基于一维空洞卷积的甲烷浓度实时检测装置设计  

Design of a real-time methane-concentration detection device based on one-dimensional cavity convolution

作　　者：阚玲玲[1] 叶炀 苗凯 梁洪卫[1] KAN Lingling;YE Yang;MIAO Kai;LIANG Hongwei(School of Electrical and Information Engineering,Northeast Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学电气信息工程学院,黑龙江大庆163318

出　　处：《实验技术与管理》2025年第1期191-197,共7页Experimental Technology and Management

基　　金：教育部春晖计划项目(HZKY20220304)。

摘　　要：可调谐激光吸收光谱(TDLAS)技术需要人工筛选二次谐波后拟合数据,且人为因素会影响拟合曲线的预测结果,因此该文设计了一种基于空洞卷积的甲烷浓度检测装置。首先,利用一维空洞卷积提取二次谐波信号特征;其次,利用反残差模块提高模型收敛速度,减少训练时间;再次,加入宽卷积核感受野模块,在保证检测准确率的前提下扩大感受野,减少模型参数;最后,将模型部署在Jetson Nano上实时检测甲烷浓度。实验结果表明,该方法检测准确率为99.89%,模型参数为12768个。与其他模型相比,该模型检测精度高、参数量少、实时性强,适用于实验室甲烷浓度实时检测。[Objective]Methane is a highly flammable and explosive gas,posing significant risks of explosions and safety hazards in industrial environments.Accurate and efficient detection of methane concentrations is critical for maintaining safety,particularly in high-risk areas,such as natural gas pipelines,storage facilities,and industrial areas.Although tunable diode laser absorption spectroscopy(TDLAS)technology is widely used and effective,it depends heavily on manual intervention to identify suitable second-harmonic signals for linear fitting.This reliance reduces efficiency and limits its real-time applicability.Additionally,conventional neural network models,owing to their complex structures and numerous parameters,are typically unsuitable for real-time methane detection in embedded systems.To address these challenges,this study proposes a novel wide kernel receptive field block 1D convolutional neural network(1D-WKRFB-CNN)designed for high accuracy and real-time performance on embedded platforms like the Jetson Nano.[Methods]First,a wider convolutional kernel was used to extract the main features of methane gas data.Dilated convolution expands the receptive field range,enabling the network to handle longer data sequences from second-harmonic signals.This allows the model to capture more distant feature information,facilitating the understanding of the relationship between long sequences and methane concentrations.The design incorporates a max pooling layer to retain the primary features of each signal region,reduce the input size,and discard non-essential information,thereby reducing the computational complexity and network parameters and simplifying the model structure.Additionally,an anti-residual module is implemented to effectively capture the relationships between feature channels,accelerate network convergence,and reduce training time.Finally,a wide convolutional kernel receptive field module is introduced to optimize the size of the convolutional kernel’s receptive field.This minimizes computational requi

关 键 词：可调谐激光吸收光谱 JetsonNano 深度学习 轻量化 实验装置 

分 类 号：TE88[石油与天然气工程—油气储运工程]