Incremental bearing fault diagnosis method under imbalanced sample conditions

The fault diagnosis of the rolling bearings is crucial for the maintenance of industrial equipment. The traditional bearing fault diagnosis methods based on deep learning focus on recognizing the initial training fault category and do not have the ability for incremental fault diagnosis. Although the replay-based continuous learning method has been used to solve this problem, there exists a sample imbalance problem between new and stored samples, which affects the accuracy of this method. To address the above issues, this paper proposes a new continuous learning method based on down-sampling and model-agnostic meta-learning (CL-DMAML). Firstly, the new class of fault samples is down-sampled according to the proposed information entropy criteria, ensuring that the sample size of the new fault data is the same as the sample size of various fault data in the storage space. This method reduces the loss of important information caused by down-sampling. Then, within the framework of model-agnostic meta-learning (MAML), the known bearing fault data are used to meta train the model, enabling the model can converge quickly in the case of small samples caused by down-sampling. Finally, the extracted samples are stored in memory space, combined with the previously stored data to form a training set, which is input into the meta trained model for classification training, so as to realize the incremental diagnosis process. The Paderborn University (PU) dataset and the Jiangnan University (JU) dataset are used to validate this paper proposed method. The results show that in three different cases of imbalanced samples, the accuracy is higher than other continuous learning algorithms.