A Novel Remaining Useful Life Prognostic Framework Combining Sample Convolutional Interaction Network and Fractal Brownian Motion

Abstract

Power MOSFETs play a crucial role in power electronic systems, and accurately predicting their remaining useful life (RUL) is fundamentally important for enhancing the reliability, safety, and maintenance planning of such systems. To this end, this article develops an innovative prognostic framework for predicting the RUL of MOSFET devices. First, a power cycle accelerated aging experimental platform under constant shell temperature fluctuation is constructed to obtain the performance degradation parameters of MOSFETs. Second, a sample convolutional interaction network (SCINet) is applied to historical data, learning long-term degradation trends via multistep prediction. Subsequently, a nonlinear fractional Brownian motion (FBM) degradation model is constructed incorporating measurement uncertainties. A state-parameter joint estimation method is then developed by combining a state-space model (SSM) with Kalman filtering, particle filtering and maximum likelihood estimation (MLE). The proposed framework fuses both SCINet predictions and historical observations for self-adaptive updating of states and parameters. A Monte Carlo (MC) simulation scheme, combined with a degradation state recursive strategy, derives the RUL and probability distribution function. Validation of real MOSFET degradation data and performance comparisons against multiple advanced methods demonstrate the efficacy and superiority of this novel prognostic framework. This research meaningfully contributes to more accurate reliability evaluation and improved maintenance planning for MOSFET devices.