Integrating causal representations with domain adaptation for fault diagnosis

In practical fault diagnosis, obtaining sufficient samples is often challenging. Transfer learning can help by using data from related domains, but significant distribution differences often exist due to different working conditions. To address this issue, cross-domain fault diagnosis (CDFD) has attracted increasing attention. However, most CDFD methods rely on statistical dependencies, which restricts their ability to uncover intrinsic mechanisms and affects both performance and reliability. In this paper, a Cross-domain Fault Diagnosis model based on Causal Representation learning (CFDCR) is proposed. This method employs causal representation learning with a graph autoencoder to learn invariant representations across domains, thereby improving the robustness of the prediction model. It further employs domain adversarial networks to align feature distributions, thus mitigating conditional distribution disparities between source domain data and target fault data, ultimately enhancing model performance. Experimental results on various bearing fault datasets demonstrate that the proposed cross-domain fault diagnosis model can effectively utilize related source domain data to guide fault classification tasks in the target domain and achieve more robust fault predictions.