Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Uncovering Issues in the Radio Access Network by Looking at the Neighbors

José Suárez-Varela, Andra Lutu

TL;DR

This work targets the challenge of detecting network-related anomalies in large Radio Access Networks (RAN) where mobility introduces confounding variation in KPIs. It introduces c-ANEMON, an unsupervised, graph neural network–based contextual anomaly monitor that models each cell together with its local neighborhood to predict KPI trends over a horizon and detect deviations independent of mobility factors. The approach emphasizes interpretability and generalization, enabling a single model to cover extensive deployment regions, and includes an interpretable attention mechanism to reduce false alarms. Evaluated on real-world data from 7,890 cells over 3 months in a European metro area, the method identifies long-lasting anomalies with actionable potential (about 45.95%) and proves robust to large mobility events such as stadium crowds.

Abstract

Mobile network operators (MNOs) manage Radio Access Networks (RANs) with massive amounts of cells over multiple radio generations (2G-5G). To handle such complexity, operations teams rely on monitoring systems, including anomaly detection tools that identify unexpected behaviors. In this paper, we present c-ANEMON, a Contextual ANomaly dEtection MONitor for the RAN based on Graph Neural Networks (GNNs). Our solution captures spatio-temporal variations by analyzing the behavior of individual cells in relation to their local neighborhoods, enabling the detection of anomalies that are independent of external mobility factors. This, in turn, allows focusing on anomalies associated with network issues (e.g., misconfigurations, equipment failures). We evaluate c-ANEMON using real-world data from a large European metropolitan area (7,890 cells; 3 months). First, we show that the GNN model within our solution generalizes effectively to cells from previously unseen areas, suggesting the possibility of using a single model across extensive deployment regions. Then, we analyze the anomalies detected by c-ANEMON through manual inspection and define several categories of long-lasting anomalies (6+ hours). Notably, 45.95% of these anomalies fall into a category that is more likely to require intervention by operations teams.

Uncovering Issues in the Radio Access Network by Looking at the Neighbors

TL;DR

This work targets the challenge of detecting network-related anomalies in large Radio Access Networks (RAN) where mobility introduces confounding variation in KPIs. It introduces c-ANEMON, an unsupervised, graph neural network–based contextual anomaly monitor that models each cell together with its local neighborhood to predict KPI trends over a horizon and detect deviations independent of mobility factors. The approach emphasizes interpretability and generalization, enabling a single model to cover extensive deployment regions, and includes an interpretable attention mechanism to reduce false alarms. Evaluated on real-world data from 7,890 cells over 3 months in a European metro area, the method identifies long-lasting anomalies with actionable potential (about 45.95%) and proves robust to large mobility events such as stadium crowds.

Abstract

Mobile network operators (MNOs) manage Radio Access Networks (RANs) with massive amounts of cells over multiple radio generations (2G-5G). To handle such complexity, operations teams rely on monitoring systems, including anomaly detection tools that identify unexpected behaviors. In this paper, we present c-ANEMON, a Contextual ANomaly dEtection MONitor for the RAN based on Graph Neural Networks (GNNs). Our solution captures spatio-temporal variations by analyzing the behavior of individual cells in relation to their local neighborhoods, enabling the detection of anomalies that are independent of external mobility factors. This, in turn, allows focusing on anomalies associated with network issues (e.g., misconfigurations, equipment failures). We evaluate c-ANEMON using real-world data from a large European metropolitan area (7,890 cells; 3 months). First, we show that the GNN model within our solution generalizes effectively to cells from previously unseen areas, suggesting the possibility of using a single model across extensive deployment regions. Then, we analyze the anomalies detected by c-ANEMON through manual inspection and define several categories of long-lasting anomalies (6+ hours). Notably, 45.95% of these anomalies fall into a category that is more likely to require intervention by operations teams.

Paper Structure

This paper contains 17 sections, 1 equation, 5 figures, 2 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Time series anomaly detection
  4. Contextual anomaly detection
  5. Proposed Solution
  6. Design principles
  7. Architecture
  8. Graph Data Representation: [Fig. \ref{['fig:solution-overview']}-(1)]
  9. Contextual Predictor: [Fig. \ref{['fig:solution-overview']}-(2)]
  10. Anomaly Detector: [Fig. \ref{['fig:solution-overview']}-(3)]
  11. Evaluation
  12. Dataset
  13. Training the Contextual Predictor
  14. Anomaly Detector calibration
  15. Evaluation of the anomalies
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Schematic representation of c-ANEMON.
  • Figure 2: Distribution of the entropy of attention coefficients.
  • Figure 3: Distribution of the normalized historical error.
  • Figure 4: Distribution of the anomaly scores.
  • Figure 5: Examples of anomalies. Solid blue lines show the PRB utilization of the affected cell. Dashed green lines show the utilization of the other cells in the same sector. Red shades mark anomalous periods; gray shades indicate the weekends.