A Physiological Signal-Based Flight Risk Prediction Model Using Pre-training and Fine-tuning LSTM strategy

In aviation safety, accurately forecasting pilot flight risk remains challenging, especially when individual physiological differences are ignored. The paper proposes a pre-training-fine-tuning LSTM framework (PT-LSTM) that leverages both group-level and individual-level physiological time-series data. First, heart rate, blood pressure, body temperature, and other biosignal from a cohort of pilots pre-train the LSTM to capture common physiological patterns linked to risk. Next, the network is fine-tuned using historical data from each target pilot, adapting to their unique physiological signatures. Experiments on real-world datasets compare PT-LSTM against standard CNN, RNN, and LSTM models without pre-training. Results show PT-LSTM boosts prediction accuracy by 8.2% and F1 score by 9.4%, effectively mitigating data sparsity and personal variability. This dual-stage approach offers a novel framework for personalized flight risk assessment, advancing predictive safety tools in aviation.