Evaluating Self-Supervised Representation Learning for UE-Level Anomaly Detection in Simulated 5G Logs

Sammanfattning

Modern 5G simulators generate large volumes of log data, making manual analysis impractical at scale and motivating automated analysis. This thesis evaluates whether self-supervised time-series representation learning improves anomaly scoring and detection performance at the user equipment (UE) level in simulated 5G logs compared with a feature engineering baseline based on aggregated per-UE summary statistics. Three self-supervised methods, TimeDRL, TS2Vec, and T-Loss, were trained on normal simulator runs and evaluated in a target-adapted unsupervised anomaly detection pipeline with multiple outlier scorers and threshold selection methods. Among the learned methods, TS2Vec performed best, achieving an average precision of 0.716 with Mahalanobis scoring and a best F1-score of 0.545 with MAD-based thresholding. However, the baseline achieved the strongest overall results, with an average precision of 0.840 and a best F1-score of 0.909. These results suggest that, in the studied simulated setting, aggregated per-UE summary features separated anomalous and normal UEs more effectively than the evaluated learned temporal representations.