Prognostics of IGBT modules based on the approach of particle filtering

Abstract

In the present work, a prognostic model combining model-based and data-driven techniques was developed and validated for dynamic life prediction of insulated gate bipolar transistor (IGBT) modules under power cycling conditions. The prognostic model integrates both anomaly detection based on semi-supervised machine learning and remaining useful life (RUL) estimation based on the particle filter (PF) approach. A range of healthy failure precursor data was predefined as labeled training data and machine learning techniques including principal component analysis (PCA) for feature extraction, and k-means clustering for anomaly detection were implemented. The clustering technique partitioned the predefined healthy data points into healthy clusters using a singular-value-weighted distance measure. The safety margin between a healthy distribution of distances between healthy data points within each cluster, and a test distribution of distances between a test data point and all the healthy data points within each cluster, was calculated to determine the affiliation of a test data point to the healthy cluster. A failure precursor process model incorporating the crack propagation physics law, the Paris Equation, and a measurement model was developed facilitating a sampling importance resampling (SIR) filter for RUL estimation. The developed prognostic model was validated on the degradation data from literature sources reporting IGBT power cycling test results to demonstrate its robustness.