Real-time classification of longitudinal conveyor belt cracks with deep-learning approach

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Mountain tunnel construction relies on long conveyor belts for mucking, which are prone to frequent damage.
• Current visual inspection methods for conveyor belt damage are time-consuming and may miss critical issues.
• Automated damage detection is needed to enhance safety and operational efficiency in tunnel construction.

Purpose of the Study:

• To develop and validate a real-time deep learning model for detecting and localizing damage on conveyor belts.
• To accurately assess the size of detected damages and identify belt markings for precise localization.
• To demonstrate the model's effectiveness on edge devices for practical application in mountain tunnel sites.

Main Methods:

• A deep learning model was configured to detect cracks and measure their size on conveyor belts.
• The model was enhanced to detect numerical markings on the belt for precise damage localization.
• Performance was evaluated in real-time on an edge device, measuring detection accuracy (mAP) and processing speed.

Main Results:

• The optimized model achieved a mean Average Precision (mAP) of 0.850 for crack detection and 0.99 for number marking detection.
• The system demonstrated superior real-time performance, capturing 15 frames per second on an edge device.
• The framework successfully localized damage with respect to the conveyor belt's length.

Conclusions:

• Deep learning offers a versatile and effective solution for real-time conveyor belt damage assessment in challenging mountain tunnel environments.
• The proposed model significantly improves the accuracy and efficiency of damage inspection, reducing downtime and maintenance costs.
• This technology validates the application of AI in enhancing safety and operational continuity in tunnel construction projects.