5G Network Fault Diagnosis System (7B Model)
Overview
This project contains the solver implementation for the 7B model track. It is designed to automatically diagnose 5G network issues by analyzing drive test data, engineering parameters, and problem descriptions.
The system employs a Hybrid Neuro-Symbolic Architecture, combining a deterministic rule-based expert system with the Qwen 2.5 7B Instruct Large Language Model (LLM) to achieve high accuracy and explainability.
Key Components
1. Core Solver (run.py)
The main execution script that orchestrates the entire diagnosis process.
- Rule Engine: Implements v17+ logic for detecting specific root causes:
- Weak Coverage: Analyzes RSRP levels and antenna tilt.
- Overlap Coverage: Compares signal strengths of neighboring cells.
- Handover Issues: Checks threshold parameters (A3 Offset) and handover events.
- PDCCH Congestion: Monitors CCE usage rates.
- AI Integration: Uses the OpenRouter API to query the Qwen 2.5 7B model for complex, non-standard questions that cannot be resolved by rules alone.
- Feature Extraction: Automatically parses unstructured text to extract key metrics (RSRP, SINR, Speed, PCI, etc.).
2. File Structure
run.py: Main entry point.train.csv: Training dataset (used for Case-Based Reasoning).phase_1_test.csv/phase_2_test.csv: Test datasets.output/: Directory for generated logs and submission files.submission.csv: Final formatted submission file.solve_log.jsonl: Detailed execution logs.
Usage
To run the solver and generate the submission file:
python run.py
The script will:
- Load and cache the case library.
- Process Phase 1 and Phase 2 questions.
- Apply rules to standard telecom questions.
- Invoke the AI model for complex scenarios.
- Generate
output/submission.csvwith the required formatting (Answer in Column 3).
Output Format
The system generates answers in the specific format required for the 7B track:
- Column 3 (Qwen2.5-7B-Instruct): Contains the reasoning and the final answer boxed in LaTeX format (e.g.,
\boxed{A}). - Other Columns: Marked as
placeholder.