[A statistical wavelength selection method of infrared spectra used in exhaust gas detection]

In order to extract the quantitative features of the rare pollution components from noisy atmospheric infrared spectra and thus create calibration models, a wavelength selection method based on statistic theory is proposed in this paper. In this method, an objective function is defined based on the estimation of spectral noise level at every wavelength position. Because the size of the wavelength subset is also included in the function, the model size will not become too big during the minimization of the error of the calibration model. To test the performance of this method, the wavelength subsets of measured spectra with background noises were selected and the calibration models were then created using neural network technique for three pollution gases, respectively. The test showed that the sizes of the selected wavelength subsets accorded with the calculated results. The subset sizes were less than 2% of the total wavelength points. Meantime, the spectral noises were also restrained markedly in the calibration model because of the wavelength selection. The experimental results proved the validity of the wavelength selection method.