Tobacco Chemical Constituents And Prediction Of Tobacco Quality By Artificial Neural Network

In this paper the general status of tobacco industry at home and abroad is briefly introduced. And the current research situation of tobacco chemistry is also presented. Tobacco leaves serve as the need of consumers. Their quality is still determined by the sense of taste and smell of professional, which requires a number of trained person. And the results are often affected by many factors. The evaluation of the tobacco leaves quality is very important in the research of tobacco chemistry. This evaluation can be performed by some chemical data of tobacco leaves. It is significant to research the relation between the chemical usual item and sensory property, or to induce the ration from the tobacco chemical data so that to evaluate the tobacco leaves quality, or to carry out statistical analysis of tobacco chemical data in order to know the influence of tobacco chemical data on the sensory property. Recently, some researchers introduce the neural network to evaluation of tobacco quality, but they only used limited samples and data and did not deeply study performance of neural network in prediction of tobacco quality.In prediction of tobacco quality by neural network, at first it is need to use reliable data. In this paper the analytical methods recommended by CORESTA is used. The chemical items include total sugar, reducing sugar, chloride, nicotine, ammonia, nitrate, total nitrogen and total volatile bases. These eight chemical items are used as the usual chemical items of tobacco leaves. And author has collected 827 samples with the eight items and 95 samples with the ten items, which are total sugar,reducing sugar, total nitrogen, total nicotine, total volatile bases, organic acid, total phenols, petroleum ether extract, chloride and potassium oxide. The analytical methods of amino acid are also studied. The eighteen kinds of amino-acid in 95 samples above mentioned are analyzed. In addition, author has developed the quantitative analytical method of volatile components in tobacco leaves and carried out the analysis of 30 samples. Finally, these samples data are employed as input of network.It is well known that the chemical compositions in tobacco is very complex. There are 5868 kinds compounds in tobacco, which have been identified. Among them 1872 kinds compounds exclusively belong to tobacco leaves, and 2824 kinds compounds only come from smoke, but 1172 kinds compounds are jointly possessed by tobacco leaves and smoke. The quality of tobacco vary markedly with growth, environment, soil, climate, culture, curing. Tobacco leaves were classified according to locality of production, position on the stalk, from which the leaves have originated, and factors such as their color, quality, time and ripeness at harvest. Accordingly if tobacco quality would be predicted by neural network, it is essential that the database of tobacco leaves will be set up, in which there are bijection relations between sensory property and chemical items in tobacco leaves. The database of tobacco leaves quality in this paper has developed with Delphi 7.0 and SQL Server 2000 and with C/S model due to the practical need of manufactory. On the one hand, the database can serve as the management of data in tobacco leaves and the research & development of products. On other hand, it can build a foundation for the artificial prediction of tobacco quality. The database was given by the following remarks from the retrieve system of science and technology outcome: "The database has been build with object oriented programming Delphi and SQL Server. It is not reported in below list literature ".Tobacco leaves are classified roughly with prior information(e.g. locality, position, color). And the model recognition system of tobacco leaves quality with backpropagation neural network has been build. The system has the ability of the variable learning rate and the momentum modification. The predictive results of thesystem are better than that reported by literature. The influences of chemical items on predictive results are discussed. The effect order is as follows: The influences of volatile constituents( I ) > the influences of(II) that were given by the addition of amino-acid and common chemical items. ( II ) > the influences of the amino-acid(ni).(III) > the influences of common chemical items(IV)(it contains ten chemical items). (IV) > the influences of usual chemical items(it is eight chemical items)( V) .In this paper the problem of BP network and the differences between BP and RBF has been explored. The optimizing methods for the parameters of RBF has been brought forth. The test calculating result has showed that the time calculated by RBF is shorter than that by BP, but the predicting result by RBF is consistent with BP network. The better predicting results are obtained, the more samples are required. In this paper it has been pointed out that when the tobacco quality would be predicted, it is not sufficient only to use feedforward neural network. Furthermore, the parameters of network need to further optimize so that to make network convergent and robust.According to the theory of genetic algorithms, the advantages of simple genetic algorithms have been discussed. And the simple genetic algorithms have been employed to assist and design weight and bias of BP neural network. The result predicted by GA-BP network is noticeably superior to that by BP network. The relative error is small. And the convergent and robust of the network is available.In this paper author has gone into the shortcomings and some further measures in genetic algorithms. To avoid prematurity, author has proposed the GA with self-adaptive crossovers/ mutations and elitist model, optimized the weight and bias of BP network and set up the model based on GA with self-adaptive crossovers/ mutations and elitist model and the ability of the partial best search. Experimental results show that the model can satisfactorily predicted the tobacco quality with stable performance and robustness, The relative error is within 6%. Thus the network can be used to the prediction of tobacco quality.In conclusion, in this paper the prediction of tobacco quality has been researched from a new view. With the development of the analytical technology and deeply studying of the quantitative analysis of tobacco chemical constituents and theoptimization of artificial neural network, it is undoubted that the tobacco quality will not only be predicted efficiently by neural network, but also resolve the puzzle of evaluating tobacco quality only by person's sense for long time. The above method can be generalized to use in the relevant industry, e.g. wine and tea industry.