Near Infrared Model Maintenance Method Based on Spectral Projection Discrimination

Abstract

The present invention provides a near-infrared model maintenance method based on spectral projection discrimination, which is characterized in that it includes the following steps: Step 1. After the near-infrared spectrum model is constructed, perform a principal component analysis of the near-infrared spectrum on the corrected spectrum in the model, and use the principal The component spectrum is used to calculate the Mahalanobis distance of each sample, step 2, calculate the average Mahalanobis distance of all samples; step 3, calculate the average Mahalanobis distance of the sample set, and set the threshold of the Mahalanobis distance in the modeling set; step 4 1. Calculate the similarity between the predicted sample and the sample in the model; step 5, judge whether the near-infrared model needs to be maintained according to the value of the similarity. The near-infrared model maintenance method based on spectral projection discrimination of the present invention avoids frequent update and maintenance of the model, and makes the update management of the near-infrared model more scientific.

Description

technical field [0001] The invention relates to a near-infrared model maintenance method based on spectrum projection discrimination, which belongs to the field of smoke near-infrared model analysis and evaluation. Background technique [0002] The near-infrared spectrum of tobacco contains a large amount of conventional chemical value information, physical information, and appearance information of tobacco leaves. Therefore, near-infrared plays a very important role in the field of routine chemical value detection, tobacco leaf stability evaluation, and tobacco leaf quality evaluation. As far as the online near-infrared model is concerned, the background of the near-infrared instrument is relatively complex, involving factors such as temperature and humidity changes in the near-infrared spectrum, instrument vibration, and light source weakening. , process control, guiding production enterprises to process and other applications, if the online near-infrared model is inaccura...

AI Technical Summary

Problems solved by technology

[0002] The near-infrared spectrum of tobacco contains a large amount of conventional chemical value information, physical information, and appearance information of tobacco leaves. Therefore, near-infrared plays a very important role in the field of routine chemical value detection, tobacco leaf stability evaluation, and tobacco leaf quality evaluation. As far as the online near-infrared model is concerned, the background of the near-infrared instrument is relatively complex, involving factors such as temperature and humidity changes in the near-infrared spectrum, instrument vibration, and light source weakening. , process control, guiding production enterprises to process and other applications, if the online near-infrared model is inaccurate, it may mislead the enterprise's processing, bring corresponding semi-finished cigarette storage, and follow-up cigarette formula design to mislead, affecting product formula quality , bringing unnecessary losses to cigarette companies; therefore, how to quickly judge whether the near-infrared model is suitable for the product to be tested is particularly important

[0003] For conventional near-infrared model maintenance and discrimination methods, there are generally the following methods: (1) According to the label attributes in the near-infrared modeling set and the sample to be tested, after the near-infrared model is established, for the sample to be tested, If the label attribute of the sample to be tested is inconsistent with the label attribute in the modeling set, it is considered not suitable for the test sample, otherwise, it is considered to be suitable for the test sample. The label attributes of the sample mainly include place of origin, grade, brand, variety, and process. , this method is more intuitive, easy to operate and easy to implement, but it is also relatively extensive, because the label attributes of national tobacco leaves are relatively complicated, which will greatly limit the scope of application of the near-infrared tobacco leaf model, and cause a large number of misjudgments for online near-infrared ; (2) According to the spectrum in the modeling set, remove outliers, after preprocessing, perform PCA projection on the spectrum of the modeling set samples, calculate the average Mahalanobis distance (MD), global Mahalanobis distance, set the model's Threshold, for the sample to be tested, compare the Mahalanobis distance between the sample to be tested and the modeling set, and then compare it with the threshold of the model. If it is greater than the model threshold, it is judged as a sample that is not suitable for near-infrared model detection, and if it is less than the model threshold, it is judged as suitable for model detection. However, because PCA is essentially a linear projection method, PCA projection is heavily dependent on the amount of modeling samples. When the modeling sample size changes, the value of the Mahalanobis distance will often change. On the other hand, the PCA projection only depends on the linear projection of the sample corresponding to the maximum variation direction of the spectrum, which does not have the specific information of the substance to be tested in the tobacco leaf, and the Mahalanobis distance is often too large, but it is detected The value is relatively normal, and the detected value may be abnormal if the Mahalanobis distance is small. In the actual specific application process, it cannot meet the rapid infrared detection under the complex background system; (3) regular sampling comparison, due to each In the process of comparison, firstly, the samples must have statistical significance, which requires that there are generally more samples for comparison, and it is also necessary to do flow basic data. Yes, the comparison is also performed when the model prediction results are poor, which brings a lot of continuous maintenance and updating of the model

Images