Rockburst Prediction via Multiscale Graph Convolutional Neural Network

Rockburst prediction has important significance for individual safety in deep geologic engineering. To accurately predict the rockburst-intensity level, this paper proposes a rockburst prediction method via multiscale graph convolutional neural network. In the method, six rockburst indicators of each rockburst are utilized as a rockburst sample, and data segmentation oversampling of rockburst samples solves the unbalanced class-distribution problem of rockburst-data samples. The similarity perception for capturing similarity structure relationships of rockburst samples in Euclidean space is given, and a novel indicator correlation constraint of the similarity relationships with multiple rockburst criteria is designed, which can reduce the structure distortion of Euclidean space. By integrating the perception and the constraint, criterion-constrained graph structure data are constructed under the graph neural network framework. On the basis of the graph data, a multiscale graph convolutional neural network is further proposed, and multiscale information with the topology structure and criterion-constrained relationships can be effectively captured by the network, which can achieve the accurate prediction of the rockburst-intensity level. The effectiveness of the proposed method is demonstrated by extensive experiment analysis. Give similarity perception to measure the rockburst similarity for similar structural relationships of rockburst samples in Euclidean space.Design a novel indicator correlation constraint with multiple rockburst criteria to reduce the structure distortion of Euclidean space.Construct criterion-constrained graph structure data by integrating the perception and the constraint.Propose a multiscale graph convolutional neural network based on the data, which achieves accurate prediction of rockburst-intensity levels.