Knowledge-based and data-driven underground pressure forecasting based on graph structure learning.

Area of Science:

• Geotechnical Engineering
• Artificial Intelligence

Background:

• Accurate rock pressure prediction is crucial for underground excavation safety and industrial automation.
• Classical machine learning methods often overlook the causal links between pressure triggers and changes.

Purpose of the Study:

• To develop an advanced machine learning model that integrates causal inference for enhanced underground pressure prediction.
• To address the limitations of existing methods by incorporating causal logic into pressure prediction.

Main Methods:

• Construction of a causal graph representing relationships between pressure inducements and manifestations, informed by prior knowledge.
• Development of a prediction network utilizing graph convolutional networks and long short-term memory.
• Design of a reinforcement learning algorithm to optimize the causal graph based on prediction performance.

Main Results:

• The proposed Reinforced and Causal Graph Neural Network (RC-GNN) demonstrates superior performance in underground pressure prediction.
• Experimental results show an 18-60% increase in performance compared to six representative methods on real-world data.

Conclusions:

• The RC-GNN effectively models causal relationships, leading to more accurate underground pressure predictions.
• This causal approach offers a significant advancement for safety and intelligentization in underground engineering projects.