馮奇 張君毅 陳麗 劉芳
摘 要:為了解決非合作通信情況下,具有特定幀結(jié)構(gòu)的復(fù)雜信號(hào)難以重構(gòu)問題,設(shè)計(jì)了一種利用深度無悔分析生成對(duì)抗網(wǎng)絡(luò)(deep regret analytic generative adversarial networks,DRAGAN)重構(gòu)信號(hào)的方法。首先利用無悔算法(no-regret algorithms)對(duì)判別器損失函數(shù)進(jìn)行約束,判別器的梯度被迫向更加穩(wěn)定的方向變化;其次通過生成器與判別器的對(duì)抗學(xué)習(xí),生成器的分布逐步擬合到目標(biāo)數(shù)據(jù)的潛在分布;最后構(gòu)建具有特定幀的復(fù)雜信號(hào)模型,并據(jù)此進(jìn)行DRAGAN方法的實(shí)驗(yàn)驗(yàn)證。仿真實(shí)驗(yàn)結(jié)果表明,在信噪比為9 dB及以上的條件下,生成信號(hào)不僅學(xué)習(xí)到了樣本信號(hào)的調(diào)制樣式、符號(hào)速率和頻率帶寬等特性,還能較準(zhǔn)確還原出特定幀部分的符號(hào)信息。相較于傳統(tǒng)方法,利用DRAGAN生成信號(hào)具有相關(guān)性高、重構(gòu)流程簡(jiǎn)易和泛化能力強(qiáng)等特點(diǎn),所設(shè)計(jì)的網(wǎng)絡(luò)模型在電磁環(huán)境構(gòu)建等場(chǎng)景中具有實(shí)用價(jià)值。
關(guān)鍵詞:無線通信技術(shù);信號(hào)重構(gòu);生成對(duì)抗網(wǎng)絡(luò);無悔算法;電磁環(huán)境構(gòu)建
中圖分類號(hào):TN975 ? 文獻(xiàn)標(biāo)識(shí)碼:A ? DOI: 10.7535/hbgykj.2022yx01001
Abstract:In order to solve the problem that complex signals with a specific frame structure were difficult to reconstruct in the case of non-cooperative communication,a method of reconstructing signals by using Deep Regret Analytic Generative Adversarial Networks (DRAGAN) was designed.Firstly,no-regret algorithms were used to constrain the loss function of the discriminator,and the gradient of the discriminator was forced to change in a more stable direction.Secondly,through the confrontation learning between the generator and the discriminator,the distribution of the generator was gradually fitted to the potential distribution of the target data.Finally,a complex signal model with a specific frame was constructed,and the experimental verification of DRAGAN method was carried out.The simulation results show that when the signal-to-noise ratio is 9 dB or above,the generated signal not only learns the modulation style,symbol rate and frequency bandwidth of the sample signal,but also accurately restores the symbol information of a specific frame.Compared with the traditional methods,the signal generated by DRAGAN has the characteristics of high correlation,simple reconstruction process and strong generalization ability.The designed network model has practical value in the construction of electromagnetic environment and other scenes.
Keywords:wireless communication technology;signal reconstruction;generative adversarial networks;no-regret algorithm;electromagnetic environment construction
復(fù)雜電磁環(huán)境構(gòu)建[1-2]是無線通信領(lǐng)域的一個(gè)重要研究方向,尤其當(dāng)今電磁環(huán)境十分復(fù)雜,無論是在空間、空中、海上和陸地,所有的通信信號(hào)都伴隨著日趨復(fù)雜多樣的人為干擾、無意串?dāng)_或者大自然產(chǎn)生的雷暴等信號(hào)。為了提高自身通信系統(tǒng)適應(yīng)未來戰(zhàn)場(chǎng)電磁環(huán)境能力,各軍事強(qiáng)國、區(qū)域大國和糾紛地區(qū)國家都增強(qiáng)了復(fù)雜電磁環(huán)境構(gòu)建的關(guān)注度。通信信號(hào)生成[3-4]技術(shù)是復(fù)雜電磁環(huán)境構(gòu)建中的關(guān)鍵一環(huán),對(duì)此開展研究具有重要意義。
對(duì)于空間中非合作方的通信信號(hào)生成,有兩種傳統(tǒng)解決方式:一種是基于參數(shù)測(cè)量分析,通過捕獲目標(biāo)信號(hào),對(duì)其碼速率、調(diào)制樣式、載頻等參數(shù)進(jìn)行估計(jì)后,通過信號(hào)重構(gòu)方式完成通信信號(hào)的生成;另一種是基于盲偵察、盲干擾的方式,對(duì)目標(biāo)信號(hào)進(jìn)行稀疏采樣后,偵知其稀疏特性,再施以信號(hào)重構(gòu)。然而,現(xiàn)代戰(zhàn)場(chǎng)上的電磁環(huán)境中必然存在大量特殊結(jié)構(gòu)的新型通信信號(hào),憑借傳統(tǒng)方式已經(jīng)無法準(zhǔn)確勾畫出目標(biāo)信號(hào)特征,亟需一種新型信號(hào)生成方式的問世。
隨著計(jì)算機(jī)計(jì)算能力的提升,許多基于深度學(xué)習(xí)[5]的智能化模型和算法被提出,解決了傳統(tǒng)通信技術(shù)無法解決的大量難題,其中,生成對(duì)抗網(wǎng)絡(luò)(generative adversarial networks,GAN)[6-8]展現(xiàn)出非常強(qiáng)大的能力,GAN是一種隱式的生成模型,可以在不知道目標(biāo)樣本先驗(yàn)知識(shí)的情況下,學(xué)習(xí)樣本在空間中的分布,生成符合目標(biāo)特征的數(shù)據(jù)。
在GAN生成信號(hào)方面,前人進(jìn)行了許多研究,秦劍[9]利用改進(jìn)的CGAN生成了AM和CPFSK信號(hào);楊鴻杰等[10]利用BEGAN生成的BPSK和8PSK信號(hào)波形具有很好的質(zhì)量;SHI等[11]使用GAN 生成的QPSK無線欺騙信號(hào)可以使欺騙率達(dá)到76.2%;趙凡等[12]利用非常簡(jiǎn)單的GAN網(wǎng)絡(luò)生成了BPSK,QPSK,16QAM和2FSK調(diào)制的通信信號(hào),并驗(yàn)證了網(wǎng)絡(luò)的泛化性。然而對(duì)于具有特定幀結(jié)構(gòu)的復(fù)雜通信信號(hào)重構(gòu),目前仍未提出解決方式。
針對(duì)以上問題,本文通過Matlab構(gòu)建了具有特定幀結(jié)構(gòu)的通信信號(hào)樣本,提出了利用DRAGAN重構(gòu)信號(hào)的算法模型[13],設(shè)計(jì)了GAN的網(wǎng)絡(luò)結(jié)構(gòu),并對(duì)樣本進(jìn)行了訓(xùn)練,實(shí)現(xiàn)了對(duì)不同信噪比下具有特征幀結(jié)構(gòu)的信號(hào)重構(gòu),對(duì)比樣本信號(hào),生成的信號(hào)取得了較高的相關(guān)性,表現(xiàn)出較好的結(jié)果。
1 DRAGAN原理及數(shù)據(jù)集準(zhǔn)備
1.1 GAN的基本原理
GAN是最近幾年發(fā)展最快的深度學(xué)習(xí)模型之一,它采用對(duì)抗的方式對(duì)目標(biāo)特征分布進(jìn)行學(xué)習(xí),模型主要由兩部分網(wǎng)絡(luò)架構(gòu)組成:一個(gè)是生成器G,用于生成數(shù)據(jù)并欺騙判別器;另一個(gè)是判別器D,用于判斷數(shù)據(jù)是生成數(shù)據(jù)還是真實(shí)樣本。隨機(jī)噪聲z輸入到生成器網(wǎng)絡(luò)中,生成器會(huì)輸出生成數(shù)據(jù),生成的數(shù)據(jù)和真實(shí)的樣本作為判別器的輸入輸進(jìn)判別器,判別器會(huì)對(duì)輸入的數(shù)據(jù)進(jìn)行分辨,判斷數(shù)據(jù)是來自生成器的生成數(shù)據(jù)還是真實(shí)樣本,而生成器力求學(xué)習(xí)真實(shí)樣本分布來欺騙判別器,2個(gè)網(wǎng)絡(luò)采用這種梯度交替更新策略(alternating gradient updates procedure,AGD),雙方通過不斷動(dòng)態(tài)博弈的過程,最后達(dá)到納什均衡,此時(shí),判別器無法分辨數(shù)據(jù)是來自真實(shí)樣本還是生成數(shù)據(jù)。GAN的結(jié)構(gòu)如圖1所示。
3 實(shí)驗(yàn)仿真
根據(jù)第2章所設(shè)計(jì)的生成器和判別器的網(wǎng)絡(luò)架構(gòu),本節(jié)將1.2節(jié)生成的目標(biāo)信號(hào)作為訓(xùn)練樣本數(shù)據(jù),對(duì)GAN進(jìn)行訓(xùn)練生成。
在SNR=20 dB時(shí),樣本信號(hào)波形如圖6所示,生成信號(hào)波形如圖7所示,可以看出,生成信號(hào)在特定幀部分與樣本信號(hào)波形十分相似,而有效負(fù)載部分與樣本信號(hào)則不盡相同,生成信號(hào)的星座如圖8所示,生成信號(hào)的調(diào)制樣式為QPSK,符合樣本信號(hào)的調(diào)制特征。
當(dāng)SNR=20 dB時(shí),在生成的1 500個(gè)信號(hào)中,有0.999 9的概率認(rèn)為特定幀部分的實(shí)部與樣本的特定幀實(shí)部具有很強(qiáng)的相關(guān)性,有0.999 9的概率認(rèn)為特定幀部分虛部與樣本的特定幀虛部具有很強(qiáng)的相關(guān)性,有0.925 9的概率認(rèn)為有效負(fù)載實(shí)部與樣本的有效負(fù)載實(shí)部沒有相關(guān)性,有0.952 2的概率認(rèn)為有效負(fù)載虛部與樣本有效負(fù)載虛部沒有相關(guān)性。
可以說明,網(wǎng)絡(luò)學(xué)習(xí)到了前導(dǎo)碼、后導(dǎo)碼和導(dǎo)頻塊中的符號(hào)信息;隨機(jī)部分生成信號(hào)和樣本信號(hào)二者幾乎沒有相關(guān)性,符合樣本信號(hào)中生成負(fù)載部分符號(hào)的隨機(jī)性。
從SNR=7 dB到SNR=20 dB分別選取1 500個(gè)生成信號(hào),統(tǒng)計(jì)這些生成信號(hào)與樣本信號(hào)在特定幀部分和有效負(fù)載部分的相關(guān)系數(shù)均值,相關(guān)系數(shù)如圖11所示,可以得出,生成信號(hào)在特定幀的實(shí)部和虛部隨著樣本信噪比的升高相關(guān)性逐漸增大,其中樣本SNR=7 dB到SNR=11 dB之間時(shí)生成信號(hào)和樣本信號(hào)的相關(guān)性快速增長,在SNR=11 dB到SNR=16 dB之間時(shí)相關(guān)性增長緩慢,在SNR=16 dB到SNR=20 dB之間時(shí)相關(guān)性增長再次加快,該現(xiàn)象表明,隨著樣本信號(hào)的信噪比增加,網(wǎng)絡(luò)更易于學(xué)習(xí)到樣本的特征,但是網(wǎng)絡(luò)對(duì)樣本信噪比的區(qū)間敏感程度不同;在有效負(fù)載部分,生成信號(hào)和樣本信號(hào)的相關(guān)系數(shù)分布在0~0.2,兩者不相關(guān)。
4 結(jié) 語
通信信號(hào)生成是復(fù)雜電磁環(huán)境仿真構(gòu)建的關(guān)鍵一步,對(duì)于空間中非合作方的具有特定幀結(jié)構(gòu)的復(fù)雜通信信號(hào),傳統(tǒng)方式很難對(duì)其進(jìn)行重構(gòu),本文提出了一種利用DRAGAN生成具有特定幀結(jié)構(gòu)的QPSK信號(hào)方法,僅對(duì)目標(biāo)信號(hào)采樣后做簡(jiǎn)單樣本處理,將樣本輸入到設(shè)計(jì)好的網(wǎng)絡(luò)中進(jìn)行訓(xùn)練,利用卷積網(wǎng)絡(luò)提取樣本的特征并使用DRAGAN的懲罰項(xiàng)使訓(xùn)練更加穩(wěn)定,通過GAN的對(duì)抗學(xué)習(xí)即可重構(gòu)出具有目標(biāo)信號(hào)特征的生成信號(hào)。實(shí)驗(yàn)結(jié)果表明,當(dāng)樣本信號(hào)信噪比為9~12 dB時(shí),生成信號(hào)在特定幀部分與樣本的相關(guān)系數(shù)可達(dá)0.7~0.8,當(dāng)樣本信號(hào)信噪比為12 dB以上時(shí),特定幀部分的相關(guān)系數(shù)可達(dá)0.8以上。
與傳統(tǒng)方式相比,該算法不需要復(fù)雜的參數(shù)測(cè)量和特征分析,縮短了研發(fā)時(shí)間,減少了人力成本;并且對(duì)于具有其他類型特定幀結(jié)構(gòu)的信號(hào),只需要適當(dāng)修改網(wǎng)絡(luò)架構(gòu)就可以重新再利用,模型泛化能力強(qiáng)。該算法在實(shí)際工程使用中具有重要意義。
利用GAN生成的信號(hào)能夠?qū)W習(xí)到目標(biāo)信號(hào)的調(diào)制樣式、符號(hào)速率和頻率特性等,對(duì)于特定幀部分的前導(dǎo)碼、后導(dǎo)碼和導(dǎo)頻塊部分,網(wǎng)絡(luò)不但能學(xué)習(xí)到它們的持續(xù)時(shí)間,存在位置,還能較準(zhǔn)確地學(xué)習(xí)到這些部分包含的符號(hào)信息;對(duì)于樣本信號(hào)信噪比低于9 dB的情況下,網(wǎng)絡(luò)無法有效學(xué)習(xí)到目標(biāo)的特征分布,這也是日后重構(gòu)帶有特定幀的低信噪比信號(hào)尚待解決的問題。
參考文獻(xiàn)/References:
[1] 周波,申緒澗,戴幻堯,等.基于逼真度的戰(zhàn)場(chǎng)電磁環(huán)境構(gòu)建要素分析[J].電子信息對(duì)抗技術(shù),2014,29(5):17-20.
ZHOU Bo,SHEN Xujian,DAI Huanyao,et al.Analysis of battlefield electromagnetic environment constructing element based on fidelity[J].Electronic Warfare Technology,2014,29(5):17-20.
[2] WEN Qi.An overview of the study of the complexity of the complex electromagnetic environments[C]//2015 8th International Symposium on Computational Intelligence and Design (ISCID).Hangzhou:IEEE,2015:245-250.
[3] 李楠,王標(biāo),底明曦,等.海戰(zhàn)場(chǎng)復(fù)雜電磁環(huán)境構(gòu)建研究[J].艦船電子對(duì)抗,2021,44(2):11-17.
LI Nan,WANG Biao,DI Mingxi,et al.Research into the construction of complex electromagnetic environment for sea battlefield[J].Shipboard Electronic Countermeasure,2021,44(2):11-17.
[4] 孫翱.復(fù)雜電磁環(huán)境下武器裝備試驗(yàn)與評(píng)估[J].艦船電子工程,2017,37(4):81-85.
SUN Ao.Test and evaluation on the armament based on complex electromagnetic environment[J].Ship Electronic Engineering,2017,37(4):81-85.
[5] NIRANJANI V,SELVAM N S.Overview on deep neural networks:Architecture,application and rising analysis trends[C]//Business Intelligence for Enterprise Internet of Things.Cham:Springer,2020:271-278.
[6] GOODFELLOW I,POUGET-ABADIE J,MIRZA M,et al.Generative adversarial nets[C]//Proceedings of the 27th International Conference on Advances in Neural Information Processing Systems.Boston:MIT Press,2014:2672-2680.
[7] ALQAHTANI H,KAVAKLI-THORNE M,KUMAR G.Applications of generative adversarial aetworks(GANs):An updated review[J].Archives of Computational Methods in Engineering,2021,28(2):525-552.
[8] 奚祥品,陳筱,朱向冰.生成對(duì)抗網(wǎng)絡(luò)介紹及應(yīng)用[J].無線電通信技術(shù),2021,47(3):338-343.
XI Xiangpin,CHEN Xiao,ZHU Xiangbing.Introduction and application of generative adversarial networks[J].Radio Communications Technology,2021,47(3):338-343.
[9] 秦劍.基于生成對(duì)抗網(wǎng)絡(luò)的信號(hào)重構(gòu)[D].西安:西安電子科技大學(xué),2018.
QIN Jian.Signal Reconstruction Based on Generative Adversarial Networks[D].Xi′an:Xidian University,2018.
[10] 楊鴻杰,陳麗,張君毅.基于生成對(duì)抗網(wǎng)絡(luò)的數(shù)字信號(hào)生成技術(shù)研究[J].電子測(cè)量技術(shù),2020,43(20):127-132.
YANG Hongjie,CHEN Li,ZHANG Junyi.Research on digital signal generation technology based on generative adversarial network[J].Electronic Measurement Technology,2020,43(20):127-132.
[11] SHI Yi,DAVASLIOGLU K,SAGDUYU Y E.Generative adversarial network for wireless signal spoofing[C]//ACM Workshop on Wireless Security and Machine Learning.New York:Association for Computing Machinery,2019:55-60.
[12] 趙凡,金虎.基于GAN的通信干擾波形生成技術(shù)[J].系統(tǒng)工程與電子技術(shù),2021,43(4):1080-1088.
ZHAO Fan,JIN Hu.Communication jamming waveform generation technology based on GAN[J].Systems Engineering and Electronics,2021,43(4):1080-1088.
[13] KODALI N,ABERNETHY J,HAYS J,et al.On convergence and stability of GANs[EB/OL].(2017-12-10).https://arxiv.org/abs/1705.07215v5.
[14] ARJOVSKY M,CHINTALA S,BOTTOU L.Wasserstein generative adversarial networks[C]//Proceedings of the 34th International Conference on Machine Learning.Sydney:MLResearch Press,2017:214-223.
[15] CHEN Xi,DUAN Yan,HOUTHOOFT R,et al.Infogan:Interpretable representation learning by information maximizing generative adversarial Nets[C]//Proceedings of the 30th International Conference on Neural Information Processing Systems.Red Hook:Curran Associates Inc.,2016:2180-2188.
[16] CHAVDAROVA T,F(xiàn)LEURET F.SGAN:An alternative training of generative adversarial networks[C]//Proceedings of the IEEE international conference on Computer Vision and Pattern Recognition.New York:IEEE.2018:9407-9415.
[17] WU Jiqing,HUANG Zhiwu,THOMA J,et al.Wasserstein divergence for GANs[C]//Proceedings of the European Conference on Computer Vision.Cham:Springer,2018:673-688.
[18] NOWOZIN S,CSEKE B,TOMIOKA R.F-gan:Training generative neural samplers using variational divergence minimization[C]//Proceedings of the 30th International Conference on Neural Information Processing Systems.Red Hook:Curran Associates Inc.,2016:271-279.