近红外光谱是热门的食品检测方法之一, 对于这种高维光谱数据的分析常常需采用数据降维算法提取其中的特征, 然而绝大多数算法都只能针对单个数据集进行分析。 虽然已有基于对比学习的对比主成分分析成功应用于不同水果表面农残的近红外光谱检测中, 但是该方法只能以线性的方式组合原有特征, 特征提取效果存在局限性, 并且需要调节对比参数来控制背景集影响, 需要消耗更大的时间成本。 cVAE(contrastive variational autoencoder)是一种基于对比学习和变分自编码器的改进算法, 被用于图像去噪和RNA序列分析中, 它仍然具备分析多个数据集的特点, 同时因为组合了神经网络的概率生成模型而具备了提取非线性隐含特征的能力。 将cVAE算法应用于近红外光谱分析, 建立了准确的近红外光谱数据降维模型。 在实际验证中, 使用cVAE算法对购买的不同品牌和批次纯牛奶中掺假三聚氰胺进行检测。 结果表明, 使用VAE算法只能区分出不同品牌和批次的纯牛奶, 而其中是否掺假三聚氰胺这一重要信息无法表现出来; 而使用cVAE算法进行数据分析时, 由于添加了背景数据集分离了无关变量, 能够清晰的将有无掺假三聚氰胺的样本分类。 这说明了, cVAE不仅具备了cPCA(contrastive principle component analysis)在近红外光谱数据降维中的优势, 而且具备提取非线性特征的能力, 同时不需要调节可变参数, 能够更方便地建立近红外光谱降维模型。

Near-infrared spectroscopy is one of the popular food detection methods. Analyzing such high-dimensional spectral data often requires using data dimensionality reduction algorithms to extract features. However, most algorithms can only analyze a single data set.Although the contrastive principal component analysis based on contrastive learning has been successfully applied to the near-infrared spectroscopy detection of pesticide residues on the surface of different fruits, this method can only linearly combine the original features, and the feature extraction effect has limitations and needs to be adjusted. Comparing parameters to control the influence of the background set requires more time and cost.cVAE (contrastive variational autoencoder) is an improved algorithm based on contrastive learning and variational autoencoder, which is used in image denoising and RNA sequence analysis. It still has the characteristics of analyzing multiple data sets, and at the same time, it can extract nonlinear hidden features because of the combination of the probability generation model of the neural network. The cVAE algorithm is applied to near-infrared spectroscopy analysis, and an accurate dimensionality reduction model of near-infrared spectroscopy data is established. In actual verification, the cVAE algorithm was used to detect melamine adulteration in pure milk purchased by different brands and batches. The results show that when analyzing whether melamine is adulterated in pure milk purchased by different brands and batches, the VAE algorithm can only distinguish between different brands and batches of pure milk, and the important information about whether or not melamine is adulterated is not available. When using the cVAE algorithm for data analysis, because the background data set is added to separate irrelevant variables, it is possible to classify samples with or without adulterated melamine clearly.This shows that cVAE has the advantages of cPCA (contrastive principle component analysis) in dimensionality reduction of near-infrared spectroscopy data and can extract nonlinear features without the need to adjust variable parameters.The dimensionality reduction model of the near-infrared spectroscopy can be established more conveniently.