Performance Comparison Analysis Research Of PLS And ANN Research On Tobacco Quantitative Model

With the increasing competition in the product market, cigarette companies arealso expanding in the requirement of production automation; the product quality hasbecome the fundamental way of the tobacco business existence and development. Ifenterprises want to seek to develop in the fierce market competition, it is not onlyneed to constantly enhance the capability of independent research, innovation andimproving service quality, but also need to strive to ensure the quality of the flagshipbrand. Only in this way can we make the quality of the product better, therebymaintain customer loyalty to the brand. Therefore, it is essential to improveforecasting performance of the tobacco index quantitative model for improving thequality of tobacco products.Near-infrared spectroscopy as a green analysis technology, has obtained the rapiddevelopment and wide application in the tobacco industry in the recent years. As anindirect analysis technique, near-infrared spectroscopy analyzes mainly by means ofchemo metrics. Currently in the domestic tobacco industry, people generally use themethod of partial least squares in quantitative analysis of tobacco, it also made a lot ofgood progress and achievements in the application. But along with the furtherapplication, PLS has been found not suitable for the analysis of all the ingredients intobacco, and when there are many external factors, the stability of model that iscreated by PLS will produce large fluctuations. As a non-linear modeling method,artificial neural network has better universality, whether it can make up the deficiencythat appears in the above modeling process, which will be the focus of this paper.To solve the above problem, this paper contra poses the indicators commonlyused to measure quality of tobacco and the influence of external factors from theperspective of practical application tobacco, and studies the differences in quantitativemodel for performance between least squares method and the artificial neural networkmethod, so that it can provide a reference for the establishment of future tobacco quantitative models, and then, it can make the model possess higher accuracy androbustness.First, this paper introduces the basic theory of partial least squares and artificialneural networks. The select of the main factor directly relates to the actual predictiveability of PLS quantitative model, the component number’s influence on the stabilityof quantitative model is mainly studied in article. This paper improves the existingmethod, and compares with two methods that currently used in the modelperformance and the stability of forecast. It turned out that the improved method canreduce the complexity of the model and increase the accuracy of the model. Then thepaper introduces BP algorithm that the most widely used in artificial neural networks,later BP algorithm will also be an example, contrast the differences of artificial neuralnetworks and PLS in the performance of tobacco quantitative model.Next, the paper contrasts and analyzes the differences of artificial neuralnetworks and PLS in estimated performance of tobacco quantitative model from theperspective of the tobacco components. This paper embarks from the differentindicators of tobacco data, according to that the different indicators data has its ownunique characteristics, taking total sugar that the index itself the difference of thenature is bigger and potassium, nicotine and chloride for example, using the sametraining and test sets, respectively using PLS and artificial neural network study theaccuracy and applicability of these four indicators’ model. The results show that,contrapose different indicators of tobacco, PLS and artificial neural networks havedifferent applicability.Finally, the paper contrasts and analyzes the differences of artificial neuralnetworks and PLS in estimated performance of tobacco quantitative model from theperspective of the spectrum reproducible. Firstly, this paper describes theaccuracy of near-infrared spectrometer, the concept of machine noise and spectralresolution and pattern of manifestation that impact on spectrum, on this basis, withtobacco as experimental subjects, using the same data set, simulating that theinstrument impact on spectrum, respectively using partial least squares and artificialneural network establish a quantitative model, contrast the difference between the two models in predicting the performance to solve the problem of model applicability ofdifferent situations.