TransKS: An Anomaly Detection Method for Telecommunication Networks Based on Deep Learning

As the scale of telecommunication networks continues to grow, the structure of network elements becomes increasingly complex. Consequently, the data service has experienced a significant boom, leading to a substantial increase in network failures and equipment alarms. Effectively detecting hidden anomalies within massive network time series presents a significant challenge for network operators during their operation and maintenance tasks. This paper proposes a multi-dimensional sequence anomaly detection method named TransKS that integrates Transform and K-SIGMA on dynamic variable thresholds and applies it on the VoLTE and 5G core network. By designing the Transform multi-dimensional time series predictor, the method carries out parallel modeling and forecasting of the index sequence. It may thoroughly investigate the time dimension correlation relationship of the indicators as well as the linking effect of the indicators. And then we designed an adaptive sliding window approach on validation score by dynamically updating the value K and the threshold interval to detect context abnormalities and evaluate the abnormality of network element indicators quantitatively and reasonably. Experiments show that compared the performance with other algorithms, the TransKS has a higher accuracy rate of anomaly detection. When MAPE=0.01 and k is 4, the accuracy rate is the highest, up to 96.92%, which is at least 2.10% higher than other combined models, and its false positive rate and false positive rate have decreased, only 3.06% and 3.25%.