A novel interpretability paradigm based on semantic features of time–frequency images for trustworthy cross-machine fault diagnosis

Recently, domain generalization-based fault diagnosis (DGFD) has garnered significant attention for recognizing faults without the accessibility of the target domain. However, most DGBD methods concentrate on enhancing generalization through improved inductive bias or learning bias, while overlooking observational bias. This oversight leads to limited interpretability and subpar cross-machine diagnostic performance, which hinders the practical deployment of intelligent fault diagnosis in industry. In response to the above issues, the core idea of this paper is to reduce observational bias by defining cross-domain consistent semantic features for different health states in time–frequency images based on fault mechanisms, then training the model to rely exclusively on these defined semantic features for diagnosis, ensuring both interpretability and generalization. To realize this idea, this paper conducts a detailed analysis of various time–frequency analysis methods, evaluating their effectiveness in extracting cross-domain consistent semantic features, and proposes improvements to enhance these features, ensuring the model generalization. Furthermore, the model’s behavior is thoroughly visualized using the class activation map (CAM), confirming that the model relies solely on the defined semantic features as its decision basis, thereby ensuring interpretability. Finally, the model generalization and interpretability are tested using a single-source domain training and multi-target domain testing approach. Notably, although this study uses rolling bearing as an example, it is applicable to other fault diagnosis scenarios, such as for gears and motors. Source code: https://github.com/kangxin8/TF_based_CAM_beaeing_fault_detection.