Research On Autoencoder Modeling And Calibration Transfer In Near-infrared Spectroscopy

The pharmaceutical is entwined with people’s lives, but the market is overcrowded with counterfeit-drugs. If people accidentally take the counterfeit-drugs, not only could the drug not effect to treat disease, but alse it may be a health hazard to human. So it is significant to detect the counterfeit-drugs. The Near-infrared(NIR) spectroscopy techniques have the advantages of fast analysis, low cost of detection and it do not destroy the samples. It is very suitable for the counterfeit-drugs detection. Using Near-infrared(NIR) spectroscopy techniques and combined with the machine learning algorithms are efficient way to discriminate the counterfeit pharmaceutical.The Autoencoder is one of the typical deep learning models. Compared with the traditional surface learning algorithm, the Autoencoder has greater modeling capability. In this paper, firstly, the preprocessing step and pre-whitening was used to the Near-infrared spectroscopy data, then the Sparse Denoising Autoencoder(SDAE) was used to build the classification model for the identification of counterfeit pharmaceutical. And compared with the BP Neural Networks and SVM algorithm for the classification accuracy and mean absolute difference(MAD). Experimental results show that the pre-whitening improved the Autoencoder classification accuracy effectively. The SDAE algorithm performed better than SVM and BP Neural Networks when the train datasets achieve a certain amount. And the generalization performances of SDAE algorithm is better than SVM and BP Neural Networks.To address the cost-sensitive problem of the NIR-based approach to counterfeit-drug detection. In this paper, two kinds of approaches was used to build the cost-sensitive Autoencoder model. The first approach is that combined the SMOTE algorithm with Autoencoder. Another is to add the cost factor to the cost function of the Autoencder. Three groups of imbalance sample set were used to test the two kinds of Cost-Sensitive Autoencoder. And compared with the Cost-Sensitive SVM. The experiment show that this two kinds of approach both can make the Autoencoder to deal with the cost-sensitive problem, They are effect solutions to the cost-sensitive problem of counterfeit-drug detection.To address the problem of model sharing between multiple spectrometers. This paper presents a new model transfer method base on Dynamic Time Warping(DTW). Firstly, calculate the correlation distance of each pair of the master machine spectroscopy and the slave machine spectroscopy. Using the Dynamic Time Warping to find the best match wavelength point relationship between two spectroscopy. Building the regression model base on the relationship of wavelength point. The experiment build the transfer model using the Near-infrared spectroscopy dataset about drug and corn. The experiment validated the effectiveness of DTW calibration transfer algorithm, and compared with the traditional algorithm such as DS and PDS method. The results show that the transfer the spectrum using DTW calibration transfer algorithm has less ARMS and root mean square error of prediction(RMSEP) which get from the PLS model of the master machine.It is superior to the DS and PDS algorithm.