An auto-detection network to provide an automated real-time early warning of rock engineering hazards using microseismic monitoring. (April 2021)

• Record Type: Journal Article
• Authors: Wang, J.X.; Tang, S.B.; Heap, M.J.; Tang, C.A.; Tang, L.X.
• Is Part Of: International journal of rock mechanics and mining sciences. Volume 140(2021)
• Publication Date: 2021-04
• Subjects: Rock engineering -- Microseismic monitoring -- Deep learning -- Self-attention mechanism -- Long signal processing -- Real-time early warning; Rock mechanics -- Periodicals; Soil mechanics -- Periodicals; Mining engineering -- Periodicals; Roches, Mécanique des -- Périodiques; Sols, Mécanique des -- Périodiques; Technique minière -- Périodiques; 624.151305
• DOI: 10.1016/j.ijrmms.2021.104685
• Languages: English
• ISSNs: 1365-1609

Abstract

Abstract: The hazards associated with the extraction of ores and the construction of underground geoengineering structures should be mitigated to prevent the loss of human life. Microseismic monitoring (MM) has been used extensively in underground engineering around the world to assess the condition of underground structures and to improve personnel safety. However, MM cannot be currently used as a real-time early warning system because arrival detection still requires manual processing. Although studies have forwarded automated algorithms to perform this task, they are far less accurate than a human expert and therefore provide inaccurate source locations. In this study, we propose a workflow to specifically obtain the state-of-the-art performance of arrival detection. First, to avoid severe class-imbalance problems and speed up convergence in training, a new label form is proposed. We then propose an innovative self-attention mechanism in deep learning to efficiently model distant correlation ("segmentation-integration self-attention"). This new self-attention mechanism can effectively encode the distant inner dependencies by processing signals in several segments and by building correlations between the corresponding descriptors for the segments. Additionally, a new neural network architecture ("extractor-encoder-generator") for automated arrival picking is proposed based on the above self-attention. This architecture extracts functional features of the signals, models correlations by an attention-related sublayer and generates a score map defining the confidence of the first break. The combination of the above technology is referred to as the "auto-detection" network. The new architecture outperforms all other algorithms by a considerable margin, provides a similar accuracy to that of a human expert, and is well balanced in terms of inference speed and accuracy. The auto-detection network allows for a practical and automated MM technique that can be used to provide an accurate real-time early warning system.