Research On Machine Learning Algorithms For Discrimination Of Medicines Using Near Infrared Spectroscopy

Near Infrared Spectroscopy(NIRS) has been used widely in the health care agency and the department of supervision and administration due to its characteristics such as rapidity, simplicity and nondestructive measurements. Owning to the presence of the relatively weak and highly overlapping spectral bands in NIRS, machine learning combined with NIRS is introduced to identify counterfeit drugs. However, two issues of class-imbalanced data sets and Cost-sensitive usually are ignored in the discrimination(classification) of pharmaceutical drugs. Likewise, it is a bottleneck of the large-scale promotion that the problem of instability, coefficients correlation and slow computation time restrict the identification of medicines using NIRS.To address some existed problems in discrimination of medicines, machine learning algorithms such as SVM, ELM and CRC are improved in this paper. We propose three novel classification models to detect the genuine and counterfeit medicines, and the related work is proposed as follows:(1) To address the class-imbalanced datasets and Cost-Sensitive problem, a suitable selection of the reduction factor of the SCHs generated by the two classes of drug samples, respectively, the maximal margin classifier between SCHs can be constructed which can obtain good classification performance. To optimize the time-consuming in SCH model, an optimization of the parameters involved in the modeling by Cuckoo Search to obtain an excellent model for classification of medicines. So the SCH(CS) classification model is proposed in this paper. The problem of Cost-Sensitive in discrimination of medicines poses a challenge for classification model, the Cost-Sensitive SCH model is proposed to reduce average misclassification costs on real big class-imbalanced datasets.(2) Owing to to the problems of slow convergence and instability of ELM network, the dual activation functions constitute a kernel framework by extracting signal features and signal simultaneously, which can obtain a good accuracy and generalization performances. To address the problem of randomly assigned network parameters in the original ELM model, the CS algorithm is adopted to optimize ELM’s parameters. Stability and Convergence of classification are improved by this method. So the SWELM(CS) has a good performance to detect the genuine and counterfeit drug.(3) The overdetermined and nonsparsity problem of L2-morn model is difficult even to predict the categories in test NIR signals, the Gabor filter is adopted to obtain the more relevant factor vectors; to distinguish the multi-label classification in CRC and SRC models as usual, Local KNN is adopted for concentrating on two classes. It improves the classification accuracy for the binary classification. The CRC_GRLS model with low time-consuming and classification errors could be obtained because of CR model and coefficients distribution.The experiment adopts three novel classification models to identify the genuine and counterfeit medicines, and uses three NIRS data sets as the main objects for binary classification. The conclusion proves that the proposed method has more stable performance, higher classification accuracy and lower time-consuming than the existing ones. Finally, GUI-HMI is designed to obtain a good application in this paper.