| Microseismic monitoring technology is an important method to monitor the hydraulic fracturing process and evaluate the fracturing effect.Microseismic monitoring methods can be divided into downhole monitoring and near-surface monitoring according to the location of receiver.Most microseismic events caused by shear faulting will produce different positive and negative polarity characteristics in the microseismic signal received by the geophone.For the surface monitoring system,the P-wave polarity can directly and quickly invert the source mechanism solution,and the polarity correction can improve the imaging accuracy of diffraction stacking.Therefore,the determination of P wave polarity of microseismic events is significance for surface microseismic monitoring.Convolutional neural network has been widely used in image classification and so on.In recent years,convolutional neural network has also been used in P-wave polarity classification.Different from the traditional polarity classification methods,CNN can extract the characteristic information of original waveform for polarity classification,and does not need to pick up the P-wave arrival times.In this research,the cross-correlation method and two neural network methods are compared to classify the initial polarity of microseismic P waves.In view of the surface monitoring,regular observation system is often used,and the records of the target track and its adjacent receivers are used as input samples to construct a multi-trace P wave polarity classification model based on convolutional neural network.The main work of this research is as follows:(1)Preprocessing of coalbed methane fracturing data in Heshun working area,Shanxi Province,and obtaining clear and identifiable microseismic events by means removal,bandpass filtering and elevation correction.Firstly,the P wave was determined by the energy ratio method of the short time and long time Windows,and then the waveform was truncated by selecting the appropriate time window to preserve the complete P wave information.Finally,the truncated waveform data are manually marked and data augmented to generate data set.(2)Cross-correlation,convolutional neural network,and residual network are used to classify the P-wave polarity of single-trace microseismic data.Compared with the three methods,the training speed of neural network algorithm is slower,but the classification accuracy is significantly higher than that of cross-correlation method.(3)Construct multi-trace microseismic data set to train the convolutional neural network classification model.The accuracy of the model is highest when the number of detectors in different adjacent traces is determined to be 5.Compared with the convolutional neural network regression model based on multi-trace data,the accuracy of the two models is similar but the convergence speed of the classification model is slightly faster.(4)The classification model based on single trace data and the classification model based on multi-trace data are used to predict the polarity of P wave for 34 microseismic events respectively.The polarities of P wave are corrected according to the predicted results and the diffraction stacking algorithm is applied to locate the event locations.By calculating the source mechanism solution and source imaging results,it is proved that the accuracy of the training model with multi-trace data as input is better than that of the single trace data model. |
