Data-driven Bearing Fault Diagnosis And Prediction

Rolling bearings are one of the most widely used important components in rotating machinery and equipment,so their health status plays a decisive role in the safe and reliable operation of the equipment system.In-depth research on the fault diagnosis and remaining useful life prediction technology of rolling bearings can reduce the economic cost while effectively guaranteeing the safety of personnel and machinery,which has great academic significance and engineering application value.Therefore,this thesis investigates the fault diagnosis and remaining useful life of bearings based on a datadriven approach.(1)The bearing fault data set of Case Western Reserve University was selected as the research object,and the fault diagnosis models based on convolutional neural network(CNN)and long and short-term memory network(LSTM)were built for different fault types of bearing vibration signals,and the models were directly applied to the original vibration signals for diagnosis,with an accuracy of 97.8% and 98.8%,respectively.The experimental results show that both of them can realize the "end-toend" mode from the original data input to the fault classification output,which demonstrates the feasibility of the models.(2)Although the fault diagnosis models based on CNN network and LSTM network have achieved high diagnostic accuracy,the CNN model ignores the timing of vibration signals and the back-and-forth correlations compared with LSTM,while the LSTM model has large oscillations in the early training period and it converges slowly,which are difficult to meet the requirement of efficient diagnosis.Therefore,considering the fault feature information of vibration data,CNN-LSTM model is designed,in which the input data are passed through two network channels,CNN and LSTM,respectively,and on this basis,the feature information extracted from the two parts is fused through the fusion layer to complete the feature fusion for diagnosis,with an accuracy of 99.8%.In addition,when tested without data augmentation,the average recognition accuracy of the CNN-LSTM model for 10 tests was 99.3%,which indicates that the CNN-LSTM model has a higher diagnostic accuracy and faster convergence rate.The CNN-LSTM model was also applied to the fault dataset under different working conditions,and the diagnostic accuracy reached 98.2%,which also obtained good diagnostic results,further illustrating the superiority of the CNN-LSTM model.(3)Based on the advantages of the CNN-LSTM model,the model was used for the remaining useful life prediction of bearings,and experiments were conducted using the PHM2012 bearing full-life dataset.The CNN-LSTM model is also compared with the prediction results using a single CNN network and LSTM network.The experimental results show that the CNN-LSTM model has the lowest root-mean-square error,which is 0.0426,and the predicted life curve is closer to the real life line,and the prediction is better.There are 47 figures,11 tables,and 65 references.