Kolmogorov-Arnold Network for Machine Learning-based Motor Condition Diagnosis

Minimizing downtime in industrial production processes is paramount, and swift, accurate machine fault detection is crucial in achieving this goal. Recent advancements in artificial intelligence (AI) have facilitated significant improvements in fault detection through pattern recognition and predictive analytics. Traditional machine learning models, such as the Multi-Layer Perceptron (MLP), have been effective but have computational complexity and interpretability limitations due to their reliance on fixed activation functions. This paper introduces the Kolmogorov-Arnold Network (KAN) as a novel approach for machine learning-based motor condition diagnosis. Unlike conventional MLPs, KAN employs learnable activation functions parameterized by splines, eliminating the need for separate weight matrices and enhancing the model’s capacity to learn complex, non-linear relationships. Our methodology involved training the KAN model on a motor current signals dataset collected from five 2-HP induction motors, each tested under five different load conditions, comprising 3,750 signals. Experimental results demonstrate that KAN outperforms traditional MLP models, achieving a validation accuracy of 88.27% compared to MLP’s 84.10%, even without hyperparameter tuning. Furthermore, KAN provides greater interpretability by converting learned splines into symbolic functions, offering a transparent understanding of the model’s decision-making process. Despite its longer training time, KAN’s enhanced accuracy and interpretability present a compelling case for its application in motor fault detection and predictive maintenance. This study underscores the potential of the KAN architecture in advancing AI-driven diagnostics in industrial settings.