A $k$ -Nearest Neighbor Algorithm-Based Near Category Support Vector Machine Method for Event Identification of $\varphi$ -OTDR

Abstract

In order to reduce the nuisance alarm rate (NAR) for the phase-sensitive optical time-domain reflectometer (φ-OTDR), we propose an event identification method based on the near category support vector machines (NC-SVM), which extends the current binary SVM classifier to multiclass problems by using k-nearest neighbor (kNN) algorithm. Five kinds of disturbance events, including watering, climbing, knocking, pressing, and false disturbance event, can be effectively identified for 25.05 km long φ-OTDR system. The experimental results demonstrate that the average identification rate of five disturbance events exceeds 94%, the identification time is 0.55s, and the NAR is 5.62%. Compared with the one against one multiclass SVM classifier, our proposed method has the distinguished advantage of higher identification rate, shorter identification time, and lower NAR.