Voltage-Only Modal Diagnostics of Internal Short-Circuit Faults in Inverter-Cable-Machine Systems Using Transient Response Analysis

Abstract

Wide-bandgap (WBG) inverter drives impose repetitive high-frequency voltage stress, accelerating insulation degradation and increasing the probability of short-circuit faults in cables and machine windings. However, existing online diagnostic approaches face significant deployment barriers due to their sensitivity to electromagnetic interference (EMI), the requirement for ultra-high-bandwidth current measurements, or the limited interpretability of purely data-driven models. This paper proposes a non-invasive diagnostic methodology formulated in the modal domain and developed within a high-fidelity inverter-cable-machine multi-physics environment. An experimentally validated multi-conductor cable model, including shield-to-core coupling, together with a high-resolution permanent-magnet synchronous machine (PMSM) winding model, is used to synthesize diverse cable faults and internal short-circuit (ISC) faults with varying spatial locations and severities. A transient voltage stimulus is injected at the inverter terminals, and the terminal voltages are analyzed during the subsequent zero-input decay to isolate the intrinsic eigen-dynamics of the system. Continuous-time modal parameters are estimated using the Total Least Squares Estimation of Signal Parameters via Rotational Invariance Techniques (TLS-ESPRIT) algorithm, from which spectral modal persistence metrics and cross-channel coupling coefficients are derived. The robustness of the proposed framework is evaluated through Monte Carlo simulations incorporating stochastic sensor noise and electromagnetic interference. The resulting feature set forms well-separated clusters in a principal component analysis (PCA) feature space and enables reliable multi-class fault classification using a support vector machine (SVM) with a radial basis function (RBF) kernel within an error-correcting output code (ECOC) framework. Cross-validation results demonstrate a classification accuracy of 89.12 percent. The proposed method provides a promising pathway toward real-time structural health monitoring through diagnostic signal injection integrated into pulse-width modulation (PWM) operation.