Driving behavior in transit operations evolves through short-duration maneuver states shaped by roadway geometry, traffic conditions, and operational constraints. High-frequency telematics data provide detailed observations of vehicle motion, creating opportunities to characterize these patterns in a systematic and data-driven way. This paper develops a framework to identify and interpret electric bus driving maneuver regimes from vehicle trajectory data while simultaneously detecting anomalous behavior. Using 1 Hz telematics observations of speed, longitudinal acceleration, and directional change collected from electric buses in real-world operation, we construct window-based motion representations and learn compact latent features using a deep autoencoder. A Gaussian mixture model is then applied in the latent space to uncover distinct maneuver regimes, which are interpreted using one-vs-rest XGBoost classifiers with SHAP explanations and examined in their spatial context along observed trajectories. The framework identifies four maneuver regimes—acceleration, braking, high-angle maneuvering, and cruising—characterized by distinct patterns in speed dynamics, acceleration behavior, and directional variability, with clear spatial alignment to roadway features, such as intersections, turning points, and terminal areas. Anomalous observations, identified through reconstruction error, occur across all regimes but are more frequent in transitional states. The results show how telematics data can be translated into interpretable behavioral indicators that reflect both temporal dynamics and roadway context, providing a transferable approach to support operational monitoring, anomaly detection, driver training, and infrastructure assessment in transit systems.
Hydrogen is increasingly recognized as a component of sustainable energy systems as it offers a pathway to decarbonize various sectors. Its transport by rail can fill a gap between high-capacity pipeline-based transportation and low-capacity flexible road-based transportation but poses safety risks to the railway system and its surroundings. While storage solutions have reached a high level of technical maturity, the catastrophic consequences of a hydrogen tank explosion, i.e., the rare worst case, can include loss of life and infrastructure damage, underscoring the urgency of developing strategies to minimize exposure to these risks. The overall objective of this research is the development of an approach driven by open geodata that estimates the societal and monetary risk of compressed gaseous hydrogen transport by rail, specifically of rupture events of transported hydrogen tanks. It provides a geospatial workflow that estimates societal risk, as well as damages to railway infrastructure elements and the built environment in the hazard areas around potential tank rupture sites. Based on this, a penalty-based optimization approach is proposed to identify risk-minimizing hydrogen supply routes in the German rail network. Case studies demonstrate the effectiveness of the developed methodology to find safer routes for hydrogen transport and point out the role of tank design and routing decisions on risk mitigation. The methodology can be adapted to other dangerous goods and contributes to the ongoing development of resilient transportation systems.
Crash severity remains a critical issue despite rapid advances in vehicle automation. This study contributes to automated vehicle (AV) safety research by analyzing a Texas crash dataset that spans multiple automation levels, unlike most studies that focus on California. Using Association Rules Mining (ARM), it identifies key relationships among crash factors, such as speed, road type, collision type, lighting, weather, and driver characteristics. Crashes were grouped by automation level: assisted driving (Level 1), partial automation (Level 2), and conditional–full automation (Levels 3–5). In assisted driving, severe injuries were frequently linked to high speeds on narrow roads and older drivers. For partial automation, moderate speeds combined with poor traffic control or complex layouts heightened injury risk. Crashes involving Levels 3–5 were not observed in the KA category, suggesting a lower observed severity profile for higher-automation crashes in this dataset. An interactive tool was also developed to visualize crash locations and factors by severity and automation level. Policy recommendations include improving roadway design in high-risk areas, enhancing driver training, and advancing AV system safety. These strategies can reduce crash severity and support safer AV deployment in Texas and beyond.
Emergency response time is a vital indicator of public safety effectiveness in dense urban settings like New York City, where even small delays can significantly affect survival rates. To support data-driven decision-making and enhance emergency response strategies, this project developed a Traffic Digital Twin (TDT): a high-fidelity, simulation- and AI-enhanced framework that replicates Emergency Medical Vehicle (EMV) operations under real-world, transient conditions. Developed in collaboration with the New York City Fire Department (FDNY), the TDT integrated diverse data sources and employed advanced calibration methods to accurately model EMV behavior and surrounding driver responses as detailed in a separate paper. This paper focuses on a complementary AI model, EMVAID, based on a Traffic Simulation-Informed Explainable Boosting Machine (TSI-EBM), designed to interpret key factors influencing EMV speeds. As an illustration of TDT decision support, it is then applied to optimize ambulance Cross Street Locations (CSLs), achieving up to 14.5
This study develops a systematic analysis of Level Crossing configurations combining road and rail factors. It provides a decision-support framework for prioritizing safety equipment deployment at railway level crossings using explainable machine learning and robust clustering. Accident risk at level crossings emerges from the interaction of heterogeneous infrastructural, operational, and environmental variables, whose combined effects cannot be captured through single-factor analysis. Rather than focusing solely on prediction, the methodology combines dimensionality reduction, clustering algorithms, supervised classification, and SHAP-based explainability to identify homogeneous crossing typologies and classify them according to structural risk profiles. Separate analyses of road and rail characteristics enable clearer identification of their interdependencies and improve interpretability. Results show that the highest-risk clusters are characterized by the joint presence of high volumes of road and rail traffic, dense surrounding road environments, confirming the multifactorial nature of accident risk. The cluster-based approach mitigates the stochasticity of rare events and reveals robust comparative patterns that support comparative risk profiling of level crossing infrastructure. The study demonstrates how interpretable machine learning can be structured as a decision-aid framework for exploratory risk priorisation for infrastructure management.
