水声JANUS信号的分数低阶时频谱迁移学习识别方法

非合作第三方水下标准协议信号识别在水声通信信号识别中具有重要研究意义。针对浅海水声JANUS信号的特征提取因易受脉冲噪声和多径效应等复杂水声环境影响而导致识别率低下的问题,提出一种分数低阶时频谱和ResNet18 （Residual Network 18）相结合的迁移学习识别方法。首先,选取JANUS固定前导作为识别对象,设计分数低阶傅里叶同步压缩变换（FLOFSST）,以分数低阶操作抑制脉冲噪声,以时频重排特性增强时频集中性。其次,将基于ImageNet的ResNet18预训练模型微调,迁移至JANUS信号和常见水声信号时频图集。仿真表明所提算法在信噪比为-10 dB时JANUS信号的识别率为96.15%,能够有效抑制脉冲噪声并减小多径效应影响,比传统算法识别性能好。海试中JANUS信号识别率达90.00%,证明算法识别准确率和网络的泛化性较高。

Fractional lower order time-frequency spectrum and transfer learning method for underwater JANUS signal recognition

Non-cooperative and third party underwater standard protocol signal recognition has important research significance in the field of underwater acoustic communication signal recognition.Feature extraction of shallow sea JANUS signal is easily affected by complex underwater acoustic interferences such as impulsive noise and multipath effect,which lead to low recognition rate.To solve this problem,a recognition method based on fractional lower order time-frequency spectrum and Residual Network 18 is proposed.First,the JANUS preamble signal is selected as the recognition object,a fractional lower order fourier synchrosqueezing transform method is designed to suppress impulsive noise by operation of fractional lower order and to improve time-frequency concentration by characteristics of timefrequency rearrangement.Secondly,ResNet18 pre-training model based on ImageNet is fine-tuned,then we train the time-frequency image sets of JANUS signal and other common underwater acoustic signals on this network,and the time-frequency domain features are extracted for recognition.The results of the simulation shows that the proposed algorithm has a high recognition rate of 96.15% when the SNR is-10 dB,it can suppress impulsive noise and reduce the influence of multipath effect,and it has better recognition performance than traditional algorithms.In the sea test,the recognition rate of JANUS signal is 90.00%,confirming that the recognition accuracy of the algorithm and the generalization of the network are relatively high.