A robust deep learning approach for rock discontinuity identification from large scale 3D point clouds

Area of Science:

• Geological Engineering
• Artificial Intelligence
• Computer Vision

Background:

• Rock discontinuities critically influence rock mass behavior, impacting slope stability, underground excavation, and blasting.
• Traditional point cloud analysis methods struggle with fine-scale feature representation and parameter sensitivity.
• Accurate discontinuity characterization is essential for reliable geotechnical engineering applications.

Purpose of the Study:

• To introduce RL-JointNet, an end-to-end deep learning model for robust discontinuity segmentation in rock masses.
• To enhance local feature extraction for improved spatial relationship and neighborhood geometry representation.
• To provide a reliable automated approach for detailed characterization of large-scale rock masses.

Main Methods:

• Developed RL-JointNet, an end-to-end deep learning framework for discontinuity segmentation.
• Implemented an enhanced local feature extraction module with relative position encoding and multi-path feature fusion.
• Validated the model on high-resolution point cloud datasets from two rock slopes.

Main Results:

• RL-JointNet achieved high performance metrics: Global Accuracy (GA) up to 98.7% and mean Intersection over Union (mIoU) of 98.1%.
• Individual discontinuity class recognition accuracy consistently exceeded 95%.
• Hyperparameter sensitivity analysis demonstrated RL-JointNet's superior robustness compared to conventional models.

Conclusions:

• RL-JointNet offers a reliable and robust solution for automated rock discontinuity analysis.
• The model significantly improves the characterization of complex, large-scale rock masses.
• Enhanced local feature extraction is key to the model's improved performance and stability.