Research On Sonar Signal Feature Enhancement Technology Based On Deep Learning

The enhancement of sonar signal feature has important research value for sonar detection,tracking,recognition and so on.In recent years,machine learning has provided a new solution for sonar signal feature enhancement.In this paper,the representative LFM signal and ship radiated signal are selected as the research objects respectively,and the application of sonar signal feature extraction,enhancement processing and target recognition are studied.The main research contents include:In the aspect of active sonar signal feature enhancement,the signal enhancement algorithm based on Stack Automatic Encoder(SAE)is studied.The algorithm takes the time-domain signal as the input feature,uses the characteristics of Denoising Automatic Encoder to enhance the signal as a whole,and combines the characteristics of full Convolution Denoising Automatic Encoder to optimize the local details of the signal to realize the LFM signal enhancement.The simulation results show that SAE can effectively enhance LFM signal when the ambient noise is Gauss white noise in the fixed sea area and the parameters of active sonar signal are known;when the target signal is single frequency pulse signal,even secondary frequency modulation signal and other active sonar signals,or when the ambient noise is the environmental noise of the port of Vigo in Spain and the environmental noise of the Arctic,SAE also has good applicability.In the aspect of ship radiated signal feature enhancement,we study the enhancement algorithm based on the combination of logarithmic power spectrum and Deep Neural Network(DNN),and the enhancement algorithm based on the combination of time-frequency characteristics and Denosing Convolutional Neural Networks(Dn CNN),and evaluate its performance;explore the classification performance of different signal-to-noise ratio time-frequency characteristics based on Goog Le Net.The simulation results show that DNN and Dn CNN can effectively enhance the relevant features,while Dn CNN can enhance the overall time-frequency features better in the case of fixed sea area,known ship type,displacement and speed,when the environmental noise is Gauss white noise,Spanish Vigo port environmental noise and Arctic environmental noise;in the classification experiment based on Goog Le Net,the ship radiation signal when the SNR is 7.5db,the target recognition rate is more than 80%.The final field test results show that SAE can effectively enhance LFM signal,and the signal-to-noise ratio increases nearly 14 d B;for non-stationary environmental noise,Dn CNN can effectively eliminate noise features in time-frequency features,and the target recognition rate increases from 72.99% before enhancement to 86.57%.