Partial discharge diagnosis in electric motor with digital twin model-enhanced ensemble learning

Partial Discharge (PD), a localized dielectric breakdown of an electrical insulation system, is considered the main cause of aging, insulation deterioration, and failure in medium and high-voltage systems; directly affecting the reliability of the operating equipment. Therefore, it is imperative to determine a real-time PD monitoring system, that can consider relevant sensor data which is continuously collected and analyzed. In this study, hall-effect sensors are chosen to detect changes in the magnetic flux, this signal is processed to detect and localize PD within a permanent magnet synchronous motor (PMSM). Traditionally, methods for PD detection require invasive current sensors, or expensive optical sensors for accurate localization. In this study, we show that the use of hall effect sensors with advanced machine learning techniques can accurately detect and localize PD in a PMSM. The proposed framework combines time series modeling, machine learning for data generation, digital twin modeling, and ensemble learning for detection and localization of PD. This methodology is generalizable for discrete-time faults within various engineered systems. For our application, experimental data is very limited, so a finite element (FE) simulation model is implemented to create initial datasets. From this data, a time-series model is utilized to indicate the instance of PD. This model is able to detect PD within a certain range around the hall-effect sensor, and ultimately an optimal network of sensors can be determined. Next, features are determined based on physics and first principles, which are then inputted into the ensemble learning classifier for localization of PD within the PMSM. This technique is enhanced by a digital twin model, namely a simulation of the healthy motor running in parallel. The difference between the healthy magnetic flux and the PD magnetic flux enables the classifier to make accurate localization. Since this machine learning model requires an adequate amount of training data, a Gaussian Process Regression model is utilized as a surrogate model to generate the necessary data. PD diagnostics is a big challenge in recent literature, and our proposed method has shown effectiveness for not only detecting PD, but also localizing it.. The first stage stacking classifier, which localizes the PD to the winding section, has 100% accuracy and a Brier score of 0.02. The second stage bagging classifier, which localizes with more precision within the winding, has an accuracy of 83%, with a Brier score of 0.21. This is a satisfactory performance for PD diagnostics in PMSMs, especially with a noninvasive hall effect sensor. In addition, the proposed framework has the potential to be a generic monitoring system for discrete-time faults in various new engineering applications, where there may exist very little data and information for analysis.