Pattern Recognition And Frequency Measurement Of ?-OTDR Based On Convolutional Neural Network

Machine learning has been applied to the pattern recognition of fiber-optical distributed vibration sensor,which effectively improves the vibration event recognition ability of fiber-optical distributed vibration sensor in perimeter security,oil and gas pipeline monitoring,structural health monitoring,rail transit internet of things,traffic flow monitoring,seismic wave monitoring and other applications.However,in engineering applications,the efficiency of existing vibration signal feature extraction methods and the generalization ability of pattern recognition algorithms still have limitations.The training of existing pattern recognition algorithms still needs a large number of experimental samples and labels.On the other hand,the current fiber-optical distributed vibration sensor has the bottleneck of complex method and high device cost in vibration frequency measurement.Transforming the time signal into a two-dimensional image can not only retain the time dependence of the signal,but also excavate the spatial correlation of the signal and the time correlation of different time intervals,which can significantly amplify the information characteristics.The vibration event pattern recognition based on twodimensional image can achieve higher anti-interference characteristics by virtue of the technical advantages of deep feature extraction.However,the existing two-dimensional image pattern recognition technology applied to fiber-optical distributed vibration sensors still has the problems of noise reduction preprocessing of the original signal,complex two-dimensional image construction and long processing time.Aiming at the problem of pattern recognition of fiber-optical distributed vibration sensor based on Phase-Sensitive Optical Time Domain Reflectometer(?-OTDR),the one-dimensional time domain signal is converted into two-dimensional image,and the vibration event pattern recognition is realized based on Convolutional Neural Networks(CNN),which effectively reduces the dependence on the number of samples and markers in the training stage,eliminates the dependence on data preprocessing,improves the accuracy of pattern recognition and reduces the recognition time.At the same time,the vibration of different vibration frequencies is regarded as different classifications,and the vibration frequency measurement problem is transformed into the vibration mode identification problem.The vibration frequency measurement based on full software method is realized,which effectively reduces the complexity and hardware cost of vibration frequency measurement.The main research work of this paper includes the following four parts:(1)Three methods based on Gramian Angular Field(GAF),Recurrent Plot(RP)and Markov Transition Fields(MTF)are studied to convert the time waveform of ?-OTDR fiber-optical distributed vibration sensor signals into two-dimensional images.Different two-dimensional images are taken as samples to construct the sample space required by supervised learning.Different vibration events are used as markers to construct the marker space.(2)The pattern recognition method of vibration events based on CNN algorithm is studied.CNN suitable for two-dimensional image sample space is built.Vibration event pattern recognition and vibration frequency measurement are realized by supervised learning classification task.(3)The vibration event identification method based on CNN and image processing was studied.The training set and test set are constructed based on the experimental data,and the algorithm training is completed.The performance of the classification model is measured based on the performance metrics such as training accuracy,test accuracy,confusion matrix,Receiver Operating Characteristic(ROC)and Area Under ROC Curve(AUC).The effective identification of five vibration events(blowing,raining,direct knocking,indirect knocking,pseudo vibration)is realized.The optimal classification accuracy reached 98.48%,96.63%,100.00%,95.00% and 100.00%,respectively,and the average recognition time was 0.3128 s.The research results verify the effectiveness of data processing methods and machine learning algorithms.(4)The method of vibration frequency measurement based on supervised learning classification task is studied.The vibration frequency is measured by classification task,and the optimal average accuracy of frequency measurement is 99.38%.