Research On Neural Network Recognition Technology Of Underwater Acoustic Signal Based On Denoising

The current marine environment is becoming more and more complex.The complex underwater environment greatly reduces the effect of traditional underwater target recognition technology.With the development of denoising technology,signal feature extraction technology and artificial intelligence technology,technological research of underwater target recognition has also achieved remarkable results.Firstly,according to the characteristics of the marine environment,the unprocessed underwater acoustic time domain signal is not input directly during the research,but the feature of the signal is extracted first.this thesis studies the preprocessing methods of underwater target signals,and uses two different types of feature extraction methods.One is the relatively simple and easy-to-implement short-time Fourier transform,and the other is the wavelet transform with multi-resolution characteristics.The advantages and disadvantages of the two feature extraction methods are different,the application scenarios are different,and the ability to extract different types of signal features is also different.Secondly,according to the actual underwater target signal,the noise characteristics are studied,and a denoising method based on singular value decomposition and a wavelet threshold denoising method are proposed.Through experimental comparison,the wavelet threshold denoising method is better,so we use wavelet threshold to denoise.The noise method denoises the underwater target signal.In the wavelet threshold denoising method,two different methods are proposed: sym8 and db2 wavelet denoising.Experiments show that the two different wavelet threshold denoising methods can suppress noise.The effects are similar,both of which can greatly improve the signal-to-noise ratio of the measured underwater acoustic signal,and sym8 wavelet denoising method has the best effect.Finally,in the target signal recognition stage,this thesis adopts a simple,efficient and easy-to-implement new underwater target neural network recognition structure based on the convolutional neural network recognition framework: MFA-Coformer.The design of MFA-Conformer is mainly inspired by the recent end-to-end speech recognition network(Conformer)and speaker recognition network(ECAPA-TDNN): it first uses a convolutional downsampling module to downsample the input acoustic features,thereby reducing the amount of model computation;then use multiple different Conformer blocks to learn local features and global features;finally,splicing the output between different Conformer blocks,and extracting speech features through an attention statistics pooling layer(Attentive Statistics Pooling).Then,by observing the equivalent error rate and minimizing the cost function,the results of the input network of the underwater acoustic signal without denoising processing and after denoising are compared.The experiment shows that through the MFACoformer framework,the signal error rate of denoising processing reaches 1.35%,and the error rate of the signal without denoising processing is 3.02%,which shows that the denoising preprocessing of the signal is effective.