Automatic Detection Of Collapse Events Using A Cost-Sensitive Deep Residual Convolutional Neural Network

Huating,Gansu is the largest coal mining area in Gansu Province.In recent years,large-scale coal mining and small earthquake clustering activities have triggered destructive mine earthquakes and collapses,which seriously threaten the safety of people's lives and property in the mining area and surrounding areas.The scope of coal mining induced collapse is relatively concentrated,and there is a big difference between the collapse waveform records and natural earthquake's.It is difficult to complete the accurate identification and classification with traditional methods,and the detection threshold need to be set specifically for the detection of collapse.Therefore,the accuracy of collapse event recognition depends heavily on human experience,and it is urgent to develop software for automatic recognition.Seismic data analysis and earthquake warning depend on the accurate detection and classification of seismic events.With the increasingly intensive seismic monitoring network,a large amount of waveform data is produced.The seismologists have been trying to develop methods to pick up the earthquake automatically and classify it accurately.With the rapid development of computer technology,deep learning has made great progress in natural language processing,target detection and other fields.In recent years,the automatic detection algorithm based on deep learning has also been applied to seismic event detection.The deep convolutional neural network algorithm has the great advantage of easier training and automatic learning of different level features according to feedback.The shallow layer can learn local features,and the deep layer network has a wider receptive field and can learn longer scale feature information.In order to avoid the degradation problem after the depth becomes deeper,the convolutional neural network with depth residual is used in this paper.Compared with natural earthquakes,the number of collapse data sets is relatively small,which will lead to the imbalance problem of data sets and make the training of deep learning model more difficult.Based on this,this paper introduces an automatic collapse detection method "Respicker" based on residual convolutional neural network.At the data level,we enhance the collapse data,and at the algorithm level,we use the cost sensitive loss function to improve the learning of the residual convolutional neural network model for the minority class,which improves the detection ability of the model for collapse events.After the first long training,the Respicker can quickly detect the events from continuous waveform,and can use GPU to accelerate the operation.The results show that the model can effectively and accurately distinguish the small magnitude collapse from the background noise,and detects collapses in the Huating area of Gansu at a higher degree of precision,recall,and F1 score than that found in traditional methods.This provides a new idea for the identification and detection of collapse events in complex environment,and also puts forward a new method for the monitoring and early warning of non natural earthquake events in mining area,which is of great significance to serve the economic and social development of mining area and help improve the safety of coal mining.