A TRANSFORMER-BASED GRAPH NEURAL NETWORK AGGREGATION FRAMEWORK FOR 5G RADIO LINK FAILURE PREDICTION

Abstract

The prediction of Radio Link Failures (RLF) in Radio Access Networks (RANs) is crucial to ensure smooth communication and meet the demanding requirements of high data rates, low latency, and improved performance in 5G networks. However, weather conditions like precipitation, humidity, temperature, and wind have a significant impact on these communication links. Typically, RLF prediction uses a learning-based model to capture the relationships between historical radio link Key Performance Indicators (KPIs) and nearby weather station data. However, existing models often lack the capability to effectively encode context-aware time series sequences and fail to be generalized for unseen radio links. To address these issues, this thesis proposes a new RLF prediction framework that employs a state-of-the-art time series transformer model as a temporal feature extractor and incorporates a graph neural network (GNN) based dynamic aggregation method for surrounding weather stations' data to achieve better model generalization. The proposed aggregation method can be integrated into any existing prediction model to enhance its generalizability. The framework was evaluated in rural and urban deployment scenarios with 2.6 million KPI data points, demonstrating significantly higher F1 scores compared to previous methods (0.93 for rural and 0.79 for urban).