IGBT Remaining Useful Life Prediction Using Optimized ELM

Abstract

Abstract To achieve accurate remaining useful life (RUL) estimation of Insulated Gate Bipolar Transistors (IGBTs), this study presents an enhanced Extreme Learning Machine (ELM) model optimized using an improved Nutcracker Optimization Algorithm with mixed strategies (NOAM). Experimental data obtained from NASA’s accelerated aging tests of IGBTs are employed, where the peak collector–emitter voltage is identified as the key degradation indicator for failure characterization. Traditional NOAM suffers from premature convergence during the early optimization stage and insufficient precision in later iterations. To overcome these limitations, the proposed approach increases population diversity by initializing candidate solutions using a Bernoulli mapping sequence, thereby strengthening global exploration. In addition, a novel adaptive weight calculation scheme is introduced to improve local search performance, while Cauchy mutation is incorporated to further enhance solution refinement and convergence accuracy. The optimized NOAM is then applied to tune the parameters of the ELM model, resulting in improved prediction performance. Comparative studies with conventional ELM, PSO-ELM, and SSA-ELM models, along with ablation analyses of the proposed improvement mechanisms, demonstrate that the NOAM-ELM approach achieves superior accuracy in IGBT remaining life prediction. These results indicate that the proposed method provides an effective reference framework for RUL prediction of IGBT devices.