Railway Accident Early Warning Using a Feature Selection Enhanced RF-MLP-RF Model

Railway accidents modeling and early warning are crucial objectives that demand a reliable model. While existing research predominantly focuses on accident numbers or types, the exploration of risk warning models for railway based on machine learning and accident causes remains relatively limited. Additionally, the problem of imbalanced accident data types further complicates predictive performance. To tackle these issues, our approach employs random forest (RF) to rank the importance of features, leading to the removal of relatively unimportant ones. Subsequently, multi-layer perceptron (MLP) is utilized as a feature extraction tool, enabling the generation of refined features. These extracted features are then used to retrain the RF. To validate the effectiveness of our proposed model, experiments using real-world data from railway accidents in Wuhan city, China, were conducted. The results demonstrated that our model achieves high accuracy and performs well in predicting accident risks, thereby offering valuable insights for early warning systems.