CLC number: TP212

Distributed Fiber Optic Vibration Sensing Event Recognition Method Based on CNN-LSTM-Transformer Net

基于CNN-LSTM-Transformer网格的分布式光纤振动传感事件识别方法

¹ College of Optoelectronic Information and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
² School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China
³ Key Laboratory of Space Active Optical-Electro Technology, Chinese Academy of Sciences, Shanghai 200083, China

Received: 24 January 2025

Abstract

Phase-sensitive Optical Time-Domain Reflectometer ($\phi$-OTDR) technology facilitates the real-time detection of vibration events along fiber optic cables by analyzing changes in Rayleigh scattering signals. This technology is widely used in applications such as intrusion monitoring and structural health assessments. Traditional signal processing methods, such as Support Vector Machines (SVM) and K-Nearest Neighbors (KNN), have limitations in feature extraction and classification in complex environments. Conversely, a single deep learning model often struggles with capturing long time-series dependencies and mitigating noise interference. In this study, we propose a deep learning model that integrates Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and Transformer modules, leveraging $\phi$-OTDR technology for distributed fiber vibration sensing event recognition. The hybrid model combines the CNN's capability to extract local features, the LSTM's ability to model temporal dynamics, and the Transformer's proficiency in capturing global dependencies. This integration significantly enhances the accuracy and robustness of event recognition. In experiments involving six types of vibration events, the model consistently achieved a validation accuracy of 0.92, and maintained a validation loss of approximately 0.2, surpassing other models, such as TAM+BiLSTM and CNN+CBAM. The results indicate that the CNN+LSTM+Transformer model is highly effective in handling vibration signal classification tasks in complex scenarios, offering a promising new direction for the application of fiber optic vibration sensing technology.

摘要

相位敏感光时域反射计($\phi$-OTDR)技术通过分析瑞利散射信号的变化，实现对光缆沿线振动事件的实时检测，广泛应用于入侵监测、结构健康评估等领域。传统的信号处理方法(如 SVM、KNN 等)在特征提取及复杂场景分类上存在局限，而单一深度学习模型在长时序依赖捕捉及噪声干扰方面效果不佳。本文基于 $\phi$-OTDR技术，提出了一种融合 CNN、LSTM 和 Transformer 模块的深度学习模型，用于分布式光纤振动传感事件的识别。提出的混合模型结合 CNN 提取局部特征、LSTM 建模时序动态、Transformer 捕获全局依赖关系，显著提升了识别的精度和鲁棒性。在对 6 类振动事件的实验中，该模型验证准确率稳定在0.92，验证损失在 0.2左右，优于其他对比模型(如 TAM+BiLSTM 和 CNN+CBAM)。研究表明，CNN+LSTM+Transformer 通过其全局建模与特征融合优势，能够有效应对复杂场景下的振动信号分类任务，为光纤振动传感技术的应用提供了新的方向。