A Novel IGBT Health Evaluation Method Based on Multi-Label Classification

Abstract

The IGBT health evaluation of power semiconductor devices is usually based on the threshold evaluation method, which is usually a single characteristic parameter evaluation system. This kind of evaluation method cannot reflect the internal correlation of the change of multiple characteristic parameters in the deep level. Multi-label classification plays an important role in machine learning and can truly reflect the internal correlation principle of multi-feature parameters. Many studies have proved that multi-label classification (mlc) can effectively increase the actual classification effect of the clustering algorithm. In this paper, a clustering algorithm based on multi-label learning is applied to the health evaluation of IGBT. There are many characteristic parameters that affect each other in the actual work of IGBT, so it is difficult for a single label to reflect its actual health status. At the same time, multi-label data often belong to multiple classifications. Multi-label learning can improve the feature dependence ability of clustering method and improve the accuracy of classification. In this paper, we propose a multi-label classification learning model based on ISODATA for the multi-feature parameters of power semiconductor device IGBT, which can comprehensively consider the multi-level correlation effect of internal parameters in the multi-feature parameter extraction. The experiment results show that the algorithm model can better adapt to the IGBT health classification evaluation compared with the general clustering algorithm.