Hybrid Transfer Learning with Uncertainty Quantification for IGBT Remaining Useful Life Prediction in High-Speed Trains

As a core component of high-speed train traction drive systems, predicting the Remaining Useful Life (RUL) of Insulated Gate Bipolar Transistor (IGBT) modules is critical for ensuring operational safety. To address key challenges in practical scenarios, including insufficient multi-source data and significant dynamic uncertainties, this paper innovatively proposes a hybrid prediction framework integrating transfer learning and uncertainty quantification. The framework first decouples degradation features into low-frequency trend components and high-frequency fluctuation components via low-pass filtering. Subsequently, a dual-channel deep learning model is constructed: a Long Short-Term Memory (LSTM) network is employed to capture deterministic degradation trends, while a Bidirectional LSTM combined with Kernel Density Estimation (KDE-BiLSTM) is designed to quantify dynamic uncertainties. Transfer learning with fine-tuning is implemented to facilitate knowledge transfer from source domain full-lifecycle data to target-domain small-sample data. Finally, a probabilistic fusion strategy is adopted to generate RUL prediction intervals. Validation based on accelerated aging experiments demonstrates that the proposed method achieves higher prediction accuracy and enhanced engineering applicability in data-scarce scenarios.