A New Index for Reliability Assessment of Power Semiconductor Devices: IGBTs

Abstract

Mechanical and thermal stresses in harsh environments makes reliability assessment of high power converters more crucial. Power semiconductor devices are the most susceptible components in power converters and any reliability assessment can be done using condition monitoring of these components with high accuracy and simplicity. In this paper, a novel reliability index will be defined for Predictive Maintenance (PM) applications using data-driven algorithms. The best performance precursors to monitor the conditions of the power semiconductor devices will be selected. Then, Replicator Neural Network (RNN), as a semisupervised machine learning algorithm, will be used to develop a normal behavior model of the power Insulated Gate Bipolar Transistor (IGBT). Finally, real-time monitored data is feed into the model to calculate the Reconstruction Error (RE) in real-time. In steady state and dynamics operating conditions, proposed reliability index will be calculated using two indexes named as Risk of Anomaly (RoA) and Anomaly Rate (AR). This reliability index does not need prior failure or repair data (frequency and duration) and can contain any uncertainty in different operating conditions of the converter.