Study On The Key Technology Of Electronic Nose In Intelligent Tobacco Barn Systems

Electronic nose is an intelligent olfactory-simulating machine, which is based on the theory of neural olfactory system and developed by using advanced intelligent information processing technique. It takes advantage of electronic technique, applies theories and methods in mathematics and information science to analyze the features of chemicals. In recent years, the applications of E-nose in the tobacco industry was rising. It is promising to use E-nose in identifying the counterfeit cigarette and controlling tobacco quality. In this thesis, E-nose is used to study the variation rules of the tobacco leaf odours during its flue-curing process.Flue-curing is an important phase in the tobacco production. Tobacco quality heavily depends on this phase. In recent years, bulk curing barn is used to conduct tobacco flue-curing nationwide, which has the advantage of high density of tobacco capacity, high use ratio of energy and ease to automation. However, in the practical process of flue-curing using bulk curing barn, the controlling behavior is based on presupposed baseline, which lacks of flexibility and needs specialized person to make adjustment frequently. Besides, traditional controlling of tobacco flue-curing is mainly relying on the color variation, ignoring the variation of tobacco leaf odours and its influence on the fragrance of the afterward curied tobacco. Therefore, a tobacco flue-curing predicting and controlling algorithm based on the odours detecting by using E-nose is proposed in this thesis. This algorithm is a part of the intelligent curing barn, aiming to realize the intelligentization of tobacco flue-curing which can intelligently make adjustment in curing phase according to different ambient conditions.First, the dataset sampled by E-nose during the flue-curing process is analyzed and the odour variation rules are obtained. Then the curing predicting and controlling model to predict the process of tobacco flue-curing according to the obtained odours variation rules is established. The implementation of the proposed algorithm contains techniques as data acquisition, data preprocessing, data dimensionality reduction, interference rejection and model training.Data acquisition is performed by using the odour-detecting E-nose sub-system to collect data during the flue-curing process according to the designed experiment scheme. Data preprocessing is conducted to extract useful information and suppress sensor noise and drift. PCA technique is used to realize the data dimensionality reduction. In order to meet the requirement of on-line detecting, two on-line PCA algorithms are compared with off-line PCA in terms of convergence rate and errors. As a result, stochastic gradient algorithm is chosen as on-line algorithm.Features obtained by PCA still contain a lot of interference. The independent component analysis(ICA), usually used for blind source separation, is used to separate interferential odours from the tobacco odours. For both on-line operating and off-line analysis, the on-line ICA algorithm and batch ICA algorithm are discussed respectively. To identify the sepatated signal as interferential component, multiple correlation is introduced. The discrimination and separation of interferential component are implemented by calculating the multiple correlations of each input/output component of ICA, temperature, humidity and air pressure. Meanwhile,comparing multiple correlations of input/output components of ICA may indicate the interference suppression effect to some degree. As to the odours of burning coal in the curing system, filters are designed to suppress it. Finally, the odour variation curve is obtained after the processing mentioned above.Aiming to achieve the flue-curing controlling and predicting through odours variation, traditional three-stage-curing for flue-cured tobacco is studied and the odorous substances released during the curing and its variation principle is analyzed, so as the feasibility is proven theoretically. The Back Propagation-Neural Network(BP-NN) is used to establishe tobacco flue-curing model based on tobacco odorous features. Model parameters are optimized by using genetic algorithm(GA), and the predicting model has a good performance. Besides, this thesis realize the practical on-line application. The experimental result validates the intelligent controlling scheme of system.