in ogy, vol. 64, pp. 4356–4359, Sep. 2015.