CLC number: TP391.4

Few-Shot Recognition of Fiber Optic Vibration Sensing Signals Based on Triplet Loss Learning

基于三元组损失学习的光纤振动感知信号小样本识别

¹ Northeast Sichuan Gas Mine, PetroChina Southwest Oil and Gasfield Company, Dazhou 635000, Sichuan, China
² Department of Quality, Health, Safety and Environmental Protection, PetroChina Southwest Oil and Gasfield Company, Chengdu 610051, Sichuan, China
³ School of Optical Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
⁴ School of Electronic Information Engineering, North China Institute of Science and Technology, Langfang 065201, Hebei, China

^† Corresponding author. E-mail: liudequan@ncist.edu.cn

Received: 15 November 2024

Abstract

The distributed fiber optic sensing system, known for its high sensitivity and wide-ranging measurement capabilities, has been widely used in monitoring underground gas pipelines. It primarily serves to perceive vibration signals induced by external events and to effectively provide early warnings of potential intrusion activities. Due to the complexity and diversity of external intrusion events, traditional deep learning methods can achieve event recognition with an average accuracy exceeding 90%. However, these methods rely on large-scale datasets, leading to significant time and labor costs during the data collection process. Additionally, traditional methods perform poorly when faced with the scarcity of low-frequency event samples, making it challenging to address these rare occurrences. To address this issue, this paper proposes a small-sample learning model based on triplet learning for intrusion event recognition. The model employs a 6-way 20-shot support set configuration and utilizes the KNN clustering algorithm to assess the model's performance. Experimental results indicate that the model achieves an average accuracy of 91.6%, further validating the superior performance of the triplet learning model in classifying external intrusion events. Compared to traditional methods, this approach not only effectively reduces the dependence on large-scale datasets but also better addresses the classification of low-frequency event samples, demonstrating significant application potential.

摘要

分布式光纤感知系统以其高灵敏度和广泛的测量能力在地下燃气管道监测中得到了广泛应用，主要用于感知外部事件引起的振动信号，并有效提供潜在入侵活动的预警。由于外部入侵事件的复杂性和多样性，传统的深度学习方法在事件识别中的平均准确率可以超过90%。然而，这些方法依赖于大规模数据集，导致数据采集过程需要耗费大量时间和人力成本。此外，当面临低频事件样本稀缺时，传统方法表现较差，难以有效应对这些稀少事件。为了解决这一问题，本文提出了一种基于三元组学习的小样本学习模型，用于入侵事件的识别。该模型采用6类20样本的支持集配置，并使用KNN聚类算法评估模型性能。实验结果表明，该模型的平均识别准确率达到91.6%，进一步验证了三元组学习模型在外部入侵事件分类中的优越性能。与传统方法相比，该方法不仅有效减少了对大规模数据集的依赖，还更好地解决了低频事件样本分类问题，显示出显著的应用潜力。