Networked Control Systems (NCSs) are increasingly deployed in critical infrastructure, yet their performance is often compromised by communication impairments such as time-varying delays and packet loss. These disturbances can undermine stability, reliability, and overall efficiency. To address these challenges, this study proposes an AI-augmented supervisory control framework that integrates communication metrics directly into the control strategy rather than relying solely on conventional error signals. The architecture employs machine learning classifiers—specifically K-Nearest Neighbors (KNN) and Random Forest (RF)—to assess real-time network conditions (normal, degraded, and emergency) and dynamically adjust the parameters of a core PID controller. Under degraded conditions, PID gains are attenuated to enhance robustness against latency, while in emergency scenarios, a conservative safe-mode configuration is activated to preserve plant stability under severe packet loss. Comparative analysis demonstrates that Random Forest, due to its ensemble-based design, achieves superior situational awareness, smoother mode transitions, and more reliable global stability than KNN. Moreover, RF reduces volatile parameter updates, thereby protecting physical actuators from excessive stress and extending component lifespans. The findings highlight the potential of ensemble learning approaches to enhance resilience and operational safety in delay-prone NCS environments.
Despite strict industry commitments to network uptime, engineers have yet to settle on a single, universally accepted method for computing the Network Unavailability Rate (NUR). This paper derives and comparatively evaluates five NUR formulations — the classical uptime ratio, MTBF/MTTR-based reliability model, failure-rate (λ) approximation, Erlang queuing model, and the telecom cell-based model — within a unified analytical-simulation environment applied to LTE/5G networks. A Monte-Carlo framework was constructed over a synthetic topology of 1,000 cells across 50 replications to isolate each model's sensitivity to MTTR, failure rate, and spatial burst failures. Results show the classical model underestimates true downtime by roughly 18%, whereas the cell-based telecom formulation achieves an error of approximately 3% by capturing both temporal and spatial failure dynamics. The MTBF/MTTR, λ, and Erlang models occupy an intermediate error band of 9–12%. These findings point toward cell-level KPI aggregation as the preferred basis for SLA monitoring and reliability engineering in contemporary and forthcoming 5G deployments.