Research On Micro-seismic Localization Method Based On Deep Learning

With the development of unconventional oil and gas,hydraulic fracturing method has been widely used.Microseismic monitoring is one of the important technologies in the hydraulic fracturing process,the performance of hydraulic is affected by the localizing accuracy and efficiency.The algorithm based on traveltime has poor noise immunity and requires precise manual picking of the microseismic travel time,which is not suitable for low signal-to-noise ratio in practical engineering applications.The most widely used algorithm for microseismic source location is the time-reverseimaging.However,this algorithm is still affected by many factors during the operation process,in order to be more widely used,the accuracy and efficiency of the algorithm needs to be further hanced.In microseismic location tasks,deep learning methods typically convert them into a mapping problem from time-series waveform signals to spatial positions.Traditional numerical simulation methods are prone to produce large errors when studying the movement of waveforms,which can affect the establishment of mapping relationships.Limited features make the model more ambiguous,and there are problems such as insufficient feature extraction and poor generalization ability.This paper aims to address these issues by conducting in-depth research on deep learning-based methods to improve the accuracy and efficiency of numerical simulation and microseismic location.The main research content is as follows:(1)Based on the features of microseismic data,we analyze the factors that affect the processing effect of earthquake source location.We introduce Gaussian noise,velocity errors,and other methods to augment the training samples and construct a sample dataset for verifying the experimental effect of the model.(2)In response to the problem that waveforms are easily interfered by dispersion phenomenon during the forward propagation process of the reverse-time migration algorithm,based on the principle of solving the wave equation using the finite difference method,a network model based on the joint learning of space-frequency domains is designed to suppress dispersion artifacts.This network combines the residual learning idea to minimize the joint error and fully extracts the features of pseudo wave textures and energy loss caused by dispersion phenomenon.The experimental results show that the network has good suppression effect on dispersion phenomenon while protecting finite signals well,and achieves highprecision numerical simulation at a low computational cost,laying a solid foundation for improving the accuracy and efficiency of the localization algorithm.(3)In response to the problem that the accuracy of the localization algorithm is limited by multiple factors,a microseismic source localization method based on a U-Net network with an encoder-decoder structure is proposed.The method utilizes a fusion attention mechanism to eliminate the influence of background noise information and irrelevant waveforms on the localization effect,and uses atous spatial pyramid pooling(ASPP)to deeply excavate the high-level semantic features obtained in the encoder.The improved network is used to complete the mapping from microseismic data to the probability distribution of the source location.The experimental results show that the proposed method has high accuracy in localizing the source coordinates,and the model has a certain robustness and generalization ability,which is superior to the reverse-time migration algorithm in terms of performance and efficiency.