Identifying Typical Pre‐Crash Scenarios for Autonomous Vehicles: An Integrated Approach Using Sequence Alignment and Association Rules

ABSTRACT

With autonomous driving systems (ADS) moving from closed‐course testing to operation in mixed traffic, reported crash data provide an empirical basis for identifying recurring pre‐crash patterns. A remaining limitation is that structured fields and state‐free event chains may not fully represent each agent's pre‐crash state evolution and the surrounding environmental context. Using 1136 ADS‐engaged autonomous‐vehicle (AV) crash records from the National Highway Traffic Safety Administration (NHTSA) Standing General Order incident reports, this study develops a large language model (LLM)‐assisted framework that converts unstructured crash narratives into standardised state‐aware action sequences. In the main state‐aware clustering, pre‐crash state and interaction‐action tokens define the clustering input, whereas collision‐mechanism labels are reserved for post hoc interpretation and validation. Optimal matching with the OMlev cost setting, k‐medoids clustering, and semantic merging identify ten collision pattern groups. Compared with a state‐free event‐chain baseline, the state‐aware and state‐free partitions show only moderate agreement (Adjusted Rand Index = 0.443; Normalised Mutual Information = 0.628). The differences are concentrated in state‐dependent mechanisms, including rear‐end while stopped, rear‐end while decelerating, sideswipe, and low‐speed parking conflicts. Association rules then describe how the ten groups co‐occur with report‐based environmental variables. Intersections, parking lots, and dark conditions with street lighting are linked in the reported sample to stationary passive‐impact, low‐speed parking, and static‐object‐contact patterns, whereas weather, surface condition, and daylight mainly form shared operating backgrounds. The groups are finally translated into tiered, process‐orientated safety test scenario skeletons. Together, the sequence and rule analyses offer a sample‐bounded workflow from reported ADS crash narratives to scenario construction.