Enhancing Predictive Maintenance in Energy Systems Using a Hybrid Kolmogorov-Arnold Network (KAN) with Short-Time Fourier Transform (STFT) Framework for Rotating Machinery

This study proposes a hybrid predictive maintenance framework that integrates the Kolmogorov-Arnold Network (KAN) with Short-Time Fourier Transform (STFT) for intelligent fault diagnosis in industrial rotating machinery. The method is designed to address challenges posed by non-linear and non-stationary vibration signals under varying operational conditions. Experimental validation using the FALEX multispecimen test bench demonstrated a high classification accuracy of 97.5%, outperforming traditional models such as SVM, Random Forest, and XGBoost. The approach maintained robust performance across dynamic load scenarios and noisy environments, with precision and recall exceeding 95%. Key contributions include a hardware-accelerated KAN architecture, adaptive feature selection, and integration of explainable AI for interpretability. This framework enables real-time, transparent diagnostics in energy-critical, resource-constrained environments, supporting improved asset lifecycle management and reduced downtime. The study advances AI-based condition monitoring, bridging theoretical innovation with practical reliability in the context of sustainable industrial energy systems.