A fault diagnosis method with continuous frequency-band indicator embedding and class-aware interpretable feature weighting

Neural networks have been extensively applied in mechanical fault diagnosis due to their strong capabilities in feature extraction and classification. However, their limited interpretability and unknown credibility of decision hinder deployment in high-reliability scenarios. To address this issue, a frequency band multi-indicator feature embedding (FIE) method based on physical information embedding is proposed. A filter bank integrated into the convolutional layer is employed to simulate the spectrum, upon which multifrequency domain indicators are computed to extract multichannel physical features that vary continuously with frequency bands. A class-aware weight mask, enabling interclass distribution differentiation, is generated to assign distinct channel weights for different faults. This facilitates the extraction of key features for decision analysis and enables credibility evaluation based on a distance metric. Experimental results on multiple fault datasets demonstrate that the FIE module exhibits strong noise robustness and enhances diagnostic performance under variable-speed conditions. Furthermore, the class-aware mask supports reliable credibility assessment and feature interpretation, thereby supporting feature-level interpretation and decision credibility assessment while maintaining internal transparency.