Gateway-level Early Warning of LoRaWAN Link Degradation Using LoED Measurements

LoRaWAN gateways observe packet-level metadata that may support reliability monitoring without modifying end devices or the protocol stack. We formulate short-term link degradation as a next-window warning task and aggregate LoED records into five-minute gateway windows described by CRC success rate, RSSI, SNR, spreading factor, traffic, frequency diversity, and historical features. A degradation event is a 10% or 20% decrease in the following window's CRC success rate. Class-balanced logistic regression (LR), random forest, XGBoost, and rule baselines are evaluated chronologically and with leave-one-gateway-out (LOGO) testing. LR obtains chronological PR-AUC values of 0.563 and 0.367 and exceeds both non-linear models in date-clustered randomization tests after Holm correction. Its LOGO means are also the highest, but the differences are not statistically significant, as only seven and five gateways contain both classes. Removing current CRC alone has a negligible effect, whereas removing all CRC history substantially reduces performance. Performance is further characterized across four window durations, probability calibration schemes, and explicit sequence models. A strictly chronological offline replay processes held-out windows with P99 scoring latency below 15 ms. The results support risk ranking within the LoED deployment, while gateway heterogeneity and calibration drift limit broader transfer claims.