| Generally,there are two parts in underwater target recognition of passive sonar: feature extraction and classification.In the earlier research,the above tasks are finished by sonarman.However,the way of artificial feature extraction is easy to lose information and can not guarantee the recognition accuracy and efficiency.Therefore,automatic recognition methods with higher efficiency,such as machine learning,are gradually proposed.Firstly,this dissertation studies the traditional machine learning methods,then establishes the underwater target recognition framework of the traditional machine learning method,and carries out simulation verification.According to the formation mechanism of ship radiated noise,the simulation model of noise is built.The simulated signals of mixed ship radiated noise under different signal-to-noise ratio are obtained by mixing the simulated ship radiated noise with the actually measured marine environment noise.The Mel frequency cepstral coefficients(MFCC)of signals are extracted and the support vector machine is used to classify.The results show that when the SNR is-2dB,the recognition accuracy of can reach 98.6%,but when the SNR decreases,the recognition accuracy drops rapidly.When the SNR is-10 dB,the recognition accuracy is only 55.9%.Secondly,this dissertation studies the feature preprocessing method of underwater target recognition,and proposes an underwater target recognition framework based on deep learning.The signal decomposition algorithm based on resonance is used to extract high resonance components of the measured ship radiated noise and simulated signal of mixed ship radiated noise.The spectral correlation coefficient(SCC)between the high resonance components of the two kinds of signals is 0.7074,which is smaller than the SCC between the original signals.Obviously it’s proved that the algorithm enhances the differentiation between the two kinds of signals which is beneficial to improve the recognition accuracy.The short-time Fourier transform is used to calculate the LOFAR spectrum for the high resonance component,and the line spectrum enhancement algorithm based on multi-step decision is used to detect and enhance the line spectrum which is closely related to underwater target recognition.The results show that all the line spectrum in the LOFAR spectrum with the SNR of-5dB can be detected,and the algorithm can complete the vacancy(“breakpoint”)of the line spectrum which emerged due to the interference of the marine environment.Finally,the performance of the traditional machine learning method is compared with that of the underwater target recognition based on the classical deep learning network structures.The convolutional neural networks(CNN)and long short term memory(LSTM)and traditional machine learning methods are respectively used to extract features of the line spectrum enhanced LOFAR spectrum which is generated by the measured ship radiated noise signals and the simulated signals of mixed ship radiated noise,and classification is done then.Several evaluation indexes like recognition accuracy,confusion matrix,receiver operating characteristic curve,AUC value are used to compare the recognition performance of the three methods.For the measured ship radiated noise,the best accuracy is obtained by CNN which reaches 95.22%.For the simulated signals of mixed ship radiated noise,the accuracy of all three methods is over 98% at a SNR of-2dB,and the two deep learning methods can both achieve nearly 80% accuracy when the SNR is-10 dB. |
