A tobacco leaf substitution judgment method based on multi-modal hybrid fusion

Abstract

This invention discloses a method for determining tobacco leaf substitution based on multimodal hybrid fusion, comprising the following steps: (1) obtaining the "static" quality feature matrix X of tobacco leaves in the cigarette formula and the "static" quality feature matrix Y of stock tobacco leaves; (2) calculating the Mahalanobis distance value d1 and performing normalization processing to obtain the "static" quality similarity value a1; (3) obtaining the "dynamic" quality feature matrix U of tobacco leaves in the cigarette formula and the "dynamic" quality feature matrix V of different stock tobacco leaves; (4) calculating the Mahalanobis distance value d2 and performing normalization processing to obtain the "dynamic" quality similarity value a2; (5) combining a1 and a2 and substituting them into f(a) for calculation, and finally determining whether the stock tobacco leaves can replace the tobacco leaves in the formula and the probability of replacement risk.

Description

Technical Field

[0001] This invention belongs to the field of cigarette technology, specifically relating to a method for determining tobacco leaf substitution based on multimodal hybrid fusion. Background Technology

[0002] The quality and style of cigarettes are primarily determined by product designers through blending tobacco leaves from different origins, varieties, and grades. When a certain type of tobacco leaf is exhausted, it's necessary to rely on blending experience and sensory evaluation to manually select similar-quality tobacco leaves from hundreds of grades in stock for blending substitution. This process is highly subjective, labor-intensive, and lacks objective, scientific, and efficient technical means, making it difficult to comprehensively guarantee the stability of cigarette quality and style. If instrumental testing methods could be used, utilizing objective data and scientific techniques to assist in blending substitution, it would be of great significance in improving the stability of cigarette quality.

[0003] The quality and style of cigarette products are a manifestation of the effects of the chemical components produced during the combustion of tobacco leaves on the human sensory organs. Therefore, the composition and content of the intrinsic chemical components of tobacco leaves, to a certain extent, determine the sensory quality of the product. For example, compounds such as sugars, nitrogenous substances, and nicotine in tobacco leaves are closely related to aroma, smoke, and taste. Domestic and international researchers and industry experts have consistently devoted considerable effort to researching the chemical composition indicators and formulation maintenance of tobacco leaves, achieving certain research and application results.

[0004] In recent years, near-infrared spectroscopy (NIRS) technology has been widely used in tobacco quality analysis and evaluation, as well as formulation design and maintenance, due to its speed, efficiency, and rich quality information. It effectively characterizes the intrinsic chemical composition of tobacco leaves. However, NIRS only considers the correlation between tobacco quality and leaf blend formulation under "static conditions," neglecting the heat transfer behavior under combustion conditions during cigarette consumption. Therefore, it cannot truly represent the smoking quality of tobacco leaves, i.e., the quality characteristics of the smoke. Consequently, formulation substitution technologies based on NIRS have certain limitations.

[0005] Thermal analysis is one of the most widely used methods in the industry for establishing a bridge between tobacco leaves and flue gas. Thermogravimetric analysis (TG) is a technique that measures the weight (mass) of a sample in relation to its temperature or time under programmed temperature control and in different atmospheres. It can indirectly reflect the pyrolysis and combustion behavior of different tobaccos, and is a concrete manifestation of the "dynamic" quality characteristics of tobacco leaves under combustion conditions. Reports have documented the use of thermal analysis to identify the thermal conversion characteristics of tobacco, conduct tobacco leaf substitution experiments in formulations, and evaluate the batch quality stability of reconstituted tobacco and cigarette paper.

[0006] Multimodal fusion, also known as multi-source information fusion or multi-sensor fusion, refers to the process of combining information from two or more modalities for prediction. In the prediction process, a single modality typically cannot contain all the effective information needed to produce accurate prediction results; multimodal fusion combines information from two or more modalities, supplementing the information, broadening the coverage of information contained in the input data, improving the accuracy of the prediction results, and enhancing the robustness of the prediction model. Multimodal fusion methods include early fusion, late fusion, and hybrid fusion. Early fusion addresses the inconsistency between the original data in each modality by first extracting feature representations from each modality and then fusing them at the feature level, i.e., feature fusion. Late fusion, also known as decision-level fusion, trains different models on different modalities first and then fuses the outputs of multiple models. Hybrid fusion combines early and late fusion, integrating the advantages of both but also increasing the complexity of the model structure and the difficulty of training.

[0007] To comprehensively characterize the smoking quality of cigarette formulations and raw tobacco leaves, this invention replaces the traditional method of tobacco leaf substitution in cigarette formulations, which relies on sensory evaluation and expert experience, by combining chemical composition indicators, near-infrared spectral data, and thermogravimetric analysis data, and employing objective data and scientific techniques. Summary of the Invention

[0008] This invention employs a multimodal hybrid fusion algorithm to comprehensively measure the quality similarity between tobacco leaves in the formula and tobacco leaves in the stock, based on three dimensions: intrinsic chemical composition indicators, near-infrared spectroscopy, and cigarette combustion thermogravimetric analysis. This enables objective judgment and intelligent recommendation of whether tobacco leaves in the stock can replace tobacco leaves in the cigarette formula.

[0009] The technical solution of the present invention is as follows:

[0010] A method for determining tobacco leaf substitution based on multimodal hybrid fusion, such as Figure 1 As shown, it includes the following steps:

[0011] (1) The intrinsic chemical composition index data and near-infrared characteristic spectrum data of tobacco leaves in cigarette formula are fused using a multimodal feature fusion algorithm based on convolutional neural network to obtain the "static" quality feature matrix X of tobacco leaves in cigarette formula; the intrinsic chemical composition index data and near-infrared characteristic spectrum data of different stock tobacco leaves are fused using a multimodal feature fusion algorithm based on convolutional neural network to obtain the "static" quality feature matrix Y of stock tobacco leaves.

[0012] (2) Calculate the Mahalanobis distance d1 between matrices X and Y, and normalize it to obtain the "static" quality similarity value a1 between the tobacco leaves in the formula and the stock tobacco leaves.

[0013] (3) Based on the main heat transfer and weight loss stages during the combustion of tobacco leaves, the characteristic spectrum segments of the thermogravimetric curve of tobacco leaves were screened to form the "dynamic" quality characteristic matrix U of tobacco leaves in cigarette formula and the "dynamic" quality characteristic matrix V of tobacco leaves in different stockpiles.

[0014] (4) Calculate the Mahalanobis distance d2 between matrices U and V, and normalize it to obtain the dynamic quality similarity value a2 between the tobacco leaves in the formula and the stock tobacco leaves.

[0015] (5) Combining the similarity values ​​of the "static" quality characteristics a1 and the "dynamic" quality characteristics a2 of the tobacco leaves, the decision fusion method based on fuzzy comprehensive evaluation is used to determine whether the tobacco leaves in the inventory can replace the tobacco leaves in the formula, and the probability of replacement risk.

[0016] Preferably, the chemical composition data of the tobacco leaves selected in step (1) consists of 70 types, and the near-infrared characteristic spectral band of the tobacco leaves is 4350-4450 cm⁻¹. -1 4500-4700cm -1 4800-4900cm -1 5050-5100cm -1 5220-5300cm -1 6050-6250cm -1 6400-6600cm -1 6950-7220cm -1 Table 1 shows the data of 70 chemical components in tobacco leaves. These 70 intrinsic chemical components were categorized and statistically analyzed according to alkalis, sugars, acids, and Amadori compounds. The correlation between the chemical components and near-infrared spectra was extracted by combining the absorption peak positions of functional groups in the near-infrared spectrum, and characteristic near-infrared spectral bands of tobacco leaves were screened. Figure 2 As shown.

[0017] Preferably, the steps for fusing using a multimodal feature fusion algorithm based on a convolutional neural network are as follows: Figure 3 As shown: First, the fully connected layer f, which is an index of intrinsic chemical composition... m and the fully connected layer f in the near-infrared characteristic spectral band n Multiplying them together forms a two-dimensional feature fusion transformation matrix A. m×n Then, for A m×n Max pooling is performed to form m feature nodes, which serve as the new fusion feature layer B. m Finally, a softmax layer is applied to B. mLogarithmic normalization is performed to compress variables and normalize them to the range [0,1], thus obtaining the final fusion feature layer; where X is the fusion feature matrix of tobacco leaves in the formula and Y is the fusion feature matrix of stock tobacco leaves.

[0018] Preferably, the formula for calculating the Mahalanobis distance d1 in step (2) is: Where X is the fusion feature matrix of "static" quality of tobacco leaves in cigarette formulation, Y is the fusion feature matrix of "static" quality of stored tobacco leaves, and Σ is the covariance matrix of X and Y; the normalization formula is: After normalization, the data is transformed into a range of 0-1, and the distance value (difference value) is transformed into a similarity value.

[0019] Preferably, the characteristic spectrum of the tobacco leaf thermogravimetric curve in step (3) is: 190℃-240℃, 320℃-360℃; which respectively form the dynamic quality characteristic matrix U of tobacco leaves in the formula and the dynamic quality characteristic matrix V of tobacco leaves in the stock.

[0020] Preferably, step (4) uses the Mahalanobis distance calculation formula: Where U is the "dynamic" quality characteristic matrix of tobacco leaves in the cigarette formula, V is the "dynamic" quality characteristic matrix of stored tobacco leaves, and Σ is the covariance matrix of U and V; the normalization formula is: After normalization, the data is transformed into a range of 0-1, and the distance value (difference value) is transformed into a similarity value.

[0021] Preferably, the judgment method in step (5) is as follows: Substitute a1 and a2 into the formula r = 0.3 * f(a1) + 0.7 * f(a2) for calculation. If r > 0.5, then the stockpiled tobacco leaves can replace the tobacco leaves in the cigarette formula, where the replacement risk probability is 1 - r; if r ≤ 0.5, then the stockpiled tobacco leaves cannot replace the tobacco leaves in the cigarette formula; where f(a) is a predefined membership function:

[0022] Table 1. Indicators of intrinsic chemical components in tobacco leaves

[0023] Serial Number Indicator Item Serial Number Indicator Item Serial Number Indicator Item Serial Number Indicator Item 1 Total alkaloids % 21 malonic acid mg / g 41 Cystine μg / g 61 Fru-Asnμg / g 2 Reducing sugar % 22 Succinic acid mg / g 42 Methionine μg / g 62 Fru-Asp μg / g 3 Total sugar % 23 malic acid mg / g 43 Isoleucine μg / g 63 Fru-Glnμg / g 4 Total nitrogen % 24 Citric acid mg / g 44 Leucine μg / g 64 Fru-Gluμg / g 5 K％ 25 vanillic acid mg / g 45 Tyrosine μg / g 65 Fru-Ileμg / g 6 chlorine% 26 Myristic acid mg / g 46 Phenylalanine μg / g 66 Fru-Leuμg / g 7 pH 27 palmitic acid mg / g 47 4-Aminobutyric acid μg / g 67 Fru-Tyrμg / g 8 starch% 28 Linoleic acid mg / g 48 Lysine μg / g 68 Fru-Phe μg / g 9 Dichloromethane extract % 29 Oleic acid + linolenic acid mg / g 49 histidine μg / g 69 Fru-Trp μg / g 10 Solanine mg / g 30 stearic acid mg / g 50 tryptophan μg / g 70 Neophytadiene mg / g 11 sulfate mg / g 31 Eicosanoids mg / g 51 Arginine μg / g 12 phosphate mg / g 32 Aspartic acid μg / g 52 Proline μg / g 13 Mg% 33 Threonine μg / g 53 Glu-Anμg / g 14 Ca% 34 Serine μg / g 54 Fru-Ambμg / g 15 Neochlorogenic acid mg / g 35 Asparagine μg / g 55 Fru-Hisμg / g 16 chlorogenic acid mg / g 36 Glutamic acid μg / g 56 Fru-Proμg / g 17 cryptochlorogenic acid mg / g 37 Glutamine μg / g 57 Fru-Valμg / g 18 Hyoscyamine mg / g 38 glycine μg / g 58 Fru-Thrμg / g 19 Rutin mg / g 39 alanine μg / g 59 Fru-Glyμg / g 20 oxalic acid mg / g 40 Valine μg / g 60 Fru-Ala μg / g

[0024] The beneficial effects of this invention are:

[0025] This invention presents a multimodal hybrid fusion-based method for determining tobacco leaf substitution. It employs a multimodal hybrid fusion algorithm to objectively characterize the complete quality features of tobacco leaves in cigarette formulations and those in stock, considering three dimensions: intrinsic chemical composition indicators, near-infrared spectroscopy, and cigarette combustion thermogravimetric analysis. This method performs a comprehensive quality similarity measurement of tobacco leaves in the formulation and those in stock, enabling objective judgment and intelligent recommendation of whether stocked tobacco leaves can substitute for those in the cigarette formulation. Furthermore, it provides a probability of formulation replacement risk for replaceable stocked tobacco leaves. This invention's method is objective, scientific, and efficient, enabling the determination of whether stocked tobacco leaves can substitute for those in the cigarette formulation and providing intelligent substitution recommendations, thus comprehensively ensuring the stability of cigarette quality and style. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a method for determining tobacco leaf substitution in cigarette formulations based on multimodal hybridization.

[0027] Figure 2 This is a screening diagram of near-infrared spectral characteristic bands of tobacco leaves according to the present invention.

[0028] Figure 3 This is a schematic diagram of the tobacco leaf "static" quality feature fusion algorithm of the present invention. Detailed Implementation

[0029] The present invention will be further described in detail below through specific embodiments, but it should not be construed as limiting the scope of the invention to the following examples. Various substitutions and modifications made based on ordinary technical knowledge and conventional methods in the art without departing from the spirit of the invention should be included within the scope of the invention.

[0030] Example: Determining and recommending whether stockpiled tobacco leaves can substitute for tobacco leaves in cigarette formulations.

[0031] The information on tobacco leaves and stockpiled tobacco leaves used in the formulation selected in this example is shown in Table 2.

[0032] Table 2 Information on tobacco leaves and stockpiled tobacco leaves used in the formulation selected in this embodiment.

[0033]

[0034]

[0035] The specific steps for determining whether stored tobacco leaves can substitute for tobacco leaves in cigarette formulations are as follows:

[0036] (1) Divide the tobacco leaf sample in the cigarette formula into two parts evenly; one part is used for near-infrared spectroscopy detection and analysis, and the other part is used for thermogravimetric analysis.

[0037] (2) Near-infrared spectral analysis of tobacco leaves in the formula and stored tobacco leaves, specifically:

[0038] The tobacco leaf sample was dried in a 40℃ oven for 2 hours, then pulverized by cyclone milling and passed through a 40-mesh sieve. 3g of the tobacco leaf powder was weighed and placed in a sample cup, and the sample was scanned at 4000-9000 cm⁻¹. -1 For each sample, the spectrum was calculated by repeating the operation 10 times and taking the average value of the spectra.

[0039] By combining the absorption peak positions of functional groups of various molecules in the near-infrared spectrum, the correlation between the chemical composition indicators of tobacco leaves and the near-infrared spectrum is extracted, and the characteristic near-infrared spectral bands of tobacco leaves are screened. For example... Figure 2 As shown, the near-infrared spectral characteristic bands of tobacco leaves were screened: 4350-4450 cm⁻¹. -1 4500-4700cm -1 4800-4900cm -1 5050-5100cm -1 5220-5300cm -1 6050-6250cm -1 6400-6600cm -1 6950-7220cm -1 .

[0040] By using the near-infrared prediction model for 70 intrinsic chemical components of tobacco leaves, the intrinsic chemical component data of tobacco leaves were obtained, as shown in Table 3.

[0041] Table 3. Data on 70 intrinsic chemical components of tobacco leaves (taking samples 1#-5# as examples).

[0042]

[0043]

[0044] (3) Targeting the "static" quality characteristics of tobacco leaves, a multimodal feature fusion algorithm based on convolutional neural networks is used. The convolutional neural network algorithm is employed to fuse the intrinsic chemical composition index data and near-infrared feature spectrum data of tobacco leaves in the formula and stored tobacco leaves. First, a fully connected layer f of the near-infrared feature spectrum is constructed. m and the characteristic spectral segment of the thermogravimetric curve of the fully connected layer f n Multiplying them together forms a two-dimensional feature fusion transformation matrix A. m×n Then, for A m×n Max pooling is performed to form m feature nodes, which serve as the new fusion feature layer B. m Finally, a softmax layer is applied to B. m Logarithmic normalization was performed to compress the variables and normalize them to the range [0,1], thus obtaining the final tobacco leaf fusion feature matrix X and the stock tobacco leaf fusion feature matrix Y in the formulation, as shown in Table 4.

[0045] Table 4 shows the blending feature matrix X and the blending feature matrix Y of tobacco leaves in the formulation.

[0046]

[0047] (3) Calculate the Mahalanobis distance d1 between matrices X and Y, and after normalization, use it as the similarity value a1 between the tobacco leaves in the formula and the stock tobacco leaves. The results are shown in Table 5.

[0048] Table 5 shows the similarity values ​​(a1) of the fusion matrix of the "static" quality characteristics of tobacco leaves and stored tobacco leaves in the formulation.

[0049]

[0050]

[0051] (5) Thermogravimetric analysis of tobacco leaves: A TGA / DSC 1LF thermogravimetric analyzer from Mettler Toledo, Germany, was used. (5.00±0.05) mg of sample was weighed and placed in a thermogravimetric alumina crucible. The sample was heated from 30℃ to 900℃ at a nitrogen flow rate of 30 mL / min and a heating rate of 10℃ / min to obtain the thermogravimetric curves of the tobacco leaves. Based on the main heat transfer and weight loss stages during tobacco combustion, characteristic spectral segments of the tobacco leaf thermogravimetric curves were selected: 190℃-240℃ and 320℃-360℃. The characteristic spectral segment data of the thermogravimetric curves of tobacco leaf samples in the formulation and those of stored tobacco leaf samples were selected to form the "dynamic" quality characteristic matrix U of the tobacco leaves in the formulation and the "dynamic" quality characteristic matrix V of the stored tobacco leaves, respectively.

[0052] (6) Calculate the Mahalanobis distance d2 between matrices U and V, and after normalization, use it as the similarity value a2 between the "dynamic quality" of tobacco leaves in the formula and the stock tobacco leaves. The results are shown in Table 6.

[0053] Table 6 shows the similarity values ​​(a2) of the fusion matrix of the "dynamic" quality characteristics of tobacco leaves and stored tobacco leaves in the formulation.

[0054]

[0055] (7) For each stock of tobacco leaves, the "static" quality characteristic similarity value a1 and the "dynamic" quality characteristic similarity value a2 are used. A decision fusion method based on fuzzy comprehensive evaluation is employed to calculate the final judgment value r of whether the stock of tobacco leaves can replace the tobacco leaves in the formula. A formula replacement risk probability 1-r is given for replaceable tobacco leaves. A list of all stock of tobacco leaves that can be replaced in the formula is provided, sorted from largest to smallest r value, and recommended to the formula maker, as shown in Table 7. That is: Substitute a1 and a2 into the formula r = 0.3 * f(a1) + 0.7 * f(a2) for calculation. If r > 0.5, the stock of tobacco leaves can replace the tobacco leaves in the cigarette formula, where the replacement risk probability is 1-r; if r ≤ 0.5, the stock of tobacco leaves cannot replace the tobacco leaves in the cigarette formula; where f(a) is a predefined membership function:

[0056] Table 7: List of tobacco leaves in stock that can be substituted for other products in the formula.

[0057]

Claims

1. A method for determining tobacco leaf substitution based on multimodal hybrid fusion, characterized in that, Includes the following steps: (1) The intrinsic chemical component index data and near-infrared characteristic spectrum data of tobacco leaves in cigarette formula are fused using a multimodal feature fusion algorithm based on convolutional neural networks to obtain the "static" quality feature matrix X of tobacco leaves in cigarette formula; the intrinsic chemical component index data and near-infrared characteristic spectrum data of different stock tobacco leaves are fused using a multimodal feature fusion algorithm based on convolutional neural networks to obtain the "static" quality feature matrix Y of stock tobacco leaves; the steps of the fusion using the multimodal feature fusion algorithm based on convolutional neural networks are as follows: First, the fully connected layer f of the intrinsic chemical component index is... m and the fully connected layer f in the near-infrared characteristic spectral band n Multiplying them together forms a two-dimensional feature fusion transformation matrix A. m×n Then, for A m×n Max pooling is performed to form m feature nodes, which serve as the new fusion feature layer B. m Finally, a softmax layer is applied to B. m Logarithmic normalization is performed to compress variables and normalize them to the range [0,1], thus obtaining the final fusion feature layer; where X is the fusion feature matrix of tobacco leaves in the formula and Y is the fusion feature matrix of stock tobacco leaves; (2) Calculate the Mahalanobis distance d1 between matrices X and Y, and normalize it to obtain the "static" quality similarity value a1 between the tobacco leaves in the formula and the stored tobacco leaves; the formula for calculating the Mahalanobis distance d1 is: Where X is the fusion feature matrix of "static" quality of tobacco leaves in cigarette formulation, Y is the fusion feature matrix of "static" quality of stored tobacco leaves, and Σ is the covariance matrix of X and Y; the normalization formula is: (3) Based on the main heat transfer and weight loss stages during the combustion of tobacco leaves, the characteristic spectrum segments of the thermogravimetric curve of tobacco leaves were screened to form the "dynamic" quality characteristic matrix U of tobacco leaves in cigarette formula and the "dynamic" quality characteristic matrix V of tobacco leaves in different stockpiles. (4) Calculate the Mahalanobis distance d2 between matrices U and V, and normalize it to obtain the "dynamic" quality similarity value a2 between the tobacco leaves in the formula and the stock tobacco leaves; Mahalanobis distance calculation formula: Where U is the "dynamic" quality characteristic matrix of tobacco leaves in the cigarette formula, V is the "dynamic" quality characteristic matrix of stored tobacco leaves, and Σ is the covariance matrix of U and V; the normalization formula is: (5) Based on the similarity values ​​of the "static" quality characteristics of tobacco leaves (a1) and the "dynamic" quality characteristics (a2), determine whether the stockpiled tobacco leaves can replace the tobacco leaves in the formula, and the probability of replacement risk. The determination method is as follows: substitute a1 and a2 into the formula r = 0.3 * f(a1) + 0.7 * f(a2) for calculation. If r > 0.5, the stockpiled tobacco leaves can replace the tobacco leaves in the cigarette formula, where the probability of replacement risk is 1 - r; if r ≤ 0.5, the stockpiled tobacco leaves cannot replace the tobacco leaves in the cigarette formula. Wherein, f(a) is:

2. The tobacco leaf substitution judgment method based on multimodal hybrid fusion according to claim 1, characterized in that, The chemical composition data of the tobacco leaves selected in step (1) consisted of 70 types, and the near-infrared characteristic spectral band of the tobacco leaves was 4350-4450 cm⁻¹. -1 4500-4700cm -1 4800-4900cm -1 5050-5100cm -1 5220-5300cm -1 6050-6250cm -1 6400-6600cm -1 6950-7220cm -1 .

3. The tobacco leaf substitution judgment method based on multimodal hybrid fusion according to claim 1, characterized in that, The characteristic spectrum of the tobacco leaf thermogravimetric curve in step (3) is: 190℃-240℃, 320℃-360℃.

Images