A deep learning approach for imbalanced crash data in predicting highway-rail grade crossings accidents

Accurate accident prediction for highway-rail grade crossings (HRGCs) is critically important for assisting at-grade safety improvement decision making. Numerous machine-learning methods were developed focusing on predicting accidents and identifying contributing physical and operational characteristics. A more advanced deep learning-based model is explored as a more accurate means of predicting HRGC crashes compared to machine learning-based approaches. In particular, the prediction performance of the convolution neural network (CNN) model is compared to the most commonly used machine learning methods, such as decision tree (DT) and random forests (RF). A 19-year HRGCs data in North Dakota (ND) is used in this study. Training a machine learning model on an imbalanced data (e.g., unequal distribution of labeled data in accident and no-accident classes) introduce unique challenges for accurate prediction especially for minority class. In this paper, a resampling approach was used to address the imbalanced data issue. Various performance measurements are used to compare the models' prediction performance. The results indicate that resampling the imbalanced dataset significantly improves the recall rate. The results also show that the proposed deep learning-based approach which deepens the layer levels and adapts to the training dataset has better prediction performance compared to other machine learning-based methods.

Automated summary

The study explores the use of deep learning for accurately predicting accidents at highway-rail grade crossings, which outperforms traditional machine learning methods, especially in handling imbalanced data. By utilizing a resampling technique, the study addresses issues with imbalanced datasets and validates the prediction performance of the model using various performance evaluation metrics.