Functional safety approval of an AI-based person-detecting safety product

Abstract

The use of artificial intelligence (AI) across different sectors has rapidly increased in recent years, yet current functional safety standards are built based on deterministic systems and do not directly align with machine learning (ML). In collaboration with RISE Research Institutes of Sweden AB an investigation was carried out concerning how an AI-based person detection system could be developed and evaluated in accordance with existing safety standards. A convolutional neural network (CNN) based person identification system was developed in the project according to principles of ISO 12100, ISO 13849-1 and ISO 8800, where traceability, repeatability and documentation was prioritized. The CNN was trained on different datasets consisting of a set of individuals in a static environment. The main goal of the model was to classify each detected individual as authorized, unauthorized, or unknown. Three different versions of the system were developed, each building upon the previous. The final version achieved consistent performance, with the highest accuracy of 97%. This demonstrated that an AI system can behave predictably under controlled conditions and could potentially be integrated into safety applications. However, the development also highlighted several limitations of the system. The dataset was limited in size and diversity, and the evaluation was conducted within a confined and controlled environment. This ultimately meant that testing and validation outside of the predefined conditions were lacking, which restricts the generalization of the results. In conclusion, the study presented a conceptual development of an AI system in accordance with safety standards and that the AI system could potentially support the case of AI integration into safety applications, but to a certain extent. More extensive work regarding expansion and quality of dataset, wider testing and further adaptation of standards is a necessity for a full accommodation of AI in safety systems.