基于Resnet-Mamba算法的光纤传感事件检测

Fiber Optic Sensing Event Detection Based on Resnet-Mamba Algorithm

摘要: 相位敏感光时域反射仪(φ-OTDR)技术在光纤传感领域具有重要应用，可用于监测振动事件。然而，传统方法在复杂环境下的事件识别准确率有限。为提高OTDR事件检测的性能，本文提出了一种融合残差网络(ResNet)与状态空间模型Mamba的ResNet-Mamba算法。该算法通过ResNet模块提取局部时空特征，并利用Mamba的全局状态空间建模能力捕获长程依赖关系，构建端到端的分类模型。实验基于包含背景噪声、挖掘、敲击、浇水、摇动及行走6类事件共15,419个样本的OTDR数据集结合早停机制与动态学习率调整进行模型优化。结果表明，ResNet-Mamba在测试集上达到99.74%的准确率，相较于ResNet模型提高了4.8%，相较于CNN和SVM等模型提高了6.51%~16.5%。本研究为φ-OTDR系统的实时事件监测提供了高效解决方案，在智慧安防与基础设施监测领域具有应用潜力。

Abstract: φ-OTDR (phase-sensitive optical time-domain reflectometer) technology has important applications in fiber optic sensing for monitoring vibration events. However, traditional methods have limited accuracy in event recognition in complex environments. To improve the performance of OTDR event detection, this paper proposes a ResNet-Mamba algorithm that fuses the residual network (ResNet) with the state space model Mamba. The algorithm extracts local spatio-temporal features through the ResNet module and uses the global state space modeling capability of Mamba to capture long-range dependencies and construct an end-to-end classification model. The experiments are based on the OTDR dataset containing six types of events with 15,419 samples in total, including background noise, digging, knocking, watering, shaking and walking combined with the early stopping mechanism and dynamic learning rate adjustment for model optimization. The results show that ResNet-Mamba achieves 99.74% accuracy on the test set, which is an improvement of 4.8% compared to the ResNet model and 6.51% to 16.5% compared to models such as CNN and SVM. This study provides an efficient solution for real-time event monitoring in φ-OTDR systems, which has potential applications in the field of smart security and infrastructure monitoring.