Research On Microseismic Signal Identification And Source Localization Methods Based On Deep Learning Framework

In order to satisfy national energy security requirements,the exploration and development of unconventional oil and gas resources such as shale oil and gas,have received more and more attention.Hydraulic fracturing is an important tool for reservoir modification in dense sandstones and shales.Accompanied by rock rupture,a large number of microseismic events are generated.Microseismic monitoring technique monitors fracture development in real time by observing and analysing the microseismic signals generated by hydraulic fracturing,and then evaluates the effect of fracturing.This technique is the most effective fracture monitoring technique available and plays an important role in improving unconventional oil and gas production.Due to the long duration of microseismic monitoring,data segments containing microseismic events first need to be detected from a large number of continuous records,so rapid identification of microseismic events and picking up their first arrivals is the basis of microseismic monitoring data processing.Microseismic source localization is the core of microseismic monitoring.By solving the parameters such as the source location of microseismic events,we can get the information such as the length,width,and orientation of the fractures.As the degree of development of unconventional oil and gas resources continues to deepen,higher demands are placed on microseismic monitoring technique.Microseismic identification and source localization methods using deep learning have the advantages of high efficiency and high accuracy,which have attracted extensive attention from researchers.However,these methods usually belong to supervised learning and rely on a large number of microseismic first-arrival labels or source location labels.There are often large amounts of unlabeled microseismic data in practical applications,and labeled data are difficult to acquire.Network training for supervised learning usually relies excessively on labeled data for guidance and is prone to overfitting.If there are errors in the labels of the microseismic data,the network may not learn properly.In addition,in terms of using deep learning to locate the source,the direct regression of the source coordinates using fully connected layers lacks spatial generalization ability,while the use of convolutional layers to generate heat maps restricts the location of the source to the grid point,which is not sufficiently accurate.Moreover,the low signal-to-noise characteristics of microseismic events have an important impact on the performance of deep learning models,limiting their application in real production.In conclusion,although conventional supervised learning methods have good performance,their drawbacks limit the scope of application.Therefore,for different microseismic identification as well as source localization tasks,appropriate deep learning algorithms need to be selected to cope with various challenges.In this paper,four methods and strategies are proposed to solve the above problems in the application of deep learning methods for microseismic identification and source localization.The main methods proposed and the main research results achieved in this paper can be summarized as follows:(1)Aiming at the low signal-to-noise characteristics of microseismic data and the fact that conventional deep learning-based microseismic identification and first-arrival pickup methods rely on a large number of microseismic first-arrival labels,a microseismic identification and first-arrival pickup first-arrival pickup method based on joint deep clustering is proposed.The residual shrinkage dense network(RSDN)is first constructed.The performance of the network in low signal-to-noise ratio microseismic data is further enhanced by adding densely connected hybrid null convolution and improving the threshold learning module.The identification of microseismic events and the picking up of first-arrival times are based on the differences between microseismic active signals and noise in terms of amplitude,frequency,and statistical characteristics,etc.The excellent performance of RSDN that can extract features from noisy data is utilized to improve the adaptability to low signal-to-noise microseismic data while automatically extracting features of microseismic events.Based on the RSDN,a joint deep clustering method for microseismic identification and first-arrival time pickup is proposed.Firstly,the monitoring data are automatically divided into two categories according to the presence or absence of microseismic events by simultaneous deep clustering to achieve the purpose of recognizing microseismic events.The identified microseismic events are then clustered a second time,using multi-stage deep clustering to classify the sampling points of a single-channel microseismic record into two categories:valid signal and noise,which in turn automatically picks up the first-arrival time of the microseismic event.The whole process does not require manual production of microseismic first arrivals labels,which substantially reduces the cost.Finally,numerical examples and field data tests demonstrate that the microseismic identification and first-arrival time pickup method based on joint deep clustering not only has higher accuracy,but also has certain generalization.The network trained with synthetic data also has good performance in real data applications.(2)To address the lack of sufficient source locations as training labels in supervised learning-based microseismic localization,a microseismic localization method based on the improved semi-supervised generative adversarial network(ISGAN)is proposed.ISGAN consists of a generator,a discriminator and a regressor.It utilizes RSDN to further enhance the stability of network training within the framework of gradient-penalized Wasserstein generative adversarial networks.By increasing the generative capacity of the generator,the discriminator is forced to improve its discriminative capacity,and then the discriminator shares shallow features with the regressor in order to improve the prediction accuracy on the source location by the regressor.The use of massive unlabeled microseismic records with a few labeled microseismic records to train ISGAN can improve the source localization accuracy of the network under the limited number of labels,effectively solving the problem of insufficient source location labels in microseismic localization.Numerical experiments and actual hydraulic fracturing data tests demonstrate that the ISGAN-based microseismic localization method is able to obtain fairly high-precision source localization results even in the absence of sufficient source location labels.(3)To address the fact that microseismic localization by supervised classification or regression requires the guidance of microseismic source location labels during the training process,a microseismic localization method based on deep reinforcement learning is proposed.We propose an optimization strategy for microseismic signal processing and source localization across deep learning paradigms from unsupervised feature learning to reinforced source localization.The problem of microseismic source localization is modeled as a Markov decision process,which provides an important foundation for applying deep reinforcement learning algorithms to solve for the source location.The raw microseismic recordings are preprocessed using an autoencoder constructed by RSDN,and the extracted microseismic features are used as environment states to reduce training difficulty and improve noise immunity.In addition,a task decomposition strategy is proposed to reduce the complexity of the microseismic localization task,and the source localization strategy is learned independently by the three intelligences in the environment interaction.Based on the weak feedback mechanism of reward,the intelligences are able to discover the deep differences between different microseismic events in environmental interactions,which improves the generalization and versatility of the localization model while obtaining high-precision microseismic localization results.(4)Aiming at the problems of lack of generalization in the direct prediction of source coordinates based on fully connected layers and loss of accuracy in the generation of heatmap based on fully convolutional networks,a microseismic source localization method based on Transformer coupled cascaded pyramid network(TCPN)is proposed.First,a method to represent the location of the source by combining Gaussian heatmap and offset map is proposed,which is used as a training label for microseismic events,and the source localization is converted into image-to-image pixel-level prediction,which improves the spatial generalization ability while avoiding the source location being restricted to grid points.Moreover,with the assistance of offset maps,the method is able to locate the microseismic sources with considerable accuracy without the need for a fine grid,which helps to reduce the network parameters and hence the computational effort.TCPN utilizes Transformer-coupled RSDN as an encoder to extract global and local features,and then utilizes a cascade pyramid network based on RSDN as a decoder to predict the Gaussian heatmap with offset maps.TCPN improves the accuracy of earthquake source localization by fusing the global and local features extracted by the coupled network as well as the multi-scale features of different layers.In addition,the method does not require a predetermined maximum number of sources to define the network architecture and has the potential to localize multiple sources simultaneously.In this paper,based on RSDN,the advantages of different deep learning paradigms are reasonably utilized to overcome the limitations of conventional deep learning methods for microseismic identification and source localization under different conditions.Numerical experiments and hydraulic fracturing microseismic tests show that the microseismic identification and source localization method under the deep learning framework proposed in this paper can effectively identify microseismic signals and locate the microseismic source.This will lay a solid foundation for the later work of determining the distribution range of artificial fractures,dynamic change process and rock damage mechanism.