军事知识服务平台
基于MS-TvsNet模型的MI-SI光纤振动传感系统信号分离方法研究
Research on Signal Separation Method for MI-SI Fiber Optic Vibration Sensing System Based on MS-TvsNet Model
摘要 针对分布式光纤传感系统中多源混叠振动信号分离难题,本文提出一种基于深度学习的新型多尺度时域振动信号分离网络(MS-TvsNet)模型.该模型结合迈克尔孙-萨格纳克(MI-SI)混合干涉结构的移位相减信号特性,可实现多事件振动信号的高效分离.通过引入多模态膨胀卷积模块和长短期记忆网络增强的分离单元,模型在时域上协同捕获瞬态冲击与准周期振动特征,结合注意力机制强化长程依赖建模能力.实验基于 40 km 光纤传感系统,采集剪切、攀爬、锯切及摇晃 4 类入侵事件信号构建混叠数据集,验证了 MS-TvsNet 模型在 15.2 dB尺度不变信噪比(SI-SNR)下的优异分离性能,较时域音频分离网络(TasNet)和全卷积时域音频分离网络模型(Conv-TasNet)提升 5.4 dB 和 3.1 dB.同时,该模型在噪声干扰(ΔP=-20 dB)及跨位置泛化测试中展现出强鲁棒性,分类准确率较单一 MI 结构提升 4.6%,为复杂环境下光纤传感系统的智能化处理提供了新的思路.
Abstract:
To address the challenge of separating multi-source mixed vibration signals in distributed fiber-optic sensing systems,this paper proposes a novel deep learning-based multi-scale time-domain vibration signal separation network(MS-TvsNet)model.By leveraging the shift-subtraction signal characteristics of the MI-SI structure,the model achieves efficient separation of multi-event vibration signals.The introduction of multi-modal dilated convolutional modules and LSTM-enhanced separation units enables synergistic capture of transient impacts and quasi-periodic vibration features in the time domain,while the attention mechanism strengthens long-range dependency modeling.Experiments conducted on a 40 km fiber-optic sensing system,with mixed datasets constructed from four types of intrusion events(shearing,climbing,sawing,and shaking),demonstrate the superior separation performance of MS-TvsNet,achieving a scale-invariant signal-to-noise ratio(SI-SNR)of 15.2 dB.This represents improvements of 5.4 dB and 3.1 dB compared to TasNet and Conv-TasNet model respectively.Additionally,the model exhibits strong robustness under noise interference(ΔP=-20 dB)and cross-position generalization tests,with classification accuracy surpassing that of the single MI structure by 4.6%.This work provides a new approach for intelligent signal processing in complex fiber-optic sensing environments.
年,卷(期): 2026, 65(5)
分类号: TP212
机标分类号: TP212 TN911.7 TP301.6
在线出版日期: 2026年11月4日