Study On Image Recognition And Bearing Fault Diagnosis Based On Spiking Neural Network

Human brain simulation has become more in-depth research due to the rapid advancement of artificial intelligence technology.The development of brain-like computer techniques is growing,particularly since the concept of spiking neural network(SNN),which has propelled the field to new heights.Nowadays,spiking neural networks have produced superior results in the areas of image recognition,speech recognition,reinforcement learning and other areas,which combine the workings of biological systems and machine learning algorithms.The third generation of artificial neural network,known as spiking neural networks,exhibits higher bio-interpretability than conventional deep learning.In neuromorphic designs,it exhibits exceptionally low energy consumption since it transmits information via discrete pulse sequences.The thesis simulates and constructs an spiking neural network structure based on LIF neurons to take use of the benefits of spiking neural networks.And the algorithm is effectively applied to rotating machinery bearing fault diagnosis to investigate new applications of spiking neural networks in processing sound signals,fault identification,etc.The following list represents the major points of this thesis.(1)Comparison of the model structure and performance of two types of neural networks,spiking neural networks and artificial neural networks.The synaptic structure of neurons and the transmission of inter-synaptic spikings are simulated based on the fundamental features of biological synapses.The main effects of synaptic weights and synaptic delays on the neural connection process are discussed.The larger the synaptic weight,the shorter the time to accumulate voltage to the threshold and to deliver the pulse;the addition of synaptic delays causes each neuron to lag behind the corresponding time while transmitting signals.The MNIST dataset is used in conjunction with the foundation of the previous simulation for the simulation research of digital image recognition.(2)A brand-new neural network transition algorithm is put forth to allow the conversion of spiking neural networks to artificial neural networks.The training test results are achieved by adjusting the activation and synaptic values in the network to be as efficient as those of the artificial neural network.It is shown that the proposed transformation algorithm has the same high accuracy as the artificial neural network,while the network structure is more simple and uses less power by using the improved spiking algorithm to the practical problem of Fashion MNIST image multi-classification.(3)A multilayer spiking neural network based on the LIF neuron model is proposed and successfully applied to the bearing fault diagnosis problem.The original bearing vibration signal is decomposed by the CEEMD method,and the features are extracted from the decomposed bearing signal.After that,a model which uses a vast quantity of training data to identify faults in the bearing signal for fault diagnosis is built.Experiments demonstrate that the proposed model,which is compared and examined with six bearing diagnosis methods on the CWRU dataset,achieves 99.17% test accuracy.