A Surveillance System for Urban Utility Tunnel Subject to Third-Party Threats Based on Fiber-Optic DAS and FPN-BiLSTM Network

• Published in: IEEE transactions on instrumentation and measurement Vol. 73; pp. 1 - 9
• Main Authors: He, Tao, Li, Hao, Zhang, Shixiong, Zeng, Zhichao, Yan, Zhijun, Sun, Qizhen, Liu, Deming
• Format: Journal Article
• Language: English
• Published: New York IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Acoustics; Bidirectional long short-term memory (Bi-LSTM); distributed acoustic sensing (DAS); Feature extraction; Feature maps; feature pyramid network (FPN); Fiber optics; Field tests; Geological mapping; Monitoring; Nodes; Optical fiber networks; Optical fiber sensors; pattern recognition; spatiotemporal–spectral (STS) feature; Surveillance systems; Target recognition; Three dimensional models; Time-frequency analysis; Underground structures
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3369087

Inevitably influenced by the complicated underground geological structure in practical applications, the received signal response of the same disturbance event is inconsistent at different sensor nodes, which is an enormous challenge in large-area safety monitoring applications based on the distributed acoustic sensing (DAS) technology. Thus, in this article, the combination of the feature pyramid network and a bidirectional long short-term memory (FPN-BiLSTM) network is first introduced to perform an accuracy and efficiency identification of third-party threats under the complicated underground geological structure. First, the comprehensive spatiotemporal-spectral (STS) 3-D feature map of the signal target is formed from some adjacent sensor nodes. In order to alleviate the high computational burden in the FPN model, the 3-D feature map is segmented into <inline-formula> <tex-math notation="LaTeX">{n} </tex-math></inline-formula> time sequences. Then, the sequences of data are sequentially transmitted to the FPN model for feature extraction. Subsequently, a Bi-LSTM network is applied to further extract the tandem contextual information among the sequences of time-frequency spatial feature vectors obtained by the FPN model. After that, the comprehensive STS multidimension feature vectors are extracted by the FPN-BiLSTM network for the identification of the target events. Finally, the field test results prove that the proposed method can achieve a high recognition accuracy rate of 93.96% for five typical events with a fast response time of 0.463 s, which is superior to the traditional network models.