Multi-Class Disturbance Events Recognition Based on EMD and XGBoost in φ-OTDR

Abstract

A novel pattern recognition method based on Empirical Mode Decomposition (EMD) and extreme gradient boosting (XGBoost) is proposed to recognize the disturbance events in phase sensitive optical time-domain reflectometer ( $\varphi $ -OTDR) to reduce nuisance alarm rate (NAR) and improve real-time performance in this paper. Eleven typical eigenvectors are extracted from components obtained by EMD of the disturbance signals and XGBoost is selected as a classifier to identify different type of disturbance signals. Five kinds of disturbance events, including watering, knocking, climbing, pressing and false disturbance event, can be identified, effectively. Experimental results show that NAR is 4.10% and identification time is 0.093 s. The recognition accuracy for the five patterns is 97.96%, 95.90%, 91.10%, 94.84% and 99.69%, respectively. The effectiveness of the proposed method is evaluated by using confusion matrix and decision boundary visualization. Experimental results demonstrate that our proposed pattern recognition method based on XGBoost has better performance in recognition rate and recognition time than other commonly used methods, such as support vector machine (SVM), Gradient Boosting Decision Tree (GBDT), Random Forest (RF) and Adaptive Boosting (Adaboost).