A Grey-Box FDI Framework Combining Kolmogorov-Arnold Networks and Decision Trees for Interpretable Fault Diagnosis

Abstract

A novel grey-box framework  is presented for fault detection and isolation (FDI), this work combines interpretable Kolmogorov-Arnold Networks (KANs) with Classification and Regression Trees (CART). The proposed methodology addresses the interpretability limitations of traditional black-box approaches while maintaining competitive detection performance. KANs are trained using the Satin Bowerbird Optimization algorithm to learn explicit polynomial representations of system dynamics. Fault detection is achieved through adaptive thresholding based on the Median Absolute Deviation of residual signals, while isolation is performed using CART on statistical features extracted from residual windows. This framework is applied to the DAMADICS actuator benchmark, the framework achieves 93.3\% isolation accuracy while providing transparent decision rules that can be directly inspected by domain experts. The combination of interpretable modeling, metaheuristic optimization, and rule-based isolation offers a practical solution for safety-critical industrial applications where model transparency is essential for inspection.