Citation: ZHANG Wei-Dong, LI Ling-Qiao, HU Jin-Quan, FENG Yan-Chun, YIN Li-Hui, HU Chang-Qin, YANG Hui-Hua. Drug Discrimination by Near Infrared Spectroscopy Based on Stacked Sparse Auto-encoders Combined with Kernel Extreme Learning Machine. Chinese Journal of Analytical Chemistry, 2018, 46(9): 1446-1454. doi: 10.11895/j.issn.0253-3820.171343 [复制]
Drug Discrimination by Near Infrared Spectroscopy Based on Stacked Sparse Auto-encoders Combined with Kernel Extreme Learning Machine
提出一种基于堆栈稀疏自编码融合核极限学习机（Stacked sparse auto-encoders combine kernel extreme learning machine，SSAE-KELM）的近红外药品鉴别方法，通过引入核极限学习机代替SSAE的Softmax分类和BP微调阶段，减少了模型的训练步骤、训练参数以及训练时间，提高了深度学习网络的实际应用能力，核函数的引入提高了模型的分类能力。其中，SSAE用于初始化整个网络模型，并且从输入数据中学习到有用的特征，KELM用于实现分类任务。研究了SSAE-KELM模型对不同厂商生产的同一包装形式（铝塑或非铝塑）药品鉴别的预测能力、稳定性及训练时间，以实现药品的二分类和多分类的无损鉴别。同时，与ELM、SSAE、BP、SVM及随机隐退深度信念网络（Dropout-DBN）进行对比。结果表明，无论是二分类还是多分类，SSAE-KELM不仅具有更优的分类能力和稳定性、还减少了训练时间。因此，SSAE-KELM是一种有效的光谱分类建模工具。
A method for drug discrimination with near infrared spectroscopy based on stacked sparse auto-encoders combined with kernel extreme learning machine (SSAE-KELM) was developed. By introducing the KELM instead of the SSAE's Softmax classification and BP fine-tuning stage, the training steps, training parameters and training time of the SSAE were reduced, and the practical application of the deep learning network was improved, as well the classification ability of the model was improved by introduction of kernel function. Among which SSAE was used to initialize the entire network model and learn useful features from the input data and KELM was used to perform the classification. To identify binary-classification and multi-classification of drugs, the predictability, stability and training time of SSAE-KELM model for the same package (Aluminum-plastic or non-Aluminum-plastic) drug by different manufactures were investigated. At the same time, SSAE-KELM was compared with ELM, SSAE, SVM, BP and Dropout-DBN, and it was found that SSAE-KELM not only reduced the training time but had higher classification accuracy and stability in binary and multi-class classification. Therefore, SSAE-KELM is an effective spectral classification modeling tool.