A Study On The Auditory-Based Algorithms For Underwater Target Radiated Noise Identification

Feature extraction and recognition for underwater target radiated noise is one of the crucial problems in passive sonar system.Based on auditory features,this paper applies Mel frequency cepstrum coefficients(MFCC)to target radiated noise feature extraction.Image texture features of MFCC are analysed and a Markov model for MFCC is studied.One-frame recognition algorithm is designed.The concept of modified gray-scale level co-occurance matrix and histogram is proposed.A multi-frame time accumulating algorithm is analysed.The effectiveness of these algorithms is verified by practical data.In chapter 2,firstly,concept of Mel frequency is introduced.Then,methods to design a Mel filter bank and to calculate MFCC are analysed.The MFCCs of five kinds of practical signals are calcualted.In chapter 3,the method using image texture features to describe the way MFCC changes in one frame is proposed by observing the experimental results in chapter 2.The problem how to quantify MFCC is firstly analysed and solved.Then image texture features are studied and those for five kinds of signals based on histogram and gray-scale level co-occurance matrix(GLCM)are calculated.In chapter 4,the Markov model for MFCC is obtained by analysing the experimental results in chapter 3.Then one-step transition probability matrix and initial probability distribution are calculated using GLCM and histogram.Then the forward-backward(FB)algorithm in hidden Markov model(HMM)is applied to the one-frame recognition.Steps and process of one-frame recoginition algorithm are also analysed.To improve the robustness and stability of recognition,the concept of modified GLCM and histogram is proposed.Computational complexity is also analysed.A multi-frame time accumulating algorithm is designed to enhance the recognition rate.The effectiveness of the algorithms in this chapter is verified by the experimental results of 5 kinds of signals' recognition experiment.