Elimination of abnormal spectra in coal-NIRS analytical model

For eliminating small amount of the abnormal spectra in modeling data set, the iterative clipping method, regarded the Euclidean distance as the criterion, has been put forward. The gradual clipping rules for abnormal spectra have been formulated with the characteristics of outlier, which deviate the subject distribution of the data set obviously. The spectra have different noise ratio for the same coal sample with different coal particle sizes, and the analytical results have different accuracy with different modeling methods. So, to verify the feasibility and stability of iterative clipping method, 3 kinds of spectral data set are collected under the sizes of 0.2 mm, 1 mm and 3 mm, and 2 kinds of analytical models are established by PLS method and BP-NN. The experimental result shows that: after eliminating the abnormal spectra in modeling data set, the predicted qualities by the analytical models are much better than before, the absolute value of error is reduced by about 0.3% times and the average decline is almost 20%. The iterative clipping method is able to accurately eliminate the outlier in the modeling data set, effectively improve the model learning accuracy and has the strong applicability.