Adaptive neuron-fuzzy inference system combined with principal components analysis for determination of compound thiamphenicol powder on near-infrared spectroscopy

The adaptive neuron-fuzzy inference system (ANFIS) is an effective modeling tool developed recently. It has gained much interest in solving classification and function approximation. In this paper, a new application based on ANFIS was presented for nondestructive determination of thiamphenicol powder drug with near-infrared (NIR) spectroscopy. The principal component analysis (PCA) technique was applied to extract relevant features from a number of spectral data in order to reduce the input variables of the ANFIS. The generated scores of the principal components (PCs) subsequently were used as the input variables of the ANFIS instead of the spectra data and constituted the principal component analysis-adaptive neuron-fuzzy inference system (PCA-ANFIS) model. A hybrid-learning algorithm which combined the least squares method and the gradient descent method was applied to optimize the parameters of PCA-ANFIS. Various optimum PCA-ANFIS models based on the conventional spectra and pretreated spectra (standard normal variate (SNV), multiplicative scatter correction (MSC) and the first-derivative) were established and compared. Experiment results indicated that the PCA-ANFIS model obtained from data sets achieved satisfactory accuracy, and the PCA-ANFIS approach with MSC pretreated spectra was found that it provided the best results. In order to present the advantages of PCA-ANFIS, the principal component regression (PCR) was also used, which was compared with PCA-ANFIS. Experiment results showed that the proposed PCA-ANFIS was more efficient than PCR.

► Adaptive neuron-fuzzy inference system (ANFIS) was used for nondestructive determination of thiamphenicol powder.
► Principal component analysis (PCA) was applied to reduce redundant variables from the original NIR spectra.
► The cross-validation technique was used for PCA-ANFIS model selection criterion.