The Aim of the paper is to consider approaches to the analysis of a safety model of complex multi-loop transportation systems comprising not completely supervised subsystems. Method. For the description of a safety model, the paper uses systems theoretic process analysis (STPA) methods and the principles specified in ISO/PAS 21448:2019 (SOTIF). Result. The paper shows drawbacks of the FTA and FMEA local risk analysis methods and demonstrates a demand for some universal approach based on the combination of STPA and control theory. It gives an overview of the major stages of such analysis for the safety model of complex transportation systems exemplified by the Moscow Central Circle, which provide a feedback for safety evaluation of a transport control system under development. The paper analyzes the feasibility of using a virtual model for control purposes in the form of a so-called “supervised artificial neural network”.Conclusion. Today, railways are actively testing autonomous systems (with no driver onboard) that apply as their subsystems automatic perception modules using machine learning. The introduction of the latter into the control loop complicates the task of hazard analysis and safety evaluation of such systems using conventional FTA and FMEA methods. The construction of a safety model of such complex multi-loop transportation systems comprising not completely supervised subsystems that use machine learning methods with not completely predictable behavior requires the application of a systems approach to the analysis of unsafe scenarios along with the compilation of a scenario library and the formalization of a hazard model’s description, pertaining to the boundaries of various control loops as well, in order to reduce the regions of unknown unsafe scenarios for autonomous transportation systems under development.

No abstract