Deep Learning and Intelligent Perception

CLC number: TP212

An Interpretable Inception-ResNet-Based Method for Intrusion Event Recognition in Distributed Optical Fiber Sensing Systems

一种可解释的基于 Inception-ResNet 的分布式光纤传感入侵事件识别方法

¹ School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China (上海理工大学　光电信息与计算机工程学院，上海　200093)
² Beijing Zhaoxin-Electronic-Technology Co., Ltd, Beijing 100000, China (北京兆芯电子科技有限公司，北京　100000)
³ Northeast Sichuan Gas Mine, PetroChina Southwest Oil and Gasfield Company, Dazhou 635000, Sichuan, China (中国石油西南油气田分公司　川东北气矿，四川　达州　635000)
⁴ The Haixi Extension Service Center for Agricultural and Animal Husbandry Technology, Delingha 810000, Qinghai, China (青海省德令哈市海西州畜牧兽医科技推广服务中心，青海　德令哈　810000)

^† Corresponding author. E-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.

Received: 28 June 2025

Abstract

Distributed optical fiber sensing (DOFS) technology has been widely applied in pipeline monitoring, seismic detection, and security protection due to its broad coverage, high sensitivity, and strong anti-interference capability. However, the acquired signals are typically noisy, exhibit complex temporal-spatial patterns, and contain high-dimensional categorical features, posing significant challenges for robust classification. To address these issues, this paper introduces an Inception-ResNet-based model for intrusion event recognition in DOFS systems. The Inception architecture extracts multi-scale features from complex vibration patterns, while the residual optimization of ResNet enables efficient deep feature propagation and stable training. Furthermore, to enhance model interpretability, a Grad-CAM-based mechanism is integrated to visualize class-discriminative regions in the vibration signals, revealing the patterns that most strongly influence the network's decisions. Extensive experiments demonstrate the effectiveness of the proposed approach, achieving an average classification accuracy of 92.6%, outperforming traditional deep learning networks even with significantly reduced training data. These results indicate that the interpretable Inception-ResNet framework not only accurately classifies complex one-dimensional sensing signals but also provides transparent and reliable support for practical DOFS applications.

摘要

分布式光纤传感(DOFS)技术因其监测范围广、灵敏度高、抗干扰能力强等优势，已广泛应用于管道监测、地震探测和安防预警等领域。然而，系统采集的信号通常噪声强烈、时空模式复杂，并包含维度较高的类别特征，使得准确识别入侵事件面临显著挑战。为此，本文提出了一种基于 Inception-ResNet 的 DOFS 入侵事件识别模型。Inception 模块能够从复杂振动信号中提取多尺度特征，而 ResNet 的残差优化结构则保证了深层特征的高效传播与稳定训练。 此外，为提升模型的可解释性，本文引入了基于 Grad-CAM 的可视化机制，用以突出对分类结果具有关键贡献的信号区域，从而揭示模型决策背后的判据。大量实验结果表明，所提出方法在入侵事件分类任务中取得了 92.6% 的平均准确率，在显著减少训练数据量的条件下仍优于传统深度学习模型。研究结果表明，该可解释的 Inception-ResNet 框架不仅能够有效分类复杂的一维光纤传感信号，也为 DOFS 的实际应用提供了透明、可靠的技术支撑。