IGBT Remaining Useful Life Prediction Based on Multi-Head Attention Mechanism and Adaptive Hybrid Pied Kingfisher Optimization Algorithm

Accurate remaining useful life (RUL) estimation of insulated gate bipolar transistors (IGBTs) is challenging because degradation records are nonlinear, noisy, and sensitive to model configuration. This study proposes AH-PKO-MHA-BiLSTM, in which a two-layer bidirectional long short-term memory (BiLSTM) network encodes each observed turn-off peak of the collector–emitter voltage (VCE) history window, four-head multi-head attention (MHA) reweights the resulting hidden-state sequence, and an adaptive hybrid pied kingfisher optimizer (AH-PKO), an enhanced variant of the pied kingfisher optimizer (PKO), searches the hidden size, learning rate, dropout, and L2 weight decay. The protocol uses a chronological 305/38/39 split, causal smoothing, and min–max scaling fitted on the training windows only, with validation-only model selection. The supervised target is an offline-defined normalized degradation index-based RUL proxy. Across ten paired final seeds, AH-PKO-MHA-BiLSTM achieved a root mean square error (RMSE) of 0.013509 ± 0.000356, a mean absolute error (MAE) of 0.010744 ± 0.000645, and a coefficient of determination (R2) of 0.8900 ± 0.0058. Its mean RMSE was 62.20% lower than that of a validation-selected VCE cumulative-trend baseline, 32.30% lower than that of unoptimized MHA-BiLSTM, and 14.36% lower than that of standard PKO-MHA-BiLSTM. All paired neural-model contrasts remained significant after the Holm correction. Under an equal budget of 600 fitness evaluations, AH-PKO also attained the lowest mean final validation RMSE and the shortest search time across five search seeds. These findings support improved RUL-proxy prediction within a single-device accelerated-aging trajectory; cross-device, variable-load, and online generalization require further validation.