Fault Diagnosis Method For Rotating Machinery Based On Multi-source Domain Transfer Learning Framework

In order to meet the requirements of the development of modern industrial productivity,the rotating machinery system is becoming larger and more complex.As the core components of rotating machinery,rolling bearings work at high speed and heavy load for a long time,so they are prone to failure and cause economic losses,and even casualties.Fault diagnosis technology helps to avoid the occurrence of such accidents.Therefore,the research on fault diagnosis of key components of mechanical system has important social and economic value.Fault diagnosis methods based on deep learning need to ensure a sufficient amount of labeled data,and assume that the training data and test data follow the same distribution,which is difficult to hold in reality.Transfer learning-based fault diagnosis methods overcome this problem by transferring knowledge from the source domain to the target domain.However,these fault diagnosis methods mainly focus on single-source domain transfer learning,ignoring data from multiple source domains that are helpful for knowledge transfer in the target domain,thus failing to make full use of data resources.Funded by the National Natural Science Foundation of China and other projects,the traditional fault diagnosis methods based on transfer learning cannot make full use of data resources from multiple source domains to solve the problem of fault diagnosis tasks under variable working conditions.This paper starts from the core idea of transfer learning,combine deep learning with multi-source domain transfer learning,use deep learning to automatically extract fault features,and use multi-source domain transfer learning to fully reduce the distribution difference between source and target domain data in the feature space,so as to achieve multiple source domains to the target domain knowledge transfer to better complete the task of bearing fault diagnosis under variable working conditions.The specific research contents are as follows:First,in order to comprehensively utilize more complete fault data information from multiple source domains,a fault diagnosis based on intra-class multi-source domain adaptive network is proposed.The method first utilizes a feature learner to map data from multiple source and target domains into the same feature space.Next,a multi-source domain distance metric based on moment matching is used to evaluate the distribution differences between multiple source and target domain features,and is embedded in the loss function to reduce iteratively.In addition,during the training process of the model,an intra-class alignment training strategy is introduced to achieve simultaneous alignment of the edge and conditional distributions of multiple source and target domain data.Finally,a joint weighted classifier is used to comprehensively evaluate the samples to be tested to give the final fault prediction.Secondly,in order to realize the fault diagnosis of complex situations and working conditions in which multiple factors change simultaneously(see Figure 4-1(b)),a multisource domain teacher-student learning network based on optimal transmission is proposed.The network applies the basic theory of optimal transmission to the multi-source domain transfer learning method,and realizes the fault diagnosis of variable working conditions with good effect.The method first uses a feature learner for feature mapping,and constructs a multi-source fault diagnosis teacher using multiple source domain classifiers and a domain discriminator.Then,using the correlation method of optimal transmission,the student classifier of the target domain learns knowledge from the multi-source fault diagnosis teacher,and continues to optimize with the iteration,so as to obtain a good target domain classifier for fault diagnosis.In addition,this work also introduces a clustering optimization term in the loss function of the model to help the model get better classification performance.