Simultaneous Rapid Analysis of Multiple Nitrogen Compounds in Polluted River Treatment Using Near-infrared Spectroscopy and a Support Vector Machine

An intermittent aerobic process has been developed to effectively remove nitrogen from polluted rivers. In addition, a chemometric model was developed to achieve simultaneous rapid analysis of total nitrogen, ammonia nitrogen, and nitrite nitrogen based on near-infrared spectroscopy data combined with a support vector machine. An intermittent aeration process showed that total nitrogen decreased from 54.25 mg·L-1 to 0.64 mg·L-1. Ammonia nitrogen decreased significantly in the aeration stage, but increased in the non-aeration stage. Eventually, ammonia nitrogen decreased from 45.04 mg·L-1 to 0.57 mg·L-1. Nitrite nitrogen and nitrate nitrogen increased in the aeration stage, but decreased in the non-aeration stage. The concentration ranges of nitrite nitrogen and nitrate nitrogen were, respectively, 0.05~31.40 mg·L-1 and 0~0.38 mg·L-1. The 138 water samples were collected during the intermittent aeration process, of which 116 samples were used as the calibration set and the remaining 22 samples were used as a test set in modeling. The actual concentration values and the near-infrared spectroscopy data were used as input of the models. Then the corresponding calibration values and predication values were output by the models. All the samples were scanned with near-infrared spectroscopy from 4,000~12,500 cm-1 and measured by chemical methods. Principal component analysis of raw near-infrared spectral data showed that the matrix dimension of spectral data was significantly reduced, which decreased from 2,203×106 to 6×106. Support vector machine models of total nitrogen, ammonia nitrogen, and nitrite nitrogen showed that the calibration correlation coefficient (R2) of calibration values and actual values were, respectively, 0.9561, 0.9661, and 0.9702, with the root mean square error of cross validation (RMSECV) being 0.09372, 0.04749, and 0.03187. The test results of support vector machine models of total nitrogen, ammonia nitrogen, and nitrite nitrogen showed that the predication correlation coefficient (R2) of prediction values and actual values were, respectively, 0.9616, 0.9410, and *e-mail: [email protected] DOI: 10.15244/pjoes/70002 2014 Huang J., et al.