Background. Existing analytical methods do not take into account nonlinear interactions between heterogeneous components of the system and do not reflect the dynamics of the system state. Thus, it is important to design a comprehensive system to assess the reliability of equipment that combines the characteristics of hardware, software and ergatic components.
Materials and methods. The training set has been generated based on a model of independent failures using Monte Carlo simulation of 100,000 scenarios, including rare events and simultaneous failures of components.
Results. The problem of assessing the reliability of automated enterprise control systems operating at nuclear power plants (NPP) is considered. The authors have proposed a hierarchical multicomponent neural network architecture consisting of structural, functional, and information security branches. It is shown that the developed model approximates the numerical reliability indicator in scaled units. However, in real-world conditions when accurate reliability calculations are essential, the relative error becomes a factor requiring further optimization of the model. This is especially relevant for NPP automated control system (ACS) reliability assessment, when target failure rates are very low, increasing the sensitivity of the assessment to small absolute errors.
Conclusions. The developed concept eliminates the limitation of analytical models to assess the reliability of ACS systems, which consists of the assumption of independent component assessment and the linear nature of their interactions. The use of neural networks allows for the consideration of nonlinear dependencies between subsystem state parameters that cannot be formalized using analytical methods. For further enhancement of accuracy, boosting, ensemble methods, and complex neural network architectures are proposed.