Abstract
As a typical dynamic disaster in underground engineering, the occurrence mechanism of rockburst is jointly influenced by various spatially heterogeneous factors and potential long-range coupling relationships. Although traditional graph neural networks can process local structural information, significant limitations remain in modeling long-range dependencies, cross-type semantic interactions, and dynamic data variations. In response, this paper proposes an intelligent rockburst risk prediction method based on Heterogeneous Graph Transformer (HGT), which innovatively incorporates the Transformer architecture to enhance high-order semantic modeling and long-range feature dependency expression among different node types. In the modeling process, information such as rock physical parameters, geological structures, microseismic events, stress monitoring points, and construction disturbances is embedded as multi-type nodes, constructing a heterogeneous graph with multiple types of relational edges (such as geological adjacency, temporal evolution, structural coupling, etc.). The proposed HGT model employs type-specific projection mechanisms and attention weight allocation strategies, enabling efficient interaction among different types of nodes in a unified representation space. Additionally, Structural Positional Encoding is introduced to enhance the model's comprehension of topological structure and global node positions. The training process incorporates a multi-task loss function, enabling simultaneous prediction of rockburst occurrence probability and output of its spatial location distribution and potential impact areas, thereby enhancing the interpretability of prediction results. Validated on multiple real-world datasets from tunnel and mining engineering projects, experimental results demonstrate that the proposed method outperforms existing GCN, GAT, and LSTM-based multimodal models in terms of prediction accuracy, early warning capability, and generalization performance under complex structural scenarios, demonstrating broad prospects for engineering applications.
Full Text
Preamble
Title: Intelligent Rockburst Risk Prediction Based on Heterogeneous Graph Transformer
Authors: Xiangdong Ran¹, Yu Zhang¹,², Junchao Wang¹
Affiliations:
¹ School of Intelligence Science and Technology & Beijing Key Laboratory of Super Intelligent Technology for Urban Architecture, Beijing University of Civil Engineering and Architecture, Beijing 102616, China
² State Key Laboratory for GeoMechanics and Deep Underground Engineering, China University of Mining & Technology, Beijing 100083, China
Abstract
Rockburst, as a typical dynamic hazard in underground engineering, is governed by complex mechanisms involving multiple spatially heterogeneous factors and potential long-range coupling relationships. While traditional Graph Neural Networks (GNNs) can capture local structural information, they exhibit significant limitations in modeling long-range dependencies, cross-type semantic interactions, and dynamic data variations. To address these challenges, this paper proposes an intelligent rockburst risk prediction method based on Heterogeneous Graph Transformer (HGT), which innovatively leverages the Transformer architecture to enhance high-order semantic modeling and long-range feature dependency expression across different node types.
In the modeling process, information including rock physical parameters, geological structures, microseismic events, stress monitoring points, and construction disturbances is embedded as multi-type nodes to construct a heterogeneous graph with diverse relational edges (e.g., geological adjacency, temporal evolution, structural coupling). The proposed HGT model employs type-specific projection mechanisms and attention weight allocation strategies, enabling efficient interaction between different node types within a unified representation space. Additionally, Structural Positional Encoding is introduced to enhance the model's comprehension of topological structures and global node positioning. During training, a multi-task loss function is employed to simultaneously predict rockburst occurrence probability and output its spatial distribution and potential impact zones, thereby enhancing the interpretability of prediction results.
Validation on multiple real-world tunnel and mining engineering datasets demonstrates that the proposed method surpasses existing GCN, GAT, and LSTM-based multimodal models in prediction accuracy, early warning capability, and generalization performance in complex structural scenarios, exhibiting promising prospects for engineering applications.
Keywords: rockburst prediction; heterogeneous graph neural network; Transformer; structural positional encoding; rock mechanics; underground engineering