Research On Time-varying Signal Classification And Face Recognition Technology Based On Deep Wavelet Neural Network

Wavelet analysis is a discipline developed on the basis of applied mathematics.Because wavelet function has good ability of time-frequency local characteristic analysis,wavelet analysis is widely used in time-domain signal and image feature extraction,representation and analysis,which is an important research and application field of basic science and information science.Wavelet neural network is a model based on wavelet analysis,which combines the self-learning and self-adaptive characteristics of neural network.The model uses wavelet basis functions to discover and highlight the time-frequency characteristics of the signal and optimize the parameters and error space of the neural network.The weighted inner product of the wavelet base and the input vector is used to complete the feature extraction of the input layer,so that the network has a strong nonlinear fitting ability and a fast convergence speed,and has achieved rich results in signal analysis and image processing.With the rapid development of modern science and technology and the continuous expansion of production scale,the research object becomes increasingly complex.One dimensional signals are mostly time-varying signals,which mainly show time-varying,nonlinear,strong coupling,uncertainty and so on.The two-dimensional image has the characteristics of randomness and real-time.It is easy to be interfered by brightness,clarity,angle and other aspects,which makes the research very difficult.Based on the wavelet neural network,which has better capability of micro analysis and feature representation for both one-dimensional signal and two-dimensional image,this paper proposes two kinds of improved wavelet neural network models for time-varying signal classification and face image recognition.The main innovations and research work include the following two aspects:1.Aiming at the problem of multi-channel time-varying signal classification,a sparse autoencoder deep wavelet process neural network(SAE-WPNN)is constructed.By constructing a multi-input/multi-output wavelet process neural network(WPNN),multi-scale decomposition of time-varying signals and preliminary extraction of process distribution features are realized;By superimposing a SAE depth network after the WPNN hidden layer,the extracted signal features are synthesized and represented at a high level,and the time-varying signals are classified based on the softmax classifier.The model combines multi-scale feature decomposition and representation based on wavelet basis,the classification mechanism and learning ability of PNN for time-varying signals,and the high-level comprehensive ability of SAE deep networks for signal features.The model has a good adaptability to the classification of complex time-varying signals in mechanism.Experiments were conducted on the classification and diagnosis of seven cardiovascular diseases based on 12-lead ECG signals in the Chinese Cardiovascular Disease Database.The results verified the effectiveness.of the model and algorithm.2.Aiming at the problem of face image recognition,a recursive wavelet neural network(RWNN)based on wavelet multi-resolution analysis is established.The model is based on wavelet multi-resolution analysis,using scale function and wavelet function as the hidden layer excitation function of the network to extract low-frequency contour features and high-frequency detail features of the image.The model realizes multi-scale analysis of two-dimensional face images from coarse to fine,reducing the redundancy of hidden layer representation.At the same time,combined with the feature memory capability and classification mechanism of the recurrent neural network,the face image features can be identified and recognized more accurately.Experiments were performed on the image recognition of LFW face image data set,which verified the feasibility of the method.