Fault Diagnosis Using Hybrid Transfer Learning for VNF-based 5G Core Networks

In recent years, fault diagnosis in the 5G core network (5GC) has been widely investigated using machine learning (ML) due to the increasing requirement for adaptability and robustness in 5GC management algorithms. However, ML-based solutions often encounter challenges such as limited fault data availability and low adaptability to changing environments. Transfer learning (TL) offers a promising solution by transferring knowledge from a source domain to a target domain. This paper presents a hybrid transfer learning-based network fault diagnosis approach for the 5GC network, involving model-based TL, instance-based TL, and feature-based TL. Unlike the traditional single model-based TL, our method utilizes the inherent knowledge in 5GC network data and prevents negative transfer in features and data instances. We build a scale-out 5G testbed to generate sufficient source domain data while collecting a small amount of target domain data from a real 5GC network setting. We conduct network fault classification among different virtualized network functions (VNFs) in 5GC. We found that, compared to methods without TL and conventional model-based TL, hybrid TL provides a significantly higher accuracy in terms of F1 score. Furthermore, for each network label, 70 % of the labels achieve a F1 score higher than 0.9.