Beyond Closed-Set Accuracy: Constituent-Level Generalization in Induction Motor Fault Diagnosis
Year: 2026
Type: article
Abstract: Reported accuracies for data-driven motor fault diagnosis are frequently near ceiling, yet they rarely measure what industrial deployment requires, namely recognizing the elementary mechanisms of a compound fault whose exact combination was never seen in training. This paper evaluates that ability on a public multimodal dataset of a 2.2-kW induction motor with 24 diagnostic states under 12 speed and load conditions, including nine compound faults pairing a bearing defect with a second mechanical or electrical mechanism. Nine elementary mechanisms are detected independently from curated mechanism-specific vibration and current features, so that unseen combinations remain expressible. A crossed protocol separates three levels of compound-context exposure (recombination, one-sided, and two-sided zero-shot withholding) from operating-condition novelty, using run-level and severity-aware splits with calibration confined to training data. Although closed-set accuracy reaches 0.913, exact constituent-set recovery by the locked logistic-regression probe is only 0.083–0.157 with $F_{1}$ -calibrated thresholds and 0.157–0.287 with a post-hoc $F_{0.5}$ objective. The failures are strongly mechanism-specific. Dynamic eccentricity is almost perfectly discriminable in isolation, yet its compound recall falls to 0.017. Static eccentricity retains discriminative information but depends markedly on the classifier family, whereas the winding fault is already less specific in isolation. Cross-partner transfer is interpreted as context-dependent score behavior rather than as causal evidence of physical signature inversion. The context-by-condition interaction is directionally positive but unresolved for micro- $F_{1}$ at the composition level, and no interaction remains resolved after matching training-set size. Constituent-level evaluation thus exposes mechanism-specific generalization failures that closed-set accuracy conceals.
Source: IEEE Access
Authors: Evin Şahin Sadik, Hüseyin Tayyer Canseven
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7223667357