Research On Underwater Acoustic Signal Recognition And Separation Based On Generative Adversarial Network

Underwater acoustic signal recognition is a technology that uses sonar or other collection equipment to collect and process underwater sounds,and identify and classify them;Underwater acoustic signal separation is to estimate and separate target signals from multi-source mixed underwater acoustic signals technology.In the actual underwater acoustic environment,there are many external interferences and noises,and there are still big obstacles to improve the accuracy of underwater signal recognition and better separate the target signal.With the development of deep learning,data-driven deep learning methods are gradually applied to underwater acoustic signal recognition and separation tasks.Based on the deep learning model,this paper applies the idea of generative adversarial idea to the task of underwater acoustic signal recognition and separation.The specific work is as follows:The first is the resarch of single signal recognition and mixed signal recognition for underwater acoustic signal recognition tasks.Two datasets of single underwater acoustic signal and mixed underwater acoustic signal are constructed.At the feature level,it is verified that in a single signal recognition task,the recognition accuracy based on the frequency domain feature is better than the recognition accuracy based on the time series feature.In the model structure,the mixed structure of the convolution structure and the recurrent neural network is obtained.The highest recognition accuracy rate proves the superiority of using the deep convolution structure to extract the features in the frequency domain and time dimension and then feeding the recurrent neural network structure in the recognition task;in the mixed signal recognition task,the same,the frequency domain feature recognition The accuracy rate is significantly higher than that based on time domain features,and the attention mechanism is introduced in the model structure to increase the recognition accuracy.Secondly,aiming at the problem of unbalanced or missing underwater acoustic signal data set,the research on data enhancement of underwater acoustic signal based on generation adversarial network was launched.Realize data enhancement from the perspective of time domain and frequency domain: From the perspective of time domain,it is verified that the time-series convolution used has better generation results than other one-dimensional time-series generation structures;from the perspective of frequency domain,deep convolution is used The combined structure of the generated confrontation network and the conditional generation confrontation network.The category loss is introduced on the discriminator side,and the self-attention module is introduced on the generator side to obtain a better generation effect.At the same time,after fine-tuning the original recognition network with the generated samples,The recognition rate on the test set has been improved,which proves the effectiveness of data enhancement using the generative confrontation network.Finally,the methods of underwater acoustic signal separation based on generative adversarial network are studied.In terms of the overall separation structure,the category condition restriction is introduced at the input of the separation model,and the discriminator loss is introduced at the output to reduce the distance between the estimated target distribution and the true target distribution,and improve the separation quality and authenticity of the estimated target.In terms of learning separation targets,the applicability of spectrum mapping and time-frequency masking methods are verified,and the applicability of each target is summarized;in terms of the model structure: separate structures such as autoencoders are used to achieve better results Separation performance,and for the first time proposed a normalized separation structure based on adaptive examples,and achieved the best separation performance.