A Shapley Value-Based Method for Formulating Physical Mechanism Semantics of Signal Sequences in Interpretable Fault Diagnosis

Abstract

Despite significant advancements in deep learning (DL) for fault diagnosis, the black-box nature of DL models hinders their reliable deployment in industrial applications. Interpretability methods have emerged to address this opacity, yet their effectiveness remains limited due to the lack of unified semantic guidance. This semantic gap not only constrains their practical application but also creates a disconnect between post hoc model explanations and ante hoc model guidance. In addition, the absence of quantitative metrics makes it challenging to evaluate the trustworthiness of interpretability methods. To address these challenges, this article proposes a signal semantic evaluation strategy (SSES), establishing a unified semantic framework. SSES employs Shapley values to evaluate significant signal components in the frequency domain. By integrating physical mechanisms, SSES enhances evaluation accuracy and formulates interpretable fault semantics. Furthermore, adversarial training and model ensemble strategies are employed to enhance the evaluation stability. To assess the reliability of interpretability methods, we introduce two metrics that quantify the consistency between constructed semantics and actual semantics. Experiments on two public datasets demonstrate that SSES accurately identifies significant signal components, while the proposed metrics effectively quantify interpretation reliability. Experiments on the XJTU-SY and Case Western Reserve University (CWRU) datasets demonstrate that SSES accurately identifies significant signal components and achieves the highest diagnostic accuracy under noise interference, reaching 86.7% and 99.1% at 0 dB noise level, respectively. In addition, the proposed reliability metrics effectively quantify interpretation reliability, showing that models with higher reliability scores exhibit superior robustness to noise.