Study On Real-time Seismic Monitoring And Early Warning Based On Fiber Distrbuted Acoustic Sensing Technology

Earthquake monitoring is one of the methods of earthquake early warning.At present,the most monitoring means of seismic stations in our country use seismometers,which are sensitive to the earthquake situation.The most common seismometers are active devices,which are easy to be affected by electromagnetic interference.In recent years,the Distributed Acoustic Sensing(DAS)uses passive common single-mode fiber as the sensing element,which has the advantages of antielectromagnetic interference,anti-corrosion,easy laying,etc.,and has realized the application of DAS in vibration monitoring and recognition.There are many challenges in earthquake early warning based on DAS,among which the most important is that the seismic wave transmission speed is fast(3～7 km/s),and an effective earthquake early warning needs to identify the earthquake within a short period of time(less than 10 seconds).At present,DAS combined with deep learning can recognize signals,but there is a contradiction between recognition speed and accuracy.Therefore,there is a contradiction between rapid recognition and recognition accuracy of earthquake signals in the method combined with DAS and deep learning.In this paper,the contradiction between rapid identification and accuracy of earthquake signals in earthquake monitoring and early warning based on DAS technology is studied.Analyze the spectral characteristics of earthquake signals,visualize one-dimensional signals,The Network structure of DSR-Net(DAS Seismic Recognition Network),built by Convolutional Neural Network(CNN),is used to identify the earthquake signals.The recognition of 5s effective earthquake signals and 20 s effective earthquake signals with more than 90% accuracy is realized,which provides an early warning method for DAS real-time earthquake monitoring and warning,and DAS makes a contribution to the earthquake monitoring.The main research content of this paper includes the following parts:(1)Through the background of earthquake early warning and the advantages of optical fiber sensing in earthquake early warning,the research background and significance of this paper are introduced.The development process of distributed optical fiber sensing technology and its development status in recent years are introduced.Finally,the development status of DAS in seismic monitoring is introduced.(2)The scattering in optical fiber,sensing principle and system structure of DAS system are introduced.The method and principle of signal processing are introduced,including quadrature demodulation of phase signal,Smooth filtering,Short-Time Fourier Transform(STFT),one-dimensional signal visualization and the principle of CNN.(3)The preprocessing process of DAS monitoring signal is introduced,including quadrature demodulation of phase signal,phase difference decomposition winding of phase signal,Smooth detrending term and low pass filtering.Fourier transform and STFT are used to analyze the spectrum of the preprocessed signal.The process of converting two-dimensional images by Gramian Angular Field(GAF)and Markov Transition Field(MTF)is introduced.Finally,the architecture of DSR-Net neural network and two classical image classification networks,VGG-16 and ADE-NET,are introduced.(4)The laying of experimental environment,the laying of optical fiber and the location where 6 natural earthquakes were detected are introduced.The correlation between DAS monitoring earthquake signals and official seismic station signals is analyzed.Three onedimensional signal imaging methods are used to transform DAS one-dimensional signals and analyze the image characteristics.Three methods are used to build the basic data set to lay the foundation for the training network.Then,the basic data set is used to train and test DSR-Net network,and the test accuracy of STFT,GAF and MTF is 97.51%,77.22% and 83.62%,respectively,which proves that STFT image has greater advantages in seismic signal recognition.(5)The method of constructing STFT data sets containing four earthquake signals at different time scales is introduced,and DSR-Net,VGG-16 and ADE-NET neural networks are used to train the four data sets,and the three trained models are evaluated and tested.The confusion matrix,test accuracy,accuracy,recall rate and F1 values were obtained to evaluate the performance of the model.The time spent by DSR-Net network to identify a single sample is about 1/20 of that of VGG-16 networks.Finally,under the training of the data set of 5s effective earthquake signals,the test accuracy of DSR-Net is 87.37%,and the F1 value is80.57%.With the increase of earthquake duration,the accuracy and F1 value are improved significantly.In the data set test of 20 s effective earthquake signals,the test accuracy reaches94.22%,and the F1 value reaches 91.96%.