Research On Ship Radiated Noise Recognition Technology Based On Machine Learning

Ship radiated noise recognition technology is an important branch of underwater acoustic target recognition.It has high civil and military value.It has gradually become one of the hot research technologies in underwater acoustic signal processing.Affected by the changeable ship navigation conditions,complex ocean transmission channel,scarcity of sample data and other factors,ship radiated noise data often have problems such as small sample data and unbalanced number,which brings great challenges to ship target recognition.This paper verifies the application of two machine learning theories of statistical learning and deep learning in ship radiated noise recognition,and solves the problem of ship radiated noise recognition with small samples and unbalanced number of samples,in order to promote the research progress of ship radiated noise intelligent recognition technology.This paper introduces from the following four aspects:1.This paper studies the formation mechanism of ship radiated noise source,and establishes a ship radiated noise simulation model with clear physical significance according to the spectrum characteristics of ship radiated noise;This paper discusses the feasibility of five underwater acoustic signal processing methods: LOFAR spectrum analysis,demon spectrum analysis,3 / 2-dimensional spectrum analysis,wavelet transform and Mel cepstrum analysis in the field of ship radiated noise feature extraction,so as to provide feature vector for ship target recognition.2.Combined with the machine learning theory of statistical learning,the line spectrum characteristics,modulation spectrum characteristics and capability scale characteristics of ship radiated noise are extracted by feature fusion technology,and the comprehensive feature vector of ship radiated noise is established.Combined with k-fold cross validation and network search method,the ship radiated noise recognition model based on support vector machine is optimized;In the problem of small sample ship target recognition,the least squares support vector machine ship radiated noise recognition model is established to improve the accuracy;Under the condition of unbalanced number of ship target samples,a class of weighted support vector machine ship radiated noise recognition model is proposed to improve the recognition accuracy.3.Combined with the machine learning theory of deep learning,this paper uses LOFAR spectrum analysis and line spectrum enhancement technology as preprocessing means to establish the ship radiated noise recognition model based on alexnet convolutional neural network,and obtains the optimal recognition performance by adjusting the model parameters.In this paper,Mel cepstrum analysis and demon spectrum axis frequency estimation technology are used as preprocessing methods to establish a ship radiated noise recognition model based on long-term and short-term memory neural network,which still has high recognition accuracy under the condition of unbalanced samples;The deep learning model based on transfer learning greatly improves the recognition accuracy of ship radiated noise recognition model in small sample and unbalanced target recognition.4.According to the shipsear design experiment,which is a network public ship radiated noise data set,the performance of different classifiers in ship radiated noise multivariate classification,small sample ship radiated noise recognition and sample unbalanced ship radiated noise recognition are discussed respectively.The convolution neural network based on transfer learning has the highest recognition accuracy in the recognition of small sample ship radiated noise.The long-term and short-term memory neural network based on transfer learning is the least affected by the unbalanced number of ship samples,and the recognition accuracy can reach more than 90%.