A Systematic Approach to Enhancing ISO 26262 With Machine Learning-Specific Life Cycle Phases and Testing Methods

This paper presents a systematic approach to enhancing ISO 26262, a widely adopted standard for automotive functional safety, by integrating Machine Learning (ML)-specific life cycle phases and testing methods for Automotive Safety Integrity Level (ASIL) A/B. With the increasing incorporation of ML techniques in automotive systems, the current ISO 26262 framework reveals significant gaps in addressing ML-specific safety requirements. While ISO/DPAS 8800 provides an approach for developing AI systems that meet some safety properties, it does not provide a mapping concept for ASIL classification of ML systems. Furthermore, given the complexity of ML techniques in automotive systems, issues such as interpretability—critical for transparency and accountability—along with robustness and uncertainty handling, pose significant challenges that are not fully addressed by ISO 26262 and ISO/DPAS 8800. This study identifies and addresses these gaps by defining three additional life cycle phases: prepare data, train ML model, and deploy ML model. For each life cycle phase, we establish desired properties such as robustness, uncertainty handling, and interpretability, and propose suitable methods to achieve these properties. We adopt a rigorous evaluation framework inspired by IEC 61508 to assess the effectiveness of these methods. Since the method recommendations of ISO 26262 for ML-based products are incomplete, the approach presented in this paper provides critical guidance and room for expert assessment and independent certification, ensuring solid and reliable recommendations. This systematic, clear, uniform development procedure not only supports product teams in achieving their safety goals but also facilitates the certification process, reducing ambiguity and enhancing the overall safety and reliability of ML-based automotive systems.