Search-Process Benchmarking of GA, PSO, DE, ABC, GWO and WOA for PMSM Inverter Fault Diagnosis in Electric Vehicle Drives

Metaheuristic-assisted permanent magnet synchronous motor (PMSM) inverter fault diagnosis is often compared using endpoint accuracy alone, while search dynamics, computational demand, and validation dependence remain under-reported. This study presents a systematic search-process benchmark—not a hybrid optimizer—for the Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Differential Evolution (DE), Artificial Bee Colony (ABC), Grey Wolf Optimizer (GWO), and the Whale Optimization Algorithm (WOA). The analysis uses 10,892 observations, 28 electrical–thermal candidate features, and nine operating classes. Each optimizer independently searches the same feature-gating and Gaussian Naive Bayes variance-smoothing space under five class-wise contiguous outer folds, three blocked inner folds, a population of 20, 100 iterations, and 30 runs per fold, yielding 900 searches and 90,000 trajectory records. A 10-sample exclusion gap is used to reduce short-range dependence; post hoc sensitivity checks over 0–100 samples show negligible change in a deterministic all-feature baseline. Relative to the non-optimized 28-feature GaussianNB baseline (macro-F1 of 0.9241), GWO produced the numerically highest mean macro-F1 of 0.9452, balanced accuracy of 0.9475, and MCC of 0.9435 while selecting 5.89 features on average. GWO reached 95% of its achieved fitness improvement earliest, at a median of 14 iterations, whereas ABC retained the greatest terminal population diversity, at 0.973. The fold-clustered Friedman test detected an overall difference (p = 0.0162), but none of the five Holm-adjusted pairwise comparisons against GWO was significant. The contribution is therefore a reproducible evaluation framework that links diagnostic performance to search dynamics, stability, calibration, and actual evaluation demand rather than a claim of pairwise optimizer superiority.