Adaptive Thresholding Using Gaussian Process Regression for Robust Inter-Turn Fault Detection in Six-Phase Induction Motor Drives

Reliable detection of stator inter-turn short-circuit (ITSC) faults is critical for the safe and efficient operation of six-phase induction motor drives. However, early ITSC signatures are weak and strongly affected by operating conditions, making fixed-threshold methods prone to missed detections or false alarms. This article proposes a confidence-aware adaptive thresholding scheme based on Gaussian process regression (GPR) trained using only healthy operating data for early fault detection. The fault indicator is defined as the magnitude of the current vector in the xy harmonic subspace. A GPR model captures the nonlinear dependence of this indicator on speed and torque while providing predictive uncertainty. During online monitoring, the residual between the measured and predicted indicators is compared with a variance-dependent dynamic threshold, enabling uncertainty-aware fault detection under varying operating conditions. The proposed method requires no faulty training data, additional sensors, or high-fidelity machine models. Experimental results demonstrate that the GPR model achieves lower prediction error than linear regression and artificial neural network (ANN) models, thereby improving the reliability of the proposed detector for low-severity ITSC faults. The method is validated in real time and achieves over 97% detection accuracy, demonstrating an effective trade-off between detection performance and computational complexity.