Improving Motor Reliability for Nuclear Systems by Sensorless Temperature Estimation

In nuclear power plants, numerous electrical devices operate continuously, making real-time monitoring essential for reliable operation. However, installing sensors on each device is expensive and can disrupt performance, particularly for electric motors widely used in pumps, valves, and actuators. Here, inference-based virtual sensing provides a viable alternative. This study intro-duces a data-driven deep learning approach to estimate critical parameters-in this case, the internal temperature of motors, which is crucial in preventing failures like insulation breakdown. Inferring these parameters non-intrusively requires a complex model that links real-time data to these critical variables. Conventional deep learning models often need retraining whenever initial conditions or boundary conditions change, limiting their real-time utility. A neural operator-based model is used to address this. An induction motor serves as the case study. Steady-state temper-ature distributions under various sets of boundary conditions are generated via ANSYS Maxwell and Icepak. These data are then employed to train a model that predicts temperature distribution based on boundary conditions in real time. Finally, the predicted results are compared for valida-tion.