Search

The rockburst risk prediction based on microseismic (MS) data is an important research task in deep mine safety prevention. However, the lack of systematic research on explicit prediction indexes and the waste of a large amount of unlabeled data are still two main problems that hinder the development of rockburst prediction. In this paper, the acoustic emission (AE) event distribution at each coal rock deformation and failure stage is studied based on the laboratory experiment. The spatial-temporal evolution of rockburst in MS data of coal mine fields is explored. Based on systematic research of the AE and MS distribution features considering the physical logic of coal rock mass failure, nine different rockburst prediction indexes are employed to describe the MS data features before rockburst. Then, according to the rockburst prediction indexes, a new self-supervision rockburst risk prediction algorithm is constructed, consisting of the pre-trained model and fine-tuning model with the same encoder and decoder structure. The pre-trained model is trained with unlabeled MS data to automatically learn rockburst prediction index features by reconstructing the masked indexes. Based on the pre-trained encoder and decoder parameters, the fine-tuning model is trained with the labeled MS data to predict rockburst risk. A large number of experiments show that the proposed rockburst prediction self-supervision algorithm is far superior to previous algorithms, by effectively utilizing unlabeled data. The ablation experiment also proves the validity of the studied rockburst prediction indexes. [ABSTRACT FROM AUTHOR]

The rockburst risk prediction based on microseismic (MS) data is an important research task in deep mine safety prevention. However, the lack of systematic research on explicit prediction indexes and the waste of a large amount of unlabeled data are still two main problems that hinder the development of rockburst prediction. In this paper, the acoustic emission (AE) event distribution at each coal rock deformation and failure stage is studied based on the laboratory experiment. The spatial-temporal evolution of rockburst in MS data of coal mine fields is explored. Based on systematic research of the AE and MS distribution features considering the physical logic of coal rock mass failure, nine different rockburst prediction indexes are employed to describe the MS data features before rockburst. Then, according to the rockburst prediction indexes, a new self-supervision rockburst risk prediction algorithm is constructed, consisting of the pre-trained model and fine-tuning model with the same encoder and decoder structure. The pre-trained model is trained with unlabeled MS data to automatically learn rockburst prediction index features by reconstructing the masked indexes. Based on the pre-trained encoder and decoder parameters, the fine-tuning model is trained with the labeled MS data to predict rockburst risk. A large number of experiments show that the proposed rockburst prediction self-supervision algorithm is far superior to previous algorithms, by effectively utilizing unlabeled data. The ablation experiment also proves the validity of the studied rockburst prediction indexes.