Supervised and unsupervised general framework for rockburst risk prediction based on feature contrast of long-term and short-term microseismic data

Short-term rockburst risk prediction based on microseismic (MS) data is a significant research task to overcome the rockburst challenge during the excavation stage. By reviewing previous short-term rockburst risk prediction methods based on MS data, this paper discovers three problems that hinder the development of this task, including poor generalization ability of explicit prediction indexes, lack of specially designed algorithms, and waste of unlabeled data. Therefore, based on the typical rockburst risk events, the geological and mining quantitative information models are constructed. The relationship among the temporal-spatial distribution features of MS data, rockburst main controlling factors, and rockburst progress stages are studied. The paper discovers that long-term MS data can estimate the stress propagation path and active main controlling factors, which is significant for estimating the stability state of the coal rock mass around the working face. In contrast, short-term MS data can estimate whether order and dense fracture development in specific areas. The difference between long-term and short-term MS data features may represent an increased rockburst risk. Then, inspired by these above insights, a novel feature extraction encoder integrating the cross-attention mechanism and prompt engineering is designed to automatically extract the implicit rockburst risk prediction (IRP) indexes from the MS data within different terms. Based on the novel encoder, the unsupervised and supervised general framework specifically designed for rockburst risk prediction is presented, dubbed the Long-term and Short-term Feature Contrast method. Abundant experiments on MS datasets and field deployments demonstrate the performance and generality superiority of the proposed method compared with previous methods.