关键词: 辐射源识别, 稀疏降噪自编码, 时频特征, 核映射, 批量随机梯度下降法, dropout正则化

Abstract: To enhance the classification performance and noise sensitivity of emitter signal recognition,a recognition system based on kernel space time-frequency feature and stacked sparse denoising auto-encoders (SSDAE) is proposed.Firstly,the noise interference and feature redundancy reduced by time-frequency transform,sparse-domain denoising and kernel space dimensionality reduction.Then,it is based on the idea of sparse auto-encoder (SAE) and denoising auto-encoder (DAE),an SSDAE based recognition network is constructed.Experimental results show that the system has the best comprehensive performance in recognition rate and time efficiency,which can significantly reduce noise sensitivity and improve low SNR environment adaptability.When the SNR is -12dB,the overall average recognition rate of the system for the 8 types of emitter signals reaches 96.75%.

Key words: emitter signal recognition, sparse denoising auto-encoder, time-frequency feature, kernel mapping, mini-batch stochastic gradient descent method (MSGD), dropout regularization