A novel fault detection and diagnosis approach for three-phase full-bridge inverter in permanent magnet synchronous motor based on optimal information gain threshold determination and supervised machine learning algorithms

• Random Forest demonstrates superior real-time performance with 99.17 % accuracy and fastest diagnosis time of 14.5 μs per sample, while KNN provides optimal balance with comparable accuracy, sensor efficiency and competitive diagnosis time of 25.1 μs. • Nested cross-validation prevents data leakage during feature selection and hyperparameter tuning, providing unbiased performance estimates for three-phase inverter fault detection. • Information Gain-based feature selection achieves 37.5 % sensor reduction (from 8 to 5 features) while maintaining >99 % classification accuracy. The establishment of numerous Fault Detection and Diagnosis (FDD) methods, particularly for the Three-phase Full-bridge Inverter (TFI), aims to improve the reliability of Motor Drive Systems (MDS). The growing need for power electronic systems to be reliable and readily available has received particular attention in various industrial applications. The use of Machine Learning (ML) has enabled the replacement of traditional fault diagnosis approaches with novel systems that can detect faults at an early stage by analyzing vast amounts of sensory data. Based on Data Acquisition (DA) and Information Gain (IG) feature selection with Supervised Machine Learning (SML), this paper proposes a novel FDD approach for the TFI in a three-phase Permanent Magnet Synchronous Motor (PMSM). The approach combines nested cross-validation to prevent data leakage, optimal IG threshold determination for sensor reduction, and temperature sensor integration as the highest discriminative indicator for short-circuit and overheating fault detection. The obtained results demonstrate that this approach can precisely detect and classify existing faults with remarkably high accuracy.