Research On The Disturbance Events Classification Of Φ-OTDR Based On Modified SVM Alogrithm

Distributed optical fiber sensing system can realize long-distance continuous detection of disturbances along the optical fiber.Among them,phase-sensitive optical time domain reflectometers(Φ-OTDR)has been widely used in safety monitoring fields such as oil pipelines,railway lines,and high-voltage line tower icing.In this thesis,the classification and recognition of disturbance events in the Φ-OTDR has carried out theoretical and experimental research,and found that it is more practical to learn intrusion rules from small samples.Therefore,this thesis chooses support vector machine(SVM)is used as a basic classifier to identify intrusion signals in optical fibers,and the following improvement schemes are proposed for the insufficiency of SVM:(1)Aiming at the problem that the number of samples of the intrusion signals in optical fiber is small and the dimension is high.In order to achieve real-time and accurate disturbance warning,an SVM classification algorithm based on feature selection is proposed.Firstly,extracting features on the sample data before and after the signal differential,including 32 features such as root mean square,energy,shape factor,and impulse factor.Then,use the three feature selection algorithms of Relief F,Fisher,and Laplace to evaluate the input features and determine the features subset with high discrimination.Compared with the traditional SVM solution,the accuracy is improved by 2.29%,and the false alarm rate is reduced by 0.95%.That is,this solution can effectively reduce the data dimension and learning training time,and improve the information integration and system recognition accuracy.(2)When dealing with specific intrusion events,the false alarm rate is high.In the phase of test classification,the K-Nearest Neighbors(KNN),AdaBoost.M1-KNN are combined with SVM algorithm.Firstly,the data after feature selection is taken as a new dataset.On this basis,four parameters optimization algorithms in SVM are compared by experiments,and the best classification parameters are obtained through grid search method.Then KNN-SVM algorithm is used for event recognition.This method uses KNN to select the first two kinds of sample labels with the largest number of K nearest neighbors around the sample to be tested,and uses SVM binary classifier to determine the final category of the sample to be tested.Compared with traditional SVM,the false alarm rate is reduced by 4.83%,and the F1 score is increased by 2.36%.Finally,on this basis,AdaBoost.M1-KNN-SVM algorithm is proposed to improve the sample misclassification near the hyperplane of SVM classification.Compared with KNN-SVM algorithm,the false alarm rate is further reduced by 1.13%.It is shown on the above experiments that the algorithm based on Laplace feature selection and AdaBoost.M1-KNN proposed in this thesis can effectively improve the traditional SVM classifier.The algorithm has a classification accuracy rate of 95.43%and a false alarm rate of 7.33%when processing the real disturbance signals in the optical fiber system.It can make a more accurate classification and early warning of disturbance events in the Φ-OTDR.