Imbalanced domain generalized fault diagnosis via semantic augmentation and causality

Abstract In order to address the issues of class imbalance and domain generalization in intelligent fault diagnosis, we provide a causal separation technique based on semantically coherent data augmentation. In the domain of manufacturing, the identification of faults in mechanical equipment is of paramount importance. However, conventional deep learning methodologies encounter difficulties in handling imbalanced data, resulting in an inadequate recognition of fault categories. A novel data augmentation strategy is proposed to ensure semantic consistency of the generated samples by synthesizing samples from source domains that are similar yet distinct. We develop a causal separation model with causal encoders and domain encoders to separate fault-related causal variables from domain-related non-causal elements. According to the experimental data, the suggested approach completes four cross-machine bearing problem diagnostic tasks with an average accuracy of 79.18%, thereby markedly enhancing the capacity to identify fault samples. The class imbalance issue in intelligent defect detection is effectively resolved by the research presented in this work, which has important application value.