Research On Feature Extraction And Classification Of Ship Noise And Whale Sound

Detection, classification and localization are among the most important andchallenging goals of underwater acoustic signal processing and classification of underwateracoustic signal has been a research highlight in the field of underwater acoustic signalprocessing for a long time. It is well known that the features presented to a classifier play acrucial role on its performance. Indeed, the classifier architecture is also very important tothe performance. The research of this paper focused on the algorithm of feature extractionof whale calls and ship noises and the design of classifier. The signals which are processedin the paper are the ship noises and whale calls. The aim of the research of this paper is toimprove the performance of classification in the field of passive sonar signal processing.The following contents are included in the paper.1、The fundamental theory of classification of underwater acoustic signal is discussed,including the composing of classification system and the function of each part. It alsofocused on the existing feature extraction algorithms and their advantages anddisadvantages, as well as the influence for recognition performance of classifier itself.2、With the advantages of wavelet packet transform in time-frequency analysis, thetheory and algorithm of wavelet packet transform are studied. An algorithm of extractingenergy features based on wavelet packet transform is proposed and the influence ofclassification performance using different mother wavelet and different layer number arediscussed. It has been shown that using a wavelet representation of acoustic signals canachieve improved classification performance.3、The Hilbert-Huang Transform (HHT) presents a completely new approach to theanalysis of time series data. Its essential feature is the use of an adaptive time-frequencydecomposition that does not impose a fixed basis set on the data, and therefore it is notlimited by the time-frequency uncertainty relation characteristic of Fourier analysis orwavelet analysis. This leads to a highly efficient tool for the investigation of transient andnonlinear signals. This article describes the application of the HHT to underwater signaldata analysis. The results show that the method has improved performance and it can beapplied to feature extraction of whale calls and ship noises.4、Try to use the techniques of speaker recognition to classify the underwater acousticsignal. Mel-frequency cepstrum coefficients (MFCC) expressing human ear’s hearingfrequency characteristic are based on the non-linear relation of Mel-frequency and Fourierfrequency and also perform a high recognition rate in speaker recognition as the features of classification. In theory, the recognition principles of human’s ears are quite similar with theprocessing of underwater acoustic signal classification. Therefore MFCC can be applied tothe recognition of ship noises and whale calls as the extracting features. The effect of thedimensions of MFCC and the combination of the MFCC features for recognitionperformance are discussed. The results show that using MFCC as features for classificationof ship noises and whale calls has higher recognition rates and this method is efficient forunderwater acoustic signal classification.5、Two kinds of classifier, Artificial Neural Network and Gaussian Mixture Models,used in this paper are discussed. The concept, principle and algorithm are discussed as wellas the parameters setting.6、Experimental data processing and the analysis of classification results verify thatthe proposed feature extraction methods and classifiers are efficient.