Detecting Gray Failures in Edge Networks with Lightweight Unsupervised Learning

Abstract

Gray failures degrade service quality without necessarily triggering binary availability monitors, leaving a gap between normal operation and complete outage. This paper investigates whether lightweight unsupervised learning can serve as an early-warning layer for proactive failover in edge networks. We present a six-node Raspberry Pi testbed that collects multiprotocol telemetry from ICMP, DNS, and HTTP probes and uses per-node anomaly calibration to account for heterogeneous wired and wireless baselines. Across five controlled fault-injection experiments and 1.09 million telemetry rows collected over 27 days, the results show that per-node calibration improves robustness across heterogeneous nodes, scoring cadence strongly affects early-warning latency, and automated failover can reduce HTTP latency by 65–72% after command delivery. A comparison with an EMA Z-Score baseline further shows that adaptive univariate detection reduces false alerts, while Isolation Forest is more sensitive to sustained multivariate degradation. Overall, the findings suggest that lightweight unsupervised learning can augment conventional observability stacks by triggering bounded mitigation actions during the gray-failure window.