基于Swin Transformer特征提取的微地震初至拾取方法

Microseismic first break picking based on Swin Transformer feature extraction

微地震初至拾取是微地震监测中基础且重要的环节,其拾取精度对于后续数据处理起到了至关重要的作用.微地震监测是长期持续的过程,人工拾取效率低下.传统的自动拾取方法依赖算法的相关特征函数,存在无法充分利用数据信息的缺点;而现有基于一维卷积神经网络或长短期记忆网络的微地震初至自动拾取方法只在单通道内进行拾取,并没有考虑相邻通道初至的相关性,从而导致在低信噪比环境下拾取误差较大.针对上述问题,将多道微地震初至前和初至后看作是图像分割问题,应用基于自注意力机制的Swin Transformer作为主干网络提取微地震初至特征,把多道微地震数据沿着道内时间维度展开成一维序列输入至网络中,在不打乱数据时序信息的同时兼顾相邻通道初至的相关性,并以此设计出端到端微地震初至多道拾取网络,实际矿山微地震监测数据中的实验结果表明,该网络在初至拾取精度上相较于传统单道自动拾取法以及现有神经网络单道拾取法有大幅提升,在低信噪比环境下依然可以取得较高的拾取精度,具有良好的工程应用价值.

Microseismic first break picking is a fundamental and important part of microseismic monitoring,and its picking accuracy plays a crucial role in the subsequent data processing.Microseismic monitoring is a long-term continuous process,manual pickup is time-consuming and inefficient.The traditional automatic picking methods rely on the correlation feature function of the algorithm,which has the disadvantage of not fully utilizing the data information.The existing microseismic first breaks based on one-dimensional Convolutional Neural Networks(CNN)or Long and Short-term Memory Networks(LSTM)only pick up within a single channel,do not consider the correlation of first breaks of adjacent channels,which leads to a large picking error in low signal-to-noise environments.In order to solve the above problem,we consider the pre-and post-microseismic first breaks as a binary image segmentation problem,and apply the Swin Transformer based on the self-attention mechanism as the backbone network to extract the microseismic first breaks features.Expanding the multi-channel microseismic data into a one-dimensional sequence along the intra-channel time dimension into the network,the correlation of the first breaks of adjacent channels is taken into account without disrupting the data timing information.,calculate the correlation between any element of the sequence and the rest of the elements using the self-attention mechanism to obtain microseismic first break feature information,to design an end-to-end microseismic first break multi-channel picking network.Experimentation in actual mine microseismic monitoring data,the results show that this network has a significant improvement in the first break picking accuracy compared with the traditional single channel automatic picking method and existing single channel neural network method,it can still achieve high accuracy first break picking results in a low signal-to-noise environment,with excellent practical engineering applications value.