Automated Data Processing For Microseismic Monitoring Based On Deep Learning

The hazards such as rockburst associated with underground construction severely threat the safety of personnel,equipment and underground structure.Microseismic monitoring has been used extensively in underground engineering around the world to assess the condition of underground structures and to improve personnel safety.However,Microseismic monitoring cannot be currently used as a real-time early warning system because the data processing is still dependent on human operations.Although studies have forwarded automated algorithms to perform these tasks,they are far less accurate than a human expert to perform the same tasks.This study mainly achieves automated microseismic data processing,including signal recognition,arrival detection and model transferability.For each problem,the study analyzes and compares the shortcomings of previous methods,and finally determines to complete these three steps with deep learning and data mining.(1)For signal recognition,a ‘lightweight’ depthwise spatial and channel attention mechanism is proposed and combined with the one-dimensional residual blocks.When dealing with multi-channel microseismic signals,this mechanism can refine the information flow in a network and improve the representation ability with limited parameter increases.Based on this,a signal channel and a multi-channel model are constructed to perform signal recognition in different cases.(2)For arrival detection,to avoid severe class-imbalance problems and speed up convergence in training,a soft label form of gaussian score map is proposed,which can guide the optimization direction for the model training and visualize the “learning process”.We then propose an innovative self-attention mechanism to efficiently model distant correlation in ultra-long signals(“segmentation-integration self-attention”).Additionally,a new neural network architecture(“extractor-encoder-generator”)for automated arrival picking is proposed based on the above self-attention.Additionally,a new network architecture(“extractor-encoder-generator”)for automated arrival picking is proposed based on the above self-attention and the characters of microseismic signals.Comparative results show that the proposed method outperforms all counterparts and is greatly suitable for modelling ultra-long signals.(3)For model transferring,this study points out why methods widely used in computer vision are not suitable for microseismic signal recognition.Then,a two-level adversarial discriminative adaptation method is proposed,which can perform transfer learning with varying difficulties and do not need hyperparameter selection.The study also proposes the“performance sacrifice” theory to explain the algorithms.The experimental results illustrate that it delivers better performance than other existing methods.Additionally,to improve the model performance with less label cost after unsupervised transfer learning,an active learning mechanism is implemented.The results show that models can be simply deployed at the start of new projects and be well fine-tuned with the progressive excavation by the proposed two techniques.In this paper,some key technologies of intelligent microseismic monitoring are proposed,and the highly accurate and automated microseismic data processing is achieved,which lays a foundation for real-time early warning system.