5G-R Network KPIs Anomaly Detection Technology Based on DSTAD-R

To address the challenges of large data volume, complex temporal dependency, and strong multi-dimensional correlation in the anomaly detection of key performance indicators (KPIs) for 5G-Railway (5G-R), this paper proposes a dual-stage Transformer anomaly detection model (DSTAD-R) with an adversarial training style. A KPI feature engineering system tailored for railway scenarios is constructed, integrating critical 5G-R KPIs such as wireless signal quality (RSRP/SINR), FTP uplink and downlink throughput, and MAC layer uplink and downlink throughput. DSTAD-R employs a Transformer-based Encoder-Decoder architecture to handle multi-dimensional and parallel processing of KPIs, and achieves model training through a two-stage joint optimization based on the Harmonization matrix. Extensive empirical studies on 5G-R dataset demonstrate that DSTAD-R can outperform other baseline methods in detection and diagnosis performance with data and time-efficient training. Specifically, compared with baseline models, this method improves the F1-score by 14% and reduces training time by 5 times compared with the optimal method.