Railway accident prediction strategy based on ensemble learning.

Area of Science:

• Railway safety and transportation engineering.
• Machine learning applications in risk management.
• Data science for predictive modeling.

Background:

• Railway accident prediction is crucial for safety and risk management.
• Historical data analysis is key to developing effective prediction models.
• Existing models face challenges with missing and imbalanced data.

Purpose of the Study:

• To propose an ensemble learning strategy for accurate railway accident prediction.
• To address data challenges like missing values and imbalance.
• To identify critical risk factors contributing to railway accidents.

Main Methods:

• An improved K-nearest neighbors (KNN) algorithm for data imputation.
• An AdaBoost-Bagging ensemble method to handle imbalanced datasets.
• Feature importance analysis to rank accident causal factors.

Main Results:

• The AdaBoost-Bagging method demonstrated superior performance over ANN, XGBoost, GBDT, Stacking, and AdaBoost.
• Achieved high accuracy (0.879), precision (0.879), recall (0.883), and F1-score (0.881).
• Significantly reduced inference time by up to 23.38% and identified key risk factors for derailment and collision.

Conclusions:

• The proposed AdaBoost-Bagging method is effective for railway accident prediction.
• The approach successfully mines important features without prior knowledge of accident mechanisms.
• Findings provide actionable insights for railway accident prevention and management.