Research On Rolling Bearing Fault Diagnosis Method Based On Deep Transfer Learning

As one of the key parts of mechanical equipment,rolling bearing is widely used in rail transit equipment,engineering machinery,precision machine tools,instruments and other engineering fields.According to relevant literature statistics,about 30% of rotating machinery faults are caused by bearing faults.Once the bearing fails,it will seriously affect the normal operation of the equipment,and may even cause safety accidents and economic losses.Therefore,it is of great significance for bearing fault diagnosis and improving the safety and reliability of the equipment.The bearing is often operated alternately under multiple working conditions,and the data under different working conditions means that the distribution of training and test data is different.The diagnosis accuracy of deep learning method is low or even invalid.Transfer learning is considered to be one of the effective methods to solve this kind of problem.At the same time,the data collected by the bearing is easily "polluted" by environmental noise,and the characteristic information reflecting the equipment state is submerged in the noise,so it is difficult to extract effective information.Although the collected data is large,most of the data are not marked.Therefore,based on deep learning and transfer learning,this paper studies the fault diagnosis methods of working condition change,unlabeled,noise and imbalance in rolling bearing fault diagnosis.The main research contents are as follows:(1)Aiming at the lack of self adaptability and decline of diagnosis accuracy of traditional intelligent diagnosis methods,a bearing fault diagnosis method based on hierarchical multi task learning deep convolution neural network model(HMCNN)is proposed.The proposed method realizes the final fault classification by sharing and learning multiple fault diagnosis tasks in the deep network model,using the information hidden in the training signal of related tasks,and weighting the classification results of multiple tasks.Compared with the current multi task diagnosis model,HMCNN model reduces error propagation,so that the proposed model has better scalability and generalization.The fault location,fault type and fault severity of bearing are diagnosed by HMCNN model,which can provide more detailed guidance for fault maintenance.Finally,the proposed model is verified on the rolling bearing data set.(2)Aiming at the low diagnosis accuracy and insufficient adaptability of current deep learning method when the bearing working conditions change,a bearing fault transfer diagnosis method based on deep MMD constraint domain adversarial network(DMWAN)is proposed.DMWAN model extracts domain invariant features shared by source domain and target domain through domain adversarial training.At the same time,the MMD constraint is added to DMWAN model to improve the aggregation degree of same kind of features,so as to expand the feature marginal distance of different classes in source domain and guide the feature alignment of target domain.Through experimental analysis,the proposed model can better feature extraction and distribution adaptation under different working conditions.DMWAN model achieves high diagnosis accuracy under different speeds,torques and loads.Compared with other fault diagnosis models,proposed model has better fault diagnosis performance.(3)Aiming at the problem that there are easily confused and difficult to diagnose samples in bearing fault diagnosis under different working conditions,resulting in the sharp decline of diagnosis accuracy.A transfer fault diagnosis method of full convolution conditional adversarial network(FCWAN)is proposed.The difference classifier used in FCWAN model makes the classification boundary clearer.It reduces the degree of feature confusion to reduce the misclassification from target domain to source domain,so as to improve the accuracy of bearing fault diagnosis.At the same time,in the domain adaptation module of the model,the conditional adversarial mechanism is adopted to enhance the effect of domain adaptation and further improve the accuracy of diagnosis.The proposed model is verified on the test-bed bearing and locomotive bearing data sets.(4)Aiming at the problem that the existing fault diagnosis methods have insufficient ability to extract the characteristics of vibration signals with noise under different working conditions.A transfer fault diagnosis method of multi-level discrete wavelet depth adversarial network(MDWAN)is proposed.In MDWAN model,multi-level discrete wavelet decomposition is integrated into the deep adversarial learning model to denoise and extract the features of noisy signals under different working conditions.The domain adversarial network is used to reduce the difference of different working conditions.At the same time,L1 distance constraint is used for the classifier of source domain and target domain in the classifier to further reduce the difference of working condition distribution and improve the bearing fault diagnosis ability of the model with noise.The effectiveness of the proposed method is verified on the test-bed bearing and locomotive bearing data sets.(5)Aiming at the problem of bearing fault diagnosis with unbalanced samples in source domain and target domain under different working conditions.A transfer fault diagnosis method of deep weighted adversarial network(DWWAN)under sample imbalance is proposed.In this method,the domain discrimination weighted network module is introduced into the model based on domain adversarial transfer learning to solve the problem of diagnosis failure caused by the inconsistency of fault sample types between the source domain and the target domain.The DWWAN model learns the domain identification weight through the domain identification module to optimize and reduce the unbalanced distribution difference between the source domain and target domain in training process.Then,the source domain and target domain are assigned a larger weight for shared features and a smaller weight for nonshared features,so as to reduce the phenomenon of "negative transfer" in the process of domain adaptation.The proposed model is verified in the bearing data set and locomotive bearing data set.Finally,after the above research and exploration,this study summarizes and prospects the rolling bearing fault diagnosis method based on deep transfer learning.