Performance Degradation Assessment Based on Adversarial Learning for Bearing

This chapter focuses on the transfer learning based on generative adversarial network for performance degradation assessment of rolling bearing. It is crucial to monitor the health status of rolling bearings so as to ensure the safe and stable operation for mechanical equipment. After detecting and diagnosing faults, how to identify the extent of bearing failure and performance degradation becomes a key step in condition-based maintenance. However, the complex operating conditions in real industrial scenarios bring new challenges to the generalization performance of models. With the development of deep learning techniques, transfer learning based on generative adversarial network is gradually becoming an effective tool to improve the generalization performance of models. Actually, in the existing transfer learning methods based on generative adversarial network, usually the distribution of the label sets in source domain and target domain is the same, that is, source domain and target domain have the same number of categories. This is different from real scenarios in industrial practice where the set of labels in the target domain is a subset of the source domain. In other words, the knowledge of the classification is migrated from the large-scale source domain dataset to the small-scale target domain dataset of interest. To solve the class imbalance problem in traditional migration diagnosis, with the help of generative adversarial network, a novel method for assessing the bearing degradation state of crossing different service processes based on partial fault information migration is proposed. First, the time–frequency distribution of the vibration data at each moment of the bearing is used to characterize its degradation state. Next, a weighted partial transfer network model is constructed. Then, the unlabeled data in the source domain and target domain are fed into the weighted partial transfer model, respectively, to train and solve the network parameters. Finally, the XJTU-SY bearing dataset is used for analysis and validation, and the results show that the proposed method achieves good results in the experiments.