Study On Bearing Fault Diagnosis Based On Compressed Deep Neural Network

With the popularization and development of complex machinery and equipment in industrial production,the stability and safety of their work is increasingly important.Rolling bearings,as one of the core components of complex machinery,are of great significance for real-time online monitoring and diagnosis of their operation status.In recent years,intelligent diagnosis methods based on deep learning have made rapid development,but due to the problems of complex network,large number of parameters,and high requirements for software and hardware,it is difficult to apply in the industrial environment with limited resources,for this reason,this paper studies the bearing fault diagnosis method based on compressed deep neural network.To begin with,aiming at the problem of excessive number of parameters and large redundancy in deep convolutional networks,a compressed deep neural network method for bearing fault diagnosis is proposed,combining multiple network pruning,network quantization and matrix compression methods from multiple levels to achieve multi-level compression of the network model.First,structured pruning is used to remove the filters corresponding to the low-rank feature maps in the convolutional layer;then unstructured pruning is used to remove the non-important connections in the fully connected layer;finally,the number of bits required for parameter representation is reduced by parameter quantization of the weight matrix,and the amount of parameter storage in the network is further reduced by combining the compressed storage method of the weight matrix.The experimental results show that the proposed compression method greatly reduces the complexity of the model while ensuring a high diagnostic accuracy,which is a useful exploration for the practical application of deep neural network methods.Secondly,in view of the existing lightweight network where redundant channels still exist in the convolutional layers and the dense connection of fully connected layers causes a large number of parameters,an improved lightweight deep convolutional neural network for bearing fault diagnosis is proposed.The method first constructs a more lightweight feature extraction structure via trimming redundant channels in heterogeneous convolution,while achieves both a reduction of parameter volume and fast excavation of fault characteristics.Then,the convolutional attention mechanism is introduced to optimize the extracted feature weights,and the optimized features are processed by dropout to relieve overfitting.Finally,the dropout processed features are superimposed with the original features to improve model stability,subsequently a global average pooling operation is adopted to achieve feature dimensionality reduction,and Softmax is employed to output the fault recognition results.Finally,The proposed method is validated on several publicly available datasets.The paper also analyzes the effects of several parameters in the method and compares it with commonly used methods,and a large number of experimental results demonstrate the effectiveness of the research method in the paper.