A Semi-Supervised Method for Key Performance Indicator Prediction of Electrical Machines Under a Self-Training Framework

Machine learning metamodels have been frequently used for predicting key performance indicators (KPIs) of electrical machines because of their rapid and precise predicting capabilities. However, the conventional procedure of data collection using finite element analysis is time-consuming. To shorten data collection time and improve the prediction accuracy of metamodels, we propose a semi-supervised learning approach using self-training techniques. Initially, we utilize the labeled samples to train a geometry classifier and a KPI predictor. We employ the classifier to eliminate geometrically infeasible designs from unlabeled data. For the remaining feasible designs, the predictor is applied to predict their KPIs. Since not all predictions of the trained metamodels are accurate and reliable, we introduce a confidence score using Monte-Carlo dropout to obtain trustworthy predictions as pseudo-labels for unlabeled data. Finally, the generated pseudo-labels and original labeled data are jointly used for metamodel retraining. The results show that the original dataset can be iteratively expanded in a short period of time, and the performance of both metamodels is improved.