Research On Whale Call Classification Based On Convolutional Neural Network

Underwater target recognition has always been an important research content in the field of underwater acoustic signal processing,including the monitoring of marine biological signals.Marine mammalian sound research has important implications for the detection and imitation of marine ecological environment.In this paper,the four types of cetacean low frequency call signals are taken as the research object.The deep learning method is used to extract and identify the underwater targets compared with the traditional method with SVM,Random Forest as the classifier and MFCC as characteristics.These classification models are compared to analyze the characteristics and advantages of the deep learning method.The main work of the thesis is as follows:Firstly,the background and basic theoretical knowledge of underwater target recognition are introduced.The development history of underwater target recognition technology is briefly described from two aspects: feature extraction and classifier design.The development history of deep learning technology and the preliminary exploration in the field of underwater target recognition are introduced.The parameter model of the cetacean animal low frequency call signal is analyzed,and the traditional underwater signal feature extraction method is introduced: the MFCC.From the perspective of time-frequency analysis,the advantages of Chirplet Transform in analyzing bio-sound signals are discussed.A fast algorithm for fast chirp wavelet transform is introduced.Combined with the measured data of four types of whales,the MFCC feature,the Mel spectrum and the Chirplet spectrum of the cetacean call signal are respectively displayed.The principles of SVM and Random Forest algorithms are briefly introduced as a comparison model based on deep convolutional neural network classification model.Next,this thesis introduces the structural composition of one-dimensional and two-dimensional convolutional neural networks and the initialization method of network parameters.The effects of commonly used activation and loss functions on the learning of deep neural networks are studied.Then the parameters are optimized based on the gradient reduction method.The methods are compared and their performance comparison is given by simulation.The training process of convolutional neural network is introduced,and the detailed mathematical process is given.The over-fitting problem of deep learning model,regularization method and the effect of the deactivation rate on the recognition rate were studied.Finally,the evaluation system of the deep learning model is introduced.A model with Chirplet spectrum as input feature and deep convolutional neural network for classification and recognition is proposed.The accuracy and recall rate of the model are calculated by analyzing and processing the measured data.The analysis is compared with the model with the original waveform and the MFCC as inputs.The experimental results show that the proposed Chirplet-Resnet classification method can effectively improve the target recognition rate and effectively prevent the model from over-fitting.