From Component Reliability to Dynamic Predictive Reliability: A Multilayer Framework for Cyber-Physical Systems

Abstract : The reliability of cyber-physical systems (CPS) cannot be reduced to the failure probability of individual hardware components. Modern industrial, urban, educational and energy CPS combine physical processes, sensors, communication networks, software services, control algorithms, cyber-security mechanisms, human operators and environmental conditions. This paper proposes an integral-predictive view of CPS reliability based on two complementary models: the multilayer integral reliability index of a cyber-physical system (MINKFS) and the dynamic cognitive-predictive reliability model (DKPN). The framework takes different types of signals, converts them into standardized layer indices, combines them using a weighted structure, and updates the current reliability estimate by adding the predicted chance of future failure. The empirical part uses a synthetic industrial CPS dataset of 336 hourly observations and demonstrates how layer statistics, scenario analysis, operational regimes and criticality rankings can support early warning and preventive decision-making. The model transforms diverse inputs to normalized layer indexes, integrates these using a weighted integrative model, and adds the forecast probability of future failure to the present measure of reliability. The empirical part studies a fabricated data set related to industrial CPS which consists of 336 hourly measurements and provides a demonstration of how layer, scenario, regime and criticality information is used for detection and prevention. It was observed that the cyber-resilience layer had the lowest average reliability as well as the highest critical contribution, whereas communications and environment-energy layers were characterized by high volatility. Cyber event scenario resulted in the lowest predictive reliability and highest likelihood of failure. These results confirm the thesis that reliability of a CPS is better analyzed as a time-dependent, interpretable, and scenario-dependent feature instead of being just another component coefficient.