Enhancing Autonomous Vehicle Safety Through Chaid Modeling: Influential Factors, Seasonal Variations, and Systematic Limitations

The growing prominence of artificial intelligence has driven transformative innovations across sectors, with autonomous vehicles representing a salient manifestation of this technological shift. The reliability of autonomous vehicles plays a crucial role in determining their societal acceptance and large-scale deployment. Within this context, disengagement data serve as an objective indicator of system reliability. A rigorous analysis of disengagement data is essential for evaluating the real-world performance and operational reliability of autonomous vehicles. Such data circumstances necessitate human intervention, thereby revealing system vulnerabilities and opportunities for improvement. Consequently, precise and transparent disengagement analyses are vital for advancing AV technology and strengthening safety. This study investigates the determinants of disengagements and contrasts human-initiated with system-initiated events. Drawing on 17,406 reports (2021–2023), CHAID models identified key triggers including environmental context, system limitations, and operational conditions. The study identified key determinants, including planning inconsistencies, detection failures, and hardware malfunctions, and revealed clear seasonal variations, with disengagements peaking in summer and autumn and declining in winter and spring. Validated CHAID models demonstrated high accuracy, underscoring the importance of comprehensive training and testing across diverse conditions to enhance effectiveness and safety.