Recognition Of Ship Radiated Noise Based On Deep Learning

Ship target recognition is one of the key contents of underwater acoustics research,which is of great significance to ocean exploration and modern naval warfare.Recognition using the radiated noise of ships working is an important means of ship target recognition.However,the sounding mechanism of ship radiated noise and the complicated marine environment require a recognition method with stable performance and high accuracy.As a more popular classification algorithm in recent years,deep learning has been widely used in image,speech and other fields.Therefore,this thesis introduces the method of deep learning to classify radiated noise signals.This article focuses on the subject of ship radiated noise classification,and has done the following work:1.Analysis of the characteristics of ship radiated noise.The concept of ship radiated noise and the different noise sources that produce ship radiated noise are introduced.The sound source level and frequency spectrum characteristics of ship radiated noise,as well as the acoustic characteristics of radiated noise,such as directivity and passing characteristics,are discussed.2.Feature extraction of ship radiated noise.This part is mainly divided into two aspects.One is to extract MFCC and LPCC two characteristics based on auditory characteristics and the characteristics based on wavelet packet energy from ship radiated noise.The other is to use wavelet packet transform has the ability of decompose the signal by frequency band.Multi-scale decomposition of the noise signal is performed,and the signal is decomposed into sub-band signals on different frequency bands.Then the two parts are combined,and the subband signals are extracted MFCC,LPCC,and wavelet packet energy feature after the signal decomposition.3.Preprocessing of feature data.Due to the high dimension and the large amount of feature data,direct classification will lead to a large amount of calculation,longer model learning time,and poor real-time performance.Therefore,this thesis uses Fisher’s criterion and canonical correlation analysis CCA to preprocess the signal.Fisher discrimination makes the intra-class distance of the data sample smaller and the inter-class distance larger,which can improve the separability of the data.CCA achieves dimensionality reduction between different features of a signal,reduces data redundancy,this will improve model learning speed and real-time performance.4.Deep learning model recognition.This article mainly uses two deep learning models,CNN and LSTM,for data classification and recognition.And use the two models to carry out classification experiments,compare with the classification results of the support vector machine model SVM in machine learning,at the same time conduct experimental analysis on the performance of the feature extraction method and data preprocessing algorithm proposed in this thesis.These experiments verify the effectiveness of the system.