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

Rolling bearings are widely used as core components in rotating machinery,such as railway vehicles,wind turbines,and aircraft engines.The operating environment of these bearings is more complex and harsh than other components,and the contact surfaces are prone to local defects.The generation and propagation of these defects pose a certain threat to the safe operation of the entire equipment.Once mechanical system failures occur,they can cause enormous economic losses and even major accidents.Therefore,real-time condition monitoring and intelligent diagnosis of rolling bearings are of great significance.In data-driven bearing fault diagnosis,a large amount of annotated data is required for model training.However,obtaining sufficient labeled data in actual industrial scenarios is not practical,and the data used for training and testing often have different distributions.To address these issues,deep transfer learning combines the advantages of deep learning and transfer learning and compensates for their respective shortcomings,making it a hot direction in intelligent diagnosis.Therefore,this thesis first studies the network structure used for feature extraction in deep transfer learning,and then uses the network structure as the model backbone to study the domain adaptation method in deep transfer learning,which provides a new idea for intelligent diagnosis of rolling bearings.Chapter 2 of the thesis introduces the problem definition of transfer learning and the structure and basic components of convolutional neural networks in deep learning.Chapter 3 of the thesis uses one-dimensional Le Net,Alex Net,WDCNN,and Res Net-18 as the diagnostic model backbone to classify faults on single operating condition data from the publicly available dataset of Case Western Reserve University.The impact of some parameters on the diagnostic performance was explored,and the superiority of WDCNN as a feature extractor was verified.By reducing the layers of WDCNN and further improving its network structure,the improved WDCNN not only ensures diagnostic accuracy but also increases training time,providing a high-performance feature extractor for subsequent deep transfer learning research.Chapter 4 of the thesis considers the measurement of joint distribution between data and proposes a Multi-layer Joint Distribution Adaptation Network.In the feature extraction stage,the improved WDCNN is used as a feature extractor,and in the domain adaptation stage,the Joint Maximum Mean Discrepancy containing both marginal and conditional distributions is used as a measurement function.Multi-layer Joint Maximum Mean Discrepancy is also applied in the fully connected layer to enhance model stability.Cross working condition transfer experiments are conducted on the publicly available dataset from the University of Paderborn and the Single Wheelset Rolling Vibration dataset,verifying the effectiveness and superiority of this method.Chapter 5 of the thesis considers the relative importance between marginal and conditional distributions and proposes a Balanced Deep Transfer Network.In the feature extraction stage,WDCNN is also used as a feature extractor,and in the domain adaptation stage,Maximum Mean Discrepancy is used to measure the marginal distribution between data,and soft pseudo-labels are used to calculate the conditional distribution.To balance the relative importance between the two distrubutions,balance factor is introduced.Cross measurement points transfer experiments are conducted on the Case Western Reserve University dataset and Single Wheelset Rolling Vibration dataset,verifying the effectiveness and superiority of this method.