Kolmogorov-Arnold networks for industrial machine health monitoring with continual learning capability using time-series vibration data

Abstract

Effective and reliable health monitoring is critical for ensuring the operational safety, efficiency, and longevity of industrial machinery. Multi-layer perceptron (MLP)-based neural networks, referred to as “Universal Approximators,” often face challenges in handling complex, non-linear data due to high computational costs, limited continual learning, and reliance on fixed, non-learnable activation functions. To address these challenges, this article evaluates the potential of Kolmogorov-Arnold networks (KANs) to enhance efficiency and generalization in machinery anomaly detection using a triaxial vibration data across four different health states. Unlike MLPs, KAN employs spline-based, trainable activations on edges, enhancing flexibility and learning efficiency. This architecture improves classification accuracy, outperforming MLP-based neural networks, and conventional autoencoders by 5.2%, 10.64%, and 6.54% on X, Y, and Z-axis data, respectively. Beyond accuracy, KAN demonstrates faster training, superior generalization, improved continual learning, and resilience to catastrophic forgetting. This study establishes KAN as a promising framework for next-generation intelligent machine health monitoring systems for real-world engineering applications such as predictive maintenance, condition-based monitoring, and autonomous fault diagnosis in rotating equipment, spindles in machining operation, wind turbine gearboxes, and industrial fans used in steel and power plants.