应用傅里叶红外光谱(FTIR)和二维相关红外光谱(2D-IR)对8种不同种类的昭通苹果进行分析, 并运用卷积神经网络(CNN)和径向基函数(RBF)神经网络对其进行了鉴别。 在8种昭通苹果的傅里叶红外光谱中, 均在3 500~2 850、 1 650~1 400和1 250~800 cm^-1等范围表现出了较强吸收峰, 可以看出苹果含有丰富的糖类、 维生素、 氨基酸、 脂质、 有机酸、 酚类和黄酮类等物质, 且8种昭通苹果光谱非常相似, 只在吸收峰的强弱和峰位置上有非常小的差异, 依据光谱来区分鉴别8种昭通苹果, 显然是不可能的。 以温度为微扰, 采集8种苹果的动态光谱, 并对动态光谱800~1 800 cm^-1进行2D-IR分析, 在2D-IR的同步光谱中, 可以清晰看出: 随着温度的升高, 2D-IR在1 010和1 642 cm^-1附近出现了相对较强的自动峰, 说明苹果中的酯类、 酸类和蛋白质出现一定的分解, 而且酯类和酸类分解较强, 蛋白质相对较弱; 另外在8种样品的2D-IR中, 红富士片红中的二维相关红外光谱中出现的自动峰1 642 cm^-1相对其他样品来说是最强的自动峰, 出现的负交叉峰(1 006, 1 642 cm^-1)最弱; 在阿克苏苹果中只出现了1 010 cm^-1一个强自动峰; 在秦冠苹果中出现了三个自动峰; 另外在2001苹果和新世纪苹果中, 最强自动峰出现在1 020 cm^-1, 对比其他苹果有10个波数的移动, 依据2D-IR, 可以区分部分苹果样品。 进一步对8种苹果216个苹果光谱进行卷积神经网络(CNN)和径向基函数(RBF)神经网络分析, 随机选择152个样品光谱数据进行模型训练, 通过一定迭代和训练, 使得两种模型在训练集上的分类准确率达到100%的最优状态, 再对64个样品光谱进行预测, 在CNN分析中准确率为89.06%, 而在RBF神经网络分析中准确率达到90.6%, 两种神经网络模型都表现出了非常好的分类准确率。 因此FTIR、 2D-IR、 CNN和RBF神经网络分析方法在苹果分析鉴别研究中相互补充, 可以对昭通苹果进行准确分类, 并且该方法可以运用到其他物质分类鉴别分析。

The article presents an analysis of eight different varieties of Zhaotong apples using Fourier Transform Infrared Spectroscopy (FTIR) and Two-Dimensional Correlation Infrared Spectroscopy (2D-IR). Additionally, Convolutional Neural Networks (CNN) and Radial Basis Function (RBF) neural networks were utilized for their identification. The Fourier Transform Infrared Spectroscopy (FTIR) analysis of eight Zhaotong apple varieties revealed prominent absorption peaks within the spectral ranges of 3 500~2 850, 1 650~1 400, and 1 250~800 cm^-1, signifying their abundance in sugars, vitamins, amino acids, lipids, organic acids, phenols, flavonoids, and various other compounds. However, a notable similarity was observed across the spectra of these apples, with only subtle variations in peak intensities and positions. Consequently, relying solely on spectral characteristics to differentiate between these eight varieties is impractical. Using temperature as a perturbation, we collected dynamic spectra of eight distinct apple varieties and subjected them to 2D-IR analysis within the spectral range of 800~1 800 cm^-1. The synchronous spectrum derived from this analysis underscores that as temperature escalates, distinct autopeaks emerge in proximity to 1 010 and 1 642 cm^-1. These peaks serve as indicators of varying degrees of decomposition occurring within esters, acids, and proteins present in apples, with esters and acids undergoing more pronounced decomposition compared to proteins. Notably, among these eight samples, Red Fuji Apples exhibit the strongest autopeak at 1 642 cm^-1, accompanied by the weakest negative cross peak at 1 006, 1 642 cm^-1, while Aksoo Apples display only one prominent autopeak at 1 010 cm^-1. Qin Guan Apples exhibit three autopeaks, whereas both 2001 Apples and New Century Apples have their strongest autopeak shifted to 1 020 cm^-1, differing by 10 wavenumbers from those of other varieties. This 2D-IR analysis enables differentiation between certain apple samples based on their unique spectral signatures. An optimized approach to analyzing the spectra of 216 apples from eight varieties utilizing Convolutional Neural Network (CNN) and Radial Basis Function (RBF) neural networks. By randomly selecting 152 sample spectra for model training and through iterative refinement and rigorous training, both models achieved an optimal state, attaining a classification accuracy of 100% on the training set. Following this, predictions were conducted on the spectra of an additional 64 samples, yielding remarkable results: an accuracy rate of 89.06% in CNN analysis and an even higher 90.6% in RBF neural network analysis. Both neural network models have demonstrated outstanding performance in terms of classification accuracy. Hence, the integration of FTIR, 2D-IR, CNN, and RBF neural network analysis methods forms a complementary approach in the study of apple analysis and identification, facilitating precise classification of Zhaotong apples. Additionally, this comprehensive methodology holds significant potential for application in the classification and identification of diverse substances beyond apples.