Research On Graph Fourier Transformation And Neural Networks For Acoustic Beacon Signal Recognition

The aircraft black box is a key data recording instrument to solve the mystery of the air dis-aster.At the moment of contact with water,the acoustic beacon on the black box will send out a specific pulse signal.The key to the black box search for a crashed plane is whether the specific acoustic beacon signal can be effectively detected and recognized.This dissertation focuses on the recognition of marine acoustic beacon signals,and proposes an acoustic beacon signal recognition framework based on Graph Fourier Transform(GFT)and Convolutional Neural Network(CNN).First,in the preprocessing stage,the characteristics of the black box acoustic beacon signal and the noise characteristics of the receiving platform are analyzed,and the difference in the fre-quency spectrum of them is used to suppress the platform noise using bandpass filtering.In order to improve the received signal to noise ratio(SNR),beamforming is implemented to data received by a uniform linear array which lays the foundation for subsequent acoustic beacon signal feature extraction and recognition.Due to random spatial-temporal-frequency variability of underwater acoustic channels,short time Fourier transform is utilized to analyze the acoustic beacon signal and extract time frequency characteristics.Due to overcoming inaccuracy of feature extraction by traditional pattern recognition methods,convolutional neural network is utilized in recognition of the acoustic beacon signal.Considering the characteristics of the acoustic beacon signal in time frequency spectrum,more attention is fo-cused on training parameter setting and structure optimization of the convolutional neural network.Further,the channel attention mechanism is combined with the convolutional neural network for adaptively adjusting the weights of feature channels of convolution layers to improve the perfor-mance of deep convolutional neural networks.The processing results of simulation and experimen-tal data have shown that combination of the channel attention mechanism and the convolutional neural network performs well in recognition of the acoustic beacon signal.Finally,graph signal processing technology is utilized to further improve the performance of the acoustic beacon signal recognition based on the improved convolutional neural network.Not being graph structure data,the acoustic beacon data is first converted into a graph signal data.The cells of the time-frequency spectrum of the acoustic beacon signal are viewed as the graph nodes.The values in the cells are regarded as the amplitudes of graph signals.The Manhattan distance between amplitudes of neighbor frequency bins is utilized to construct an adjacency ma-trix.The Gauss kernel function of the Manhattan distance creates the edge weight of graph.After Graph Fourier Transform is carried out over the created graph signal,a feature representation can be achieved with stronger noise immunity and tighter intraclass distance in the graph frequency do-main.Simulation and experimental results have shown that the recognition rate of acoustic beacon signals is significantly improved under the GFT framework.