融合格拉姆角场的深度特征学习在痕量气体浓度识别中的应用研究

针对气体绝缘金属封闭式组合电器(gas insulated switchgear, GIS)设备中痕量气体紫外分析光谱信号易出现吸收峰重叠的问题，提出了一种结合格拉姆角场(Gram’s angle field, GAF)和VGG16改进模型的多组分痕量气体的定量检测方法。首先利用GAF将一维紫外光谱信号转换为时序图像，将光谱信号映射为具有丰富特征信息的图像形式，从而提升原始光谱信号的特征表达能力。其次将GAF特征图输入到VGG16改进模型中，实现痕量气体浓度的特征识别。最后通过不同浓度下采集到的CS2、SO2和H2S的单组分气体和混合气体，与卷积神经网络(convolutional neural network, CNN)、VGG16和SDP＿VGG16等模型进行对比实验，并结合受试者工作特征曲线下面积(area under the curve, AUC)进行验证。结果表明，该方法可以有效地检测出SF6分解所产生的CS2、SO2和H2S痕量气体，是一种行之有效的特征提取方法。

Application of deep feature learning with Gram's angle field for trace gas concentration identification

To solve the problem of overlapping absorption peaks in the UV analysis of gas insulated switchgear (GIS), this paper presents a quantitative detection method for multi-component trace gases combining Gram''s Angle Field (GAF) and an improved model of VGG16. To enhance the feature expression of the original spectral signal, GAF is used to convert one-dimensional UV spectral signals into sequential images. A large amount of feature information is included in these feature images. The GAF feature map is then fed into the modified VGG16 model for feature identification of trace gas concentrations. Finally, the models of a convolutional neural network (CNN), VGG16, and SDP_VGG16 are compared to the single and mixed gases of CS2, SO2, and H2S collected at various concentrations. From the findings of the validation and the area under curve (AUC), the proposed method can successfully detect CS2, SO2, and H2S trace gases produced by SF6 breakdown, and it is a tried-and-true feature extraction technique.