Interpretable bearing fault diagnosis based on wavelet scattering network, PCA dimensionality reduction and PLUKAN network

Aiming at the issues of poor interpretability, weak generalization under small-sample conditions, and high industrial deployment cost in conventional deep learning-based bearing fault diagnosis models, this paper presents an interpretable fault diagnosis framework by integrating Wavelet Scattering Network (WSN), Principal Component Analysis (PCA), and Piecewise Linear Unit-based Kolmogorov-Arnold Network (PLUKAN). The framework adopts a three-stage collaborative architecture: robust feature extraction, efficient dimensionality compression, and structured interpretable classification. Multi-scale time-frequency scattering features with translation invariance and noise robustness are extracted via WSN, followed by efficient dimensionality reduction of high-dimensional features through PCA, and fault classification with structured decision interpretation is implemented by PLUKAN. Comprehensive validations are carried out on three public datasets and a self-built 6001-VS variable-speed bearing dataset. The results show that the framework achieves over 99% average accuracy under an 80% training ratio, corresponding to an 8:1:1 split, and demonstrates strong resilience under extreme few-shot conditions, maintaining competitive accuracy between 87.7% and 100% across four datasets when only 10% of the samples, corresponding to a 1:4.5:4.5 split, are used for training. Compared with the traditional KAN, PLUKAN reduces both the training time and inference time to approximately one-half, satisfying the real-time computing requirements of industrial edge devices. Meanwhile, by virtue of the physical interpretability of WSN features and the explicit mapping structure of PLUKAN, the framework overcomes the "black-box" limitation of deep learning models, and clearly establishes the correspondence between feature variations and fault mechanisms. The proposed WSN-PCA-PLUKAN framework realizes the collaborative optimization of diagnostic accuracy, computational efficiency and interpretability, providing a new engineering-feasible solution for the predictive maintenance of bearings in industrial rotating machinery.