Baselines and Anomalies in Network Test Data for Degradation Detection

Abstract

Network performance degradation may occur without anomalous events in individual network measurements. This paper presents a baseline-and-anomaly framework for detecting gradual degradation in network test data. The approach establishes probability-density and time-series baselines from historical measurements and quantifies observed degradation relative to expected behavior. It considers statistical, machine-learning, hybrid, and probabilistic methods, while accounting for temporal, spatial, seasonal, and diurnal variability. A state-machine representation is also used to characterize transitions from healthy operation through minor, moderate, and severe degradation to failure. Evaluation considerations include synthetic and real-world datasets, robustness to controlled perturbations, detection metrics, operational thresholding, interpretability, and deployment constraints. The discussion highlights data quality, privacy, transferability, and adaptation to evolving traffic patterns as continuing challenges.