Prediction of rockburst occurrence time: Implications from microseismic monitoring and deep learning

Predicting the occurrence time of rockbursts is a core challenge for the safe construction of deeply buried tunnels, the complexity of which stems from the combined effects of high ground stress, complex geological conditions, and construction disturbances. To address this challenge, this study employed microseismic (MS) monitoring to collect vibration waveforms during rockburst incubation. This study analyzed the relationship between periodic drilling and blasting (D&B) method construction activities and monitoring waveforms preceding rockburst events and proposed a rockburst occurrence time prediction model integrating a convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM). This model extracts the image features of vibration waveforms and the temporal patterns of maximum amplitudes during rockburst incubation separately. These dual-modality spatiotemporal features are dynamically fused using a cross-attention mechanism, the rockburst occurrence probability is calculated on the basis of historical rockburst cases and the sigmoid function, and the rockburst time is extrapolated on the basis of the probability growth rate. The test results demonstrated that the model achieved high precision (R 2 =0.95) in predicting rockburst occurrence times, with prediction errors reduced by more than 10% compared with those of the traditional long short-term memory (LSTM) and BiLSTM models. During engineering application validation, the model successfully predicted two rockburst events occurring 14 h and 9 h later while ensuring prediction errors under 3 h, achieving a minimum prediction error of 0.3 h. The application results demonstrated that the integrated method of MS monitoring and deep learning proposed in this study could directly predict the rockburst occurrence time in D&B method tunnel construction, providing an effective new approach for predicting the occurrence time of rockbursts.