RUL Estimation of Power Semiconductor Switch using Evolutionary Time series Prediction

Electric vehicle (EV) and hybrid EV (HEV) are popular for their low fuel cost per mile and near zero carbon emission. These vehicles utilize power semiconductor switches for high efficiency power conversion. These switches experience electrical, thermal, mechanical stresses during their operation and these stresses result in degradation and subsequently, wire-bond lift-off and solder fatigue. This degradation can be identified at an early stage by monitoring the tendency of fault precursor trajectory. Moreover, remaining useful life (RUL) is estimated from prediction and projection of this trajectory. Bayesian filters such as Kalman filter (KF), extended KF and generic particle filtering (GPF) methods have been recently used for trajectory tracing and projection. These methods suffer large variance in tendency projection when trajectory has both linear and non-linear tendencies and subject to harsh measurement noise. Moreover, these methods require large number of samples for probability density function (PDF) construction. In this paper, a hybrid Auto regression integrated Moving Average (ARIMA)-Neural Network (NN) model is utilized for tendency prediction and RUL estimation. The contribution of these two models is estimated and optimized using a nature inspired Covariance Matrix Adaptation (CMA) evolutionary technique. This hybrid algorithm combines the advantages of ARIMA and NN model to precisely trace and project fault precursor trajectory even under harsh noise. Simulation results verify its effectiveness under different noise level. The experimental validation of the proposed method is shown using RUL estimation of collector-emitter on-state voltage (V_CE,ON) of IGBT. The performance of this method is compared to ARIMA model, NN, and PF model.