Study On The Year Calibration Method Of Wheat Component Near-infrared Detection Model

Duo to the influence of environment, climate and new variety, the ingredient of wheat from different years will change and the corresponding spectrum will change as well, which may lead the detection model of wheat components based on near infrared reflectance spectroscopy not apply to new samples. However, re-establish a stable and accurate model need to re-prepare modeling samples and measure data, which consumes a lot of workload and time. Therefore, research on vintage calibration methods of wheat quality detection model to improve applicability and usage time has important practical significance and application prospect.First, taking samples from 2011 as main vintage samples, effect of abnormal samples, sample set partitioning methods, different spectra preprocessing and modeling methods to the performance of the models were compared and analyzed to determine the optimal quantitative analysis model of wheat protein. By combining with Monte Carlo cross validation(MCCV) and the second discriminant method, abnormal samples were judged and deleted. Optimized set Partitioning based on joint X-Y distance(O-SPXY) method was the most suitable sample set portioning method, and the optimal model established with optimized partial least squares regression(O-PLSR) and Smoothing + the first derivative + standard normal variety(SNV) preprocessing spectra. Principal component analysis score, Mahalonabis and Fisher value were used to qualitatively and quantitatively verify the applicability of 2011 model for 2012 samples, the results indicated that 2011 detection model could not performed well for 2012 samples, which was consistent with the results of cross-test method, prediction performance was poor with only R2 and RPD of 0.908, 3.290, and RMSEP of 0.447.Then, this thesis carried out the study of model vintage calibration from three aspects: spectral value correction, model updating and predicted results correction. Direct standardization algorithm, Euclidean distance- direct standardization algorithm, selecting add samples by artificial method, Selecting add samples based on Kennerd-store(KS) algorithm and Slope/Bias method respectively were used for calibrating model. On the basis of comprehensive analysis and comparison, selecting add samples based on KS algorithm with only standard samples of 14 was considered to be the most suitable method for the vintage calibration of wheat protein detection model. R2, RPD and RMSEP respectively were 0.957, 4.818 and 0.305, and the prediction performance of calibration model for 2012 samples achieved optimal.