Evaluating the Spatiotemporal Distribution of Hazardous Driving States for Buses: An Investigation on GPS Data With a BiLSTM-CNN

Abstract

Enhancing bus safety is critical due to the severe societal and economic impacts of bus-related crashes. This study investigates hazardous driving behaviors in buses by analyzing high-frequency GPS data, focusing on kinematic parameters such as speed, acceleration, and steering angle. A hybrid convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM-CNN) model, is proposed to identify hazardous states, including overspeed, sharp acceleration/deceleration, and sharp turning. Comparative evaluations with standalone CNN, long short-term memory (LSTM), and BiLSTM models demonstrate the superior performance of the BiLSTM-CNN, achieving 93.1% recognition accuracy with rapid convergence. Using the identified hazardous state data, a statistical analysis of their spatiotemporal distribution is conducted by segmenting time periods and road sections. Hotspot road segments for dangerous operating states are identified through a spatial autocorrelation analysis. A spatial econometric model is employed to analyze the significant factors of these hazardous operating states. Spatiotemporal analysis reveals distinct risk patterns: overspeed clusters in peripheral road segments with low traffic density, while sharp acceleration and deceleration peak during off-peak hours. The study proposes targeted recommendations for early warning systems and safety management of hazardous bus operations. These recommendations aim to provide a reference for enterprises and traffic management departments in supervising bus operating conditions. This research establishes a data-driven framework for mitigating operational risks and enhancing public transportation safety.