Machine learning-based direct estimation of remaining useful life (RUL) of IGBT using multi-precursor prognostics

The reliability of motor drives plays a crucial role in all motion control applications. Due to circuit overstresses, the IGBTs used in motor drives are prone to degradation and aging, which may result in motor drives failure. The RUL prediction of IGBT gives early warning of the unexpected IGBT failures, by predicting the useful life of IGBT while it is in operating condition. The existing methods for RUL prediction of IGBT focus on the use of a single precursor which cannot effectively determine the level of degradation of IGBT, and the methods are too laborious and have prediction errors. This work employs three precursors V_CE(ON), I_C and V_GE, and uses a simple yet effective method to predict the RUL of IGBT by machine learning techniques, and the results are tested using online-accelerated IGBT aging data from the NASA's Ames Research Centre. The data preprocessing and feature selection are carried out using MATLAB toolboxes, namely Data Cleaner and Diagnostic Feature Designer. Then, the MATLAB Classification Learner toolbox is used to train the various machine learning classifiers, and a comparison is made for various cases involving single, double and multi-precursor parameters in predicting RUL of IGBT. Hyperparameter tuning of the machine learning classifiers is also done, and evaluation metrics such as Precision, Recall, F1 score and ROC-AUC are used to validate the performance of the different ML classifiers. Highest classification accuracy of 99.6% is obtained with ensemble, neural networks and KNN model classifiers after hyperparameters were optimized.