Fault Signal Diagnosis Based On Pulse Convolutional Neural Networ

Bearing fault diagnosis of electrical equipment has been a popular research area in recent years because there are often some faults during continuous operation in production due to the harsh working environment.Traditional fault diagnosis methods based on signal processing rely on highly expert experience,while fault diagnosis methods based on machine learning also suffer from problems such as low recognition accuracy.With the development of deep learning in industrial applications,the application of deep learning methods to bearing fault diagnosis has become a new research hotspot.In this paper,the spiking neural network(SNN)is applied to bearing fault diagnosis,combined with deep learning methods,and trained on a fault signal dataset to establish a model that integrates two fusion attention mechanisms and SNN.This provides new ideas and methods for the application of deep learning in bearing fault diagnosis.The main research contents are as follows:(1)Artificial neural networks and convolutional neural networks are currently used by most deep learning methods applied to bearing fault diagnosis.However,the information processing of artificial neural networks cannot imitate the information transmission mechanism of biological brain neurons,and deep convolutional neural networks also suffer from slow training speed.In this paper,SNN is introduced into bearing fault diagnosis to overcome these problems,and the gradients of SNN neurons are transformed and substituted to enable direct training on bearing fault datasets.(2)A deep spiking residual shrinkage network(DSRSN)is proposed in this paper for bearing fault diagnosis under the condition of large amounts of high background noise interference during actual production processes.DSRSN uses the spiking residual network as the backbone network and introduces attention mechanisms and soft thresholding processing in the model.During the training process,the data is preprocessed by simulating the actual production situation through the manual addition of noise.The noisy one-dimensional fault signal is then converted into a twodimensional gray-scale image as the input of the model.Through experiments on bearing fault signals with signal-to-noise ratios of-9d B to 0d B and comparison with common models,it is demonstrated that the proposed model has the highest recognition accuracy under different noise intensities.When the signal-to-noise ratio is-8d B,the accuracy of DSRSN can reach 93.2%,which is 4.8% higher than that of the second-ranked Res Net,and the training time is about three times faster than that of artificial neural networks,highlighting the high efficiency of SNN.(3)In this paper,a deep spiking residual shrinkage network with mix attention(DSRSN-MA)model based on a hybrid-domain attention mechanism is proposed,which further focuses on the features of bearing fault signals and the spiking neural output pulses in the spatial domain by fusing multiple attention mechanisms.By fusing multiple attention mechanisms,more accurate diagnosis of bearing fault signals can be achieved by the DSRSN-MA model.According to experimental comparisons,the DSRSN-MA model demonstrates superior performance in bearing fault diagnosis with a higher degree of accuracy.The recognition accuracy of the three different connection modes of the hybrid-domain attention mechanism model is respectively higher by 1.16%,0.59%,and0.24% than that of the DSRSN model.In addition,the best connection mode of the hybrid-domain attention mechanism in the DSRSN-MA model is determined according to the experimental results of the F1-score of different models,which further improves the performance of the model.