Evaluation of Robust Covariance Estimation Techniques for Anomaly Detection of Insulated Gate Bipolar Transistors (IGBT)

An approach to detect anomalies in IGBTs is to monitor the collector-emitter current and voltage in application. These current and voltage parameters can then be reduced to a univariate distance measure called the Mahalanobis Distance (MD). The MD values with the use of an appropriate threshold enable anomaly detection of these devices. Mahalanobis distances (MD) are weighted Euclidean distances; the distance of each point from the center of the distribution is weighted by the inverse of the sample variance-covariance matrix. The presence of outliers in the monitored data can lead to the overestimation of the covariance matrix that in turn affects the anomaly detection results. This issue can be addressed by the use of robust covariance estimation techniques. In this study, the minimum volume ellipsoid (MVE) estimator, the minimum covariance determinant estimator (MCD) and the nearest neighbor variance estimator (NNVE) were used for anomaly detection of IGBTs. IGBTs were aged under a resistive load until failure. The monitored collector-emitter current and voltage values were used as input parameters for the MD calculation. The three robust covariance estimation techniques were used to compute the MD values and the anomaly detection times were compared to the results obtained by the classical covariance estimation technique.