Research on Aviation Safety Prediction Based on Variable Selection and LSTM.

Area of Science:

• Aviation Safety
• Statistical Modeling
• Machine Learning

Background:

• Aviation safety prediction is crucial for incident prevention but complicated by complex causal mechanisms and temporal dynamics.
• Existing methods face challenges in accurately predicting aviation safety levels due to these complexities, increasing operational costs.
• The Aviation Safety Reporting System (ASRS) provides valuable data for analyzing aviation security events.

Purpose of the Study:

• To develop an innovative statistical method for accurate aviation safety prediction.
• To enhance the efficiency and robustness of aviation safety prediction models.
• To identify key predictors for monthly aviation accidents.

Main Methods:

• Utilized a combination of adaptive sparse group LASSO (ADSGL) for variable selection and multistep stacked LSTM (MSSLSTM) for prediction.
• Introduced group variables and a weight matrix into LASSO for adaptive variable selection.
• Compiled and analyzed 138 monthly aviation insecure events from the ASRS, using minor accidents as predictors.

Main Results:

• The MSSLSTM model demonstrated a 41.98% reduction in root mean square error (RMSE) compared to the original model.
• The ADSGL method successfully identified key variables, reducing processing time by 42.67% (13 seconds).
• The proposed ADSGL and MSSLSTM approach showed excellent generalization ability and robustness.

Conclusions:

• The integrated ADSGL and MSSLSTM method offers a significant improvement in aviation safety prediction accuracy and efficiency.
• This approach enhances the capability for early warning and prevention of aviation incidents.
• The study highlights the effectiveness of adaptive variable selection combined with deep learning for complex safety prediction tasks.