Abstract:

In this paper, we investigate a simulation-to-reality domain adaptation approach for detecting motor faults, aiming to address the data scarcity problem in data-driven fault detection. A physics-based fault model is developed to generate simulation data under various fault conditions. A DQ transformation and feature extraction step is then performed for both simulation and real data, before domain adaptation is applied to align the simulation data with the limited real measurement data. Machine learning models can then be trained on the adapted data to make predictions. We demonstrate the effectiveness of the proposed method on eccentricity fault level prediction of an induction motor using stator current signal. Results demonstrate superior prediction performance compared to baseline model with only real data, with significant error reduction under both no-load and on-load conditions. This approach offers a promising and practical solution for motor fault detection in scenarios where obtaining comprehensive real fault data is challenging, as it leverages simulated data to enhance model performance with limited real-world measurements.