Introduction to Predictive Models for Motor Dielectric Aging

Abstract

Standardized testing methods are generally utilized to assess in service dielectric material aging and over time use of this information allows one to become knowledgeable as to the condition of the motor, and ultimately, when to repair the machine prior to failure. Traditional accelerated aging experiments are performed to evaluate base dielectric materials for properties such as thermal class and the ability to withstand voltage over time. Physical models for dielectric aging have also been developed. Emulators (aka predictive models) are statistical models trained using advanced analytics and machine learning algorithms to capture the input/output relationships of an underlying system or data set. Once trained, the emulator can be used in lieu of the process that generated the training data to rapidly predict outputs for arbitrary input combinations. Emulators can be created of both deterministic physics-based simulation codes and physical or experimental processes. Using a simple physical motor model, the process of building an emulator will be illustrated. This process begins with a design of experiments to select input combinations for the experimental collection of training data from the physical model. The emulator's predictive accuracy can be iteratively improved through an adaptive design process that combines knowledge of the previously conducted experiments with the emulator's ability to assess areas of greatest uncertainty in its predictions. The final validated emulator can be used for sensitivity analyses of inputs on the output(s) of interest, uncertainty propagation, and optimization. Applications of emulators include virtual sensors, predictive maintenance, calibration of physics-based simulation models, and digital twins. Avenues of application for an emulator of a motor model include predictions of motor life and dielectric failure probabilities based on dielectric and insulation material properties.