Research On Bearing Fault Diagnosis Based On Transfer Learnin

As a key component of rotating machinery and equipment,the running state of a rolling bearing is of critical importance to maintaining the smooth operation of rotating machinery and equipment.Monitoring the health of rolling bearings in real time is of immense practical importance to prevent the failure of rotating machinery and equipment,ensure the safety of the equipment and personnel,and reduce the economic loss of enterprises.As modern mechanical equipment becomes faster and more precise,rolling bearing health monitoring must be highly reliable.In recent years,several data-driven methods have been employed for rolling bearing fault diagnosis,but most of them are based on the assumption that sufficient labeled data can be trained and the data contains the complete health information of the equipment.They ignore concerns such as the absence of labels for training samples and the lack of historical training data for fault types,which makes them impractical for rolling bearing fault diagnosis in realworld engineering scenarios.Therefore,this study aims to develop a method for rolling bearing health monitoring under the conditions of unlabeled data and zero samples based on the transfer learning theory.Two methods are proposed:(1)Fault diagnosis method based on Joint Adversarial Domain Adaptation(JADA): This method,based on deep transfer learning,starts from the principle of unsupervised domain adaptation and combines it with the countermeasure domain adaptation mechanism to perform fault diagnosis of rolling bearing under the condition of unlabeled data.In order to identify the fault characteristics of rolling bearings,JADA first uses the source domain data to train the classifier,combined with cross-entropy loss and center loss for supervised learning.Then,the domain discriminator and feature extractor are trained using confrontation to minimize the difference between the characteristics of the source domain and the target domain and to match the conditional distribution and edge distribution between domains.Finally,the classifier is used to diagnose the sample features extracted from the target domain.Through the fault diagnosis experiments of ER-16 K and SKF6205 deep-groove ball bearings,the effectiveness of the JADA fault diagnosis method is verified.The fault diagnosis experiments in 24 transfer tasks show that,compared with other migration learning methods,JADA can effectively improve the fault diagnosis effect of rolling bearings without labeled data.(2)Fault diagnosis method based on Multi-Label Zero-Shot Learning(MLZSL): MLZSL takes zero-shot learning,a special type of transfer learning,as its theoretical source.Starting from the principle of embedding sample attribute space and combining it with a multi-label training mechanism,MLZSL completes rolling bearing fault diagnosis under the condition of zero samples.First,a lightweight feature extraction model is constructed by combining depth separable convolution,nonlinear activation function,and a residual learning mechanism.Then,a multi-label attribute learning network composed of multiple attribute learners to predict the attribute vector of each unseen class sample is constructed,and the embedding of unseen class fault samples in high-dimensional attribute space is obtained.Finally,the cosine distance between the attribute vector and the fault attribute tag is calculated to realize the fault diagnosis.Experimental results show that MLZSL is feasible and efficient.Compared with other zeroshot learning methods,MLZSL achieves better diagnosis,which provides a certain theoretical and engineering application value for the research of intelligent fault diagnosis of rolling bearings under the condition of zero samples.