A method for matching tobacco leaf and leaf group formula and a related device thereof

Abstract

The application provides a tobacco leaf and leaf group formula matching method and related equipment, including: obtaining one-to-one associated n kinds of tobacco leaves and n groups of tobacco leaf attributes, and m kinds of leaf group formulas, wherein each leaf group formula of the m kinds of leaf group formulas comprises at least one of the n kinds of tobacco leaves; creating a tobacco leaf knowledge graph of the n kinds of tobacco leaves according to the n kinds of tobacco leaves and the n groups of tobacco leaf attributes, and creating a leaf group formula interaction matrix according to the m kinds of leaf group formulas, the leaf group formula interaction matrix representing a component belonging relationship between the m kinds of leaf group formulas and the n kinds of tobacco leaves; and training a tobacco leaf and leaf group formula matching prediction function according to a graph convolutional neural network algorithm, the tobacco leaf knowledge graph and the leaf group formula interaction matrix, the tobacco leaf and leaf group formula matching prediction function being used to determine a matching degree of a to-be-matched leaf group formula and a to-be-matched tobacco leaf.

Description

Technical Field

[0001] This invention relates to tobacco leaves, and more particularly to a method and apparatus for matching tobacco leaves with leaf blend formulations. Background Technology

[0002] A tobacco blend is a formula containing a specific type of tobacco leaf. Different types of tobacco leaves are mixed in a certain proportion to create shredded tobacco. Because each type of shredded tobacco is made from a different blend of tobacco leaves according to the blend, the style of each type is influenced by the properties of the tobacco leaves themselves, including physicochemical and sensory properties.

[0003] In order to centralize the production of tobacco shreds, the collected tobacco leaves are often stored separately. Then, in the tobacco shred production process, specific tobacco leaves are first obtained from the inventory according to the composition of the tobacco leaf blend formula. Then, the quantity of specific tobacco leaves is determined according to the tobacco shred composition ratio to complete the production of tobacco shreds.

[0004] In the existing production process, leaf blending engineers, combining their historical experience with the results of tobacco tasting, first identify the tobacco leaves in the inventory, then determine the matching leaf blending formula, and proceed with subsequent tobacco shred production. However, this method relies on the subjective experience of the leaf blending engineers, and tobacco tasting can cause physical harm to them. Summary of the Invention

[0005] A method for matching tobacco leaves and leaf blends is provided to address the tedious and health-damaging work process of leaf blend engineers who rely on tasting and subjective experience to match tobacco leaves and leaf blends.

[0006] The first aspect of this application provides a method for matching tobacco leaves and leaf blend formulations, characterized in that it includes:

[0007] Obtain n types of tobacco leaves and n groups of tobacco leaf attributes that are associated one-to-one, as well as m types of leaf group formulas, wherein each of the m types of leaf group formulas includes at least one of the n types of tobacco leaves;

[0008] A tobacco knowledge graph is created based on the n types of tobacco leaves and the n groups of tobacco leaf attributes, and a leaf group formula interaction matrix is ​​created based on the m types of leaf group formulas. The leaf group formula interaction matrix represents the component relationship between the m types of leaf group formulas and the n types of tobacco leaves.

[0009] Based on the graph convolutional neural network algorithm, and the tobacco knowledge graph and the leaf group formula interaction matrix, a tobacco leaf and leaf group formula matching prediction function is obtained through training. The tobacco leaf and leaf group formula matching prediction function is used to determine the degree of matching between the leaf group formula to be matched and the tobacco leaf to be matched.

[0010] In one implementation, after the method obtains the tobacco leaf and leaf blend formula matching prediction function by training according to the graph convolutional neural network algorithm, the tobacco leaf knowledge graph, and the leaf blend formula interaction matrix, it further includes:

[0011] If the tobacco leaf to be matched and the leaf blend formula to be matched are obtained, the degree of matching between the tobacco leaf to be matched and the leaf blend formula to be matched is determined according to the matching prediction function of the tobacco leaf and leaf blend formula;

[0012] If the tobacco leaf to be matched is obtained, then the leaf blend formula that has the required degree of matching with the tobacco leaf is determined according to the tobacco leaf and leaf blend formula matching prediction function.

[0013] If the leaf blend formula to be matched is obtained, then the tobacco leaves that have the required degree of matching with the leaf blend formula are determined according to the tobacco leaf and leaf blend formula matching prediction function.

[0014] In one implementation, each of the n groups of tobacco leaf attributes includes multiple different tobacco leaf attribute categories, and each group of tobacco leaf attributes has the same tobacco leaf attribute category. Then, the step of creating a tobacco leaf knowledge graph based on the n types of tobacco leaves and the n groups of tobacco leaf attributes includes:

[0015] Obtain p tobacco leaf attribute categories from the n sets of tobacco leaf attributes;

[0016] Create a tobacco knowledge graph that includes entity nodes containing the n types of tobacco leaves and the p tobacco leaf attribute categories, as well as the edge relationships between the entity nodes.

[0017] In one implementation, the step of training a tobacco leaf and leaf blend formula matching prediction function based on a graph convolutional neural network algorithm, the tobacco leaf knowledge graph, and the leaf blend formula interaction matrix includes:

[0018] Randomly obtain training tobacco leaves and training leaf group formulas;

[0019] The graph embedding features of the training tobacco leaves and the training leaf group formula are determined based on the convolutional neural network, the tobacco leaf knowledge graph, and the training tobacco leaves and training leaf group formulas.

[0020] Based on the graph embedding features and the element values ​​associated with the training tobacco leaves and the training leaf group formula in the leaf group formula interaction matrix, a tobacco leaf and leaf group formula matching prediction function is trained.

[0021] In one implementation, determining the graph embedding features of the training tobacco leaves and the training leaf group formula based on the graph convolutional neural network, the tobacco leaf knowledge graph, and the training tobacco leaves and training leaf group formulas includes:

[0022] In the tobacco knowledge graph, target adjacent entity nodes and target edge relationships corresponding to the training tobacco leaves are determined, wherein the target edge relationships are associated with the target adjacent entity nodes.

[0023] The topological features of the training tobacco leaves and the training leaf group formula are determined based on the training tobacco leaves, the training leaf group formula, the target edge relationship, and the target adjacent entity nodes.

[0024] The graph embedding features of the training tobacco leaves and the training leaf group formulation are calculated based on the graph convolutional neural network algorithm and the topological features.

[0025] In one implementation, the method includes a convolutional layer, wherein determining the topological features of the training tobacco leaf and the training leaf group formula based on the training tobacco leaf, the training leaf group formula, the target edge relationship, and the target adjacent entity node includes:

[0026] Based on each layer of the convolutional layer, the target edge relationships and target adjacent entity nodes of each layer of the training tobacco leaf are determined in the tobacco leaf knowledge graph;

[0027] Based on the target edge relationships and target adjacent entity nodes of each layer, the topological features of the training tobacco leaf and the training leaf group formula in their respective layers are determined;

[0028] The step of calculating the graph embedding features of the training tobacco leaves and the training leaf group formulation based on the convolutional neural network and the topological features includes:

[0029] Determine the initial graph embedding features of the training tobacco leaves;

[0030] Based on the graph convolutional neural network algorithm, and using the initial graph embedding features, the topological features and graph embedding features of the respective layers that are matched are iteratively aggregated and calculated to determine the graph embedding features of the training tobacco leaves and the training leaf group formulation.

[0031] In one implementation, the n types of tobacco leaves, the n sets of tobacco leaf attributes, and the m types of leaf group formulas in the method are respectively n types of tobacco leaves in Glot vectorization, and the n sets of tobacco leaf attributes and m types of leaf group formulas.

[0032] In one implementation, the tobacco leaf attribute categories include tobacco leaf origin, tobacco leaf grade, tobacco leaf year, as well as multiple physicochemical indicators and multiple sensory indicators of tobacco leaves.

[0033] A second aspect of this application provides a matching device for tobacco leaves and leaf blend formulations, comprising:

[0034] The acquisition module is used to acquire n types of tobacco leaves and n groups of tobacco leaf attributes that are associated one-to-one, as well as m types of leaf group formulas, wherein each of the m types of leaf group formulas includes at least one of the n types of tobacco leaves;

[0035] A creation module is used to create a tobacco knowledge graph of the n types of tobacco leaves based on the n types of tobacco leaves and the n groups of tobacco leaf attributes, and to create a leaf group formula interaction matrix based on the m types of leaf group formulas, wherein the leaf group formula interaction matrix represents the relationship between the m types of leaf group formulas and the n types of tobacco leaves;

[0036] The training module is used to train a tobacco leaf and leaf group formula matching prediction function based on the graph convolutional neural network algorithm, the tobacco leaf knowledge graph, and the leaf group formula interaction matrix. The tobacco leaf and leaf group formula matching prediction function is used to determine the degree of matching between the leaf group formula to be matched and the tobacco leaf to be matched.

[0037] In one implementation, after the training module, the following is also included:

[0038] The determination module is used to determine the degree of matching between the tobacco leaves to be matched and the leaf blend formula to be matched, based on the matching prediction function of the tobacco leaves and leaf blend formula, if the tobacco leaves to be matched and the leaf blend formula to be matched are obtained.

[0039] If the tobacco leaf to be matched is obtained, then the leaf blend formula that has the required degree of matching with the tobacco leaf is determined according to the tobacco leaf and leaf blend formula matching prediction function.

[0040] If the leaf blend formula to be matched is obtained, then the tobacco leaves that have the required degree of matching with the leaf blend formula are determined according to the tobacco leaf and leaf blend formula matching prediction function.

[0041] A third aspect of this application provides a computer storage medium storing one or more instructions that, when executed by one or more computers, cause the one or more computers to perform the operation of any of the methods of the first aspect.

[0042] A third aspect of this application provides a computer program product including computer-readable instructions that, when executed on a computer device, cause the computer device to perform the operation of any of the methods described in the first aspect.

[0043] Beneficial effects: After acquiring n types of tobacco leaves and n groups of tobacco leaf attributes, as well as m types of leaf blend formulas, this invention creates a tobacco leaf knowledge graph and a leaf blend formula interaction matrix. Based on the tobacco leaf knowledge graph and leaf blend formula interaction matrix, a tobacco leaf and leaf blend formula matching prediction function is trained. The tobacco leaf and leaf blend formula matching prediction function is used to determine the degree of matching between the leaf blend formula to be matched and the tobacco leaf to be matched, avoiding the need for leaf blend distribution engineers to judge whether the tobacco leaf and leaf blend formula match based on human tasting and their own historical experience, thus freeing up labor. Attached Figure Description

[0044] Figure 1 A flowchart of a method for matching tobacco leaves with leaf blend formulation;

[0045] Figure 2 Another flowchart of a method for matching tobacco leaves and leaf blend formulations;

[0046] Figure 3 Another flowchart of a method for matching tobacco leaves and leaf blend formulations;

[0047] Figure 4 It is a knowledge graph of tobacco leaves;

[0048] Figure 5 A matching device for tobacco leaves and leaf blend formulations;

[0049] Figure 6 This is another matching device for tobacco leaf and leaf blend formulation. Detailed Implementation

[0050] It should be noted that the accompanying drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention. It should be understood that relative terms such as "above," "below," "top," "bottom," and "over" shown in the drawings are used to describe the relationships between various elements. These relative terms are intended to cover different orientations of elements other than those depicted in the drawings. For example, if the device is inverted relative to the view in the drawings, an element described, for example, as being "above" another element would now be below that element.

[0051] Currently, the matching of tobacco leaves and blending formulas is done by blending engineers who first taste the leaves to obtain tasting results. They then combine this with their historical experience to find suitable blending formulas, determining the proportions of each tobacco leaf in the blending formula to produce the tobacco shreds. However, before determining the proportions of each tobacco leaf in the blending formula, the matching of tobacco leaves and blending formulas is based on the subjective determination of the blending engineers. This lacks interpretability regarding the reasons for the matching, and the excessive tasting during long-term work damages the health of blending engineers.

[0052] Currently, tobacco production generally consists of two steps: first, selecting suitable varieties of tobacco leaves to form a leaf blend formula; second, adjusting the proportions of the tobacco leaf blends to form tobacco shreds. Based on the first step of tobacco shred production, this invention provides a method for matching tobacco leaves with leaf blend formulas. It should be emphasized that the method in this application does not involve proportions, but only the components of the leaf blend formula. The leaf blend formula mentioned below specifically refers to the constituent components of the leaf blend formula. Please refer to [link to relevant documentation]. Figure 1 :

[0053] 101. Obtain n types of tobacco leaves and n groups of tobacco leaf attributes that are associated one-to-one, as well as m types of leaf group formulas, wherein each leaf group formula of the m types of leaf group formulas includes at least one of the n types of tobacco leaves;

[0054] Obtain n types of tobacco leaves, m types of leaf group formulations, and n groups of tobacco leaf attributes, wherein each of the m types of leaf group formulations includes at least one of the n types of tobacco leaves.

[0055] It should be noted that each type of tobacco leaf has multiple corresponding tobacco leaf attributes, and each of these multiple tobacco leaf attributes constitutes a set of tobacco leaf attributes. Moreover, each set of tobacco leaf attributes has the same tobacco leaf attribute category but different tobacco leaf attribute values. Therefore, the n sets of tobacco leaf attributes obtained are the tobacco leaf attribute values ​​of the associated tobacco leaves under their different tobacco leaf attribute names. In other words, each type of tobacco leaf corresponds to a set of tobacco leaf attributes.

[0056] 102. Create a tobacco knowledge graph of the n types of tobacco leaves based on the n types of tobacco leaves and the n groups of tobacco leaf attributes, and create a leaf group formula interaction matrix based on the m types of leaf group formulas, wherein the leaf group formula interaction matrix represents the component relationship between the m types of leaf group formulas and the n types of tobacco leaves;

[0057] After obtaining n types of tobacco leaves, m types of leaf group formulas, and n groups of tobacco leaf attributes, a tobacco leaf knowledge graph about the n types of tobacco leaves is created based on the n types of tobacco leaves and the n groups of tobacco leaf attributes, and a leaf group formula interaction matrix is ​​created based on the m types of tobacco leaf formulas.

[0058] For example, each row of the leaf blend interaction matrix represents a leaf blend formula, and each column represents a type of tobacco leaf. It can be seen that any column in any row indicates whether the tobacco blend formula in that row contains the tobacco leaf in that column. Therefore, the leaf blend interaction matrix represents the component relationship between m leaf blend formulas and n types of tobacco leaves. The leaf blend interaction matrix can be a matrix composed of 0 or 1 values, where 0 indicates that the leaf blend formula does not contain tobacco leaves, and 1 indicates that the leaf blend formula contains tobacco leaves. Other values ​​can also represent the presence or absence of components in the leaf blend formula; this is not limited here. In other embodiments, each row of the leaf blend interaction matrix represents a type of tobacco leaf, and each column represents a leaf blend formula.

[0059] 103. Based on the graph convolutional neural network algorithm, and the tobacco knowledge graph and the leaf group formula interaction matrix, a tobacco leaf and leaf group formula matching prediction function is obtained through training. The tobacco leaf and leaf group formula matching prediction function is used to determine the degree of matching between the leaf group formula to be matched and the tobacco leaf to be matched.

[0060] Multiple training data points were obtained based on the created tobacco knowledge graph and leaf group formula interaction matrix. These training data points were then randomly divided into training and testing sets. The training set was used to train a tobacco leaf and leaf group formula matching prediction function, which is used to determine the degree of matching between the leaf group formula to be matched and the tobacco leaf to be matched.

[0061] It should be noted that the tobacco leaf and leaf blend formula matching prediction function obtained through training includes the matching degree information between tobacco leaves and leaf blend formulas recorded in the leaf blend formula interaction matrix, as well as the matching degree information between tobacco leaves and leaf blend formulas recorded in the leaf blend formula interaction matrix that are not recorded in the leaf blend formula interaction matrix.

[0062] Please see Figure 2 , Figure 2 This document provides instructions on how to apply a tobacco leaf and leaf blend formulation matching prediction function. The function yields the following information: (1) the matching degree between tobacco leaves and leaf blend formulations recorded in the leaf blend formulation interaction matrix; and (2) the matching degree between tobacco leaves and leaf blend formulations not recorded in the leaf blend formulation interaction matrix. Therefore, the matching prediction function can be used to find the degree of matching between tobacco leaves and leaf blend formulations. It should be noted that... Figure 2 In this framework, the leaf group formula to be matched is any leaf group formula from the leaf group formula interaction matrix, and the tobacco leaves to be matched are both recorded and unrecorded tobacco leaves in the leaf group formula interaction matrix. Unrecorded tobacco leaves are those determined by the tobacco leaf-leaf formula matching prediction function but not recorded in the leaf group formula interaction matrix. For tobacco leaves not recorded in the leaf group formula interaction matrix obtained by the tobacco leaf-leaf formula matching prediction function, these unrecorded tobacco leaves are added to the tobacco knowledge graph to obtain their graph embedding features. The degree of matching between the unrecorded tobacco leaves and the recorded leaf group formulas in the leaf group formula interaction matrix is ​​then determined based on the weights and bias parameters in the tobacco leaf-leaf formula matching prediction function.

[0063] 201. If the tobacco leaves to be matched and the leaf blend formula to be matched are obtained, the degree of matching between the tobacco leaves to be matched and the leaf blend formula to be matched is determined according to the matching prediction function of the tobacco leaves and leaf blend formula.

[0064] If the tobacco leaves and leaf group formulas to be matched are obtained, the tobacco leaves and leaf group formulas to be matched are input into the tobacco leaf and leaf group formula matching prediction function to determine the degree of matching between the tobacco leaves and leaf group formulas to be matched.

[0065] 202. If a tobacco leaf to be matched is obtained, then a leaf blend formula with the required degree of matching with the tobacco leaf to be matched is determined according to the tobacco leaf and leaf blend formula matching prediction function.

[0066] If a tobacco leaf to be matched is obtained, the degree of matching between the tobacco leaf to be matched and any leaf group formula in the leaf group formula interaction matrix is ​​determined, and then the leaf group formula with the required degree of matching with the tobacco leaf to be matched is determined.

[0067] It should be noted that, firstly, any leaf group formula in the interaction matrix of tobacco leaves and leaf group formulas to be matched is combined, and then the combination result is input into the tobacco leaf and leaf group formula matching prediction function to calculate the matching degree of each combination, thereby determining the leaf group formula with the required matching degree. The leaf group formula with the required matching degree is the target leaf group formula set whose calculated matching degree is greater than or equal to the first preset matching threshold, and the number of target leaf group formula sets is greater than or equal to 0.

[0068] In one implementation, after calculating the matching degree of each combination, and before determining the leaf group formulation with the required matching degree, the matching degree of each combination can be sorted to obtain leaf group formulations in descending or ascending order.

[0069] 203. If the leaf blend formula to be matched is obtained, then the tobacco leaves that have the required degree of matching with the leaf blend formula are determined according to the tobacco leaf and leaf blend formula matching prediction function.

[0070] If the leaf blend formula to be matched is obtained, the leaf blend formula to be matched and the leaf blend formula interaction matrix record and various unrecorded tobacco leaves are combined. Then, each combination is input into the tobacco leaf and leaf blend formula matching prediction function to obtain the matching degree of each combination. Then, based on the second preset matching threshold, all target matching degrees greater than or equal to the second preset matching threshold are determined. Finally, the target combination and the tobacco leaves in the target combination are determined by all target matching degrees.

[0071] The following is a detailed description of this technical solution. Steps 301 to 306 describe the acquisition of any training data, and step 307 describes the training of the tobacco leaf and leaf group formula matching prediction function based on the training set.

[0072] 301. Obtain the one-to-one vectorized n types of tobacco leaves and n sets of tobacco leaf attributes, as well as the vectorized m types of leaf group formulas;

[0073] The vectorization of n types of tobacco leaves and n sets of tobacco leaf attributes, as well as the vectorization of m types of leaf group formulations, are obtained based on the Glorot method. Each type of tobacco leaf corresponds to a set of tobacco leaf attributes, each set of tobacco leaf attributes contains the same tobacco leaf attribute category, and each tobacco leaf attribute contains multiple different tobacco leaf attribute categories.

[0074] 302. Based on the vectorized m leaf group formulas, establish a leaf group formula interaction matrix with element values ​​of 0 or 1;

[0075] Based on the vectorized m leaf group formulations, establish a leaf group formulation interaction matrix M with element values ​​of 0 or 1. m×n Where m is the number of leaf blends, n is the number of tobacco leaves, and e ij For M m×n The element values, where 0≤i≤m, 0≤j≤n.

[0076] 303. Obtain p tobacco attribute categories from the n sets of tobacco leaf attributes, and create a tobacco leaf knowledge graph containing the n types of tobacco leaves and the p tobacco leaf attribute categories, as well as the edge relationships between the entity nodes;

[0077] Since a set of tobacco leaf attributes contains multiple tobacco leaf attribute categories, and each tobacco leaf shares the same attribute category, these categories include tobacco leaf origin, tobacco leaf grade, tobacco leaf year, physicochemical indicators, and sensory indicators. The physicochemical indicators include total sugar, total reducing sugar, total nitrogen, chloride ions, nicotine, crude protein, and potassium oxide. The sensory indicators include irritation, balance, off-flavors, luster, aroma, and aftertaste. Based on this, n sets of tobacco leaves and p entity nodes corresponding to any of these n sets of tobacco leaves are created, each node representing a tobacco leaf attribute type, including tobacco leaf origin, tobacco leaf grade, tobacco leaf year, physicochemical indicators, and sensory indicators. The physicochemical indicators and sensory indicators each include at least one from the set of tobacco leaf physicochemical indicators and at least one from the set of tobacco leaf sensory indicators. Then, edge relationships are created between these entity nodes, ultimately resulting in a tobacco leaf knowledge graph.

[0078] For example, a knowledge graph is created containing 6 tobacco leaves and 3 tobacco leaf attribute categories corresponding to each leaf: tobacco leaf origin, tobacco leaf grade, and tobacco leaf year. The relationship between each tobacco leaf and the aforementioned tobacco leaf origin, grade, and year is respectively a knowledge graph of origin, grade, and year. Please refer to [link to relevant documentation]. Figure 4 It should be noted that, under normal circumstances, the tobacco category attributes in the tobacco knowledge graph are not limited to tobacco origin, tobacco grade, and tobacco year.

[0079] 304. Establish the topological features of training tobacco leaves and training leaf group formulations;

[0080] Multiple matching training tobacco leaves and training leaf group formulas are randomly obtained. Then, based on the graph convolutional neural network, the tobacco knowledge graph, and the training tobacco leaves and training leaf group formulas, the topological features corresponding to any multiple matching training tobacco leaves and training leaf group formulas are determined.

[0081] The following description uses a pair of training tobacco leaves i and training leaf group formula m as an example. Specifically, in the tobacco knowledge graph, the target adjacent entity nodes and target edge relationships corresponding to the training tobacco leaf i are determined, where the target edge relationships are associated with the target adjacent entity nodes. Then, based on the training tobacco leaf i, the training leaf group formula m, the target edge relationships, and the target adjacent entity nodes, the topological features of the training tobacco leaf i and the training leaf group formula m are determined. Calculation formula:

[0082]

[0083]

[0084] Where, N (i) Let e ​​be the set of target adjacent entity nodes corresponding to the training tobacco leaf i, and r be any target adjacent entity node corresponding to the training tobacco leaf i. ei To train the relationship between tobacco leaf i and its target adjacent entity node e, m is the training leaf group formula m. The feature weights between the training leaf group formulation and any training tobacco leaf and the target adjacent entity node of the training tobacco leaf, for example, The calculation method can be the inner product of the relationship between the training leaf group formula and the training tobacco leaf and the target adjacent entity node of the training tobacco leaf.

[0085] In the presence of convolutional layers, the target edge relationships and target adjacent entity nodes of each layer of the training tobacco leaf i are determined in the tobacco leaf knowledge graph according to the respective layers; and the topological features of the training tobacco leaf i and the training leaf group formula m in their respective layers are determined according to the target edge relationships and target adjacent entity nodes of each layer.

[0086] Understandably, topological features The mathematical transformation will follow the convolutional layer, that is... k represents the k-th convolutional layer number. The target adjacent entity nodes of each convolutional layer are generated iteratively. Specifically, the target adjacent entity nodes of each convolutional layer are determined by the target adjacent entity nodes of the layers above them. The target edge relationships of each convolutional layer are determined by the target adjacent entity nodes of the layers above them and the target adjacent entity nodes of each convolutional layer. Here, the first convolutional layer or the initial convolutional layer... The target adjacent entity nodes of the initial layer are the tobacco leaf attribute category values ​​of the training tobacco leaves, such as the tobacco leaf origin value (Sichuan), the tobacco leaf grade value (Level 1), and the tobacco leaf year value (2022). The target edge relationships of the initial layer are the relationship values ​​between the training tobacco leaves and their tobacco leaf attribute categories, such as the tobacco leaf origin, tobacco leaf grade, and tobacco leaf year, in conjunction with the previous example. Therefore, the target adjacent entity nodes of each layer in the convolutional layer also have their own target adjacent entity nodes and their target edge relationships.

[0087] Understandably, based on formula (1) and the richness of tobacco attribute categories in the tobacco knowledge graph and the preset convolutional layer, the topological features of the training tobacco leaf and the training leaf group formula will comprehensively and meticulously represent the tobacco attribute information of the training tobacco leaf i under the training leaf group formula m. It should be emphasized that, based on the tobacco knowledge graph and the convolutional layer, the training tobacco leaf i is obtained by aggregating itself and its own domain. The domain of the training tobacco leaf i includes the tobacco attribute categories of the training tobacco leaf i and the tobacco attribute categories of other training tobacco leaves. Therefore, the topological features of the training tobacco leaf and the training leaf group formula not only represent the tobacco attribute information of the training tobacco leaf i under the training leaf group formula m, but also represent the association information of the training tobacco leaf i under the training leaf group formula m with other training tobacco leaves.

[0088] 305. The graph embedding features of the training tobacco leaf and the training leaf group formulation are calculated based on the graph convolutional neural network and topological features.

[0089] After obtaining the topological features of the training tobacco leaves and the training leaf group formulation, the initial graph embedding features of the training tobacco leaves are determined. Based on the graph convolutional neural network algorithm, and using the initial graph embedding features, the matching topological features and graph embedding features of each layer are iteratively aggregated and calculated to iteratively determine the graph embedding features of the training tobacco leaves and the training leaf group formulation. The calculation formula is as follows:

[0090]

[0091]

[0092] in, Let m be the graph embedding features of the training leaf group formula m and the training tobacco leaf i in the k-th convolutional layer, where k is an integer greater than 0. To train the initial image embedding features of tobacco leaf i, i.e., to train tobacco leaf i itself, `aggregate` is the aggregation operation. The aggregation operation can be a concat aggregator or a sum aggregator; the specific operation is not limited here. σ is the activation function, and W... k B kThe weights and bias parameters of the k-th convolutional layer are respectively determined by a preset. Therefore, formula (3) represents the graph embedding features of each convolutional layer, obtained by iteratively aggregating the graph embedding features of the previous layer and the topological features of the matching convolutional layer. When the graph embedding features of the preset convolutional layer are calculated, these preset convolutional layer graph embedding features are the graph embedding features of the training tobacco leaf i and the training leaf group formula m. Thus, It includes the tobacco leaf attribute information of training tobacco leaf i under training leaf group formula m, and the association information of training tobacco leaf i under training leaf group formula m with other training tobacco leaves.

[0093] 306. Based on the graph embedding features, obtain the element values ​​associated with the training tobacco leaf and the training leaf group formula in the leaf group formula interaction matrix to obtain training data for training the tobacco leaf and leaf group formula matching prediction function;

[0094] After obtaining the graph embedding features of training tobacco leaf i and training leaf group formula m, the element values ​​of training tobacco leaf i and training leaf group formula m are obtained according to the leaf group formula interaction matrix. This determines a pair of training data for training the tobacco leaf and leaf group formula matching prediction function, namely the graph embedding features of training tobacco leaf i and training leaf group formula m, and the element values ​​in the leaf group formula interaction matrix associated with training tobacco leaf i and training leaf group formula m.

[0095] 307. Obtain a training set for training the tobacco leaf and leaf group formulation matching prediction function, and train the tobacco leaf and leaf group formulation matching prediction function based on the training set.

[0096] After obtaining multiple training data sets for training the tobacco leaf and leaf blend formulation matching prediction function based on steps 301 to 306, the multiple training data sets are randomly divided into a training set and a test set. Then, according to the training set and the defined tobacco leaf and leaf blend formulation matching prediction function, matching data sets are randomly selected from the training set. and A tobacco leaf and leaf blend formulation matching prediction function is defined, and then trained to obtain the tobacco leaf and leaf blend formulation matching prediction function. The trained tobacco leaf and leaf blend formulation matching prediction function is then tested. The tobacco leaf and leaf blend formulation matching prediction function is defined as follows: represents the element values ​​in the leaf group formulation interaction matrix that are associated with training tobacco leaf i and training leaf group formulation m. The graph embedding features are used to train tobacco leaf i and training leaf group formulation m.

[0097] During training, the tobacco leaf and leaf blend formula matching prediction function is trained by modifying the bias and weights of the tobacco leaf and leaf blend formula matching prediction function through the binary cross-entropy loss function. The hidden layer uses the ReLU activation function and the output layer uses the Sigmoid activation function.

[0098] Please see Figure 5 This application provides a matching device for tobacco leaves and leaf blend formulations, comprising:

[0099] The acquisition module 501 is used to acquire n types of tobacco leaves and n groups of tobacco leaf attributes that are associated one-to-one, as well as m types of leaf group formulas, wherein each leaf group formula of the m types of leaf group formulas includes at least one of the n types of tobacco leaves.

[0100] A creation module 502 is used to create a tobacco knowledge graph of the n types of tobacco leaves based on the n types of tobacco leaves and the n groups of tobacco leaf attributes, and to create a leaf group formula interaction matrix based on the m types of leaf group formulas, wherein the leaf group formula interaction matrix represents the relationship between the m types of leaf group formulas and the n types of tobacco leaves.

[0101] Training module 503 is used to train a tobacco leaf and leaf group formula matching prediction function based on graph convolutional neural network algorithm, tobacco leaf knowledge graph and leaf group formula interaction matrix. The tobacco leaf and leaf group formula matching prediction function is used to determine the degree of matching between the leaf group formula to be matched and the tobacco leaf to be matched.

[0102] Please see Figure 6 This application also provides a matching device for tobacco leaves and leaf blend formulations, which, after the training module, further includes:

[0103] The determining module 501 is used to determine the degree of matching between the tobacco leaf to be matched and the leaf group formula to be matched, based on the matching prediction function of the tobacco leaf and leaf group formula, if the tobacco leaf to be matched and the leaf group formula to be matched are obtained;

[0104] If the tobacco leaf to be matched is obtained, then the leaf blend formula that has the required degree of matching with the tobacco leaf is determined according to the tobacco leaf and leaf blend formula matching prediction function.

[0105] If the leaf blend formula to be matched is obtained, then the tobacco leaves that have the required degree of matching with the leaf blend formula are determined according to the tobacco leaf and leaf blend formula matching prediction function.

[0106] This application also provides a computer storage medium storing one or more instructions, which, when executed by one or more computers, cause the one or more computers to perform actions. Figures 1 to 3The operation of the method described in any of the embodiments.

[0107] This application also provides a computer program product, including computer-readable instructions that, when executed on a computer device, cause the computer device to perform... Figures 1 to 3 The method described in any of the embodiments.

[0108] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0109] Furthermore, although the present invention has been disclosed above with reference to preferred embodiments, these embodiments are not intended to limit the present invention. For any person skilled in the art, many possible variations and modifications can be made to the technical solutions of the present invention based on the disclosed technical content, or equivalent embodiments can be modified accordingly, without departing from the scope of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the present invention shall still fall within the scope of protection of the present invention.

[0110] It should also be understood that, unless otherwise specified or indicated, the terms "first," "second," "third," etc., used in the specification are merely for distinguishing individual components, elements, steps, etc., and are not for indicating logical or sequential relationships between individual components, elements, steps, etc. Furthermore, it should be recognized that the singular forms "a" and "an" used herein and in the appended claims include plural bases unless the context clearly indicates the opposite. For example, a reference to "a step" or "a device" means a reference to one or more steps or devices, and may include secondary steps and secondary devices. All conjunctions used should be understood in the broadest sense. And the word "or" should be understood as having the definition of logical "or," not the definition of logical "exclusive or," unless the context clearly indicates the opposite. Furthermore, implementation of the methods and / or devices in embodiments of the present invention may include performing selected tasks manually, automatically, or in combination.

Claims

1. A method for matching tobacco leaves and leaf blend formulations, characterized in that, include: Obtain n types of tobacco leaves and n groups of tobacco leaf attributes that are associated one-to-one, as well as m types of leaf group formulas, wherein each of the m types of leaf group formulas includes at least one of the n types of tobacco leaves; A tobacco knowledge graph of the n types of tobacco leaves is created based on the n types of tobacco leaves and the n groups of tobacco leaf attributes, and a leaf group formula interaction matrix is ​​created based on the m types of leaf group formulas. The leaf group formula interaction matrix represents the component relationship between the m types of leaf group formulas and the n types of tobacco leaves. Based on the graph convolutional neural network algorithm, and the tobacco knowledge graph and the leaf group formula interaction matrix, a tobacco leaf and leaf group formula matching prediction function is obtained through training. The tobacco leaf and leaf group formula matching prediction function is used to determine the degree of matching between the leaf group formula to be matched and the tobacco leaf to be matched. After the method obtains the tobacco leaf and leaf blend formula matching prediction function by training according to the graph convolutional neural network algorithm, the tobacco leaf knowledge graph, and the leaf blend formula interaction matrix, it further includes: If the tobacco leaf to be matched and the leaf blend formula to be matched are obtained, the degree of matching between the tobacco leaf to be matched and the leaf blend formula to be matched is determined according to the matching prediction function of the tobacco leaf and leaf blend formula; If the tobacco leaf to be matched is obtained, then the leaf blend formula that has the required degree of matching with the tobacco leaf is determined according to the tobacco leaf and leaf blend formula matching prediction function. If the leaf blend formula to be matched is obtained, then the tobacco leaves that have the required degree of matching with the leaf blend formula are determined according to the tobacco leaf and leaf blend formula matching prediction function.

2. The method according to claim 1, characterized in that, Each of the n groups of tobacco leaf attributes includes multiple different tobacco leaf attribute categories, and each group of tobacco leaf attributes has the same tobacco leaf attribute category. Therefore, creating a tobacco leaf knowledge graph based on the n types of tobacco leaves and the n groups of tobacco leaf attributes includes: Obtain p tobacco leaf attribute categories from the n sets of tobacco leaf attributes; Create a tobacco knowledge graph that includes entity nodes containing the n types of tobacco leaves and the p tobacco leaf attribute categories, as well as the edge relationships between the entity nodes.

3. The method according to claim 2, characterized in that, The step of training a tobacco leaf and leaf blend formula matching prediction function based on a graph convolutional neural network algorithm, the tobacco leaf knowledge graph, and the leaf blend formula interaction matrix includes: Obtain training tobacco leaves and training leaf group formulas; The graph embedding features of the training tobacco leaves and the training leaf group formula are determined based on the convolutional neural network, the tobacco leaf knowledge graph, and the training tobacco leaves and training leaf group formulas. Based on the graph embedding features, the element values ​​associated with the training tobacco leaves and the training leaf group formula in the leaf group formula interaction matrix are randomly trained to obtain the tobacco leaf and leaf group formula matching prediction function.

4. The method according to claim 3, characterized in that, The step of determining the graph embedding features of the training tobacco leaves and the training leaf group formula based on the convolutional neural network, the tobacco leaf knowledge graph, and the training tobacco leaves and training leaf group formulas includes: In the tobacco knowledge graph, target adjacent entity nodes and target edge relationships corresponding to the training tobacco leaves are determined, wherein the target edge relationships are associated with the target adjacent entity nodes. The topological features of the training tobacco leaves and the training leaf group formula are determined based on the training tobacco leaves, the training leaf group formula, the target edge relationship, and the target adjacent entity nodes. The graph embedding features of the training tobacco leaves and the training leaf group formulation are calculated based on the graph convolutional neural network algorithm and the topological features.

5. The method according to claim 4, characterized in that, The method includes a convolutional layer, wherein determining the topological features of the training tobacco leaf and the training leaf group formula based on the training tobacco leaf, the training leaf group formula, the target edge relationship, and the target adjacent entity node includes: Based on each layer of the convolutional layer, the target edge relationships and target adjacent entity nodes of each layer of the training tobacco leaf are determined in the tobacco leaf knowledge graph; Based on the target edge relationships and target adjacent entity nodes of each layer, the topological features of the training tobacco leaf and the training leaf group formula in their respective layers are determined; The step of calculating the graph embedding features of the training tobacco leaves and the training leaf group formulation based on the convolutional neural network and the topological features includes: Determine the initial graph embedding features of the training tobacco leaves; Based on the graph convolutional neural network algorithm, and using the initial graph embedding features, the topological features and graph embedding features of the respective layers that are matched are iteratively aggregated and calculated to determine the graph embedding features of the training tobacco leaves and the training leaf group formulation.

6. The method according to claim 1, wherein the n types of tobacco leaves, the n sets of tobacco leaf attributes, and the m types of leaf group formulations are respectively grotesquely vectorized n types of tobacco leaves, the n sets of tobacco leaf attributes, and the m types of leaf group formulations.

7. The method according to claim 2, characterized in that, The tobacco leaf attribute categories include tobacco leaf origin, tobacco leaf grade, tobacco leaf year, as well as multiple physicochemical indicators and multiple sensory indicators of tobacco leaves.

8. A matching device for tobacco leaves and leaf blend formulations, characterized in that, include: The acquisition module is used to acquire n types of tobacco leaves and n groups of tobacco leaf attributes that are associated one-to-one, as well as m types of leaf group formulas, wherein each of the m types of leaf group formulas includes at least one of the n types of tobacco leaves; A creation module is used to create a tobacco knowledge graph of the n types of tobacco leaves based on the n types of tobacco leaves and the n groups of tobacco leaf attributes, and to create a leaf group formula interaction matrix based on the m types of leaf group formulas, wherein the leaf group formula interaction matrix represents the relationship between the m types of leaf group formulas and the n types of tobacco leaves; The training module is used to train a tobacco leaf and leaf group formula matching prediction function based on the graph convolutional neural network algorithm, the tobacco leaf knowledge graph and the leaf group formula interaction matrix. The tobacco leaf and leaf group formula matching prediction function is used to determine the degree of matching between the leaf group formula to be matched and the tobacco leaf to be matched. Following the training module, the following is also included: The determination module is used to determine the degree of matching between the tobacco leaves to be matched and the leaf blend formula to be matched, based on the matching prediction function of the tobacco leaves and leaf blend formula, if the tobacco leaves to be matched and the leaf blend formula to be matched are obtained. If the tobacco leaf to be matched is obtained, then the leaf blend formula that has the required degree of matching with the tobacco leaf is determined according to the tobacco leaf and leaf blend formula matching prediction function. If the leaf blend formula to be matched is obtained, then the tobacco leaves that have the required degree of matching with the leaf blend formula are determined according to the tobacco leaf and leaf blend formula matching prediction function.

9. A computer storage medium, characterized in that, The computer storage medium stores one or more instructions that, when executed by one or more computers, cause the one or more computers to perform the operation of the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes computer-readable instructions that, when executed on a computer device, cause the computer device to perform the method as described in any one of claims 1 to 7.

Images