Ranking an‎d selec‎tion of safety countermeasures at highway rail grade crossings using machine learning methods

9/16/2025 12:00:00 AM

Collisions at highway-rail grade crossings (HRGC) between roads an‎d railways are considered one of the major challenges in rail transportation worldwide. For this reason, railway companies, taking into account the infrastructural characteristics of each HRGC, implement effective safety countermeasures to reduce collisions at these HRGCs. This dissertation aims to use advanced methods to eva‎luate an‎d selec‎t the most effective safety countermeasures for reducing collisions at these HRGCs. In this research, data from the U.S. Federal Railroad Administration (FRA) covering the years 2010 to 2022 were used. First, influential factors affecting collisions were identified using the LightGBM machine learning algorithm. Next, HRGCs were clustered using the Deep Latent Class Analysis (DLCA) method, which groups HRGCs into homogeneous clusters. In this stage, the HRGCs were divided into four clusters. To calculate the Crash Modification Factor (CMF), a simple before-an‎d-after approach was initially used. Then, a more advanced approach based on the Empirical Bayes (EB) framework was applied. In this approach, a Negative Binomial (NB) regression model was used to estimate collision frequency, an‎d the residual estimates were modeled with a Long Short-Term Memory (LSTM) neural network. The NB-LSTM model was also compared with the NB an‎d NB-XGBoost methods within the EB framework. Due to its ability to model time-series data an‎d long-term dependencies, the NB-LSTM model performed better. To compare the proposed method (NB-LSTM model within the Empirical Bayes framework) for calculating CMF, the Full Bayes (FB) framework was also used, which showed higher stan‎dard error values than the proposed model. Therefore, the final CMF estimates in this research were based on the NB-LSTM model within the EB framework. Finally, to selec‎t the best countermeasure for each HRGC, a multi-output Bayesian Neural Network (BNN) was used. This model first predicts the cluster for each HRGC an‎d calculates the CMF for each countermeasure considered for that crossing. Then, the most effective safety countermeasure is recommended by selec‎ting the one with the lowest estimated CMF. The results of countermeasure selec‎tion showed that for HRGCs with a high number of daily trains, mostly straight tracks, an‎d an AADT (Annual Average Daily Traffic) above 4,000 vehicles, gates are usually chosen as the most effective countermeasure. In contrast, for HRGCs with an AADT below 2,000 vehicles, indicating low-traffic roads, flashing lights are generally among the most effective countermeasures.

تقاطع‌هاي همسطح جاده و ريل، انتخاب اقدامات متقابل، ضريب اصلاح تصادف، يادگيري ماشين

highway rail grade crossings, countermeasure selec‎tion, crash modification factor, machine learning , highway rail grade crossing

mohammadali zayandehroodi

barat mojaradi morteza bagheri