平均分布差异最小化的NIR标定迁移方法研究

凭借高效、无损和环保的优点,近红外光谱在多个领域广泛用作物质快速分析方法的同时,仍面临着光谱标定模型生命周期短,构建仪器标定迁移方法的标准样品难以获得和保存等问题.在化学计量学文献中,迁移方法通常能够矫正主从仪器之间的光谱差异,但绝大多数方法都需要在两台仪器相同条件下测量一组迁移标准样品.虽然样品数目不必过多,但总体上...

NIR Calibration Transfer Method Based on Minimizing Mean Distribution Discrepancy

With the advantages of high efficiency, non-destructive and environmental protection, NIR is widely used in many fields to rapidly analyse substances. However, it is still faced with the problems of the short life cycle of spectral calibration model and difficulty obtaining and preserving standard samples for instrument calibration transfer method. In the stoichiometric literature, transfer methods usually correct the spectral differences between master and slave instruments. Most methods need to measure a set of transfer standard samples under the same conditions of two instruments. Although the number of samples does not need to be too much, generally speaking, it must be well selected to ensure a successful transfer. The Kennard-Stone algorithm is the main algorithm for selecting representative sample subset in the master-slave instrument. In determining the standard sample, it is assumed that the master instrument has found the standard sample, and the selected sample set needs to be measured in the slave instrument. It is only possible when the transferred sample is sufficiently stable, but this cannot be guaranteed in the near-infrared spectroscopy technology. If it is assumed that the sample of the slave instrument is used as the standard sample, the master instrument is replaced by the slave instrument in consideration of the change of the spectrum light source in the new industrial application, so it is no longer available. Based on these problems, this paper proposes a method of minimizing mean distribution discrepancy calibration transfer for NIR (MCT), without considering the standard sample (standard-free) of the slave instrument, due to the multicollinearity of NIR spectroscopy data, this method first assumes that there is a subspace of the partial least squares of the master-slave instrument, and then the spectral data of the master-slave instrument are projected to the common subspace respectively; then, the mean distribution discrepancy minimization algorithm is introduced, that is, the mean distribution (center point) representation function of the master-slave spectral data in the subspace is given Function to minimize the discrepancy between the mean distribution (center point) of the two spectra, and maximize the covariance of the main instrument spectrum after projection to derive the optimal subspace; finally, the main spectrum samples and the secondary spectrum prediction samples are projected into the partial least squares subspace respectively, and the regression model is obtained by using the main spectral data, and the modified model can be used to predict the secondary spectral concentration. Through the test and research on the corn data set and the wheat data set, it is proved that the prediction effect of this method is improved compared with SBC, PDS, CCACT, TCR and MSC. The experiment shows that MCT can achieve a lower prediction value.