Research On Analysis And Recognition Method Of Ship Passive Sonar Signal

Ship sonar signal recognition is one of the key technologies for the intelligentization of underwater acoustic equipment and weapon systems,and has important research significance.With the development of maritime combat intelligence and informatization,the battlefield environment is becoming more and more complex.Actual ship sonar signal recognition systems often face low signal-to-noise ratio,insufficient characterization capabilities,few samples,and many types of recognition.This thesis has studied from the following aspects to deal with these problems.The signal preprocessing method is researched.The passive sonar signal samples contain a large amount of marine noise,so,noise reduction algorithms are needed.According to the source and composition of the ship’s sonar signal,the singular value ratio spectrum denoising method and wavelet threshold denoising method are studied in this thesis.this thesis proposes a singular value ratio spectrum noise reduction method in wavelet transform domain,which can effectively improve the signal-to-noise ratio of the signal.In terms of feature extraction methods,this thesis studies the feature extraction methods related to auditory perception,and extracts the Mel cepstrum coefficient and Gammatone cepstrum coefficient of the ship’s sonar signal.For cepstrum coefficient features,only averaged spectral features can be expressed.and the relative information of the spectrum cannot be expressed.Therefore,this thesis proposes to use the spectral contrast feature and the cepstrum coefficient feature in series to improve the ability of expressing the feature.In the design of the classifier,this article applies deep learning theory to the research of ship sonar signal recognition,with analyzes and studies the traditional convolutional neural network and recurrent neural network.Aiming at the problem that the ship sonar signal samples are difficult to obtain and there are many types of ships,the feasibility of using similarity discrimination to replace the traditional threshold classification is analyzed.This paper constructs the sample pair as the input of the siamese network,uses the siamese network to determine the similarity,and determines the prediction result based on the similarity between the samples.Since the scale of the sample set is enlarged and the category labels are weakened,it can also have a high recognition rate in the case of small samples and multiple categories.Using the actual ship sonar signal as the data set,the method in this thesis is compared with the traditional threshold classification method.The experimental results show that the siamese network recognition effect is better.