A prototype-based rockburst types and risk prediction algorithm considering intra-class variance and inter-class distance of microseismic data

The prediction and classification of rockburst risk based on microseismic data is the premise of preventing rockbursts during deep mine excavation. By reviewing previous studies, this paper finds two problems that hinder the rockburst prediction: 1) there is a lack of research on the distribution features of monitoring data on the main controlling factors of rockbursts; 2) there is no research on the intra-class variance and inter-class gap of microseismic data. Based on the typical rockburst risk events, a quantitative information model of geology and mining is constructed. The relationship between the spatial–temporal distribution characteristics of microseismic data before a rockburst and the main controlling factors of a rockburst is studied. The results show that the distribution features may be different for the same type of microseismic (MS) and rockburst events, and different types of events may show similar distribution features. Therefore, based on the quantitative study of the relationship between the performance of a deep learning prediction algorithm and a rockburst prediction vector, a rockburst risk and type prediction algorithm based on a convolutional neural network (CNN)-gated recurrent unit (GRU) model with prototype-based prediction is proposed. The CNN-GRU model can produce prediction vectors by fusing implicit and explicit information extracted from the original MS data and early warning indicators. Cross-entropy loss, vector-prototype contrastive loss, and vector-prototype contrastive loss are proposed to automatically control the intra-class variance and inter-class gap of prediction vectors belonging to different rockburst risks and types. Many experiments show that the performance of the proposed CNN-GRU model with prototype-based prediction is superior to other algorithms in the prediction of rockburst risks and types based on MS data.