提出一种稀疏降噪自编码结合高斯过程的近红外光谱药品鉴别方法。 首先对近红外光谱数据进行小波变换以消除基线漂移, 然后用稀疏降噪自编码(SDAE)网络提取光谱特征并降维表示, 最后采用高斯过程(GP)进行二分类, 其中GP选用光谱混合(SM)核函数作为协方差函数, 记此分类网络为wSDAGSM。 自编码网络具有很强的模型表示能力, 高斯过程分类器在处理小样本数据时具有优势。 wSDAGSM网络通过稀疏降噪自编码学习得到维数更低但更有价值的特征来表示输入数据, 同时将具有很好表达力的光谱混合核作为高斯过程的协方差函数, 有利于更准确的光谱数据分类。 以琥乙红霉素及其他药品的近红外光谱为实验数据, 将该方法与经过墨西哥帽小波变换的BP神经网络(wBP)、 支持向量机(wSVM), SDAE结合Logistic二分类(wSDAL)、 SDAE结合采用平方指数(SE)协方差核的GP二分类(wSDAGSE), 以及未采用小波变换的SDAGSM网络等方法进行对比。 实验结果表明, 对光谱数据进行墨西哥帽小波变换预处理能有效提升SDAGSM网络的分类准确率和稳定性。 wSDAGSM方法无论从分类准确率还是分类结果稳定性方面, 都优于其他分类器。

In this paper, a new method for pharmaceutical discrimination by the near infrared spectrum is proposed, which is based on the sparse denoising autoencoder (SDAE) combined with Gauss process (GP). First of all, the Mexican hat wavelet transform was used to eliminate noise and baseline drift from the spectra data, then SDAE network was used to extract the feature and reduce dimension of spectrum. Finally, GP was used for binary classification, in which the GP selected the spectral mixture (SM) kernel function as its covariance function. This classification method was named as wSDAGSM. Autoencoder network has a strong ability of model representation, and GP classifier has the advantage in dealing with small sample data. The WSDAGSM network is able to obtain fewer dimensions and more valuable features by SDAE learning to represent the input data. Meanwhile, the spectral mixture kernel function which has a good expression was used as the covariance function of the GP in the WSDAGSM network. Therefore the WSDAGSM network is conducive to more accurate classification of spectral data. With near infrared spectra of Erythromycin Ethylsuccinate and other pharmaceuticals as experimental data, some classification methods were used after Mexican hat wavelet transform, they were BP neural network (wBP), support vector machine (wSVM), SDAE combined with binary classification of Logistic (wSDAL), SDAE combined with binary classification of GP selected the squared exponential (SE) kernel function (wSDAGSE). And another method was also applied, which was SDAGSM network without Mexican hat wavelet transform. All above methods were used for comparing with wSDAGSM network. Experimental results show that SDAGSM can effectively improve the classification accuracy and stability by applying the wavelet transform to the spectral data. The proposed method wSDAGSM is superior to other classifiers in terms of classification accuracy and stability of the classification results.