Adaptive contrastive learning based network latency prediction in 5G URLLC scenarios

Abstract

With the increasing demand for 5G Ultra-Reliable Low Latency Communication (URLLC), accurate network time delay analysis is crucial. However, current network latency prediction models designed for 5G URLLC scenarios overlook the multiple downstream tasks case, such as network intrusion detection, recognition of network operational states, and decision-making for network resource scheduling.Furthermore, the robustness of the models needs to be enhanced. In this paper, we propose a Contrastive Convolutional Structure for Network Delay (CCSND), an adaptive contrastive learning framework that utilizes a temporal representation extraction module for capturing temporal periodicity and statistical features. To improve robustness and adaptability in time-series representation, CCSND designs strategic data augmentation methods, including temporal segmentation and delay obscuration. Experiment results on several real-world 5G URLLC scenario datasets demonstrate the superior performance of the proposed CCSND. It outperforms the current state-of-the-art Informer model by 38.7% and the traditional LSTM model by 64.5% in terms of mean squared error (MSE). Our code can be accessed at https://github.com/MerelyBreeze/CCSND.