A Dual-Stage-Recognition Network for Distributed Optical Fiber Sensing Perimeter Security System

• Published in: Journal of lightwave technology 2023-07, Vol.41 (13), p.1-9
• Main Authors: He, Tao, Sun, Qizhen, Zhang, Shixiong, Li, Hao, Yan, Baoqiang, Fan, Cunzheng, Yan, Zhijun, Liu, Deming
• Format: Article
• Language: English
• Subjects: Acoustics; Back propagation networks; Clinics; Decision trees; Distributed optical fiber acoustic sensor (DAS); dual-stage-recognition network; Energy distribution; False alarms; Field tests; Frequency spectrum; human-animal activities discrimination; intrusion detection; Intrusion detection systems; Optical fiber sensors; Optical fibers; Optical network units; Pattern recognition; Security; Security systems; Sensors; Time-frequency analysis
• ISSN: 0733-8724; 1558-2213
• DOI: 10.1109/JLT.2022.3222472

Accurate intrusion recognition along the optical fiber is still an enormous challenge in the distributed acoustic sensing (DAS) based security system. Especially in the complex environments, various unknown disturbs such as the animal activities will lead to high false alarm rate of intrusion detection system. In this work, an accurate and effective intrusion pattern recognition using a dual-stage-recognition network is proposed and demonstrated for practical environments with various animal activities and mechanical movements. The dual-stage-recognition network consists of the pre-recognition stage for shallow classification and the sub-recognition stage for discriminating the similar events. In the pre-recognition stage, three target events of non-intrusion, human-animal activities and mechanical movements can be classified by the decision tree classifier based on the temporal energy and the frequency spectrum information. After that, in the sub-recognition stage, the target events of human and various animal activities can be further distinguished by the combination of the time-frequency analysis and BP neural network. Besides, in order to improve the computation efficiency of BP network model, the characteristics information of the time-frequency energy distribution is efficiently compressed by the proportion statistics of four energy-levels. The field test of a month proves that the proposed method can realize a high average recognition accuracy rate of 97.6% for five typical events with a fast average response time of 0.253 s, which is very promising in the intrusion events recognition in practical environments.