Research On Deep Transfer Learning And Intelligent Diagnosis Method For Rolling Bearing Based On Improved Domain Adversarial Networks

Rolling element bearing is one of the most common and easily damaged mechanical fundamental components in mechanical equipment.With the improvement of demand and the advance of industrial technology,its safety and reliability has attracted much attention in various fields.For example,in the aircraft,ship,train and large shield machine,due to their complex operating environments,their safe operations require to keep the bearings healthy and safe.As the key components of these equipment,if applying condition monitoring and fault diagnosis technology to the bearings,it not only obtains their health states in time,but also can effectively prevent sudden failure and shutdown and realize the prevention first instead of taking decisions after the accident occurred.With the development of artificial intelligence and industrial intelligence,industrial big data and its intelligent analysis provide strong guarantee for prognostics and health management of the whole equipment.On the basis of existing research results,intelligent diagnosis of bearings based on deep learning has been widely studied and obtains successful applications.However,such success greatly relies on complete fault data,which is difficult to meet in practical application to ensure continuous operation of equipment.Therefore,in this thesis,the transfer learning is introduced to make full use of game theory and deep neural networks,and the research focuses on the deep transfer learning method for various bearings under complex operation environment and aims to improves the accuracy of intelligent diagnosis for different bearings and working conditions,so that the problems of lacking labeled data and extracting fault knowledge can be solved.Main research contents and contributions in this thesis are summarized as follows:Firstly,a full convolution adversative transfer model is proposed to solve the problem of lack of labeled data.Based on the adversarial deep transfer method,the model structure and adversarial training strategy are improved.In the proposed model,the convolution pooling module is replaced by the fully connected to strengthen the feature extraction capability of the model and then reduce the risk of over-fitting.Furthermore,the improved training strategy solves the problem of slow convergence in the original network and fast achieve predetermined learning objective.The experimental data collected from the bearings under different working conditions are used to verify and compare the classification performance of the proposed model.The results show the proposed method has a faster learning speed and a higher classification accuracy for a new bearing under different rotating speeds and loads,which indicates the successful transfer between bearings when working under different conditions.Secondly,in order to solve the over-adaptation problem of target domain data,an adversarial transfer model without an independent classifier is proposed.In this model,the classifier in the classical structure is removed and its classification function is integrated into the network used for feature extraction.Meanwhile,the residual structure is embedded into the feature extractor to improve its recognition ability.Combining with the maximum mean difference algorithm,the marginal distribution adaption of the proposed model can be further enhanced.The proposed model is then verified by using experimental data sets collected from bearings by two damage generation methods(artificial damage and accelerated life test).The results show that the proposed model can extract the fault knowledge from one bearing and then dragonize the health state of another bearing with completely different damage modes.Therefore,using the proposed model,the knowledge transfer is achieved across damage sources and the corresponding classification accuracy is higher than other improved networks.Finally,the above two improved deep transfer learning models are applied to two real cases to investigate their feasibility and effectiveness for qualitative and quantitative analyses of bearing intelligent diagnosis.The first case is to apply the proposed models to wheel bearings in high-speed train.The knowledge is transferred from one small deep groove ball bearing with fault labels to another big taper roller bearing(used as wheel bearing)without labels;that is to say,the knowledge transfer between two different bearings is successfully completed.The results demonstrate the effectiveness of the proposed models and thus the problems of analyzing no labeled data and lacking fault samples in engineering applications.Moreover,considering the complexity in the operating environment,the proposed models are verified to keep good classification capability when artificially adding noise to the analyzed data.The second case is used for the performance degradation analysis of bearings.In this case,the fault knowledge is transferred from a small ball bearing to a double-row roller bearing and then used for the quantitative assessment of the latter health status.The experimental result and comparison demonstrate that the proposed model can accurately identify change points of the initial fault and fault evolution,which clearly reflects the dynamic degradation process of the bearing and provide the support for improving the prognosis accuracy.Through the data analysis and method comparison of various bearings,the experimental results indicate that two modified adversarial deep transfer models proposed in this thesis can not only identify the bearing health state when extracting the fault knowledge from one bearing or one working condition and applying to another bearing or condition,but also can be used for qualitative and quantitative diagnosis of bearings in real applications.Meanwhile,it realizes the fault knowledge transfer and intelligent diagnosis of different bearings,expands the application range of deep transfer learning,and provides a feasible method and valuable reference for intelligent detection of bearings.