Research On Improved Mine Microseismic Event Recognition Method Integrating Neural Network And Transfer Learning

The mine microseismic monitoring system can collect a variety of vibration signals with a frequency of several hertz to several kilohertz.The information contained in it is more complex.The accurate identification of microseismic events of coal and rock fractures is the most critical scientific issue for the location of microseisms and the understanding of their focal mechanisms.one.The accurate identification of microseismic events determines the timeliness and accuracy of microseismic monitoring and early warning technology.Traditional microseismic event recognition methods mostly require manual feature extraction,and cannot combine the classifier with the feature extraction process.The process is cumbersome and complicated,and most of them are shallow-structured algorithms,which have weak generalization capabilities for classification problems.Deep learning models are widely used in text,image recognition and other fields,including convolutional neural networks and recurrent neural networks.Convolutional neural network is a multi-level neural network structure based on deep learning.It has two major characteristics: local perception and parameter sharing,which greatly reduces the time complexity of the algorithm and enhances the ability to extract features;the recurrent neural network can process time series data and has memory The ability of historical information.Due to the small size of the microseismic data set,it is not enough to support the model training of deep learning.Transfer learning is one of the very effective methods to solve this problem.It uses existing knowledge to deal with different but related domain problems,so as to achieve the knowledge in the relevant field.The purpose of migration between domains.From the perspective of transfer learning,this paper proposes an improved mine microseismic event identification method that combines neural network and transfer learning based on the research of a large number of scholars,aiming at the identification of microseismic events.The main research contents and innovations of the thesis are as follows:(1)In view of the low accuracy of microseismic event recognition in the existing deep learning CNN methods,this paper improves the structure of CNN,adds similar feature layers to the convolutional neural network,and builds a Sim CNN model.Research the extraction of similar feature layers,convert image data into graph structure,use non-recursive Sim Rank algorithm to calculate similarity values,and extract similar feature layers that are similar to it.(2)Aiming at the problem that the small-sample microseismic data set is not enough to support the training of neural network models and the recognition accuracy is not high,research the current mature cross-domain migration learning technology,and transfer the features of large-sample source-domain seismic data sets to small In the model of sample microseismic data,the source domain feature transfer learning method based on improved CNN is implemented.Perform pre-training on the source-domain seismic data set to obtain a pre-trained Sim CNN model,and train parameters to realize parameter feature transfer learning.(3)In order to make the model have a better ability to extract features,this paper studies the time series feature transfer learning method based on LSTM,and transfers the time series features of microseismic time series data.First,use the long and short time window method to select effective wavelengths and perform waveform alignment operations,and use the attention mechanism evaluation method to reset the weights according to the characteristics of the P wave before and after the arrival;second,build and train the LSTM neural network model,and migrate the fully connected layer The parameter features of LSTM realize the transfer of time series features of LSTM,and the complete transfer learning model T-Sim CNN is obtained.(4)Design experiments to verify the effectiveness of the model.For seismic data sets,the experiment compares the effects of changes in the number of similar feature layers on the accuracy and loss function of the model.The experiments compare the accuracy,loss function and event recognition time of LSTM,CNN and Sim CNN models;for microseismic data sets,Experiments compare the accuracy,loss function and event recognition time of LSTM,CNN,Sim CNN and T-Sim CNN models.The experimental results show that the T-Sim CNN model performs well after transfer learning.