[Sparse Denoising Autoencoder Application in Identification of Counterfeit Pharmaceutical].

Near-infrared(NIR)As a fast and non-destructive testing technology, spectroscopy techniques is very suitable for pharmaceutical discrimination. Autoencoder network, as a hot research topic, has drawn widespread attention in machine learning research in recent years. Compared with traditional surface learning algorithm models, Autoencoder network has more powerful modeling capability as a typical deep networks model. Based on the unsupervised greedy layer-wise pre-training, autoencoder trains the network layer by layer while minimizing the error in reconstructing. Each layer is pre-trained with an unsupervised learning algorithm, learning a nonlinear transformation of the input of each layer which is the output of the previous layer. Pre-whitening process could get the inner structural features of the data more effectively. The supervised fine-tuning is followed with the unsupervised pre-training which sets the stage for a final training phase. The deep architecture is fine-tuned with respect to a supervised training criterion with gradient-based optimization. In this paper, firstly, the preprocessing step and pre-whitening transformation were used to treat near-infrared spectroscopy data of erythromycin ethylsuccinate, The pre-whitening transformation would reduce the correlation of the features, which gave each feature the same variance. Experimental results showed that the pre-whitening process had improved the classification accuracy of Sparse Denoising Autoencoder (SDAE) effectively. The SDAE with two hidden layers combined with pre-whitening was used to build the classification model for the identification of counterfeit pharmaceutical. The BP neural networks was compared with SVM algorithm for the classification accuracy and mean absolute difference (MAD). SDAE algorithm had higher classification accuracy than BP neural networks which had the same network structure with the SDAE networks, and SDAE algorithm also performed better than the SVM algorithm when the train datasets achieved a certain amount. As to the generalization performances, SDAE algorithm had less mean absolute difference of classification accuracy than SVM and BP Neural Networks. This result showed that SDAE algorithm could be effectively used to discriminate the counterfeit pharmaceutical.

[Drug discrimination by near infrared spectroscopy based on summation wavelet extreme learning machine].

As an effective technique to identify counterfeit drugs, Near Infrared Spectroscopy has been successfully used in the drug management of grass-roots units, with classifier modeling of Pattern Recognition. Due to a major disadvantage of the characteristic overlap and complexity, the wide bandwidth and the weak absorption of the Spectroscopy signals, it seems difficult to give a satisfactory solutions for the modeling problem. To address those problems, in the present paper, a summation wavelet extreme learning machine algorithm (SWELM(CS)) combined with Cuckoo research was adopted for drug discrimination by NIRS. Specifically, Extreme Learning Machine (ELM) was selected as the classifier model because of its properties of fast learning and insensitivity, to improve the accuracy and generalization performances of the classifier model; An inverse hyperbolic sine and a Morlet-wavelet are used as dual activation functions to improve convergence speed, and a combination of activation functions makes the network more adequate to deal with dynamic systems; Due to ELM' s weights and hidden layer threshold generated randomly, it leads to network instability, so Cuckoo Search was adapted to optimize model parameters; SWELM(CS) improves stability of the classifier model. Besides, SWELM(CS) is based on the ELM algorithm for fast learning and insensitivity; the dual activation functions and proper choice of activation functions enhances the capability of the network to face low and high frequency signals simultaneously; it has high stability of classification by Cuckoo Research. This compact structure of the dual activation functions constitutes a kernel framework by extracting signal features and signal simultaneously, which can be generalized to other machine learning fields to obtain a good accuracy and generalization performances. Drug samples of near in- frared spectroscopy produced by Xian-Janssen Pharmaceutical Ltd were adopted as the main objects in this paper. Experiments for binary classification and multi-label classification were conducted, and the conclusion proved that the proposed method has more stable performance, higher classification accuracy and lower sensitivity to training samples than the existing ones, such as the BP neural network, ELM and-ELM by particle swarm optimization.