Research And Application Of Small Sample Bearing Fault Diagnosis Based On Improved Relational Networ

Deep learning has received a lot of attention in recent years from academics due to its exceptional performance and widespread usage in fault diagnosis.However,such methods require a large amount of training data to ensure accuracy in practical applications.Most of the time,it is challenging to produce enough fault sample data from normally working machines in industrial scenarios such as rolling bearings,which limits the robustness and generalizability of deep learning-based approaches and lowers the accuracy of fault diagnosis.Therefore,the requirement for a high-accuracy fault diagnosis method that can be trained on small sample datasets is urgently needed to solve the problem of the deep learning application bottleneck in the fault diagnosis domain.Such a method must be able to learn enough feature information in a constrained number of data samples to raise the robustness and generalization capability of the model to increase the accuracy of fault diagnosis.To address the above issues,in this thesis,a relation network model based on an attention mechanism and category traversal module is proposed based on metric learning,and a bearing fault diagnosis system is designed and implemented.This thesis carries out the following work:The thesis proposes a relation network model based on an attention mechanism and category traversal module to address the issue of the limited number of fault samples and difficulties in extracting fault feature information in real-world industrial applications.Firstly,the relation network’s embedding and attention modules are integrated to help the neural network concentrate on significant information while ignoring irrelevant information.Secondly,the category traversal module is shown,allowing the network to utilize the support set’s global feature distribution data to further acquire more discriminative features and improve category representation capability.Finally,the relational network’s relationship module uses the feature vectors of the support and prediction samples to identify the prediction samples for fault classification based on the relationship values.The experimental results demonstrate that the method in this paper outperforms other small-sample methods in terms of fault diagnosis accuracy and generalization performance,and outperforms the pre-modified relational network and other conventional small-sample methods in terms of accuracy under small-sample conditions by about two percentage points.Based on the proposed scheme,the thesis designs and implements a bearing fault diagnosis system.It primarily consists of user management,data management,and problem diagnosis management modules.System testing demonstrates the system’s reliability and usability.