Two-round feature selection combining with LightGBM classifier for disturbance event recognition in phase-sensitive OTDR system

Abstract

To tackle with the high nuisance alarm rate in phase-sensitive optical time-domain reflectometer (φ-OTDR) system, a novel event recognition method is proposed. Taking advantage of strong interpretability and good performance in the application of multi-classification, light gradient boosting machine (LightGBM) classifier is introduced for multiple event recognition. Meanwhile, in order to improve the recognition efficiency, two rounds of feature selection are implemented to filter out redundant features. After the two-round feature selection, the number of features reduces from 40 to 14 and the corresponding recognition time reduces from 1 s to 0.3 s. Five numbers of disturbance events (watering, climbing, knocking, pressing and false disturbance) can be effectively recognized with an average accuracy rate of 97.41%.