Study Of Data Mining In Near-infrared Spectral Analysis For Coal Qualities

It’s necessary to seek a rapid and accurate method of coal proximate analysis in order to guide the study of processing properties and utilization effectively, and promote the upgrading of the coal industry. The near-infrared spectroscopy(NIRS) technology has opened up a new way for the rapid analysis of coal with the advantages of the green, high efficient and online analysis. The effectiveness of the result of coal NIRS analysis firstly depends on the modeling sample. Therefore, the data mining problems of NIRS sample is the main research content in this paper, and the evaluation the performance of the relevant methods is completed through the unified BP and PLS model. At last, a simple coal NIRS analysis system is designed based on the above research.Firstly, 146 groups of coal samples collected from different regions are divided into two parts after the completion of the coal sample preparation. One part is used for obtaining NIRS of coal, and the other part is used for standard coal quality analysis. They are respectively used for input and output of quantitative analysis model.Aiming at the filtrating of abnormal coal spectrum sample, this paper deeply studied the methods of leave-one-out(LOO) cross-validation and distance discriminant analysis. This paper proposes an improved method of LOO by Kmeans and the iteration clipping based on mahalanobis distance. The experiments show that the accuracy and stability of the prediction model is improved obviously after being processed by the improved methods. Compared with BP neural network model(BP) partial-least-square(PLS) model based on original spectral data, root-mean-square error(RMSE) based on filtered data are decreased respectively from 0.046925、0.047087 to 0.017338、0.019975.Aiming at the problems existed in original spectral data such as noise, scattering effect and so on, the original spectrogram recovery has been processed by the methods of mean centering(MC) and standard Normal Variate(SNV) correction. The experiments show that the model prediction accuracy is improved to some degree after being processed by the above two methods. The prediction RMSE are decreased from 0.037537 to 0.021438.Aiming at the problems of multicollinearity, high dimension, easily over-fitting and time-consuming, this paper studies the spectral nonlinear feature extraction based on manifold learning algorithm and kernel method, including locally multi-dimensional scaling(LMDS), locally linear embedding(LLE) and kernal principle component analysis(KPCA). Aiming at KPCA, an improvement measure based on the two-iteration(TI-KPCA) is put forward. When the number of principal component after being processed by LMDS, LLE and TI-KPCA is set to 3, 7, 5 respectively, the cumulative contribution rate are all more than 99.9%, nearly covering all the information of the original spectrum. In addition, After being processed by TI-KPCA, the models prediction performances based on BP and PLS are all obviously better than before. The prediction RMSE are decreased from 0.07886 to 0.02769.Finally, simple analysis system about coal NIRS is designed based on the study of filtrating abnormal spectrum sample, recovering the original spectrogram and extracting spectral feature. The coal NIRS analysis system mainly includes the interface design and the database management. It is the preliminary exploration to realize coal NIRS online analysis.