Deadline-aware adaptive network coding for critical telemetry: deterministic, model-informed, and contextual-bandit control in narrowband wireless mesh networks
Аннотация
Critical narrowband supervisory control and data acquisition (SCADA) telemetry must deliver alarms, control commands, and periodic measurements under unequal deadlines and scarce airtime. We compare three recovery-control philosophies under a common absolute-deadline earliest-deadline-first (EDF) scheduler, hard feasibility guard, systematic random linear network coding (RLNC) actions over the Galois field GF($$2^8$$), and a four-hop 19.2-kb/s model. EDF deadline-aware adaptive network coding (EDF-DANC) uses calibrated class/loss thresholds; EDF model-informed utility network coding (EDF-MU-NC) evaluates the same actions analytically; and EDF contextual-bandit network coding (EDF-CB-NC) uses an offline-trained upper-confidence-bound (UCB) action map frozen before testing. Calibration/training and final test seeds are disjoint. At load 0.80 and packet-erasure probability 0.10, all three adaptive controllers obtain 73.8% on the 4:2:1 criticality-weighted objective, versus 72.1% for causal one-retry EDF automatic repeat request (ARQ) and 61.6% for uncoded EDF. Near threshold boundaries, MU-NC and CB-NC each improve over DANC by 5.50 percentage points at 8% loss and 9.03 points at 22%. In the time-varying 2%$$\rightarrow $$10%$$\rightarrow $$18%$$\rightarrow $$5% experiment, feedback-driven CB-NC reaches 78.0±1.7%, versus 75.5±1.6% for DANC and 76.1±1.6% for MU-NC, but uses 85.0 ms added airtime versus 50.9 and 52.6 ms. Independent checks show that the preferred controller changes with criticality weights and is broadly stable under the tested moderate feedback-loss conditions. The results therefore establish a criticality-versus-radio-cost frontier rather than universal superiority of learning.