Comprehensive early warning of rockburst hazards based on unsupervised learning

Abstract

Intelligent early warning of rockburst hazards is critical for ensuring safe and efficient coal mining operations. The utilization of monitoring techniques, such as microseismic (MS), acoustic emission (AE), and electromagnetic radiation (EMR), has become standard practice for monitoring dynamic hazards in mining environments. However, the inherent complexity and unpredictability of the signals generated by these monitoring systems present significant challenges. While the application of deep-learning methods has gained traction in the field of coal-rock dynamic disaster management, their reliance on vast amounts of data and susceptibility to subjective labeling and poor generalization have hindered the achievement of timely, efficient, accurate, and comprehensive warning of rockburst hazards. In response to these challenges, this study applied an unsupervised learning method based on long short-term memory and an autoencoder to identify precursors of rockburst hazards and predict signals. The robustness and universality of the model were evaluated using MS, AE, and EMR data from the mine site. Then, the entropy method was used to comprehensively process the MS, AE, and EMR signals and conduct risk assessment. Finally, impressive results were achieved: the accuracy of precursor recognition reached 99.18% and the fitting rate of signal prediction reached 93%. Through on-site verification, the efficacy of this approach is evidenced by its synchronization with field records, enabling proactive responses to potential rockburst risks. This method is expected to enhance intelligent warning systems and ensure the safety of coal mine activities.