Electric motor fault detection under nonstationary operation using koopman operator spectral analysis.

Abstract

Accurate motor fault detection and flow rate estimation in industrial centrifugal pumps is essential for diagnostics, performance monitoring, and process optimization, yet traditional frequency-domain techniques such as Motor Current Signature Analysis (MCSA) often rely on fixed spectral bands and assume stationarity, making them less effective under variable-speed operation, changing loading conditions, and multiphase flow environments. This work introduces a Koopman-theoretic framework for motor fault and flow-rate classification based on Hankel-embedded Higher-Order Dynamic Mode Decomposition (HODMD), enabling extraction of physically interpretable spectral-dynamical features from multichannel current, voltage, and vibration measurements. The proposed pipeline computes Koopman eigenvalues, modal energies, growth rates, temporal coherence, and spectral entropy to construct a compact descriptor of pump electromechanical behaviour, which is subsequently evaluated through supervised and unsupervised learning pipelines incorporating SMOTE balancing, nonlinear UMAP projection, and extensive hyperparameter optimization via TPESearchCV. Using 2 laboratory datasets, the first consisting of 106 signals describing 4 motor faulty operation states with different input frequencies, its results were as high as 96.8% cross validation accuracy and 80.7% balanced test accuracy. The second dataset consists of 1304 signals spanning multiple rotational speeds, heterogeneous fluid compositions, and 11 unbalanced flow-rate range classes, the system achieves up to 97.8% cross-validation balanced accuracy and up to 92.6% balanced test accuracy while revealing coherent manifold structure aligned with physical flow-rate regimes. Although overlapping operating states and speed-dependent spectral shifts introduce classification challenges, the results demonstrate that Koopman spectral feature captures the essential dynamic structure of the pump more effectively than traditional Fourier-based approaches, highlighting their potential for robust, reliable, and generalizable monitoring and diagnostics of industrial pumping systems.