High-accuracy prediction of carbon content in semi-coke by laser-induced breakdown spectroscopy

摘要： Semi-coke, as one kind of special coal resource with relatively high concentration carbon and low volatility, plays an important role in the coal chemical industry and city clean. Laser-Induced Breakdown Spectroscopy (LIBS) has been proved as an effective way to make an online analysis for the coal products. However, the lower volatility of semi-coke makes it hard to be pressed into a slice to get a smooth surface for a uniform laser-irradiation. Therefore, it is necessary to find an effective way to realize a high-accuracy LIBS detection for semi-coke application. Herein, two feasible ways of sample preparation are tried, one easy way is directly painting semi-coke powders on a tape that suitable for online fast monitoring, and the other complicated way is to mix binder into the semi-coke powder then that the uniformly and tightly coal slices are obtained, thus to improve the repeatability of measurement. Moreover, a totally new algorithm, support vector machine (SVM) combined with partial least square (PLS) regression(SVM-PLS), is utilized to establish an effective prediction model to make a high prediction accuracy. The coefficient of determination (R2), root mean square error of prediction (RMSEP), and average relative error (ARE) are 0.944, 0.90%, and 0.80%, respectively. In comparison with the result of the traditional PLS model, the SVM residual correction greatly improves the quality of the calibration curve and makes RMSEP and ARE reduced 0.17%, thus improves the prediction accuracy, which is much better than basic PLS regression. Meanwhile, the prediction error from binder mixed semi-coke slice is significantly reduced compared to that with directly painting samples on a tape. The maximum relative errors (MRE) are 2.71% and 5.19%, and the average RSD of the characteristic peaks are 12.1% and 16.2%, respectively, indicating that the easy way with painting sample on tape has little prediction uncertainties. Finally, in a three-day random test, the average RMSEP is 1.89% and average ARE is 1.74%, which also proves the binder additive can effectively reduce the matrix effect and enhance the stability of the spectrum for semi-coke measurement. The result proposes the proper LIBS analysis on semi-coke is a feasible and promising approach for on-line prediction of such kind of coal sample.