A Framework for Systematising the Interactions Between ISO/PAS 8800:2024 and ISO 21448:2022 for AI-System-Based Advanced Driver Assistance Systems

Sammanfattning

The integration of artificial intelligence into safety-critical automotive systems such as Advanced Driver Assistance System (ADAS) introduces challenges that existing safety standards were not originally designed to address in combination. ISO 21448:2022 (Safety of the Intended Functionality, SOTIF) provides a framework for hazards arising from functional insufficiencies and foreseeable misuse, while ISO/PAS 8800:2024 extends this by addressing AI-specific risks such as data dependency, model uncertainty, and non-deterministic behavior. However, no clear guidance currently exists on how these two standards should be systematically applied together during early safety lifecycle activities. This thesis addresses that gap by proposing a framework for systematizing the interactions between ISO/PAS 8800:2024 and ISO 21448:2022 within the scope of Item Definition and Hazard Analysis for AI-based ADAS. The framework was developed through structured analysis of both standards, followed by identification of interaction points at the process and product levels. The resulting process-level framework is modeled using SysML activity diagrams, and integrates the unchanged, extended, and newly introduced activities into a unified workflow. The product-level artifacts complement the framework. A Block Definition Diagram represents a generalized AI-based ADAS architecture, and a Hazard Analysis table incorporates AI-specific triggering condition categories derived from ISO/PAS 8800:2024. The framework was applied to an Intelligent Speed Assistance system as a case study, and evaluated through survey-based validation with 16 participants evaluating the process-level framework and 12 evaluating the product-level artifacts. Validation results indicate that the Collaboration Swimlane representation is the preferred visualization of the process-level framework, and that both process-level and product-level models are perceived as relatively understandable, consistent, and useful. The main area identified for improvement is the visibility of standard-specific contributions within the product-level artifacts.