Demagnetization Fault Diagnosis of a Permanent Magnet Synchronous Linear Motor Based on Noisy-KAN on an Edge Computing System

A noise-enhanced demagnetization fault diagnosis method for permanent magnet synchronous linear motors (PMSLMs) based on edge computing is proposed, aiming to address the limitations of the existing approaches, including poor real-time performance, low noise immunity, and high computational resource consumption. First, a simulation and experimental platform for PMSLM demagnetization faults is established. Subsequently, stray magnetic field signals in the external space are collected. Finally, a Noisy-Kolmogorov– Arnold networks (Noisy-KAN) is proposed to extract the features for fault diagnosis. The proposed method is evaluated through comparative experiments with other baseline models. Training and testing are conducted on the Jetson Orin NX, while inference is performed on the Raspberry Pi (RPi) 4B. The Noisy-KAN model requires only 1.13 MB of memory occupation and 4.51 MB storage space, making it easily deployable on an ARM Cortex-A platform. The inference time per sample is 0.412 ms and 2.9 ms on Jetson and RPi, respectively. The average accuracy of the proposed method is 98.24% and 90.11% under noise-free and different noise levels, which is 2.36% and 7.86% higher than those of the baseline methods. Results demonstrate that the Noisy-KAN exhibits strong noise immunity and reduced computational overhead, enabling its efficient deployment on edge devices for accurate fault detection.