AGMFormer: An Explainable Intelligent Diagnosis Framework Based on Adaptive Gaussian Mixture Constrained Feature Aggregation Network

Significant advancements have been made in the intelligent fault diagnosis of industrial equipment, due to the transformative impact of deep learning technology. However, the limited explainability of many existing methods presents substantial challenges to their practical implementation in industrial settings. An end-to-end explainable transformer variant framework, AGMFormer, has been introduced to address this issue. First, AGMFormer incorporates a learnable adaptive Gaussian mixture model (AGMM) to constrain the dot-product attention mechanism. This innovation effectively extracts local and sparse features from vibration signals while mitigating sensitivity to irrelevant information, thereby establishing a more transparent and explainable feature extraction process. Second, a novel feature aggregation mechanism is developed to further consolidate and emphasize key features across multiple attention subspaces. This enhancement improves the robustness of the model under varying operational conditions, ensuring consistent performance in diverse industrial environments. Finally, the explainability of AGMFormer is comprehensively evaluated from both intrinsic and post hoc perspectives. Extensive experiments conducted on two public test bench datasets and a proprietary wind turbine gearbox (WTG) dataset demonstrate that AGMFormer outperforms competing methods in fault diagnosis accuracy and provides superior explainability.