冲击噪声下基于演化长短时记忆神经网络的调制信号识别  

Modulation signal recognition based on evolutionary long short-term memory neural network under impulse noise

作　　者：高洪元[1] 王世豪 程建华 郭瑞晨 张志伟[1] GAO Hongyuan;WANG Shihao;CHENG Jianhua;GUO Ruichen;ZHANG Zhiwei(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China;College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001 [2]哈尔滨工程大学智能科学与工程学院,黑龙江哈尔滨150001

出　　处：《智能系统学报》2023年第4期676-687,共12页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(62073093);黑龙江省自然科学基金项目(LH2020F017);黑龙江省博士后科研启动金项目(LBH-Q19098)。

摘　　要：为了解决冲击噪声下长短时记忆(long short term memory,LSTM)神经网络调制信号识别方法抗冲击噪声能力弱和超参数难以确定的问题,本文提出了一种演化长短时记忆神经网络的调制识别方法。利用基于短时傅里叶变换的卷积神经网络(convolution neural network,CNN)去噪模型对数据集去噪;结合量子计算机制和旗鱼优化器(sailfish optimizer,SFO)设计了量子旗鱼算法(quantum sailfish algorithm,QSFA)去演化LSTM神经网络以获得最优的超参数;使用演化长短时记忆神经网络作为分类器进行自动调制信号识别。仿真结果表明,采用所设计的CNN去噪和演化长短时记忆神经网络模型,识别准确率有了大幅度的提高。量子旗鱼算法演化LSTM神经网络模型降低了传统LSTM神经网络容易陷于局部极小值或者过拟合的概率,当混合信噪比为0 dB,所提方法对11种调制信号的平均识别准确率达到90%以上。In order to solve the problems of weak resistance against impulsve noise and difficulty in determining hyperparameters of the modulation signal recognition method with long short-term memory(LSTM)neural network under impulse noise,this paper presents a modulation recognition method based on evolutionary LSTM neural network.The convolution neural network(CNN)denoising model based on short-time Fourier transform is used to denoise the data set.Then,combined with the quantum computation mechanism and sailfish optimizer(SFO),the quantum sailfish algorithm(QSFA)is designed to evolve LSTM neural network to obtain the optimal hyper-parameters.An evolutionary LSTM neural network is used as a classifier for automatic modulating signal recognition.Simulation results show that the recognition accuracy is greatly improved by using the designed CNN denoising and evolutionary LSTM neural network model.Moreover,the evolutionary LSTM neural network model based on quantum sailfish algorithm reduces the probability that traditional LSTM neural network is easy to fall into local minimum or over fitting.When the mixed signal-to-noise ratio(MSNR)is 0 dB,the average recognition accuracy of the proposed method for 11 modulated signals is more than 90%.

关 键 词：调制信号识别 冲击噪声 卷积神经网络 量子旗鱼优化算法 长短时记忆神经网络 稳定分布 超参数 短时傅里叶变换 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TN911.7[自动化与计算机技术—控制科学与工程]