ROUTE CONTINUITY OF ANOMALY ASSESSMENT IN TELECOMMUNICATION TIME SERIES UNDER TRANSITION REGIMES BASED ON HYBRID AWRED

Abstract

The article considers the task of detecting transition anomalies in KPI time series, where KPI denotes key network performance indicators. Unlike short failures, such a regime does not always appear as a sharp outlier in a single indicator. The change may accumulate gradually. In this case, the shape of the route changes, the consistency between RSRP (reference signal received power), DL_BLER (downlink block error rate), UL_MCS (uplink modulation and coding scheme), and UL_SNR (uplink signal-to-noise ratio) is disturbed, while individual values may still remain close to acceptable limits. For this reason, reconstruction-based models often respond to local peaks of the anomaly score but cannot keep it elevated over the entire transition segment.

To address this task, a route-oriented Hybrid AWRED v6 method is proposed. It combines a Transformer reconstructor, the Normal-Support Anchor mechanism, and the route component of the anomaly score. This structure takes into account not only the magnitude of a local deviation, but also the stability of the model response along the transition. The evaluation was performed on the TelecomTS dataset for a soft regime-change scenario. The comparison was carried out in a series of five independent runs against DAGMM, TranAD, and Anomaly Transformer.

Hybrid AWRED v6 achieved mean values of ROC-AUC = 0.959 and PR-AUC = 0.890. The method also demonstrated higher F1, Recall, and route detection quality RDQ. Precision was slightly lower than that of TranAD and Anomaly Transformer, but this difference is related to broader transition coverage, rather than to a loss of ability to separate normal and anomalous windows. Full repeatability of Continuity was also recorded across five runs — 0.378 ± 0.000. The longest continuous segment of correct detection had the same relative length regardless of model initialization.

The obtained results show that Hybrid AWRED v6 forms not only a pointwise, but also a route-stable anomaly score for soft transition regimes in time series.

Keywords: Hybrid AWRED v6, anomaly detection, telecommunication time series, transition anomaly, route continuity, KPI, PR-AUC, RDQ.