Adaptive variational sampling-embedded domain generalization network for fault diagnosis with intra-inter-domain class imbalance

Domain generalization (DG)-based methods have manifested promising potential in fault diagnosis when dealing with unseen target domain. However, these methods hold the class balance assumptions within and across domains, which contradict actual industrial scenarios, where the varying difficulty of collecting fault data from various classes under the same working condition leads to intra-domain class imbalance. Moreover, changing working conditions may alter the probability of occurrence of different faults, leading to inter-domain class imbalance. For the above concerns, we propose an adaptive variational sampling-embedded domain generalization network (AVS-DGNet) for fault diagnosis with intra-inter-domain class imbalance. In this method, we first apply domain adversarial strategy to map source data to a shared feature space. Sharpness-aware minimization (SAM) is leveraged to alleviate the over-fitting problem owing to the limited data scale of minority classes. Subsequently, a generator is constructed to estimate shared latent variable of each minority class regardless of source domains, and diverse feature generation is realized by sampling from them to mitigate class imbalance. Furthermore, considering different class-wise feature generation, we design adaptive reconstruction weights (ARWs) to control the generation distribution ranges. The experimental results on three datasets verify the effectiveness and superiority of the proposed AVS-DGNet.