Cigarette making machine comprehensive quality evaluation method based on graph neural network

Abstract

The invention relates to the technical field of cigarette processing, in particular to a cigarette making machine comprehensive quality evaluation method based on a graph neural network, which comprises the following steps: dividing pre-processed related data into evaluation layers of different indexes based on hierarchical analysis decision, and finding out factors of different levels from the evaluation layers of different indexes; finding out relevance between different level factors and different index evaluation layers from different index evaluation layers based on gray correlation analysis; on the basis of an entropy evaluation method weight solving method, finding out the weights of the correlation factors corresponding to the different index evaluation layers in the corresponding index evaluation layers from the different index evaluation layers and the factors of the different levels; and based on different index evaluation layers, a correlation analysis layer, a weight analysis layer and the defined graph convolutional network layer, combining an attention mechanism, normalization operation and FC regression to construct an optimized cigarette making machine comprehensive quality evaluation graph neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of cigarette processing, and in particular to a comprehensive quality evaluation method for cigarette making machines based on graph neural network. Background Art

[0002] Quality process evaluation is a systematic and comprehensive evaluation method that aims to comprehensively consider the quality of the production process of the wrapping machine. This evaluation not only focuses on a single aspect, but involves multiple levels and dimensions. Its purpose is not only to evaluate the quality of the wrapping machine production process, but also to facilitate operators to fully understand the evaluated indicators and trace the problem to which link, thereby providing decision makers with scientific basis and comprehensive information to assist reasonable decision-making and management.

[0003] In the production process of the wrapping machine, the quality evaluation in the production process of the wrapping machine has the characteristics of multi-dimensionality and multi-level; the multi-dimensional characteristics, that is, it is necessary to consider multiple factors, including but not limited to the physical properties, performance indicators, service life and safety of the product; the multi-level characteristics, that is, the quality evaluation usually adopts a hierarchical structure, decomposing the evaluation object into factors of different levels, from the overall goal to the specific details, to form an organic structure, which helps to divide the overall problem into smaller and manageable parts, making the evaluation more systematic and hierarchical.

[0004] In the quality evaluation of the cigarette rolling and packaging machine production process, since the cigarette-related data collection results involve a wide range of areas, unclear levels and complex points, it is necessary to fully screen and sort out the various points involved and extract the logical structure of the cigarette rolling and packaging machines to achieve the effect of modeling and clarification.

[0005] Therefore, it is necessary to improve and design the method of quality evaluation in the production process of wrapping machines in order to solve the problem that the existing quality evaluation of the production process of wrapping machines cannot meet the requirements of the current quality evaluation of the production process of wrapping machines due to differences in scoring rules and customer concepts and differences in the analysis weight of the analysis method and the experience of the prophet, and the error of the quality evaluation of the production process of wrapping machines is still large.

[0006] Common evaluation methods include expert method and linear weighted method. Among them, the expert method directly assigns weights to corresponding indicators based on expert systems and knowledge graphs, which lacks robustness in dealing with some abnormal situations and will result in poor scoring results; the linear weighted method is consistent with the expert method, and its use of regression algorithms to fit weight coefficients lacks robustness. The present invention intends to use an adaptive hierarchical graph neural network modeling to assign weight coefficients, which is more universal than traditional SPC and AHP algorithms. Summary of the invention

[0007] The purpose of the present invention is to propose a comprehensive quality evaluation method for cigarette making machines based on graph neural networks, so as to solve the problem that the existing quality evaluation of the production process of cigarette rolling machines cannot meet the requirements of the current quality evaluation of the production process of cigarette rolling machines due to the differences in scoring rules and customer concepts and the differences in analysis weights of analysis methods and prophetic experience, and the error of the quality evaluation of the production process of cigarette rolling machines is still large.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A comprehensive quality evaluation method for cigarette making machines based on graph neural network includes the following steps:

[0010] Step 1, obtain the relevant data after preprocessing in the production process of the wrapping machine, divide the preprocessed relevant data into evaluation layers of different indicators based on hierarchical analysis and decision-making, find the corresponding different level factors from the evaluation layers of different indicators, and construct the evaluation layers of different indicators;

[0011] Step 2: Based on gray correlation analysis, find out the correlation between different level factors and different indicator evaluation layers from different indicator evaluation layers, and construct a correlation analysis layer;

[0012] Step 3: Based on the entropy method, find out the weights of the associated factors corresponding to different indicator evaluation layers in the corresponding indicator evaluation layers from different indicator evaluation layers and different level factors, and construct a weight analysis layer;

[0013] Step 4: Based on the different indicator evaluation layers constructed in Step 1, the correlation analysis layer constructed in Step 2, the weight analysis layer constructed in Step 3 and the defined graph convolutional network layer, combined with the attention mechanism, normalization operation and FC regression, an optimized graph neural network for comprehensive quality evaluation of cigarette machines is constructed;

[0014] Step 5. Input the real-time relevant data after preprocessing in the production process of the cigarette rolling machine into the optimized cigarette machine comprehensive quality evaluation graph neural network to obtain the real-time evaluation of the quality of the cigarette rolling machine production process.

[0015] Preferably, the Step 1, obtaining the relevant data after preprocessing in the production process of the wrapping machine, dividing the preprocessed relevant data into evaluation layers of different indicators based on hierarchical analysis decision, finding the corresponding different level factors from the different indicator evaluation layers, and constructing the evaluation layers of different indicators, includes the following steps:

[0016] Step 1.1, collect the production-related data of the wrapping machine in the past two months through the data acquisition system, and then use the filtering technology to reduce the noise of the collected production-related data to obtain the pre-processed related data;

[0017] Step 1.2, based on hierarchical analysis and decision making, the relevant data pre-processed in Step 1.1 are divided into evaluation layers of different indicators in the order of first-level indicator evaluation (overall goal), second-level indicator evaluation (sub-goals of each layer), evaluation criteria and finally specific alternative plans, and evaluation layers of different indicators are constructed;

[0018] Step 1.3, weight, moisture, draw resistance, total ventilation, tobacco filling status, impurities in cigarettes, appearance inspection and small box sealing degree are used as the first-level factors in the first-level index evaluation (overall goal);

[0019] Deviation value, rejection rate, rejection limit, vehicle speed and cigarette output are used as secondary factors in the secondary indicator evaluation (sub-targets at each level).

[0020] Preferably, in Step 1.3, weight, moisture, draw resistance, total ventilation, tobacco filling state, impurities in cigarettes, appearance inspection and small box sealing are used as the first-level factors in the first-level indicator evaluation (overall goal); deviation value, rejection rate, rejection limit, vehicle speed and cigarette output are used as the second-level factors in the second-level indicator evaluation (sub-goals of each layer), and the weights between the factors in the evaluation layers of different indicators need to be determined to evaluate the importance of each plan under a certain criterion, including the following steps:

[0021] Step 1.31, taking the target layer of the first-level indicator evaluation (overall goal) analysis as an example, based on the first-level factors in the first-level indicator evaluation (overall goal), construct a judgment matrix about the weights between the factors in different indicator evaluation layers. The judgment matrix is ​​as follows:

[0022]

[0023] Among them, the horizontal direction represents the influence of the first factor on N factors; the vertical direction represents the influence of N factors on the first factor.

[0024] Step 1.32, use the arithmetic mean weighting method to find the weights between the factors in different indicator evaluation layers for the judgment matrix constructed in Step 1.31. The calculation formula is as follows:

[0025]

[0026] in, Indicates that the judgment matrix is ​​summed over each column; a ij Indicates the influence of the jth factor on the ith factor; represents the relationship weight between each factor and the jth factor; It means to find the average on the rows of the matrix after column normalization; ω iRepresents the average influence relationship between the i-th factor and all factors.

[0027] Preferably, the Step 2, finding the correlation between factors of different levels and different indicator evaluation layers from different indicator evaluation layers based on grayscale correlation analysis, and constructing a correlation analysis layer, includes the following steps:

[0028] Step 2.1, normalize the data corresponding to the first-level indicator evaluation (overall goal) and the second-level indicator evaluation (sub-goals of each layer) in different indicator evaluation layers in Step 1.2 to obtain standardized data;

[0029] Step 2.2, based on the standardized data obtained in Step 2.1, construct a correlation matrix between factors of different levels and different indicator evaluation layers; where rows represent factors and columns represent data points;

[0030] Step 2.3: For the correlation matrix constructed in Step 2.2, calculate the correlation coefficient and correlation degree between factors of different levels and different indicator evaluation layers. The specific formula is as follows:

[0031]

[0032] Among them, x i (k) is the data sequence element of factor i; x0(k) represents the data sequence element of the indicator to be analyzed; |x0(k)-x i (k) represents the absolute difference between factor i and the indicator data; ρ·max i max k |x0(k)-x i (k) represents the maximum value of the absolute difference between factor i and the index to be analyzed; ρ is the discrimination coefficient (the smaller the correlation coefficient, the greater the difference between the correlation coefficients, the stronger the discrimination ability); min i min k |x0(k)-x i (k)| represents the minimum value of the absolute difference between the factor i and the index to be analyzed; ζ i (k) represents the correlation coefficient between factor i and the indicator to be analyzed; r 0i It represents the correlation degree, that is, the average of the correlation coefficients between factor i and the indicator to be analyzed in the sequence data.

[0033] Step 2.4: Based on the correlation coefficients and correlation degrees between the different levels of factors constructed in Step 2.3 and the different indicator evaluation layers, construct a correlation analysis layer.

[0034] Preferably, the Step 3, finding the weights of the associated factors corresponding to different indicator evaluation layers in the corresponding indicator evaluation layers based on the entropy value method, and constructing the weight analysis layer, includes the following steps:

[0035] Step 3.1, process the data corresponding to the first-level indicator evaluation (overall goal) and the second-level indicator evaluation (sub-goals of each layer) in different indicator evaluation layers in Step 1.2 by zero-mean method or normalization method to obtain standard data;

[0036] Step 3.2, based on the standard data obtained in Step 3.1, construct the original matrix Rx, the original matrix Rx is as follows:

[0037]

[0038] Assuming there are m evaluation objects and n evaluation indicators, Xij is the correlation between the i-th evaluation object and the j-th indicator (i = 1, 2, 3, ..., m, j = 1, 2, 3, ..., n).

[0039] Step 3.3, use the extreme value method to standardize the data in the original matrix Rx in Step 3.2 to obtain the standardized matrix Rx, where the conversion formula for standardizing the data in the original matrix Rx is as follows:

[0040] Positive indicators:

[0041] Negative indicators:

[0042] Step 3.4, based on the standardized matrix Rx obtained in Step 3.3, calculate the weight of the i-th sample under the j-th indicator. The calculation formula is as follows:

[0043]

[0044] in, represents the sum of the jth column in the matrix R; X ij represents the i-th row and j-th column element in R, so P ij Represents the value of each element after normalization by column

[0045] Step 3.5: Based on the weight of the i-th sample under the j-th indicator obtained in Step 3.4, calculate the entropy value of the j-th indicator. The calculation formula is as follows:

[0046]

[0047] Among them, (P ij*ln(P ij )) represents the information entropy of the association between factor i and indicator j; represents the average associated information entropy of all factors for index j; is the scaling factor used to eliminate the number of factors; e j Represents the entropy of the j-th indicator after standardization.

[0048] Step 3.6, based on the entropy value of the j-th indicator obtained in Step 3.5, calculate the difference coefficient of the j-th indicator. The calculation formula is as follows:

[0049] d j =1-e j

[0050] Among them, e j represents the entropy of the jth index after standardization; d j Represents the coefficient of difference of the j-th indicator

[0051] Step 3.7, based on the difference coefficient of the jth indicator obtained in Step 3.6, calculate the weight of each evaluation indicator. The calculation formula is as follows:

[0052]

[0053] Among them, (1-e j ) represents the coefficient of variation of the j-th indicator; represents the sum of the coefficients of difference of all indicators; w i Indicates the proportion of the difference coefficient of the ith indicator in the difference coefficients of all indicators, that is, the indicator weight

[0054] Step 3.8: Based on the weights of each evaluation index obtained in Step 3.7, construct a weight analysis layer.

[0055] Preferably, the Step 4, based on the different indicator evaluation layers constructed in Step 1, the correlation analysis layer constructed in Step 2, the weight analysis layer constructed in Step 3 and the defined graph convolutional network layer, combined with the attention mechanism, normalization operation and FC regression, constructs an optimized graph neural network for comprehensive quality evaluation of cigarette machines, including the following steps:

[0056] Step 4.1, based on the weight analysis layer constructed in Step 3.8, the association analysis layer constructed in Step 2.4, the evaluation layer of different indicators constructed in Step 1.2 and the defined graph convolutional network layer, construct a preliminary graph neural network for comprehensive quality evaluation of cigarette making machines;

[0057] Step 4.2. Combine the attention mechanism, normalization operation and FC regression to optimize the preliminary cigarette machine comprehensive quality evaluation graph neural network obtained in Step 4.1 to obtain the optimized cigarette machine comprehensive quality evaluation graph neural network.

[0058] Preferably, the graph convolutional network layer after the definition is defined as follows:

[0059]

[0060] in, Represents the weighted product of the output feature of the lth layer (the input of the l+1th layer) in the graph convolutional network and the corresponding weight in the filter; represents the summation on the neighborhood graph of i nodes; σ represents a nonlinear function; Represents the output feature of the i-th node in the graph model after passing through the l+1th layer of the graph convolutional neural network

[0061] Preferably, the Step 4.2, combining the attention mechanism, normalization operation and FC regression, optimizes the preliminary cigarette machine comprehensive quality evaluation graph neural network obtained in Step 4.1 to obtain an optimized cigarette machine comprehensive quality evaluation graph neural network, including the following steps:

[0062] Step 4.21. Query all analysis data within the defined time window from the time series database, and build the analysis node H set based on the analysis data. The expression of the H set is as follows:

[0063] H={h1,h2,...,h N}, h i ∈RF

[0064] Where N is the number of nodes in the graph; F represents the characteristic dimension of the input data

[0065] Step 4.22, perform a weight matrix linear transformation on the analysis node H set constructed in Step 4.21 to obtain the updated feature representation set of the evaluation index node. The weight matrix linear transformation and the updated feature set expressions are as follows: Weight matrix linear transformation formula:

[0066] h' i =Wh i

[0067] Among them, Wh i Indicates the linear transformation of the linear matrix W; h' i The expression of the updated feature set after the linear transformation is represented by:

[0068] h'={h'1,h'2,...,h' N},h'i ∈RF'

[0069] Among them, h' represents the updated feature representation set of the evaluation index node; F' represents the transformed evaluation index feature dimension

[0070] Step 4.23, based on a shared self-attention mechanism a: RF'X RF'→R, calculate the attention correlation coefficient between the nodes in the updated feature set in Step 4.22. The calculation formula is as follows:

[0071] e ij =a(h i '||h j ')

[0072] Among them, (h' i ||h' j ) represents the mutual correlation between node i and node j; a() represents the attention mechanism calculation; e ij Represents the attention correlation coefficient between node i and node j

[0073] Step 4.24, concatenate the linearly transformed features of the target node and the neighboring nodes, and then use the weight vector a∈R 2F After parameterization and inputting the nonlinear activation function LeakyReLu, we get the unnormalized attention correlation coefficient, which is expressed as follows:

[0074] e ij =LeakyReLU(a(Wh i ||Wh j ))

[0075] Among them, a(Wh' i ||Wh' j ) represents the cross-correlation between nodes i and j after linear transformation; LeakyReLU represents the modified ReLU function; e ij Represents the attention correlation coefficient between nodes i and j after linear transformation

[0076] Step 4.25, use the softmax function to normalize the unnormalized attention correlation coefficient obtained in Step 4.24 to obtain the attention correlation coefficient a between the nodes i,j , the calculation formula is as follows:

[0077]

[0078] Among them, exp(LeakyReLU(a(Wh' i ||Wh' j))) represents the attention correlation coefficient between the i node and the j node after linear transformation, and then passes through a modified ReLu function as the final attention correlation coefficient; softmax(e i,j ) is a processing function. The whole formula represents the normalization operation of the coefficient by the softmax function, that is, converting the attention correlation coefficient into a probability, which represents the degree of attention of node i to node j.

[0079] Step 4.26, based on the attention correlation coefficient a between the nodes obtained in Step 4.25 i,j， The new node obtained by weighted linear combination of neighboring points is expressed as follows:

[0080]

[0081] in, represents the linear transformation of the neighborhood node j; Indicates a i,j The weight is used to perform linear weighted combination of the neighborhood nodes after linear transformation; σ represents a nonlinear function; Represents the i node after being updated by the l+1th layer of the graph convolutional network

[0082] Step 4.27. Based on the previous experience and knowledge, the evaluation system of the cigarette packaging machine score is calculated in the form of statistical distribution. The calculation formula is as follows:

[0083]

[0084] in, Indicates the deviation of the difference between the actual value and the average value from 3 times the historical standard deviation; Indicates the deviation of the difference between the actual value and the average value from 1.5 times the historical standard deviation

[0085] Step 4.28, based on the evaluation system of cigarette packaging machine scores in Step 4.27, establish a primary scoring sample set, input the primary scoring sample set into the preliminary cigarette machine comprehensive quality evaluation graph neural network for training, and obtain the trained cigarette machine comprehensive quality evaluation graph neural network;

[0086] Step 4.29: Based on the new nodes obtained in Step 4.26, the cigarette making machine comprehensive quality evaluation graph neural network trained in Step 4.28 is optimized to obtain an optimized cigarette making machine comprehensive quality evaluation graph neural network.

[0087] Preferably, the Step 5, inputting the real-time relevant data pre-processed in the production process of the cigarette rolling machine into the optimized cigarette rolling machine comprehensive quality evaluation graph neural network to obtain the real-time evaluation of the quality in the production process of the cigarette rolling machine, comprises the following steps:

[0088] Step 5.1, collect the real-time data of the production process of the wrapping machine through the data acquisition system, and then use the filtering technology to reduce the noise of the real-time collected data to obtain the real-time related data after preprocessing;

[0089] Step 5.2: Input the real-time related data preprocessed in Step 5.1 into the cigarette making machine comprehensive quality evaluation graph neural network optimized in Step 4.24 to obtain the real-time evaluation of the quality of the cigarette making machine during the production process.

[0090] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0091] 1. The comprehensive quality evaluation method of cigarette making machines based on graph neural networks in the present invention can evaluate the quality of the production process of the cigarette rolling machine in real time. It can evaluate the production quality of the cigarette rolling machine at this stage according to the data collected in real time, and timely warn of possible abnormalities in the production process, assist decision makers to improve production efficiency, ensure the stability of product quality and reduce the economic losses caused by abnormalities. At the same time, it can also investigate the causes of abnormalities according to the corresponding secondary and tertiary indicators, assist in maintaining production, and solve the problem that the existing quality evaluation of the production process of the cigarette rolling machine cannot meet the requirements of the current quality evaluation of the production process of the cigarette rolling machine due to the differences in scoring rules and customer concepts and the differences in the analysis weight of the analysis method and the experience of the prophet, and the quality evaluation error of the production process of the cigarette rolling machine is still large.

[0092] 2. The hierarchical analysis method in the present invention can sort out the hierarchical structure of the data; the grayscale correlation method can find the correlation between factors and indicators; the entropy method can find the weight of the corresponding indicator occupied by the factor; the graph attention network can use the attention mechanism to improve the expression ability of the graph neural network model; the normalization method can avoid the polarization of input and weight; in this way, the hierarchical analysis method, grayscale correlation method, entropy method and graph attention network are combined, so that the error of the result of real-time evaluation of the production process quality of the cigarette rolling machine based on the graph neural network comprehensive quality evaluation method is smaller. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 The figure is a flow chart of a comprehensive quality evaluation method for cigarette making machines based on graph neural network in the structure of the present invention.

[0094] Figure 2 It is a schematic diagram of the evaluation index hierarchy in the structure of the present invention.

[0095] Figure 3 It is a schematic diagram of the evaluation index hierarchy model structure in the structure of the present invention.

[0096] Figure 4 This is a schematic diagram of the neural network structure in the structure of the present invention.

[0097] Table 1 is the real-time evaluation of a machine of Protos 2C-GDX500 DETAILED DESCRIPTION

[0098] like Figure 1-4 In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0099] Example

[0100] In the quality evaluation of the production process of the rolling and packaging machine, since the results of the collection of cigarette-related data involve a wide range of areas, unclear levels and complex points, it is necessary to fully screen and sort out the points involved in each aspect and extract the logical structure of the cigarette machine and the packaging machine to achieve the effect of modeling and clarification. However, in the existing quality evaluation of the production process of the rolling and packaging machine, due to the differences in scoring rules and customer concepts and the differences in the analysis weight of the analysis method and the experience of the forefathers, it cannot meet the current requirements of the quality evaluation of the production process of the rolling and packaging machine, and the error of the quality evaluation of the production process of the rolling and packaging machine is still large, which in turn affects the stability of ensuring product quality.

[0101] Therefore, this application proposes a comprehensive quality evaluation method for cigarette making machines based on graph neural networks to solve the problem that the existing quality evaluation of the production process of cigarette rolling machines cannot meet the requirements of the current quality evaluation of the production process of cigarette rolling machines due to differences in scoring rules and customer concepts and differences in the analysis weight of the analysis method and the experience of the prophet, and the error of the quality evaluation of the production process of cigarette rolling machines is still large.

[0102] For details, please refer to Figure 1-4 , a comprehensive quality evaluation method for cigarette making machines based on graph neural network, comprising the following steps:

[0103] The first step is to obtain the relevant data after preprocessing in the production process of the wrapping machine, divide the preprocessed relevant data into evaluation layers of different indicators based on hierarchical analysis and decision-making, find the corresponding different level factors from the different indicator evaluation layers, and construct the evaluation layers of different indicators, including the following steps:

[0104] 1) Through the data acquisition system, the production-related data of the wrapping machine in the past two months are collected, and then the collected production-related data are subjected to noise reduction processing using filtering technology to obtain the pre-processed related data;

[0105] 2) Based on hierarchical analysis and decision making, the relevant data pre-processed in 1) are divided into evaluation layers of different indicators in the order of first-level indicator evaluation (overall goal), second-level indicator evaluation (sub-goals of each layer), evaluation criteria and specific alternative plans, and evaluation layers of different indicators are constructed;

[0106] 3) Take weight, moisture, draw resistance, total ventilation, tobacco filling status, impurities in cigarettes, appearance inspection and small box sealing as the first-level factors in the first-level index evaluation (overall goal);

[0107] Deviation value, rejection rate, rejection limit, vehicle speed and cigarette output are used as secondary factors in the secondary indicator evaluation (sub-targets at each level).

[0108] It should be noted that weight, moisture, draw resistance, total ventilation, tobacco filling status, impurities in cigarettes, appearance inspection and small box sealing are taken as the first-level factors in the first-level indicator evaluation (overall goal); deviation value, rejection rate, rejection limit, vehicle speed and cigarette output are taken as the second-level factors in the second-level indicator evaluation (sub-goals of each level). It is also necessary to determine the weights of the factors in the evaluation layers of different indicators in order to evaluate the importance of each plan under a certain criterion, including the following steps:

[0109] 1) Taking the target layer of the first-level indicator evaluation (overall goal) analysis as an example, based on the first-level factors in the first-level indicator evaluation (overall goal), a judgment matrix about the weights between factors in different indicator evaluation layers is constructed. The judgment matrix is ​​as follows:

[0110]

[0111] Among them, the horizontal direction represents the influence of the first factor on N factors; the vertical direction represents the influence of N factors on the first factor.

[0112] 2) Use the arithmetic mean weighting method to find the weights between the factors in different indicator evaluation layers for the judgment matrix constructed in Step 1.31. The calculation formula is as follows:

[0113]

[0114] in, Indicates that the judgment matrix is ​​summed over each column; a ij Indicates the influence of the jth factor on the ith factor; represents the relationship weight between each factor and the jth factor; It means to find the average on the rows of the matrix after column normalization; ω i Represents the average influence relationship between the i-th factor and all factors.

[0115] The second step is to find out the correlation between different level factors and different indicator evaluation layers from different indicator evaluation layers based on gray correlation analysis, and construct a correlation analysis layer, including the following steps:

[0116] 1) Normalize the data corresponding to the first-level indicator evaluation (overall goal) and the second-level indicator evaluation (sub-goals of each level) in different indicator evaluation layers to obtain standardized data;

[0117] 2) Based on the standardized data, a correlation matrix between factors of different levels and different indicator evaluation layers is constructed, where rows represent factors and columns represent data points;

[0118] 3) For the constructed correlation matrix, calculate the correlation coefficient and correlation degree between factors of different levels and different indicator evaluation layers. The specific formula is as follows:

[0119]

[0120] Among them, x i (k) is the data sequence element of factor i; x0(k) represents the data sequence element of the indicator to be analyzed; |x0(k)-x i (k)| represents the absolute difference between factor i and the indicator data; ρ·max i max k |x0(k)-x i (k) represents the maximum value of the absolute difference between factor i and the index to be analyzed; ρ is the discrimination coefficient (the smaller the correlation coefficient, the greater the difference between the correlation coefficients, the stronger the discrimination ability); min i min k |x0(k)-x i (k) represents the minimum value of the absolute difference between factor i and the index to be analyzed; ζ i (k) represents the correlation coefficient between factor i and the indicator to be analyzed; r 0i It represents the correlation degree, that is, the average of the correlation coefficients between factor i and the indicator to be analyzed in the sequence data.

[0121] 4) Based on the correlation coefficients and correlation degrees between factors of different levels and different indicator evaluation layers, a correlation analysis layer is constructed.

[0122] The third step is to find the weights of the associated factors corresponding to different indicator evaluation layers in the corresponding indicator evaluation layers based on the entropy method, and construct a weight analysis layer, including the following steps:

[0123] 1) Process the data corresponding to the first-level indicator evaluation (overall goal) and the second-level indicator evaluation (sub-goals of each level) in different indicator evaluation layers using the zero-mean method or normalization method to obtain standard data;

[0124] 2) Based on the standard data, construct the original matrix Rx, the original matrix Rx is as follows:

[0125]

[0126] Assuming there are m evaluation objects and n evaluation indicators, Xij is the correlation between the i-th evaluation object and the j-th indicator (i = 1, 2, 3, ..., m, j = 1, 2, 3, ..., n).

[0127] 3) Use the extreme value method to standardize the data in the original matrix Rx to obtain the standardized matrix Rx, where the conversion formula for standardizing the data in the original matrix Rx is as follows:

[0128] Positive indicators:

[0129] Negative indicators:

[0130] 4) Based on the standardized matrix Rx, calculate the weight of the i-th sample under the j-th indicator. The calculation formula is as follows:

[0131]

[0132] in, represents the sum of the jth column in the matrix R; X ij represents the i-th row and j-th column element in R, so P ij Represents the value of each element after normalization by column

[0133] 5) Based on the weight of the i-th sample under the j-th indicator, the entropy value of the j-th indicator is calculated. The calculation formula is as follows:

[0134]

[0135] Among them, (P ij *ln(P ij )) represents the information entropy of the association between factor i and indicator j; represents the average associated information entropy of all factors for index j; is the scaling factor used to eliminate the number of factors; e j Represents the entropy of the j-th indicator after standardization.

[0136] 6) Based on the entropy value of the j-th indicator obtained in Step 3.5, calculate the difference coefficient of the j-th indicator. The calculation formula is as follows:

[0137] d j =1-ej

[0138] Among them, e j represents the entropy of the jth index after standardization; d j Represents the coefficient of difference of the j-th indicator

[0139] 7) Based on the difference coefficient of the jth indicator, the weight of each evaluation indicator is calculated. The calculation formula is as follows:

[0140]

[0141] Among them, (1-e j ) represents the coefficient of variation of the j-th indicator; represents the sum of the coefficients of difference of all indicators; w i Indicates the proportion of the difference coefficient of the ith indicator in the difference coefficients of all indicators, that is, the indicator weight

[0142] 8) Based on the weights of each evaluation indicator, a weight analysis layer is constructed.

[0143] Step 4, please refer to 4. Based on different indicator evaluation layers, correlation analysis layers, weight analysis layers and the defined graph convolutional network layer, combined with the attention mechanism, normalization operation and FC regression, an optimized graph neural network for comprehensive quality evaluation of cigarette machines is constructed, including the following steps:

[0144] 1) Based on the weight analysis layer, correlation analysis layer, evaluation layer of different indicators and the defined graph convolutional network layer, a preliminary graph neural network for comprehensive quality evaluation of cigarette making machines is constructed;

[0145] It should be noted that the formula for the defined graph convolutional network layer is as follows:

[0146]

[0147] in, Represents the weighted product of the output feature of the lth layer (the input of the l+1th layer) in the graph convolutional network and the corresponding weight in the filter; represents the summation on the neighborhood graph of i nodes; σ represents a nonlinear function; Represents the output feature of the i-th node in the graph model after passing through the l+1th layer of the graph convolutional neural network

[0148] 2) Combining the attention mechanism, normalization operation and FC regression, the preliminary cigarette machine comprehensive quality evaluation graph neural network is optimized to obtain the optimized cigarette machine comprehensive quality evaluation graph neural network, including the following steps:

[0149] a. Query all analysis data within the defined time window from the time series database, and build the analysis node H set based on the analysis data. The expression of the H set is as follows:

[0150] H={h1,h2,...,h N}, h i ∈RF

[0151] Where N is the number of nodes in the graph; F represents the characteristic dimension of the input data

[0152] b. Perform a linear transformation of the weight matrix on the constructed analysis node H set to obtain the updated feature representation set of the evaluation index node. The linear transformation of the weight matrix and the updated feature set expressions are as follows:

[0153] Weight matrix linear transformation formula:

[0154] h' i =Wh i

[0155] Among them, Wh i Indicates the linear transformation of the linear matrix W; h' i Represents the value after linear transformation

[0156] Updated feature set expression:

[0157] h'={h'1,h'2,...,h' N},h' i ∈RF'

[0158] Among them, h' represents the updated feature representation set of the evaluation index node; F, represents the transformed evaluation index feature dimension

[0159] c. Based on a shared self-attention mechanism a:RF'XRF-R, calculate the attention correlation coefficient between nodes in the updated feature set in Step 4.22. The calculation formula is as follows:

[0160] e ij =a(h i '||h j ')

[0161] Among them, (h' i ||h' j ) represents the mutual correlation between node i and node j; a() represents the attention mechanism calculation; e ij Represents the attention correlation coefficient between node i and node j

[0162] d. Concatenate the linearly transformed features of the target node and the neighboring nodes, and then use the weight vector a∈R 2FAfter parameterization and inputting the nonlinear activation function LeakyReLu, we get the unnormalized attention correlation coefficient, which is expressed as follows:

[0163] e ij =LeakyReLU(a(Wh i ||Wh j ))

[0164] Among them, a(Wh' i ||Wh' j ) represents the cross-correlation between nodes i and j after linear transformation; LeakyReLU represents the modified ReLU function; e ij Represents the attention correlation coefficient between nodes i and j after linear transformation

[0165] e. Use the softmax function to normalize the unnormalized attention correlation coefficient to obtain the attention correlation coefficient a between nodes. i,j , the calculation formula is as follows:

[0166]

[0167] Among them, exp(LeakyReLU(a(Wh' i ||Wh' j ))) represents the attention correlation coefficient between the i node and the j node after linear transformation, and then passes through a modified ReLu function as the final attention correlation coefficient; softmax(e i,j ) is a processing function. The whole formula represents the normalization operation of the coefficient by the softmax function, that is, converting the attention correlation coefficient into a probability, which represents the degree of attention of node i to node j.

[0168] f. Based on the attention correlation coefficient a between nodes i,j， The new node obtained by weighted linear combination of neighboring points is expressed as follows:

[0169]

[0170] in, represents the linear transformation of the neighborhood node j; Indicates a i,j The weight is used to perform linear weighted combination of the neighborhood nodes after linear transformation; σ represents a nonlinear function; Represents the i node after being updated by the l+1th layer of the graph convolutional network

[0171] g. Based on the previous experience and knowledge, the evaluation system for calculating the scores of cigarette packaging machines is calculated in the form of statistical distribution. The calculation formula is as follows:

[0172]

[0173] in, Indicates the deviation of the difference between the actual value and the average value from 3 times the historical standard deviation; Indicates the deviation of the difference between the actual value and the average value from 1.5 times the historical standard deviation

[0174] h. Based on the evaluation system of cigarette packaging machine scores, a primary scoring sample set is established, and the primary scoring sample set is input into the preliminary cigarette machine comprehensive quality evaluation graph neural network for training to obtain the trained cigarette machine comprehensive quality evaluation graph neural network;

[0175] i. Based on the obtained new nodes, the trained cigarette making machine comprehensive quality evaluation graph neural network is optimized to obtain the optimized cigarette making machine comprehensive quality evaluation graph neural network.

[0176] The fifth step is to input the real-time relevant data pre-processed in the production process of the cigarette rolling machine into the optimized cigarette machine comprehensive quality evaluation graph neural network to obtain the real-time evaluation of the quality in the production process of the cigarette rolling machine, including the following steps:

[0177] 1) Through the data acquisition system, the real-time data of the rolling and packaging machine production process is collected, and then the filtering technology is used to reduce the noise of the real-time collected data to obtain the real-time related data after preprocessing;

[0178] 2) Input the preprocessed real-time related data into the optimized cigarette making machine comprehensive quality evaluation graph neural network to obtain the real-time evaluation of the quality of the cigarette making machine during the production process.

[0179] This patent is applied to the Protos 2C-GDX500 model of a certain cigarette factory, deploying and implementing 12 online production equipment to achieve real-time quality evaluation during the rolling and packaging production process. The effects are as follows:

[0180] Table 1 Real-time evaluation of a certain machine of Protos 2C-GDX500

[0181]

[0182] A comprehensive quality evaluation method for cigarette rolling machines based on graph neural networks in the present application can evaluate the quality of the production process of the rolling and packaging machine in real time. It can evaluate the production quality of the rolling and packaging at this stage based on the real-time collected data, and timely warn of possible abnormalities in the production process, assist decision makers in improving production efficiency, ensuring the stability of product quality and reducing economic losses caused by abnormalities. At the same time, it can also investigate the causes of abnormalities according to the corresponding second and third level indicators, assist in maintaining production, and solve the problem that the existing quality evaluation of the production process of the rolling and packaging machine cannot meet the requirements of the current quality evaluation of the production process of the rolling and packaging machine due to the differences in scoring rules and customer concepts and the differences in the analysis weight of the analysis method and the experience of the prophet, and the quality evaluation error of the production process of the rolling and packaging machine is still large.

[0183] The hierarchical analysis method in this application can sort out the hierarchical structure of the data; the grayscale correlation method can find the correlation between factors and indicators; the entropy method can find the weight of the corresponding indicator occupied by the factor; the graph attention network can use the attention mechanism to improve the expressive power of the graph neural network model; the normalization method can avoid the polarization of input and weight; in this way, the hierarchical analysis method, the grayscale correlation method, the entropy method and the graph attention network are combined to make the result error of the real-time evaluation of the production process quality of the cigarette rolling machine based on the graph neural network comprehensive quality evaluation method smaller.

[0184] Although the present invention is described herein with reference to a number of illustrative embodiments of the present invention, it will be appreciated that those skilled in the art may devise many other modifications and implementations that fall within the scope and spirit of the principles disclosed herein. More specifically, within the scope of the present disclosure, drawings, and claims, a variety of variations and modifications may be made to the components and / or layout of the subject combination layout. In addition to the variations and modifications made to the components and / or layout, other uses will also be apparent to those skilled in the art.

Claims

1. A comprehensive quality evaluation method for cigarette making machines based on graph neural network, characterized by: The following steps are involved: Step 1, obtain the relevant data after preprocessing in the production process of the wrapping machine, divide the preprocessed relevant data into evaluation layers of different indicators based on hierarchical analysis and decision-making, find the corresponding different level factors from the evaluation layers of different indicators, and construct the evaluation layers of different indicators; Step 2: Based on gray correlation analysis, find out the correlation between different level factors and different indicator evaluation layers from different indicator evaluation layers, and construct a correlation analysis layer; Step 3: Based on the entropy method, find out the weights of the associated factors corresponding to different indicator evaluation layers in the corresponding indicator evaluation layers from different indicator evaluation layers and different level factors, and construct a weight analysis layer; Step 4: Based on the different indicator evaluation layers constructed in Step 1, the correlation analysis layer constructed in Step 2, the weight analysis layer constructed in Step 3 and the defined graph convolutional network layer, combined with the attention mechanism, normalization operation and FC regression, an optimized graph neural network for comprehensive quality evaluation of cigarette machines is constructed; Step 5. Input the real-time relevant data after preprocessing in the production process of the cigarette rolling machine into the optimized cigarette machine comprehensive quality evaluation graph neural network to obtain the real-time evaluation of the quality of the cigarette rolling machine production process.

2. The method for comprehensive quality evaluation of cigarette making machines based on graph neural network according to claim 1, characterized in that: The Step 1, obtaining the relevant data after preprocessing in the production process of the wrapping machine, dividing the relevant data after preprocessing into evaluation layers of different indicators based on hierarchical analysis decision, finding the corresponding different level factors from the evaluation layers of different indicators, and constructing the evaluation layers of different indicators, includes the following steps: Step 1.1, collect the production-related data of the wrapping machine in the past two months through the data acquisition system, and then use the filtering technology to reduce the noise of the collected production-related data to obtain the pre-processed related data; Step 1.2, based on hierarchical analysis and decision making, the relevant data pre-processed in Step 1.1 are divided into evaluation layers of different indicators in the order of first-level indicator evaluation (overall goal), second-level indicator evaluation (sub-goals of each layer), evaluation criteria and finally specific alternative plans, and evaluation layers of different indicators are constructed; Step 1.3, weight, moisture, draw resistance, total ventilation, tobacco filling status, impurities in cigarettes, appearance inspection and small box sealing degree are used as the first-level factors in the first-level index evaluation (overall goal); Deviation value, rejection rate, rejection limit, vehicle speed and cigarette output are used as secondary factors in the secondary indicator evaluation (sub-targets at each level).

3. The method for comprehensive quality evaluation of cigarette making machines based on graph neural network according to claim 2, characterized in that: In the aforementioned Step 1.3, weight, moisture, draw resistance, total ventilation, tobacco filling state, impurities in cigarettes, appearance inspection and small box sealing degree are used as the first-level factors in the first-level index evaluation (overall goal); deviation value, rejection rate, rejection limit, vehicle speed and cigarette output are used as the second-level factors in the second-level index evaluation (sub-goals of each layer), and the weights between the factors in the evaluation layers of different indicators need to be determined to evaluate the importance of each plan under a certain criterion, including the following steps: Step 1.31, taking the target layer of the first-level indicator evaluation (overall goal) analysis as an example, based on the first-level factors in the first-level indicator evaluation (overall goal), construct a judgment matrix about the weights between the factors in different indicator evaluation layers. The judgment matrix is ​​as follows: Among them, the horizontal direction represents the influence of the first factor on N factors; the vertical direction represents the influence of N factors on the first factor. Step 1.32, use the arithmetic mean weighting method to find the weights between the factors in different indicator evaluation layers for the judgment matrix constructed in Step 1.

31. The calculation formula is as follows: in, Indicates that the judgment matrix is ​​summed over each column; a ij Indicates the influence of the jth factor on the ith factor; represents the relationship weight between each factor and the jth factor; It means to find the average on the rows of the matrix after column normalization; ω i Represents the average influence relationship between the i-th factor and all factors.

4. The method for comprehensive quality evaluation of cigarette making machines based on graph neural network according to claim 2, characterized in that: The Step 2, based on gray correlation analysis, finds the correlation between factors of different levels and different indicator evaluation layers from different indicator evaluation layers, and constructs a correlation analysis layer, including the following steps: Step 2.1, normalize the data corresponding to the first-level indicator evaluation (overall goal) and the second-level indicator evaluation (sub-goals of each layer) in different indicator evaluation layers in Step 1.2 to obtain standardized data; Step 2.2, based on the standardized data obtained in Step 2.1, construct a correlation matrix between factors of different levels and different indicator evaluation layers; where rows represent factors and columns represent data points; Step 2.3: For the correlation matrix constructed in Step 2.2, calculate the correlation coefficient and correlation degree between factors of different levels and different indicator evaluation layers. The specific formula is as follows: Among them, x i (k) is the data sequence element of factor i; x0(k) represents the data sequence element of the indicator to be analyzed; |x0(k)-x i (k)| represents the absolute difference between factor i and the indicator data; ρ·max i max k |x0(k)-x i (k)| represents the maximum value of the absolute difference between factor i and the index to be analyzed; ρ is the discrimination coefficient (the smaller the correlation coefficient, the greater the difference between the correlation coefficients, the stronger the discrimination ability); min i min k |x0(k)-x i (k)| represents the minimum value of the absolute difference between the factor i and the index to be analyzed; ζ i (k) represents the correlation coefficient between factor i and the indicator to be analyzed; r 0i It represents the correlation degree, that is, the average of the correlation coefficients between factor i and the indicator to be analyzed in the sequence data. Step 2.4: Based on the correlation coefficients and correlation degrees between the different levels of factors constructed in Step 2.3 and the different indicator evaluation layers, construct a correlation analysis layer.

5. The method for comprehensive quality evaluation of cigarette making machines based on graph neural network according to claim 2, characterized in that: The Step 3, based on the entropy value method, finds out the weights of the associated factors corresponding to different indicator evaluation layers in the corresponding indicator evaluation layers from different indicator evaluation layers and different level factors, and constructs a weight analysis layer, including the following steps: Step 3.1, process the data corresponding to the first-level indicator evaluation (overall goal) and the second-level indicator evaluation (sub-goals of each layer) in different indicator evaluation layers in Step 1.2 by zero-mean method or normalization method to obtain standard data; Step 3.2, based on the standard data obtained in Step 3.1, construct the original matrix Rx, the original matrix Rx is as follows: Assuming there are m evaluation objects and n evaluation indicators, Xij is the correlation between the i-th evaluation object and the j-th indicator (i = 1, 2, 3, ..., m, j = 1, 2, 3, ..., n). Step 3.3, use the extreme value method to standardize the data in the original matrix Rx in Step 3.2 to obtain the standardized matrix Rx, where the conversion formula for standardizing the data in the original matrix Rx is as follows: Positive indicators: Negative indicators: Step 3.4, based on the standardized matrix Rx obtained in Step 3.3, calculate the weight of the i-th sample under the j-th indicator. The calculation formula is as follows: in, represents the sum of the jth column in the matrix R; X ij represents the i-th row and j-th column element in R, so P ij Represents the value of each element after normalization by column Step 3.5: Based on the weight of the i-th sample under the j-th indicator obtained in Step 3.4, calculate the entropy value of the j-th indicator. The calculation formula is as follows: Among them, (P ij *ln(P ij )) represents the information entropy of the association between factor i and indicator j; represents the average associated information entropy of all factors for indicator j; is the scaling factor used to eliminate the number of factors; e j Represents the entropy of the j-th indicator after standardization. Step 3.6, based on the entropy value of the j-th indicator obtained in Step 3.5, calculate the difference coefficient of the j-th indicator. The calculation formula is as follows: d j =1-e j Among them, e j represents the entropy of the jth index after standardization; d j Represents the coefficient of difference of the j-th indicator Step 3.7, based on the difference coefficient of the jth indicator obtained in Step 3.6, calculate the weight of each evaluation indicator. The calculation formula is as follows: Among them, (1-e j ) represents the coefficient of variation of the j-th indicator; represents the sum of the coefficients of difference of all indicators; w i Indicates the proportion of the difference coefficient of the ith indicator in the difference coefficients of all indicators, that is, the indicator weight Step 3.8: Based on the weights of each evaluation index obtained in Step 3.7, construct a weight analysis layer.

6. The method for comprehensive quality evaluation of cigarette making machines based on graph neural network according to claim 5, characterized in that: The Step 4, based on the different index evaluation layers obtained in Step 1, the correlation analysis layer obtained in Step 2, the weight analysis layer obtained in Step 3 and the defined graph convolutional network layer, combined with the attention mechanism, normalization operation and FC regression, constructs an optimized graph neural network for comprehensive quality evaluation of cigarette machines, including the following steps: Step 4.1, based on the weight analysis layer constructed in Step 3.8, the association analysis layer constructed in Step 2.4, the evaluation layer of different indicators constructed in Step 1.2 and the defined graph convolutional network layer, construct a preliminary graph neural network for comprehensive quality evaluation of cigarette making machines; Step 4.

2. Combine the attention mechanism, normalization operation and FC regression to optimize the preliminary cigarette machine comprehensive quality evaluation graph neural network obtained in Step 4.1 to obtain the optimized cigarette machine comprehensive quality evaluation graph neural network.

7. The method for comprehensive quality evaluation of cigarette making machines based on graph neural network according to claim 6, characterized in that: The graph convolutional network layer after the definition is defined as follows: in, Represents the weighted product of the output feature of the lth layer (the input of the l+1th layer) in the graph convolutional network and the corresponding weight in the filter; represents the summation on the neighborhood graph of i nodes; σ represents a nonlinear function; Represents the output features of the i-th node in the graph model after passing through the l+1th layer of the graph convolutional neural network.

8. The method for comprehensive quality evaluation of cigarette making machines based on graph neural network according to claim 6, characterized in that: The Step 4.2, combining the attention mechanism, normalization operation and FC regression, optimizes the preliminary cigarette machine comprehensive quality evaluation graph neural network obtained in Step 4.1 to obtain an optimized cigarette machine comprehensive quality evaluation graph neural network, including the following steps: Step 4.

21. Query all analysis data within the defined time window from the time series database, and build the analysis node H set based on the analysis data. The expression of the H set is as follows: H={h1,h2,...,h N },h i ∈RF Where N is the number of nodes in the graph; F represents the characteristic dimension of the input data Step 4.22, perform a linear transformation of the weight matrix on the analysis node H set constructed in Step 4.21 to obtain the updated feature representation set of the evaluation index node. The linear transformation of the weight matrix and the updated feature set expressions are as follows: Weight matrix linear transformation formula: h' i =Wh i Among them, Wh i Indicates the linear transformation of the linear matrix W; h' i Represents the value after linear transformation Updated feature set expression: h’={h'1,h’2,...,h’ N },h’ i ∈RF’ Among them, h' represents the updated feature representation set of the evaluation index node; F' represents the transformed evaluation index feature dimension Step 4.23, based on a shared self-attention mechanism a: RF'X RF'→R, calculate the attention correlation coefficient between the nodes in the updated feature set in Step 4.

22. The calculation formula is as follows: e ij =a(h’ i ||h’ j ) Among them, (h' i ||h' j ) represents the mutual correlation between node i and node j; a( ) represents the attention mechanism calculation; e ij Represents the attention correlation coefficient between node i and node j Step 4.24, concatenate the linearly transformed features of the target node and the neighboring nodes, and then use the weight vector After parameterization and inputting the nonlinear activation function LeakyReLu, we get the unnormalized attention correlation coefficient, which is expressed as follows: e ij =Leaky ReLU(a(Wh i ||Wh j )) Among them, a(Wh' i ||Wh' j ) represents the cross-correlation between nodes i and j after linear transformation; LeakyReLU represents the modified ReLU function; e ij Represents the attention correlation coefficient between nodes i and j after linear transformation Step 4.25, use the softmax function to normalize the unnormalized attention correlation coefficient obtained in Step 4.24 to obtain the attention correlation coefficient a between the nodes i,j , the calculation formula is as follows: Among them, exp(Leaky ReLU(a(Wh' i ||Wh' j ))) represents the attention correlation coefficient between the i node and the j node after linear transformation, and then passes through a modified ReLu function as the final attention correlation coefficient; softmax(e i,j ) is a processing function. The whole formula represents the normalization operation of the coefficient by the softmax function, that is, converting the attention correlation coefficient into a probability, which represents the degree of attention of node i to node j. Step 4.26, based on the attention correlation coefficient a between the nodes obtained in Step 4.25 i,j , perform weighted linear combination of neighbor points to obtain a new node, whose expression is as follows: in, represents the linear transformation of the neighborhood node j; Indicates a i,j The weight is used to perform linear weighted combination of the neighborhood nodes after linear transformation; σ represents a nonlinear function; Represents the i node after being updated by the l+1th layer of the graph convolutional network Step 4.

27. Based on the previous experience and knowledge, the evaluation system of the cigarette packaging machine score is calculated in the form of statistical distribution. The calculation formula is as follows: in, Indicates the deviation of the difference between the actual value and the average value from 3 times the historical standard deviation; Indicates the deviation of the difference between the actual value and the average value from 1.5 times the historical standard deviation Step 4.28, based on the evaluation system of cigarette packaging machine scores in Step 4.27, establish a primary scoring sample set, input the primary scoring sample set into the preliminary cigarette machine comprehensive quality evaluation graph neural network for training, and obtain the trained cigarette machine comprehensive quality evaluation graph neural network; Step 4.29: Based on the new nodes obtained in Step 4.26, the cigarette making machine comprehensive quality evaluation graph neural network trained in Step 4.28 is optimized to obtain an optimized cigarette making machine comprehensive quality evaluation graph neural network.

9. The method for comprehensive quality evaluation of cigarette making machines based on graph neural network according to claim 8, characterized in that: The step 5, inputting the real-time relevant data pre-processed in the production process of the cigarette rolling machine into the optimized cigarette rolling machine comprehensive quality evaluation graph neural network to obtain the real-time evaluation of the quality in the production process of the cigarette rolling machine, comprises the following steps: Step 5.1, collect the real-time data of the production process of the wrapping machine through the data acquisition system, and then use the filtering technology to reduce the noise of the real-time collected data to obtain the real-time related data after preprocessing; Step 5.2: Input the real-time related data preprocessed in Step 5.1 into the cigarette making machine comprehensive quality evaluation graph neural network optimized in Step 4.24 to obtain the real-time evaluation of the quality of the cigarette making machine during the production process.

Images