红外光谱分析在自然科学、 工程技术等诸多领域发挥着重要作用。 随着计算机和人工智能技术的不断发展, 对红外/近红外光谱分析提出了更高的要求。 深度学习以人工神经网络为架构, 通过对数据进行分层特征提取完成特征/表征学习, 在解析数据细节特征方面具有独特的优势, 在计算机视觉、 语音识别、 疾病诊断等多领域得到成功应用。 尽管深度学习在图像、 音频、 文字分析方面获得了较好的效果, 但是在红外/近红外光谱数据分析中的应用还十分有限。 针对深度学习的卷积运算, 首先将一维傅里叶变换(Fourier transform infrared spectroscopy, FTIR)红外光谱数据通过对称点模式(symmetrized dot patterns, SDP)变换为二维RGB彩色图像, 然后将SDP变换得到的彩色图像数据作为VGG(oxford visual geometry group)深度卷积神经网络的输入进行深度学习, 建立基于红外光谱数据的分类识别模型。 对不同浓度甲烷(CH4)、 乙烷(C2H6)、 丙烷(C3H8)、 正丁烷(C4H10)、 异丁烷(iso-C4H10)、 正戊烷(C5H12)、 异戊烷(iso-C5H12)七种单组分烷烃及其混合气体SDP转化获得的224×224彩色(RGB)图像, 呈现出显著差别, 且更符合VGG卷积运算的数据格式。 将SDP-VGG方法应用于气测录井中甲烷浓度范围的识别: 气测录井气体为上述七组分烷烃气体的混合气体, 其中主要成分甲烷的浓度范围按照<20%, 20%～40%, 40%～60%, 60%～80%, 80%～100%分为5类, 不同七组分烷烃混合气体样本的红外光谱由红外光谱仪在波数范围为4 000～400 cm-1、 间隔12 nm的条件下扫描获得。 在未经过特殊预处理和特征提取的情况下, 采用随机选择的4 500个样本, 由SDP-VGG法建立的七组分混合气体甲烷浓度范围识别模型, 对5种甲烷浓度范围的识别准确率达到91.2%, 优于相同红外光谱数据所建立支持向量机(support vector machine, SVM)和随机森林(random forest, RF)模型的识别准确率88.7%和86.2%。 研究表明, SDP结合深度学习可以准确提取红外光谱数据的关键特征, 提高了红外光谱识别的准确率, 是一种更为有效的红外光谱分析方法, 具有广阔的应用前景。

Infrared spectrum analysis plays an important role in many fields such as natural science, engineering technology, and so on. With the continuous development of computer and artificial intelligence technology, higher requirements have been imposed on infrared/near-infrared spectral analysis. Based on artificial neural networks, the deep learning algorithm performs representation learning by extracting hierarchical features from data layer by layer. It has unique advantages in analyzing the details features of data. It has been successfully applied in many fields such as computer vision, speech recognition, and disease diagnosis. Although deep learning has achieved good results in the analysis of images, audio, and text data, its application in infrared/near-infrared spectral analysis is still very limited. A deep learning convolution operation method for infrared spectroscopic analysis is presented. Firstly, one-dimensional Fourier Transform Infrared Spectroscopy (FTIR) data are transformed into two-dimensional RGB color image data through Symmetrized Dot Patterns (SDP), and then, the transformed SDP color image data is fed into the VGG (Oxford Visual Geometry Group) deep convolutional neural network for deep learning to establish a classification and recognition model. By SDP transformation, the infrared spectra of sevensingle-component gases of different concentrations, including methane (CH4), ethane (C2H6), propane (C3H8), n-butane (C4H10), iso-butane (iso-C4H10), n-pentane (C5H12), iso-pentane (iso-C5H12), and its mixtures convert to 224×224 color images. The SDP transformed images show a significant difference in the distribution of the pattern points and are more in line with the data format of the VGG convolution operation. The SDP-VGG method is used to identify the methane concentration range in gas logging: the gas logging gas is a mixture of the above seven components of alkanes, and the concentration ranges of methane are divided into five categories: <20%, 20%～40%, 40%～60%, 60%～80%, and 80%～100%. The infrared spectra of different seven-component alkane mixed gas samples are collected by the infrared spectrometer in the wavenumber range of 4 000～400 cm-1 and scanning interval 12 nm. Without special pre-processing and feature extraction, 4 500 samples are used to establish the identification model of various methane concentration ranges by the SDP-VGG method. The recognition accuracy of the SDP-VGG model reached 91.2%, which is better than the recognition accuracy of 88.7% and 86.2% of the Support Vector Machine (SVM) and Random Forest (RF) models established by the same infrared spectral data. The research shows that SDP combined with deep learning can accurately extract the key features of infrared spectra. It is a more effective infrared spectral analysis method, which improves the recognition accuracy of the infrared spectrum and has broad application prospects.