Detection and identification of external intrusion signals from 33 km optical fiber sensing system based on deep learning

Abstract In real-world environments, it is usually hard to achieve accurate identification and classification of external vibration signals collected by optical fiber. In this paper, we have applied deep neural networks to a 33 km optical fiber sensing system to recognize and classify the signals of the external intrusion (third-party intrusion) events. It enables the fast identification and localization of the destructive events in complex environments with large amount of monitoring data. Pipeline intrusion events intelligent identification system in this paper is mainly divided into two parts: a distributed acoustic sensing (DAS) System and a pattern recognition system (PRS). DAS was utilized to monitor external intrusion signals in the real-world environment. A Deep learning model, which is called Convolutional, Long Short-Term Memory, Fully Connected Deep Neural Networks (CLDNN), is first applied in PRS to directly input the time series of data into the network for deep learning without any preprocessing, which is simpler and better than the ways used in the previous work. After training and testing with real data, the average recognition rate of the constructed model for intrusion events can reach over 97%. Finally, 33 km blind tests were carried out to verify that the model has good recognition, classification and localization applications for external intrusion signals in the real-world environment.