Empirical Assessment of Uncertainty-Aware Real-Time Decision-Making for Reliable and Resilient Cyber-Physical Systems

Abstract

Cyber-physical systems increasingly require real-time decision mechanisms capable of maintaining dependable operation under uncertainty arising from sensing errors, communication degradation, computational constraints, physical faults, and cybersecurity disturbances. This study empirically assessed uncertainty-aware real-time decision-making for improving predictive performance, reliability, safety, resilience, and recovery across heterogeneous cyber-physical environments. A quantitative experimental design was applied to 7,846,320 retained observations obtained from 15 datasets, 42 CPS environments, and 3,600 experimental runs incorporating smart-grid, industrial-control, autonomous, robotic, and networked critical-infrastructure systems. The experimental environment included 6,284 sensors, 2,176 actuators, 1,248 computational nodes, and 14,860 communication links. Uncertainty-aware and deterministic decision approaches were evaluated under increasing sensor noise, missing observations, packet loss, communication latency, component degradation, workload pressure, physical faults, cyberattacks, and combined cyber-physical disturbances. Under uncertain conditions, uncertainty-aware decision-making achieved 94.8% accuracy, 94.2% precision, 93.7% recall, 93.9% F1-score, 0.981 ROC-AUC, and 0.969 PR-AUC, compared with 88.6% accuracy and 86.8% F1-score for deterministic decision-making. Expected calibration error decreased from 0.086 to 0.031, while the deadline-miss rate declined from 6.9% to 2.8% despite increased mean inference latency from 16.4 to 21.7 ms. Successful operation increased from 86.7% to 95.1%, system availability reached 97.0%, and unsafe-action frequency decreased from 6.8% to 2.3%. Uncertainty-aware operation also reduced maximum performance degradation from 31.6% to 18.7%, increased retained functionality from 68.4% to 81.3%, and shortened mean recovery time from 18.4 to 10.7 seconds. Inferential analyses identified statistically significant performance differences with moderate-to-large effect sizes, while validation accuracy remained between 93.8% and 94.6% across alternative cross-validation procedures and reached 92.8% under external validation. Overall, the findings demonstrated that explicit uncertainty representation strengthened calibrated prediction, real-time decision quality, fault tolerance, safety, operational continuity, and resilience, particularly under severe and interacting cyber-physical disturbances.