ORCA-VFD: Field-Validated Domain Alignment and Core-8 Feature Reduction for Sensor-Less Remaining Useful Life Estimation from Standard Variable Frequency Drive Parameters

Variable frequency drives (VFDs) control approximately 40% of industrial motors, yet machine learning prognostics validated in laboratory settings routinely degrade in field deployment, with reported false-positive rates exceeding 30%. This work presents ORCA-VFD, a physics-informed synthesis and feature-reduction pipeline for remaining useful life (RUL) estimation built on standard drive parameters. The method unifies synthetic physics simulation (40 million samples), fault-injection data (10,892 samples), and long-horizon field measurements from an operational ABB ACS880 drive (20,751 samples across 720 hours) into a constructed 100,000-hour degradation lifecycle spanning infant mortality, useful life, and wearout phases. It reduces 41 raw parameters to eight universal "Core-8" signals available on nearly all commercial VFDs, enabling application without additional sensors. Two results follow. First, a Random Forest regressor recovers the synthesized degradation trajectory with high fidelity (R² = 0.9991, RMSE = 53 hours), and its ensemble spread yields calibrated prediction intervals, establishing the internal consistency of the synthesis and the sufficiency of the Core-8 feature set. Second, a field-validated discriminant analysis confirms that the genuine field measurements carry degradation texture, in noise, drift, and operational variability, that is absent from the physics and fault domains, and the ORCA alignment, validated on those field measurements, reduces cross-domain distribution shift (maximum mean discrepancy 2.84 to 0.31). An economic layer illustrates how uncertainty-aware replacement timing can lower expected lifecycle cost relative to run-to-failure under stated cost assumptions. The monitored drive operated under light load and did not reach failure during the collection window; accordingly, ORCA-VFD is presented as a deployable proof of concept for sensor-less condition-based maintenance rather than a field-validated failure predictor, and prospective validation against observed failures is identified as the necessary next step.