Domain-adaptive Anomaly Detection and Severity Prediction of Electric Machine Drives at the Point of Common Coupling
Year: 2025
Type: preprint
Abstract: This paper introduces an domain-adaptive anomaly detection framework for electric machine drives that not only identifies hard-to-detect waveform irregularities in the initial stages, but also predicts their severity in various operating conditions. Using calculated features, the method extracts informative components in the frequency domain, allowing the identification of subtle signal anomalies. Random Convolutional Kernel Transform (ROCKET)-based classifier is used which excels at handling diverse data structures. Using randomized convolutional kernels, our model achieves high accuracy with minimal training and low inference latency. The proposed scheme can achieve precision over 98% in anomaly detection and 95% in severity level classification. Compared to traditional approaches, our method is independent of the motor model and its parameters, enabling generalization to new operating conditions and ensuring consistent protection in industrial scenarios. Through extensive simulations of various types of anomalies, we demonstrate that our approach reliably detects early stage faults, mitigates damage, and preserves the operational integrity of electric machines.
Authors: Abolfazl Najar, Shushan Wu, He Yang, Tao Wang, Jin Ye, Wenxuan Zhong, Ping Ma, Wenzhan Song
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4415372012