Safety Analysis Concept and Methodology for EDDI development (Initial Version)

Aslansefat, Koorosh; Gerasimou, Simos; Michalodimi-trakis, Emmanouil; Papoutsakis, Manos; Reich, Jan; Sorokos, Ioannis; Walker, Martin; Papadopoulos, Yiannis

Abstract

Executive Summary:
This deliverable describes the proposed safety analysis concept and accompanying methodology to be defined in the SESAME project. Three overarching challenges to the development of safe and secure multi-robot systems are identified — complexity, intelligence, and autonomy — and in each case, we review state-of-the-art techniques that can be used to address them and explain how we intend to integrate them as part of the key SESAME safety and security concept, the EDDI.
The challenge of complexity is largely addressed by means of compositional model-based safety analysis techniques that can break down the complexity into more manageable parts. This applies both to scale — modelling systems hierarchically and embedding local failure logic at the component-level — and to tasks, where different safety-related tasks (including not just analysis but also requirements allocation and assurance case generation) can be handled by the same set of models. All of this can be combined with the existing DDI concept to create models — EDDIs — that store all of the necessary information to support a gamut of design-time safety processes.
Against the challenge of intelligence, we propose a trio of techniques: SafeML and Uncertainty Wrappers for estimating the confidence of a given classification, which can be used as a form of reliability measure, and SMILE for explainability purposes. By enabling us to measure and explain the reliability of ML decision making, we can integrate ML behaviour as part of a wider system safety model, e.g. as one input into a fault tree or Bayesian network. In addition to providing valuable feedback during training, testing, and verification, this allows the EDDI to perform runtime safety monitoring of ML components.
The EDDI itself is therefore our primary solution to the twin challenges of autonomy and openness. Using the ConSert approach as a foundation, EDDIs can be made to operate cooperatively as part of a distributed system, issuing and receiving guarantees on the basis of their internal executable safety models to collectively achieve tasks in a safe and secure manner.
Finally, a simple methodology is defined to show how the relevant techniques can be applied as part of the EDDI concept throughout the safety development lifecycle.

Citation

Aslansefat, K., Gerasimou, S., Michalodimi-trakis, E., Papoutsakis, M., Reich, J., Sorokos, I., Walker, M., & Papadopoulos, Y. (2023). Safety Analysis Concept and Methodology for EDDI development (Initial Version). European Comission

• Report Type: Research Report
• Online Publication Date: Dec 1, 2023
• Publication Date: Dec 1, 2023
• Deposit Date: Jul 13, 2024
• Publicly Available Date: Jul 16, 2024
• Pages: 147
• DOI: https://doi.org/10.3030/101017258
• Public URL: https://hull-repository.worktribe.com/output/4736517
• Publisher URL: https://cordis.europa.eu/project/id/101017258/results