Physics-Guided Machine Learning for Joint Fault Diagnosis in Inverter-Fed Induction Motor Drives

Abstract

In intelligent electric vehicle (EV) traction drives, pulse-width modulation and field-oriented control (FOC) introduce broadband harmonics and non-stationary start-up transients. These dynamics obscure incipient inter-turn stator faults (ITSF), particularly when inverter switch failures and measurement interference co-occur. To meet the low-latency requirements of on-board vehicle control units (VCUs), this paper proposes a data-centric, physics-guided diagnosis pipeline. Relying exclusively on current measurements, the framework jointly detects and localises ITSF (1–3% severity), open-circuit IGBT faults, and additive white Gaussian noise (5–20 dB SNR). A compact 27-dimensional feature vector is constructed by fusing db4 discrete wavelet transform sub-band energies, residualised negative-sequence phasors, αβ hodograph descriptors, and harmonic energy ratios. This targeted extraction empowers an ensemble of bagged decision trees to perform lightweight classification suitable for embedded edge hardware. In a simulation-based evaluation across four scenarios (transient and steady state, with and without AWGN), the pipeline achieves cross-validated accuracies of 91.63–98.97% and hold-out test accuracies of 93.13–98.75%. Unlike traditional threshold detectors that conflate inverter or sensing artefacts with ITSF, this approach achieves subsystem-level localisation across motor, inverter, and sensing hardware, providing the reliable condition monitoring required for cyber–physical EV digital twins.