摘要详情

ID / 提交时间

24 / 2025-05-06 20:28:41

标题

A rolling bearing fault diagnosis method based on multi-scale feature extraction and transfer learning

关键字

Fault diagnosis; One-dimensional convolutional neural networks; Bidirectional gated recurrent units; Attention mechanism; Transfer learning; Multiple Kernel Maximum Mean Discrepancy

主题及专题

4、System Monitoring, Fault Diagnosis and Prediction

状态

全文待审

作者

Siying Xie / China Aerospace Science & Industry Corp Defense Technology R&T Center

Liang Mei / China Aerospace Science & Industry Corp Defense Technology R&T Center

Shuo Li / China Aerospace Science & Industry Corp Defense Technology R&T Center

Caixia Zhang / Beijing Key Laboratory of Advanced Manufacturing Technology, School of Mechanical and Energy Engineering

Zhifeng Liu / Key Laboratory of CNC Equipment Reliability, Ministry of Education

Qiang Qin / China Aerospace Science & Industry Corp Defense Technology R&T Center

摘要

Intelligent diagnostic algorithms based on deep learning have opened up a new way for rolling bearing fault diagnosis and the effect is obvious, but it is undeniable that most of the research now focuses on the homologous data (i.e., the same working conditions, the same equipment). However, considering the normal working scenarios in actual workshops and factories, the difficulty of implementing intelligent diagnostic algorithms will increase, because in workshops and factories, working conditions and equipment change frequently, and it is difficult to realize the training of intelligent diagnostic algorithms with the same source of data. And in some cases, the collected data is difficult to be labeled resulting in insufficient labels, which requires the intelligent diagnostic algorithm to be able to diagnose faults under different working conditions and equipment. To address the above factors, this research designs a Two-Way One-Dimensional Convolutional Neural Network (2C-1DCNN) based on transfer learning (TL) technique with a Bidirectional Gated Recurrent Unit (BiGRU) and an Attention mechanism (Attention) combined fault diagnosis model (2C-1DCNN-BiGRU-Attention). The model can be directly input the vibration signals of the captured bearings and pre-trained using the source domain data, followed by using the Multiple Kernel Maximum Mean Discrepancy (MK-MMD) to measure the feature distribution distance between the layers of the network in the source and target domains, to determine the migrating and freezing layers of the model, and finally, using a small amount of labeled target domain data for re-training to complete the task of classifying the target domain fault data. labeled target domain data for re-training to complete the task of classifying the fault data in the target domain. After experimental verification, the algorithm has a high diagnostic accuracy and effectively solves the problem of increased difficulty in diagnosing faults in the target domain due to the different distribution of data in the target domain and the source domain and the fact that there is only a very small amount of labeled data due to the difficulty in labeling the data.