Abstract

Developing  reliable  motor  fault  diagnosis  systems  is challenging due to the high cost of collecting experimental data across a continuous spectrum of operating conditions. This data scarcity  leads  to  performance  degradation  when  diagnostic models  encounter  unseen  conditions.  To  overcome  this,  this paper proposes a physics-guided virtual metrology framework that generates high-fidelity motor current signals for previously unmeasured,  non-stationary  scenarios  by  leveraging  explicit physical causality. The core principle lies in axis-decoupled mapping, which models how rotational speed governs frequency- based temporal scaling, while torque dictates signal amplitude. By structuring baseline measured signals into an Operation- Fault Library, the framework enables the extrapolation of virtual current waveforms to unseen operating domains. Furthermore, a stationarization module retargets these dynamic profiles to constant conditions via point-by-point coordinate mapping to 

facilitate  robust  diagnosis.  The  method was  validated  using Interior  Permanent  Magnet  Synchronous  Motor  (IPMSM) datasets.  Results  demonstrate  that  the  framework  achieves exceptional signal fidelity, yielding a Mean Absolute Error as low as 0.071 and a Peak Signal-to-Noise Ratio of 39.83 dB, outperforming conventional generative baselines. Additionally, it ensures superior diagnostic performance with an accuracy of 99.4% under dynamic conditions. This framework establishes a new paradigm for virtual metrology that can effectively replace exhaustive physical experiments in industrial applications.

Keywords: Virtual metrology, Motor fault diagnosis, Non- stationary operating-condition, Extrapolation