electronics Article Implementation of Deep Learning-based Automatic Modulation Classifier on FPGA SDR Platform Zhi-Ling Tang ID , Si-Min Li * and Li-Juan Yu Guangxi Key Laboratory of Wireless Broadband Communication and Signal Processing, Guilin University of Electronic Technology, Guilin 541004, China; [email protected] (Z.-L.T); [email protected] (L.-J.Y.) * Correspondence: [email protected]; Tel.: +86-136-3519-1980 Received: 3 June 2018; Accepted: 16 July 2018; Published: 19 July 2018 Abstract: Intelligent radios collect information by sensing signals within the radio spectrum, and the automatic modulation recognition (AMR) of signals is one of their most challenging tasks. Although the result of a modulation classification based on a deep neural network is better, the training of the neural network requires complicated calculations and expensive hardware. Therefore, in this paper, we propose a master–slave AMR architecture using the reconfigurability of field-programmable gate arrays (FPGAs). First, we discuss the method of building AMR, by using a stack convolution autoencoder (CAE), and analyze the principles of training and classification. Then, on the basis of the radiofrequency network-on-chip architecture, the constraint conditions of AMR in FPGA are proposed from the aspects of computing optimization and memory access optimization. The experimental results not only demonstrated that AMR-based CAEs worked correctly, but also showed that AMR based on neural networks could be implemented on FPGAs, with the potential for dynamic spectrum allocation and cognitive radio systems. Keywords: wireless communication; signal recognition; cognitive radio; neural networks; reconfigurable hardware 1. Introduction Intelligent radios collect information by sensing signals within the radio spectrum, including the presence, type, and location of the signals. This capability not only identifies friendly or hostile signals in military applications, but also helps regulators to detect whether the use of the radio equipment complies with the spectrum rules. In both cases, an intelligent radio is allowed to take countermeasures corresponding to the perceived information. In recent years, the application of intelligent radio technology in commercial communications has been defined as software-defined radio (SDR) and cognitive radio (CR) [1,2]. On software and hardware platforms under software control, an SDR can select its operating parameters, including modulation format, coding scheme, antenna configuration, and bandwidth. Therefore, the receiver needs to have a robust method to identify these parameters. CR, by means of opportunities, utilizes the available spectrum of the existing users to provide a solution to the problem of insufficient spectrum utilization [2]. Therefore, one of the key tasks of CR is spectrum sensing [3], which collects information in spectrum scenarios. This information needs to be able to evaluate the possibility of interference with other users and set the corresponding operating parameters. There are plans to allocate the unused spectrum to television broadcasting services through CR. Signal recognition is a challenging and critical task for intelligent radios. First, the recognition algorithm needs to be able to flexibly sense different signal types under different conditions. For example, primary users with different sensitivity requirements, different transmission rates, and different modulation methods. Second, the recognition algorithm should minimize the requirement Electronics 2018, 7, 122; doi:10.3390/electronics7070122 www.mdpi.com/journal/electronics Electronics 2018, 7, 122 2 of 16 Electronics 2018, 7, x FOR PEER REVIEW 2 of 15 for signal preprocessing. For example, these algorithms should not rely on carrier synchronization and channelfor estimation, signal preprocessing. because the smart For example, receiver isthese not algorithms synchronized should with not the rely perceived on carrier radio synchronization signal and thereand is nochannel a priori estimation knowledge, because of the signal the smart parameters. receiver Finally, is not even synchronized at low signal-to-noise with the perceived ratios radio (SNR), thesignal recognition and there algorithm is no a priori can k providenowledge higher of the performance signal parameters. with lower Finally, complexity even at low in shorter signal-to-noise observationratios time. (SNR), the recognition algorithm can provide higher performance with lower complexity in Automaticshorter signalobservation modulation time. recognition (AMR) is a major research direction of signal recognition. At present, AMRAutomatic is implemented signal modulation using the recognition three types (AMR) of methods is a major listed inresearch Figure 1direction: (1) The of signal first methodrecognition. is based At onpresent, the likelihood AMR is implemented ratio test (LRT), using such the three as the types average of methods LRT proposed listed in Fig byure 1: (1) El-MahdyThe et al.first [4 method], the general is based LRT on proposed the likelihood by Panagiotou ratio test et (LRT), al. [5], suc andh theas the hybrid average LRT LRT proposed proposed by by El- HameedMahdy et al. [6 ].et The al. [4], disadvantage the general of LRT this proposed type of algorithm by Panagiotou is that the et assumptional. [5], and the that hybrid the symbols LRT proposed are by independentHameed and identicallyet al. [6]. The distributed disadvantage is usually of this not type true of in algorithm reality, and is thethat complexity the assumption increases that with the symbols an increaseare inindependent the number an ofd unknownidentically parameters. distributed Inis usually addition, not this true type in reality, of method and isthe more complexity sensitive increases to the mismatchwith an increase of the model; in the for number example, of unknown when a time parameters. shift occurs, In addition, the recognition this type performance of method is more deterioratessensitive considerably; to the mismatch (2) The second of the method model; isfor a feature-basedexample, when (FB) a methodtime shift that occ reliesurs, onthe prior recognition knowledge.performance The features deteriorates for classification considerably include; (2) Dobre’s The second use of first-ordermethod is cyclostationarya feature-based features (FB) method of that digital signalsrelies [on7] andprior second-order knowledge. cyclostationarity The features for [8], theclassification space–time include correlation Dobre’s matrix use features of first-order proposedcyclostationary by Marey and Choqueusefeatures of [digital9,10], wavelet signals features[7] and proposedsecond-order by Hassan cyclostationarity [11], and constellation [8], the space–time featurescorrelation proposed by matrix Mobasseri features [12 proposed]. This method by Marey can and obtain Choqueuse better results [9,10], under wavelet the features condition proposed of by a low SNR,Hassan but it[1 is1], necessary and constellation to calculate features various proposed feature by quantities Mobasseri according [12]. This to themethod mathematical can obtain better model ofre thesults modulation under the methodcondition in of advance, a low SNR and, thebut calculationit is necessary method to calculate is relatively various complicated. feature quantities In addition,according some featureto the mathematical quantities are model not complete. of the modulation For example, method the cyclic in advance, statistics and of signals,the calculation such as single-carriermethod is relatively linear complicated. digital modulation, In addition, orthogonal some feature frequency quantities division are not multiplexing, complete. For and example, single-carrierthe cyclic frequency statistics domain of signals, equalization, such as single have-carrier even-order linear cyclostationarity. digital modulation, Therefore, orthogonal there frequency are certaindivision restrictions multiplexing, when dealing and withsingle signals-carrier of frequency these unknown domain modulation equalization, modes; have (3) Theeven-order third methodcyclostationarity. includes blind Ther recognitionefore, there methods are certain that do restrictions not rely on priorwhen knowledge. dealing with For example,signals of these Migliori’sunknown method ofmodulation recurrent neuralmodes network; (3) Theautoencoder third method [13 includes] and the blind convolutional recognition neural methods network that do not method proposedrely on prior by O’Sheaknowledge [14].. SuchFor example, methods Migliori’s do not need method to know of recurrent the modulation neural mathematicalnetwork autoencoder model of[1 the3] and signal the [15 convolutional]. However, finding neural anetwork suitable method neural network proposed structure by O’Shea requires [14]. proficiencySuch methods in do not artificialneed intelligence. to know the modulation mathematical model of the signal [15]. However, finding a suitable neural network structure requires proficiency in artificial intelligence. Automatic Modulation Recognition Likelihood Ratio TEst Feature BAsed Blind REcognition (LRT) (FB) (BR) Averaged LRT Cyclostationarity Sparse Denoising Autoencoders General LRT Correlation matrices Convolutional Neural Hybrid LRT Wavelet Networks Fractal Constellation Figure 1. FigureMain implementation 1. Main implementation of modulation of modulation classification. classification. AlthoughAlthough the neural the network-based neural network modulation-based modulation classification classification