TLDR: A novel 5G Core Fault Analysis Engine uses Machine Learning (Golden Flow and AI Engine) to detect faults in PCAP files and Generative AI (LLM with RAGs) to provide detailed troubleshooting steps and root cause analysis. This system significantly reduces manual effort and improves efficiency in 5G network management, showing high accuracy in fault detection and providing actionable solutions.
The rapid expansion of 5G networks brings incredible speed and connectivity, but also introduces complex challenges in maintaining network integrity and performance. Ensuring that packet core traffic flows smoothly is crucial, as errors can significantly impact connectivity strength and handover quality. Traditionally, identifying and resolving these network faults has been a time-consuming and labor-intensive process, often requiring many hours of manual analysis of test files like Packet Capture (PCAP) files and log files.
Addressing this challenge, researchers from Ericsson India have developed a groundbreaking AI/ML-driven Fault Analysis (FA) Engine. This innovative system is designed to automatically classify successful and faulty frames within PCAP files, specifically focusing on the 5G packet core. By leveraging natural language processing (NLP) techniques, the FA engine efficiently analyzes network traffic to pinpoint anomalies and inefficiencies, drastically cutting down the time and effort needed for troubleshooting.
A key feature of this FA Engine is its integration with Generative AI, specifically a Large Language Model (LLM). This LLM is trained on extensive 5G packet core documentation, including 3GPP standards and internal user documents. When a fault is detected, the engine not only identifies the issue but also suggests concrete steps to fix it, providing detailed explanations from a domain perspective. This capability significantly enhances the troubleshooting process, making it more accessible and efficient for network administrators.
How the 5G Fault Analysis Engine Works
The proposed solution, the 5G Core Fault Analysis Engine, operates through a multi-component architecture. It takes PCAP files as input, which are data files containing captured network traffic. These files record details like source and destination IP addresses, port numbers, protocol types, and packet contents, making them essential for network diagnostics.
The engine comprises two primary modules for fault detection: the Golden Flow model and the AI Engine.
The Golden Flow model uses a knowledge graph-based approach. It builds a graph from successful PCAP test files, where individual frames are nodes and connections between frames are directed edges. During analysis, it traverses this graph using the input PCAP file’s frames. If the flow deviates from the expected “golden” path, it marks the location as an error and can even suggest the correct path, helping developers understand the problem better.
The AI Engine is a supervised machine learning model, specifically an NLP model. It processes individual frames from PCAP files, cleaning and tokenizing the text, then vectorizing it using the Bag-of-words method. These vectorized representations are then fed into a Support Vector Machine (SVM) model, which classifies frames as either faulty or correct. The model is trained using both successful and failed test run PCAP files, automatically identifying faulty message frames without requiring explicit labeling of error causes.
Once the Golden Flow and AI Engine identify a faulty frame, the information is passed to the LLM-based Troubleshooting Agent. This agent uses MistralAI as its LLM and incorporates Retrieval-Augmented Generation (RAGs). These RAGs are built from a vast corpus of 5G testing schemes, error codes, 3GPP standards, and Packet Core Controller (PCC) specifications. By providing the LLM with context from the faulty frame (e.g., “Registration Reject, Congestion”) and previous frames, the RAGs enable the LLM to access relevant domain-specific information. This allows the LLM to generate accurate, contextually rich troubleshooting steps and even reference related documents, preventing “hallucinations” often seen in untrained LLMs.
Also Read:
- MX-AI: Intelligent Automation for Future 6G Networks
- Advancing 6G Network Management with Generative AI: A Hierarchical Learning Approach
Promising Results and Future Outlook
The research demonstrated promising results. The Golden Flow model achieved 100% detection accuracy on test files. The AI Engine, particularly when frames were categorized by protocol type (e.g., NAS-5GC), also showed 100% accuracy, precision, recall, and F1-score. Overall, the Support Vector Machine model outperformed other tested models like Neural Networks and Random Forests in classification accuracy.
For instance, when presented with a “Registration reject (Congestion)” error, the system not only identified the suspicious message and expected flow but also provided detailed procedures to fix the congestion, along with references to relevant documents. This significantly streamlines the diagnostic and resolution process for network engineers.
The researchers envision future enhancements, including expanding the engine to incorporate other forms of network data beyond PCAP files, such as log dumps and network configuration files. They also plan to explore multimodal LLMs that can directly process log or PCAP files, potentially leading to faster inference times and more efficient training pipelines. This work represents a significant step towards more efficient testing and network management for advanced telecommunications systems. You can read the full research paper here.
