Application research on coal analysis of Near Infrared Spectroscopy (NIRS) by intelligent algorithms

Traditional Modeling Methods (such as PCA, PLS, Neural Network) have the disadvantages of low determination precision and long analysis time resulted by lots of wavelength points in Near Infrared Spectroscopy (NIRS). Considering the global search ability of genetic algorithm, this paper proposed a new back-propagation neural network model which selects parts of the spectroscopy wavelength points as the modeling data base on genetic algorithm. The whole spectrum range is divided into 20 subintervals, whose all probable combinations compose the searching space. The determination coefficient denoted by R2 is selected as the fitness function. Through evolving generation by generation, the combination of subintervals with best fitness is selected as the modeling data. The experiment compared the results of proposed model with traditional back-propagation neural network model whose modeling data is the whole range of spectrum, after selection with genetic algorithm, the number of wavelength points is just about 65% of the whole spectrum range; the determination coefficient R2 of two methods are 0.9312 and 0.7382, respectively. The experiment results show that, region selection with genetic algorithm before modeling of coal analysis, the precision of prediction and the speed of analysis can be improved a lot.