Data-driven predictive maintenance of induction motors using self-supervised and federated learning on noisy current and vibration signals

Résumé fourni par la source

Induction motors (IMs) sustain a vast share of industrial activity but are prone to bearing wear, rotor-bar breakage, eccentricity, and insulation defects that emerge under non-stationary, noisy conditions. While Motor Current Signature Analysis (MCSA) remains attractive for its non-intrusive sensing, its discriminative power collapses at low signal-to-noise ratios (SNR) and when labelled fault exemplars are scarce. We present a unified Self-Supervised + Federated Learning (SSL–FL) framework that (i) learns transferable, noise-tolerant embeddings from large unlabelled corpora of stator current and vibration signals, and (ii) enables privacy-preserving, cross-site training without sharing raw data. Using chronological splits, SNR stress tests (0–15 dB), fault-severity breakdowns (incipient/developing/severe), non-IID federated client simulations, and leave-one-site-out transfer across CWRU, Paderborn, IMS, and an industrial pump–IM testbed, the approach consistently outperforms strong deep baselines (CNN/LSTM/Transformer), achieving 94.2% overall accuracy (92.4% incipient), 0.92 F1, and 0.90 MCC. It delivers 11–17 percentage-point gains in low-SNR regimes and 83.5% cross-domain accuracy, while attaining ~ 91% of the centralized upper bound with ~ 58% aggregation bandwidth under secure aggregation and (ε = 2.0, δ = 10⁻ 5 ) differential privacy. Performance remains stable under Dirichlet non-IID distributions (α = 0.1), confirming practical robustness to heterogeneous multi-site data. By coupling unlabelled representation learning with confidentiality-aware collaboration and severity-aware evaluation, the method advances a practical path to scalable, noise-robust, and compliant condition monitoring for Industry 4.0 assets.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.