Research of Deep Learning and Adaptive Threshold-Based Signaling Storm Prediction and Top Cause Tracking

Malicious software or misbehaving applications have the potential to trigger signaling storms on mobile networks, leading to battery drainage on devices and causing bandwidth overuse at the cell level. Additionally, these storms may result in an excessive signaling load within the mobile operator's infrastructure. This paper uses a combination of time series prediction, adaptive threshold, and anomaly detection algorithms to predict signaling storms. Whether a signaling storm will be triggered in the future can be determined based on the fluctuation pattern of the data. Our method enables us to identify the top cause of the signaling storm in advance, so that the network optimization team can address the issues that will arise in advance, maximizing the stop-loss. The time series prediction algorithm has significant advantages over the moving average and TFT(Temporal Fusion Transformers), with a WAPE(Weighted Absolute Percentage Error) of only 0.09. Adaptive threshold can avoid treating holiday data as abnormal data, and the accuracy of anomaly detection based on the automatic adaptive threshold is higher than the traditional fixed threshold. In addition, combining the signaling conduction chain can also perform top cause localization to identify the upstream network element instance that first encountered the problem. The entire algorithm not only performs well in the current network but also performs well in artificially generated signaling storm data, pioneering the field of signaling storm prediction.