Electric Motor Fault Detection Under Nonstationary Operation Using Koopman Operator Spectral Analysis

Abstract

Early detection of faults in induction motors is fundamental for ensuring reliability, reducing maintenance costs, and preventing unexpected downtime in industrial systems. Conventional Motor Current Signature Analysis (MCSA) techniques typically rely on predefined spectral bands and stationarity assumptions, which limits their effectiveness under variable-speed operation and time-varying electromechanical loading. This work proposed a Koopman-theoretic framework for motor fault classification based on Hankel-embedded Higher-Order Dynamic Mode Decomposition (HODMD), enabling extraction of compact, physically interpretable spectral-dynamical descriptors directly from electrical measurements. The proposed feature extraction pipeline characterizes the motor dynamics through Koopman eigenvalues, modal energies, growth rates, temporal coherence, and entropy-based measures, forming a low-dimensional representation of the system’s nonlinear behaviour. The discriminative capability of these features is evaluated using a supervised learning framework with class balancing, nonlinear manifold projection, and hyperparameter optimization via stratified cross-validation. Experiments conducted on a laboratory dataset of multiple fault conditions and variable operating speeds demonstrate that the Koopman-based representation consistently outperforms raw time-domain signals and classical Fourier-based descriptors getting results as high as 96.8% cross-validation balanced accuracy and 80.7% balanced test accuracy. The results highlight the ability of Koopman spectral features to capture fault-related dynamical signatures while remaining robust to speed variations, indicating strong potential for practical condition monitoring and intelligent diagnostics of industrial electric drives.