SHapley Estimated Explanation (SHEP): A Fast Post-Hoc Attribution Method for Interpreting Intelligent Fault Diagnosis

TL;DR

SHAP-based explanations in intelligent fault diagnosis suffer from exponential subset enumeration, especially after applying domain transforms. The authors propose Patch-Wise Attribution to reduce dimensionality and SHEP to approximate SHAP with linear complexity, enabling real-time interpretations. Through simulated data and two real datasets (CWRU bearing and Helical gearbox), they demonstrate that SHEP closely matches SHAP while delivering substantial speedups. An open-source implementation is released to establish a practical benchmark for post-hoc interpretability in IFD.

Abstract

Despite significant progress in intelligent fault diagnosis (IFD), the lack of interpretability remains a critical barrier to practical industrial applications, driving the growth of interpretability research in IFD. Post-hoc interpretability has gained popularity due to its ability to preserve network flexibility and scalability without modifying model structures. However, these methods often yield suboptimal time-domain explanations. Recently, combining domain transform with SHAP has improved interpretability by extending explanations to more informative domains. Nonetheless, the computational expense of SHAP, exacerbated by increased dimensions from domain transforms, remains a major challenge. To address this, we propose patch-wise attribution and SHapley Estimated Explanation (SHEP). Patch-wise attribution reduces feature dimensions at the cost of explanation granularity, while SHEP simplifies subset enumeration to approximate SHAP, reducing complexity from exponential to linear. Together, these methods significantly enhance SHAP's computational efficiency, providing feasibility for real-time interpretation in monitoring tasks. Extensive experiments confirm SHEP's efficiency, interpretability, and reliability in approximating SHAP. Additionally, with open-source code, SHEP has the potential to serve as a benchmark for post-hoc interpretability in IFD. The code is available on https://github.com/ChenQian0618/SHEP.

Figures (21)

• Figure 1: The illustration of SHAP attribution in the time and target domains. (a) Traditional time-domain SHAP produces attribution results as components with varying impulse timings, offering limited clarity. (b) By applying the domain transform, the input is preprocessed into the target domain, and the model integrates the inverse transform, enabling clear and interpretable attributions within the target domain.
• Figure 2: The process of patch transform for 1D and 2D samples. (a) 1D sample (i.e., Freq and Env domains). (b) 2D sample (i.e., TF and CS domains).
• Figure 3: The illustration of SHEP-Remove and SHEP-Add. (a) SHEP-Remove calculates the contribution by "removing" the feature $\tilde{\boldsymbol{x}}_i$ from the input sample $\tilde{\boldsymbol{x}}$. (b) SHEP-Add calculates the contribution by "adding" the feature $\tilde{\boldsymbol{x}}_i$ to the data distribution $X$.
• Figure 4: The flow chart of applying SHEP for interpretability analysis of IFD model.
• Figure 5: The domain representations of three classes in the simulation dataset. (a) Frequency domain. (b) Envelope domain. (c) Time-frequency domain. (d) Cyclic-spectral domain.