Study On Bearing Fault Diagnosis Algorithm Base On Neural Network

With the coming of the industrial age,the level of manufacturing industry and the quality have changed greatly.The era of "industry 4.0" is getting closer to people.In China,"made in China 2025" plan is proposed for better development in the future,which guides the Chinese manufacturing industry to develop in a more refined,intelligent,rapid and secure direction.Generally speaking,intelligent diagnosis technology has accurate diagnosis results and short working time when performing tasks.However,existing methods are vulnerable to the size of bearing datasets and do not care enough about generalization performance,what is more worrying is that these methods are basically helpless in the face of noisy environments.To solve these issues,fault diagnosis of rolling bearing based on neural network is proposed.This paper includes the following aspects:(1)Firstly,a new algorithm for multi-scale convolutional neural network(MSCNN)classification of bearing health status was firstly proposed in this paper.This methodology takes the original vibration timing signal of bearing as input,uses convolutional neural network to obtain the bearing fault feature information,and finally the classification function implements the final diagnosis classification.For best results,the model uses sample data with a certain repetition rate,which significantly expands the number of samples,and obtains features that are more representative of fault types through synchronously extracting vibration signals using convolution kernels with different sizes.MSCNN not only achieves nearly 100% fault recognition accuracy with a small amount of data but also has excellent generalization performance in the case of different data sets.(2)While MSCNN has good property of integration,the fault diagnosis rate of MSCNN is greatly reduced in noisy environment.To solve this problem,self-attention deep neural network fault diagnosis algorithm(SA-DNN)is proposed.Compared with MSCNN,the way of vibration feature extraction method of SA-DNN is replaced by self-attention.The input of SA-DNN is still the original diagnosis signal,but it needs to be further processed by the technology of "feature extension" which achieved by residual learning before feature extraction;then,the different feature weights with respect to different feature vectors were generated by using self-attention network.Finally,the whole diagnosis process is completed by carried out classification function.In this method,the residual network is used to expand the characteristics of vibration signals,and the attention network can screen out more representative eigenvectors than CNN.The most noteworthy of SA-DNN is the superior performance in noisy and complex environment.