| Rolling bearing is one of the most important parts in many industrial instances.It is crucial for operation of mechanical equipments,but is prone to fail.Therefore,the identification of rolling bearing faults is important to ensure the safety of industrial production.In this paper,several methods were proposed based on deep learning for fault diagnosis of rolling bearings.Considering the ability to process non-stationary signals in wavelet packet decomposition and the ability to extract features in auto-encoder,the paper realized a rolling bearing fault identification based on wavelet packet decomposition and auto-encoder.Wavelet decomposition and auto-encoder was used to extract features of rolling bearing faults.Then,the fully connected layer were utilized to identify the bearing faults.This method was used in data produced in Western Reserve University.The accuracy rate reached 92.3%.However,the above method requires a step of manual feature extraction which makes it less applicable.Considering the advantages of LSTM in processing time series data,a method based on LSTM was proposed for diagnosis of bearing faults because of the ability of LSTM to process time series data.The method abolished manual feature extraction and established an end-to-end fault state identification,and achieved 94.8% in accuracy for the same bearing faults data.In order to improve the accuracy of fault identification,a rolling bearing fault diagnosis method was further designed with CNN-LSTM by the extracting local information capability of CNN.In the method,convolution layers and pooling layers of CNN replaced the input layer of LSTM.With the method,classification accuracy of 97.7% was achieved and the training time was shortened by more than half.Subsequently,the CNN-LSTM model for fault identification on the same fault data was verified under different conditions and with different noises.The results showed good generalization of the model.Finally,with the CNN-LSTM bearing fault diagnosis model,a method was designed to analyze degradation process of rolling bearing on NASA bearing data.The experiment showed that the method was competent to sense the bearing fault in the life cycle.Comparing with traditional methods,the designed method was found to predict bearing faults at least 6 days earlier. |
