Edge Intelligence-Based Rail Transit Equipment Inspection System.

Area of Science:

• Intelligent transportation systems
• Edge computing in critical infrastructure
• Computer vision for industrial automation

Background:

• Rail transit safety depends on reliable equipment maintenance, but manual inspections are inefficient and error-prone.
• Traditional methods struggle with the high costs, low efficiency, and human error associated with manual rail equipment checks.
• Automating inspection processes is crucial for enhancing the safety and operational efficiency of rail transit.

Purpose of the Study:

• To develop and evaluate an automated rail transit equipment inspection system using Edge Intelligence (EI) and 5G technology.
• To enhance the safety, real-time performance, security, and robustness of rail transit maintenance.
• To address the limitations of traditional manual inspection methods in rail transport.

Main Methods:

• A cloud-edge-end collaborative architecture integrating Computer Vision (CV) for automated inspection.
• Utilized a fine-tuned YOLOv8 model for object detection and a ResNet-18 network for equipment status classification.
• Implemented an ETSI MEC-compliant framework on edge servers (NVIDIA Jetson AGX Orin) within a 5G private network.

Main Results:

• YOLOv8 achieved 92.7% mAP@0.5 for equipment detection; ResNet-18 attained 95.8% accuracy for status classification.
• Reduced average end-to-end latency for anomaly detection by 45% (28.5 ms vs. 52.1 ms) compared to cloud-centric systems.
• Lowered daily bandwidth consumption by approximately 98.1% (40.0 GB to 0.76 GB) via event-triggered evidence uploads.

Conclusions:

• The EI and 5G-based system offers superior real-time performance, enhanced security, and improved robustness for rail transit equipment inspection.
• The automated system significantly outperforms cloud-centric approaches in terms of latency and bandwidth efficiency.
• The deployed system on Beijing Metro Line 6 demonstrates practical viability and effectiveness in real-world scenarios.