Bus crash patterns in the United States: a clustering approach based on self-organizing maps

Accident taxonomy is widely used by researchers and practitioners worldwide as a tool for understanding accident risks and designing effective policy measures to lessen these risks. Interestingly, despite the usefulness of accident taxonomy for recognising accident risks and the growing interest in improving bus safety operations, information regarding the taxonomy of bus accidents is limited. The current study analyses the risk-factors underlying bus accidents in the United States by uncovering prevailing typologies and analysing their severity. Data from the General Estimates System (GES) crash database were clustered by means of a two-stage clustering method consisting of self-organizing maps (SOM) followed by neural gas, Bayesian classification and unified distance matrix edge analysis. A multi layer perceptron (MLP) neural network was employed to confirm the correctness and usefulness of the SOM-based clustering process. Five clusters were identified: (i) multi-vehicle collisions at intersections: vehicle encroaching or travelling; (ii) multi-vehicle collisions with school bus at an intersection: distracted drivers; (iii) multi-vehicle collisions in road sections: infrastructure and traffic; (iv) single-vehicle bus accidents off-road: bus travelling and bus driver distraction at low speeds; (v) single-vehicle collisions with non-motorists: pedestrian and cyclists. The analysis pointed out conflicts among buses and other road users and indicated possible cluster-driven directions towards the enhancement of bus safety.