Functional safety architectural patterns for AI-based critical systems

Abstract

Embodied Artificial Intelligence (EAI) systems are transforming industries by enabling complex interactions with dynamic environments. However, their integration into safety-critical applications presents challenges, particularly in meeting functional safety requirements. Existing standards such as IEC 61508 and ISO 26262 provide guidance for risk mitigation but lack explicit solutions for AI-driven systems. The stochastic nature of Artificial Intelligence (AI) models, combined with increasing hardware and software complexity, introduces risks such as model insufficiencies and failures in heterogeneous computing architectures. This paper proposes a modular reference architecture for safety-critical EAI systems that aligns with functional safety and AI standards, including ISO/IEC TR 5469 and ISO/PAS 8800. Inspired by established safety patterns like the E-gas Concept in the automotive domain, the proposed reference architecture is applied to three incremental safety patterns on an NVIDIA JETSON ORIN platform, progressively addressing safety requirements while promoting reusability and compliance. This work provides a structured pathway for integrating AI into critical systems by defining a set of architectures, safety techniques, and measures that form the foundation of safety-critical systems, ensuring reliability, predictability, and compliance with functional safety frameworks.

The research leading to these results has received funding from the SAFEXPLAIN project of the European Union’s Horizon Europe programme under grant agreement number 101069595 and from CAPSUL-IA project PLEC2023-010240 funded by the Spanish Ministry of Science and Innovation (MICIU/AEI/10.13039/501100011033).

Peer Reviewed

Postprint (author's final draft)