Explainable bearing fault diagnosis using Shapley additive explanations: Cross-dataset insights from time, frequency, and chaotic domains

This work presents an explainability-driven study of bearing fault diagnosis using the Tabular Prior-Data Fitted Network (TabPFN), a transformer-based foundation model. The study is validated across two public vibration datasets: the Case Western Reserve University (CWRU) dataset and the Hanoi University of Science and Technology (HUST) dataset. A comprehensive set of time-, frequency-, and chaotic-domain features was extracted from the vibration signals, and Shapley Additive Explanations (SHAP) was applied to generate transparent explanations of the models’ predictions. The results show that each dataset produces a distinct feature landscape, yet two feature interactions consistently generalize across datasets: Impulse Factor combined with Root Mean Square (RMS) or Standard Deviation, and Skewness combined with Spectral Roll-off. These interactions capture fundamental properties of healthy and faulty behavior, while other dependencies remain dataset-specific but physically meaningful. A broader analysis reveals that the models evaluate three key aspects of the vibration signals: their overall power, the distribution of that power, and the degree of structure or randomness present. These findings highlight the value of Explainable AI for validating model reliability in real-world engineering settings.