Multi-bearing fault diagnosis method based on convolutional autoencoder causal decoupling domain generalization

Currently, most transfer learning (TL) methods for solving bearing fault diagnosis primarily focus on diagnosing a single bearing. However, in practical engineering scenarios, multi-bearing co-occurrence is often encountered, and joint diagnostic research under such conditions has received less attention. Additionally, it is challenging to collect sufficient fault samples for training, and due to changes in working conditions, the collected signals often follow different distributions. These factors limit the generalization ability of transfer learning models. Furthermore, most fault diagnosis methods for domain generalization primarily model the statistical correlation between time-series data and labels without effectively predicting the actual data, resulting in a relatively superficial description of the data generation process, which may introduce personal bias. To address these challenges, this paper proposes a causal decoupling network based on convolutional autoencoder (CDN-CAE) to diagnose and locate faults in multi-bearing systems using vibration signals from auxiliary bearings. First, maximum entropy optimization and mutual information neural estimation are introduced to decouple the time-series data into causal and non-causal factors, thereby reconstructing the data generation process comprehensively. On this basis, an aggregation loss is proposed to separate causal and non-causal factors, enabling better aggregation of the same fault type and improving the generalization ability of the method. Finally, a reconstruction loss is introduced to ensure the completeness of the information within the decoupled factors and to enhance the network's robustness. Experiments conducted on a multi-bearing device demonstrate that the proposed method achieves high accuracy and overall stability.