Automated rock mass condition assessment during TBM tunnel excavation using deep learning

Area of Science:

• Geotechnical Engineering
• Artificial Intelligence
• Computer Vision

Background:

• Rock mass condition assessment is crucial for intelligent tunnel boring machine (TBM) control.
• Existing methods often lack automation and real-time capabilities.
• A large in-situ rock mass image dataset was collected from a water conveyance channel project.

Purpose of the Study:

• To develop an automated visual assessment system for rock mass condition.
• To transform rock mass assessment into a fine-grain classification task.
• To design and validate a novel deep learning network for this purpose.

Main Methods:

• A Self-Convolution based Attention Fusion Network (SAFN) was designed.
• The network incorporates Self-Convolution based Attention Extractor (SAE) and Self-Convolution based Attention Pooling (SAP) modules.
• SAE detects intact rock regions, while SAP fuses attention maps to enhance classification.

Main Results:

• SAFN demonstrated superior performance compared to state-of-the-art models on the rock mass assessment dataset.
• Evaluations included interpretability, ablation studies, accuracy, and cross-validation.
• Dynamic field tests confirmed the system's accuracy and efficiency for automated rock mass classification.

Conclusions:

• The developed SAFN model provides an accurate and efficient solution for automated rock mass condition assessment.
• The visual assessment system integrated with SAFN enhances TBM operational intelligence and safety.
• This approach represents a significant advancement in applying deep learning to geotechnical engineering challenges.