Pyrolysis of biomass is a promising method to produce a variety of gases, liquids (bio-oil), or solid materials (bio-char) that can then be used for fuel production. The product compositions depend largely on the variability of different proportions of protein, triglycerides, hemicellulose, cellulose, lignin etc., in the original biomass [1, 2]. Therefore, many researches focus on the study of the association between the biomass and the bio-products [3–5]. In this study, tobacco was chosen as the object for biomass. As a commercial product, the cigarette smoke (including water, tar and gases) released by tobacco pyrolysis reactions can satisfy the consumer's demand, not the tobacco itself, which is similar to the use of biomass as a renewable resource through the pyrolysis process.
Tobacco leaves cultivated in different areas have different styles, and their grades are based on the positions they grow on the stalk. The classification of tobacco style and grade is great important in the processes of tobacco blend design and product maintenance of cigarette . Current evaluation of tobacco style and grade mainly relies on artificial sensory analysis, which is subjective and relatively unstable . It is therefore necessary and urgent in tobacco industry to develop a new rapid and convenient method to automatically evaluate the tobacco style and grade.
Artificial intelligence has opened a new page in the field of data analysis. Many efforts have been devoted to develop automatic evaluation methods by using the advance of the machine learning (ML) algorithm with the data from the tobacco leaves and smoke. Early works mainly focused on the classification of tobacco cultivation area and growing position using near infrared spectra (NIR) due to its high efficiency and non-destructive characteristic. Hana et al. employed artificial neural networks (ANNs) to classify whether the burley tobacco grows in the USA or outside USA, and obtained high prediction accuracy . For the classification of tobacco style and grade, Ni et al. developed an improved and simplified K-nearest neighbors algorithm (IS-KNN) to discriminate more than 1000 Chinese flue-cured tobacco leaf samples with moderate accuracy . Their results suggest that it is better to establish classification model of tobacco grade from the same cultivation fields to get better classification results. By applying a combined random-forest (CRF) based on gas chromatography (GC) fingerprinting, Lin et al. managed to classify three different grades of “Furong” series cigarettes with accuracy up to 93.74% . Based on image processing on tobacco color, texture and shape, Zhang and Zhang implemented a two-level fuzzy comprehensive evaluation (FCE) and classified the tobacco leaves into three grades, but accuracy is achieved just 72% for the non-trained tobacco leaves . Recently, Gu et al. successfully built a relationship between chemical compounds and the aromatic quality of flue-cured tobacco leaves, by using support vector machine (SVM) algorithm with 22 chemical compounds selected by Relief-F-particle swarm optimization (R-PSO), and obtained high accuracy of 90.95% .Very recently, Wang et al. employed genetic algorithm (GA) to optimize the performance of SVM for data analysis of NIR spectroscopy sensors. They demonstrated that the GA could indeed improve the performance of SVM for tobacco classification based on NIR spectra although the accuracy is just 83% . All previous works have focused on the relationship of tobacco style and grade with either the components of the reactant (tobacco) or the product (smoke). In this study we choose to pay attention on the tobacco pyrolysis reaction process, which can be visually expressed by the thermogravimetric analysis (TGA). To the best of our knowledge, the automatic classification of tobacco planting area and growing position based on thermogravimetric analysis has not yet been reported.
TGA has been proven to be a useful tool to study the pyrolysis behavior and kinetics of pyrolysis process since it provides precise measurement depending on temperature and other experimental conditions that are well-known and well-controlled [15–17]. Investigations on biomass have shown that the differences in pyrolytic characteristics are mainly caused by the differences in the constituent and physical structure [18–20]. Studies on the pyrolysis of tobacco have also demonstrated that the tobacco pyrolysis DTG curve can be divided into different Gaussian peaks representing the thermal decomposition of individual components [21, 22]. For instance, the mass loss below 373K represents the evaporation of water ; The peaks between 373-473K corresponds to the thermal decomposition of sugars, nicotine, pectin and some other volatile species [24, 25]; and in the temperature of 474-873K the mass loss would be attributed to the pyrolysis of hemicellulose, cellulose and lignin, respectively [26–28]. Moreover, Baker and Bishop have demonstrated that the thermogravimetric analysis spectra of tobacco pyrolysis is highly reproducible under well-defined conditions . The thermogravimetric analysis data not only represent the tobacco pyrolysis characteristics, but also supply the information of the tobacco composition. Hence, it can be taken as an important index to evaluate tobacco planting area and growing position.
Recently, we demonstrated that thermogravimetric analysis data in conjunction with the normalized root mean square error (NRMSE) can be used to quantitatively evaluate the pyrolysis difference between tobacco of different stalk positions, planting areas and crop years . On this basis, we proposed a tobacco leaves substitute scheme in tobacco blend maintenance, and the results showed that this substitute scheme could achieve artificial substitute level . In this work, we further extended previous investigations and introduced SVM to the thermogravimetric analysis for the first time. Using the thermogravimetric analysis data as the information source, we demonstrated that automatic classification of tobacco planting area and growing position can be achieved with high accuracy as well as high efficiency by applying SVM. In our opinions, during the biomass-to-bio-oil conversion process, this new analysis method would be a promising way to exploring bio-oil quality prior to biomass pyrolysis production.