Enhanced Estimation of IGBT Remaining Useful Life Using an Optimized LSTM Model

To prevent insulated gate bipolar transistor (IGBT) failures from affecting the reliability and safety of electronic power systems. This study presents an enhanced long short-term memory (LSTM) network model specifically designed for predicting the lifespan of IGBT. It analyzes the collector-emitter spike voltage from the IGBT aging dataset, selecting it as the primary aging feature, which is then extracted using the Empirical Mode Decomposition (EMD) method. The model also utilizes particle swarm optimization (PSO) to globally optimize both the network and LSTM prediction model hyperparameters. Through the integration of EMD feature extraction technology and the particle swarm optimization algorithm, the LSTM network prediction model is enhanced. The effectiveness and superiority of the model are validated using real IGBT degradation data. Relative to conventional approaches, this method demonstrates a reduction in mean absolute error (MAE) and root mean square error (RMSE) by about 85%, leading to a more accurate estimation of the remaining lifespan of IGBT.