Remaining Useful Life Prediction and Uncertainty Quantification with Multi-Branch Deep Networks and Probabilistic Calibration

Accurate remaining useful life (RUL) prediction is a key enabler of safe equipment operation and effective predictive maintenance. In practice, operating variability often induces distribution mismatch between training and deployment data, undermining the accuracy and transferability of existing models. We propose a multi-branch framework that hierarchically fuses CNN, BiLSTM, and Transformer components to learn robust degradation representations across diverse conditions. The model is further augmented with an uncertainty module that combines snapshot ensemble and Monte Carlo dropout to estimate epistemic and aleatoric uncertainties, and employs conformal prediction to construct prediction intervals with statistical coverage guarantees. Experiments on the C-MAPSS turbofan and A123 lithium-ion battery datasets show improved performance on challenging multi-condition subsets, achieving lower RMSE and Score than strong baselines while maintaining high interval coverage.