Autonomous vehicles (AV) are starting to be deployed on public roads.  Identification of operational patterns within AV traffic accidents data under real world conditions can be used for future regulatory guidance, definition of safety operations and AV system improvements. This study analyses the records of NHTSA AV crash database. This repository contains an unlabeled mix of variables like precrash speed, weather, lighting, road surface and manufacturer. Using statistical methods does not allow to identify correlations in such mixed data. This study applies unsupervised machine learning to 892 NHTSA autonomous vehicle crash records (2019-2024). It was used K-Means clustering combined with Principal Component Analysis (PCA) to segment distinct kinematic and environmental AV collisions. Also,  Apriori Association Rule Mining was applied to extract if-then patterns that show correlations between specific vehicle manufacturers, environmental factors and crash data. The results of this analysis show that kinematic behavior is the main differentiator in AV collisions. The clustering model identified three distinct crash types. One involves a minority high-speed profile averaging around 49 mph. The other two dominant are occurring at very low speed profiles (around 3.7 and 4.5 mph), which suggests frequent problems in stop-and-go or intersection situations. The association rules mining show a tendency for specific manufacturers to experience collisions under distinct conditions. Cruise vehicles show a strong correlation (69.2% confidence, Lift 2.04) with accidents in dark but lighted environments. Waymo vehicles present strong associations with daytime fog. These patterns provide verifiable data that can be used for safety audits and AV improvement.