一种基于TM-UNet模型的光纤振动传感系统的信号降噪方法  

A Signal Denoising Method of Optical Fiber Vibration Sensing System Based on TM-UNet Model

作　　者：唐亮 熊灵奇 贾波[1] 肖倩[1] 吴红艳[1] TANG Liang;XIONG Lingqi;JIA Bo;XIAO Qian;WU Hongyan(Department of Materials Science,Fudan University,Shanghai 200433,China)

机构地区：[1]复旦大学材料科学系,上海200433

出　　处：《复旦学报(自然科学版)》2025年第1期39-49,共11页Journal of Fudan University:Natural Science

摘　　要：分布式光纤振动传感技术具备灵敏度高、响应速度快、监测范围大、抗干扰能力强等一系列优点,已广泛应用于周界监控、环境探测等多种类目标事件的复杂应用场景。然而,真实振动信号数据总是受到各种环境噪声的污染,这极大地影响了振动信号质量。为了滤除环境噪声并提高信噪比,本文提出了一种改进的UNet模型TM-UNet。基于UNet网络架构,通过在瓶颈层加入Transformer机制进一步提高特征提取和表示的能力,使模型性能比一般的UNet有显著的提升,训练好的TM-UNet模型对各种常见入侵振动信号都具有较好的降噪能力,能使信噪比提高约18 dB。本文介绍了基于Sagnac结构的光纤振动传感系统的工作原理,通过在隔音室和户外工作环境下采集各种入侵行为的振动信号,构造了高真实性的输入-目标信号数据对并生成降噪数据集。此外,本文使用了一种基于结构相似性和尺度不变信噪比的加权复合损失函数,能够有效提高模型的各项评估指标。实验结果证明TM-UNet模型比传统降噪方法信噪比增益高,计算开销小。Distributed fiber optic vibration sensing technology has a series of advantages such as high sensitivity,fast response speed,large monitoring range,strong anti-interference ability,etc.,and has been widely used in complex application scenarios such as perimeter monitoring,environmental detection,and other types of target events.However,the real vibration signal data is always polluted by various environmental noises,which greatly affects the vibration signal quality.In order to filter out the environmental noise and improve the signal-to-noise ratio,this paper proposes an improved UNet model TM-UNet.Based on the UNet network architecture,the ability of feature extraction and representation is further improved by adding the Transformer mechanism in the bottleneck layer,which makes the model performance significantly better than that of the general UNet.The trained TM-UNet model has good denoising capability for various common intrusion vibration signals,increasing the signal-to-noise ratio by approximately 18 dB.This paper describes the working principle of the fiber optic vibration sensing system based on the Sagnac structure,and constructs highly realistic input-target signal data pairs and generates datasets by collecting vibration data of various intrusion behaviors in a soundproof room and an outdoor working environment.In addition,a weighted composite loss function based on structural similarity and scale-invariant signal-to-noise ratio is used in this paper,which can effectively improve the various evaluation indexes of the model.The experimental results demonstrate that the TM-UNet model has a higher signal-to-noise ratio enhancement and less computational overhead than the traditional denoising methods.

关 键 词：光纤振动传感器 深度学习 信号降噪 

分 类 号：TP212[自动化与计算机技术—检测技术与自动化装置]