Revolutionary Approaches to Functional Safety in AI-Enabled Embedded Systems

Abstract

The artificial intelligence in safety-critical embedded systems has required a redesign of functional safety concepts. This article discusses revolutionary approaches to the problem of the probabilistic nature of AI that introduces unprecedented challenges in application areas, where classical deterministic standards of safety are inapplicable. It examines novel architectures of safety, such as independent oversight systems and redundancy inference systems that offer life-saving safety measures to the AI components. The specialized verification techniques, which adapt formal methods to neural networks, and runtime monitoring methods, which can identify possible security breaches during operation, are reviewed. The article examines how to extend an existing set of standards, such as ISO 2626,2 to be compatible with the specific use of AI and how to create safety-aware training processes, which introduce safety constraints into the training process. The most promising directions to achieve certification are illustrated through explainable AI methods, which provide visibility into safety validation and hybrid systems, considering rule-based systems and AI capabilities in combination. Applying case studies in the automotive, aerospace, and medical sectors, this article describes how the use of such complementary methods can allow safe application of AI in highly regulated sectors to ensure that strict safety requirements are met, even where intrinsic limitations on verification are acknowledged.