Research On Fault Diagnosis Method Of Rolling Bearing Based On Time-Frequency Analysis And CNN

Rolling bearings are widely used in major industrial production areas.As one of the core components of mechanical equipment,their status has a major impact on the operation of the entire system.In modern industry,the number of bearings needed to be monitored is large.It takes time and manpower to diagnose by traditional signal analysis.In addition,the bearing fault signal is small sample data,and the fault type is variable,so it is difficult to obtain sufficient data to train a robust diagnosis model,and the generalization ability of existing intelligent bearing fault diagnosis algorithms is insufficient.Aiming at the above problems,a fault diagnosis model based on timefrequency analysis and convolution neural network is proposed for single-point damaged rolling bearings.On the basis of this model,domain adaptive optimization is accomplished by using transfer learning technology,and cross-domain intelligent fault diagnosis of rolling bearings is realized.Aiming at simultaneously diagnosing the fault location and severity of rolling bearings,the comparison of time-frequency analysis of fault signals are carried out,and a dual-input(images and energy)convolution neural network model based on fine-tuned Alexnet is proposed.This model uses fine-tuning technology to greatly shorten the training time,and the highest accuracy of 100% has been achieved on the CWRU bearing dataset.Aiming at the problem that the fault signal of rolling bearings changes and the diagnostic performance of rolling bearings decreases under the condition of variable load and variable bearing dataset,the transfer learning technology is introduced to optimize the model.By adding CORAL loss layer in the network to calculate the difference of data distribution between the training domain and the diagnostic domain,the domain adaptation is realized,and the cross-domain diagnostic ability of the model is improved.The validity of the proposed method is verified by experiments,and the influence of hyper-parameters on the recognition accuracy is further explored.In this paper,an adaptive optimization method based on joint distribution adaption is proposed to solve the problem that the class information of diagnosis domain is not utilized under unsupervised training,and the above model only adapts to the edge distribution of data without considering the intra-class characteristics.By introducing pseudo-label learning to predict the label in the diagnosis domain,the CORAL losses of Intra-class are calculated.Experiments on variable load and variable bearing dataset show that the method has higher accuracy than only adapting the edge distribution of data.In order to balance the classification loss and adaptive loss with hyperparameters when adding the adaptive layer,there is a problem of manual parameter adjustment.The adversarial training method is introduced to complete the automatic domain adaptation of the two domains.Experiments show that the model has good cross-domain diagnosability under variable load and variable bearing dataset.