Research On Bearing Intelligent Fault Diagnosis Method Based On Self-Supervised Knowledge And Transfer Learning

In modern industry,mechanical equipment is gradually developing towards intelligence,precision,and complexity.As a key component in the mechanical system,detecting and diagnosing the health status of rolling bearings is of great significance to the intelligent operation and maintenance of the mechanical system.The intelligent diagnosis method based on deep learning has become a hot spot for research due to its excellent performance.In most studies,the data for training and testing the performance of its fault diagnosis models are in the same distribution,while in actual production,the complex and variable operating conditions can drastically reduce the diagnostic accuracy of the fault diagnosis models.To solve the problem of cross-conditions fault diagnosis,this thesis carries out research on intelligent fault diagnosis of bearings based on self-supervised learning and transfer learning.Firstly,this thesis proposes an improved method based on self-supervised learning combined with adversarial domain adaptation to solve the problem that it is extremely difficult to obtain sufficient labeled fault data due to the complex and variable actual working conditions,and there are large distribution differences between the signals in the target and source domains,which leads to the difficulty of feature extraction and classification and weak generalization of the model for cross-working condition domain adaptation recognition in the depth model.The method takes into account the existence of a large amount of unlabeled data in the target domain,creates an auxiliary task by using its signal characteristics,and motivates the model to learn the sample common features through a large amount of unlabeled data to establish the intrinsic connection between the source domain and the target domain fault categories;then transfers the sample knowledge from the source domain to the target domain through conditional adversarial domain adaptation,combines the auxiliary task to optimize the difference between the two domains,and achieves the intelligent accurate fault classification in the target domain.Secondly,because it is difficult to obtain the target domain labeling knowledge in advance in actual industrial production,and the diagnosis effect is greatly affected when the traditional migration domain adaptation model is faced with scenarios that do not match it,this thesis proposes an unsupervised general domain adaptation fault diagnosis method based on self-supervised knowledge.To achieve better feature extraction in the model framework,a multi-scale feature extraction deep residual convolutional neural network combined with attention mechanism is designed.And by constructing a nonadversarial domain discriminator,the transferability of the samples is weighted to select the samples suitable for transfer.Then the identification of unknown class samples in the target domain is achieved based on the transferability weights output from the nonadversarial domain discriminator to separate unknown class faults,and the feature alignment of common classes in the source and target domains is achieved based on the transferability weights under conditional adversarial training.To achieve general domain adaptive bearing intelligent fault diagnosis.In this thesis,the effectiveness of the proposed two methods is verified using the bearing dataset of Case Western Reserve University and the bearing dataset of Paderborn University,Germany.The influence of hyperparameter selection on the diagnostic effect of the fault diagnosis model is analyzed experimentally,appropriate parameters are selected,and experiments are conducted to compare the current commonly used methods with the method studied in this thesis,and the results show that the method studied in this thesis is superior in various cross-service fault diagnosis situations.