Improved Probabilistic Spiking Neural Network Method And Its Application On Mechanical Fault Diagnosis

With the development of the artificial intelligence,the research of human brain has been changed with great speed.How to use a computer to simulate the operation mechanism of the brain neurons has attracted widespread attention,then the artificial neural network is proposed.The artificial neural network is still inadequate,because the artificial neural network only simulates the neuron connection mode,and it cannot simulate the mechanism of the brain neuron which transmits information with spike.The research of the Spiking Neural Network(SNN)came into being.The SNN which provides a potential computing paradigm for simulating the complex information processing mechanism of the brain is a research direction formed by the intersection of computer science and biological neuroscience.Due to the non-differentiable character of the firing function of most spiking neuron models,it is not possible to use the backpropagation algorithm to iteratively optimize the synaptic efficacies like the deep neural network for the SNN.Thus,the development of the deep SNN is limited.Aiming at the problem,a deep SNN supervised learning algorithm based on the improved PSRM(Probabilistic Spike Response Model)neuron model is proposed in this thesis.Meanwhile,the proposed algorithm is applied to the rotating machine fault diagnosis for exploring the capability of the SNN in terms of fault diagnosis.The main research content of this thesis are as follows:(1)An improved PSRM neuron model has been proposed.By defining the probabilistic expression form of the spike train,the improved PSRM neuron model which solves the shortcomings of the weak scalability of the PSRM neuron model is proposed in this thesis.Based on the important property of continuous differentiability of the improved PSRM neuron model,the SNN can iteratively optimize the synaptic efficacies through the back-propagation algorithm.In order to follow up the derivation of the supervised learning algorithm,a recursive relation of the improved PSRM neuron model membrane voltage on time series is derived.(2)A supervised learning algorithm based on the improved PSRM model has been proposed.The single-layer and multi-layer learning algorithms based on the improved PSRM model are proposed,and the effectiveness of the two algorithms is verified by the experiments.The multi-layer learning algorithm based on the improved PSRM model is further applied to some classification problems in this thesis,and the experimental results show that the proposed method can achieve a promising accuracy.(3)A fault diagnosis method for the rotating machine based on the improved PSRM neuron model has been proposed.The multi-layer learning algorithm based on the improved PSRM model is applied to the bearing fault diagnosis and its dataset.The features are extracted from data by the local mean decomposition algorithm,and the fault diagnosis model is designed.The multi-layer learning algorithm based on the improved PSRM model is used to train the fault diagnosis model,and the well-trained model is used for fault diagnosis.The proposed method achieves high test accuracy on the two bearing fault datasets.