Physics-Informed Lightweight Federated Learning with Explainability for Privacy-Preserving Predictive Maintenance of PMSM Drives: A Simulation Study

Physics-Informed Lightweight Federated Learning with Explainability for Privacy-Preserving Predictive Maintenance of PMSM Drives: A Simulation Study

zargham Heidari1 Nariman Heidari2

1) .Department of Electrical Engineering, National University of Skills, Bushehr, Iran Email:
2) Department of Computer Engineering Islamic Azad University, Bushehr Branch Bushehr, Iran

کد IOI مقاله : ABGG-ABKZHB
محل انتشار : یازدهمین کنفرانس بین المللی پژوهش در علوم و مهندسی و هشتمین کنگره بین المللی عمران، معماری و شهرسازی آسیا(11icrsie.ir)
چکیده :
    Abstract The proliferation of the Industrial Internet of Things (IIoT) and the increasing deployment of electric drive systems in critical infrastructures have heightened the demand for intelligent, reliable, and privacy-preserving Predictive Maintenance (PdM). However, centralized deep learning methods face significant challenges regarding generalization, data confidentiality, and their inherent black-box nature. In this paper, we propose a framework titled Physics-Aware Lightweight Federated Learning with Explainability for Privacy-Preserving Predictive Maintenance in PMSM Drives: A Simulation-Based Study. At the core of this framework, the d-q dynamic equations of the Permanent Magnet Synchronous Motor (PMSM) are integrated as regularizing constraints into the local loss function of the federated neural networks to enhance model accuracy and generalization under out-of-distribution operating conditions. To mitigate communication overhead and facilitate deployment on edge nodes, a hybrid mechanism combining weight pruning and knowledge distillation is designed. Furthermore, SHAP values are utilized to augment decision transparency, and a novel quantitative metric, termed the Physics-Fidelity Score (PFS), is introduced to evaluate the alignment of explanations with characteristic fault frequencies. The efficacy of the framework is assessed through comprehensive numerical simulations on a synthetic dataset generated from precise mathematical PMSM models under various mechanical, electrical, and magnetic faults. The results demonstrate that, compared to standard federated learning, the proposed method achieves a 2.9% improvement in F1-Score and a 69% increase in PFS, alongside a 47% reduction in communication overhead. Explainability analyses further confirm that the identified signatures align with the physical principles of the faults. This framework represents a step towards realizing transparent, secure, and reliable predictive maintenance within a simulated IIoT environment, with experimental validation on real hardware constituting the trajectory for future research.

کلمات کلیدی : predictive maintenance electric drive systems federated learning physics informed neural networks explainable artificial intelligence privacy preservation industrial internet of things model compression