Analysis Of Coal Properties By Near Infrared Spectroscopy Based On Partial Least Squares

Fast and non-destructive detection of coal’s ash, volatile and sulfur content with diffuse reflectance near-infrared spectroscopy was studied in this paper. And the on-line detection system was built for coal washery. The main work is summarized as follows.1. Collect diffuse reflectance near infrared spectra of190clean coal samples while control the temperature and humidity of laboratory. Choose150coal samples from the whole as calibration set. Calibration set was for the regression model building of ash, volatile and sulfur. And the rest were prediction set, which was used for detecting the model’s prediction precision. Select partial least squares regression (PLSR) to build the regression model after the comparison of principle component regression (PCR) and PLSR. Then different outlier detection methods, spectroscopy pre-processing and band selecting algorithms were used for model’s optimization. The best optimizing method for ash’s regression model was deleting outliers with studentized residual, selecting bands qualitatively. For volatile’s regression model, the best optimizing method was combining studentized residual with selecting bands according the relativity of spectra and coal’s volatile. For sulfur, the best method was studentized residual and orthogonal signal correction.2. Design and construct the online belt detection system based on factory environment. In order to detecting successively, the rotating pan was replaced by conveyor belt. It helps the automatic production of washery, and the coal after detection could be used for the commercial coal production. Collecting spectroscopy online can increase the scanning area. But stability of the online detection system was worse. In order to lower the noise of collecting spectroscopy, each sample was scanned15times and the average of15spectra was the near infrared spectrum of this sample. Then, build the PLSR model between the spectra online and coal’s ash, volatile and sulfur, and optimize the model. Model’s precision built by spectra collected online reached the model built by spectra collected offline.