Fault classification in electric motors using feature selection on vibration data from fiber Bragg grating accelerometers: a comparison of MLP and KAN networks

This work proposes a fault classification method for different fault types across distinct motor speeds using vibration data from two Fiber Bragg Grating (FBG) accelerometers. Features from the time, frequency, and time-frequency domains are evaluated using the overlapping coefficient between class histograms for each feature, combined with Principal Component Analysis (PCA). Supervised classifiers, including a Multilayer Perceptron (MLP) and the novel Kolmogorov-Arnold Network (KAN), are employed, with the KAN used to enhance interpretability and develop a symbolic model for the features. The results demonstrate that 99% accuracy was achievable for both MLP and KAN neural networks. The proposed methods offer direct feature selection, providing insight into the most relevant features and combining them to effectively reduce dimensionality. The results obtained with the KAN are promising; its symbolic model achieves 99% accuracy using only sets of well-known functions, while the network model reaches accuracy comparable to the MLP but with fewer parameters.