Prognostics of Insulated Gate Bipolar Transistors
Publication or External Link
Insulated gate bipolar transistors (IGBTs) are the devices of choice for medium and high power, low frequency applications. IGBTs have been reported to fail under excessive electrical and thermal stresses in variable speed drives and are considered as reliability problems in wind turbines, inverters in hybrid electric vehicles and railway traction motors. There is a need to develop methods to detect anomalous behavior and predict the remaining useful life (RUL) of IGBTs to prevent system downtime and costly failures.
In this study, a framework for prognostics of IGBTs was developed to provide early warnings of failure and predict the remaining useful life. The prognostic framework was implemented on non punch through (NPT) IGBTs. Power cycling of IGBTs was performed and the gate-emitter voltage, collector-emitter voltage, collector-emitter current and case temperature was monitored in-situ during aging. The on-state collector-emitter current (ICE(ON)) and collector-emitter voltage (VCE(ON)) were identified as precursors to IGBT failure. Electrical characterization and X-ray analysis was performed before and after aging to map degradation in the devices to observed trends in the precursor parameters.
A Mahalanobis distance based approach was used for anomaly detection. The initial ICE(ON) and VCE(ON) parameters were used to compute the healthy MD distance. This healthy MD distance was transformed and the mean and standard deviation of the transformed MD data was obtained. The μ+3σ upper bound obtained from the transformed healthy MD was then used as a threshold for anomaly detection. This approach was able to detect anomalous behavior in IGBTs before failure.
Upon anomaly detection, a particle filter approach was used for predicting the remaining useful life of the IGBTs. A system model was developed using the degradation trend of the VCE(ON) parameter. This model was obtained by a least squares regression of the IGBT degradation curve. The tracking and prediction performance of the model with the particle filter was demonstrated.