Majdi Mansouri, Khadija Attouri, Abdelmalek Kouadri
This article introduces and evaluates adaptive polynomial Kolmogorov–Arnold network (AdaptPolyKAN) architectures for intelligent fault diagnosis, addressing limitations of classical KANs in dynamic and nonlinear industrial systems. While classical KANs offer interpretability, their fixed univariate mappings lack the flexibility needed for evolving operating conditions. The proposed AdaptPolyKAN uses adaptive polynomial expansions that adjust the degree of each basis according to local reconstruction errors, enabling real-time adaptation, improved accuracy, and efficient handling of complex nonstationary fault patterns. Three variants are examined-Standard PolyKAN, SplineKAN, and Online PolyKAN-alongside the proposed AdaptPolyKAN. Their performance is benchmarked against classical KAN, artificial neural networks, support vector machines, and random forest models in both real and simulated fault scenarios. The evaluation uses datasets from a cement rotary kiln at the Ain El Kebira plant, consisting of 768 normal samples, multiple simulated sensor faults, and one real fault. Monitored variables include temperatures, pressures, motor currents, and rotational speeds from 44 sensors recorded at 20-s intervals. The dataset contains noise, class imbalance, and limited duration, reflecting realistic industrial conditions. Comprehensive metrics—including accuracy, precision, recall, F1- score, false alarm rate (FAR), and missed detection rate (MDR)—demonstrate the superiority of the proposed approach. AdaptPolyKAN achieves 98.4% accuracy, balanced precision and recall (98.3% and 98.4%), and the lowest FAR (0.0060), while maintaining competitive MDR (0.178). Online PolyKAN adapts effectively to changing fault patterns, whereas classical KAN suffers from elevated false alarms. Overall, AdaptPolyKAN provides reliable detection in nonlinear and time-varying processes, offering a practical and interpretable solution for safety-critical industrial environments.