Leveraging Meta-Learner-Based Stacking Framework for Stator Winding Interturn Fault Detection of LSPMSM

Interturn short-circuit faults (ISCF) account for more than 21\% of failures in line-start permanent magnet synchronous machines (LSPMSMs), and their early detection is essential for protecting drives that lack the current sensors typically used for winding fault diagnosis. This article presents an acoustic-only ISCF detection framework that combines three complementary base learners in a stacking ensemble: a learnable filter-bank LSTM (LFB-LSTM) that operates on the raw windowed waveform, a 1DCNN trained with an adaptive focal loss (AFL-1DCNN) driven by a data-selected subset of features, and a Random Forest over a 19-D engineered feature vector. Their class-posterior vectors are fused by a gradient boosting machine (GBM) acting as a meta-learner trained on three-fold out-of-fold posteriors. The framework is evaluated on a ten-class severity-by-load dataset from a 1 hp LSPMSM. Across five independent random seeds, the proposed framework attains 98.53 ± 0.14% test accuracy (95% confidence interval [98.35, 98.70]%) with statistically significant improvements over four acoustic baselines tested on the same data. On a recording-level validation protocol in which one full unseen recording per multirun class is routed through the trained ensemble, aggregate accuracy is 99.64%. A noise-robustness study over AWGN, pink, and Brownian noise at six SNR levels and an implementation-cost analysis complete the evaluation. The neural portion of the framework occupies under 2 MB in fp32 and end-to-end ensemble inference runs at ∼ 20μs /window on GPU.