A production line design parameter correlation relationship reasoning method and system

Abstract

The present application relates to a production line design parameter correlation reasoning method and system. The method comprises: constructing a production line design parameter correlation graph, abstracting four-dimensional parameters such as configuration and dynamic type into nodes with id and edges with labels; constructing a feature extraction model, extracting self features through multi-head graph attention and four-type features through graph attention, and outputting a first feature matrix containing node self and adjacent information and a second feature matrix containing four-type information; obtaining a correlation reasoning model through third data preprocessing, building a bidirectional triangle graph neural network model and training, reasoning parameter correlation and adding to the graph. The system contains a graph construction, feature extraction and correlation reasoning module. The present application uses graph attention mechanism and bidirectional triangle graph neural network to solve the problems of poor modeling of non-symmetrical relationship, loss of cause-effect identification and other problems in production line design parameter correlation, and can effectively analyze parameter correlation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial production line, and particularly relates to a production line design parameter correlation reasoning method and system. BACKGROUND

[0002] With the rapid development of manufacturing industry, production lines have become complex manufacturing systems integrating high-cost investment and engineering technology. The complexity mainly lies in the multi-dimensional dynamic correlation characteristics among the four dimensions of configuration design, dynamic design, control design and optimization design. In the production line design process, a large number of parameters are involved, such as equipment speed, material flow, temperature, production rhythm, etc. However, the dependency relationship between these parameters is often not fully modeled, and there are some unknown implicit correlations. Especially in the current background of the rise of multi-variety, small-batch customized production mode, production lines need to be able to quickly adjust parameter configuration, which urgently needs accurate parameter correlation to guide dynamic optimization.

[0003] In traditional production line design, engineers' experience is excessively relied on to deal with parameter correlation. This way cannot operate efficiently when facing complex parameter systems, and it is difficult to achieve large-scale application. At the same time, in complex production lines, it is difficult for manual to comprehensively and accurately analyze the nonlinear coupling relationship between parameters, which leads to a long trial and error cycle and causes a lot of resource waste.

[0004] In the aspect of reasoning the correlation between design parameters, although it has received extensive attention in recent years, many scholars have developed various models and methods to model and reason the relationship between design parameters. However, most of the methods have limitations, for example, few scholars use systematic means, such as using graph neural networks to model design parameters and verifying the relationship through a new deep learning network architecture to bidirectional reasoning to perfect the correlation of design parameters and make it more reliable.

[0005] At present, as one of the existing reasoning methods, knowledge graph driven parameter reasoning has obvious shortcomings. First, the implicit relationship mining ability is insufficient, and the modeling ability for asymmetric relationships (such as "parameter A affects B, but B does not affect A") is poor, which is difficult to capture indirect correlations across processes. Second, there are functional shortcomings in the reasoning mechanism, such as the lack of reverse causal recognition, which cannot effectively learn from data, and lack of self-validation ability, etc.

[0006] Therefore, the present application provides a production line design parameter correlation reasoning method and system to solve the above technical problems. SUMMARY

[0007] In view of the above problems, the purpose of the present application is to provide a production line design parameter correlation reasoning method and system for solving the problems of poor modeling of asymmetric relationships, loss of causal identification, etc.

[0008] The present application provides a production line design parameter correlation reasoning method, comprising: S1, constructing a correlation graph of production line design parameters; using a visualization tool to abstract the design parameters of the four dimensions of configuration, motion, control and optimization of the production line into nodes with unique ids, labeling the classification attributes of each node according to the four type categories; and abstracting the mutual correlation between each design parameter into a labeled edge; based on the node id, four type category attributes and labeled edge, constructing a production line correlation graph containing node feature attributes and edge relationship labels;

[0009] S2, feature extraction model construction, comprising:

[0010] Multi-head graph attention feature extraction:

[0011] First data preprocessing: arranging all nodes in ascending order according to node id and labeling the labels, and constructing a first adjacency matrix according to the correlation between the head node and the tail node;

[0012] First model building: adopting a multi-head attention network, converting the node id into a matrix through an embedding layer, and building a four-layer multi-head graph attention convolution layer as a first initial model; wherein the activation function of the first three convolution layers of the four-layer multi-head graph attention convolution layer is LeakyRelu, and the activation function of the fourth convolution layer of the four-layer multi-head graph attention convolution layer is Softmax;

[0013] First model training: input the sorted node id and the first adjacency matrix into the first initial model, train the first initial model based on the cross-entropy loss function and the Adam optimizer, obtain a multi-head graph attention mechanism feature extraction model, and the multi-head graph attention mechanism feature extraction model outputs a first feature matrix containing node feature and adjacent node correlation information;

[0014] Four-type feature extraction of graph attention:

[0015] Second data preprocessing: labeling the labels according to the four-type categories to which the nodes belong, and constructing a second adjacency matrix;

[0016] Second model building: adopting a general graph attention mechanism, building a five-layer graph attention convolution layer as a second initial model; wherein the activation function of the first four convolution layers in the five-layer graph attention convolution layer is LeakyRelu, and the activation function of the fifth convolution layer in the five-layer graph attention convolution layer is Softmax;

[0017] Second model training: input the first feature matrix and the second adjacency matrix into the second initial model, train the second initial model based on a cross-entropy loss function and an Adam optimizer, and obtain a graph attention mechanism type four feature extraction model, wherein the graph attention mechanism type four feature extraction model outputs a second feature matrix containing four type information;

[0018] S3, reasoning of the correlation relationship of the production line design parameters:

[0019] Third data preprocessing: rearranging the second feature matrix according to the ascending order of node id, abstracting the first adjacency matrix and the second adjacency matrix into an undirected graph; traversing the undirected graph through a breadth-first search algorithm (BFS), obtaining the shortest path between any two nodes with a path length greater than 2, determining the intermediate node based on the shortest path, extracting the second feature matrix corresponding to the first-middle-tail three-node pair and concatenating it into a feature vector in order; the feature vector containing the correlation relationship between the first node and the tail node is used as learning data, and the rest is used as reasoning data; performing label marking, feature matrix exchange and reverse label calculation on the learning data, dividing the learning data into a training set, a validation set and a test set in a ratio of 8:1:1, and simultaneously performing reverse feature matrix processing on the reasoning data;

[0020] Third model building: extracting the second feature matrix corresponding to the positions of the first node, the tail node and the intermediate node from the learning data; calculating the feature values and weight coefficients of each node based on the extracted second feature matrix; providing a bidirectional triangular graph neural network model composed of seven fully connected layers and one convergence layer as a third initial model; inputting the second feature matrix of the first node, the tail node and the intermediate node into three independent fully connected layers respectively, and the activation function connected by the three fully connected layers is LeakyRelu; inputting the matrix output by the three fully connected layers into the convergence layer, introducing the two end node correction coefficients and the intermediate node correction coefficient in the convergence layer to enhance the feature importance of the first node and the tail node, and generating a single matrix after convergence; sequentially passing the single matrix after convergence through four fully connected layers, wherein the activation function connected by the first three fully connected layers of the four fully connected layers is LeakyRelu, and the activation function connected by the fourth fully connected layer is Softmax;

[0021] Third model training: inputting the training set into the third initial model, training the third initial model based on a cross-entropy loss function, an Adam optimizer, a StepLR learning rate decay strategy, a Dropout regularization method and a LayerNorm regularization method, and obtaining a correlation relationship reasoning model;

[0022] Correlation relationship reasoning: input the reasoning data and the reasoning data processed by the inverse feature matrix into the correlation relationship reasoning model respectively, add the results, if the sum is N (N is the type of correlation relationship), then the two nodes exist the correlation relationship of the corresponding label, otherwise not; the first node id, tail node id and relationship name of the existing relationship are inferred by the correlation relationship reasoning model, and the function statement in the visualization tool is added into the graph.

[0023] Preferably, the expression of the activation function LeakyRelu is: LeakyReLU(x) = max(0, x) + a * min(0, x); wherein, a is the slope of the negative activation value, and x is the input matrix.

[0024] Preferably, the expression of the activation function Softmax is: Wherein, X i is the i-th element in the input matrix, X j is the j-th element in the input matrix, and K is the column number of the input matrix.

[0025] Preferably, the expression of the cross-entropy loss function is: Wherein, N is the total number of categories, y i is the encoding of the true label, and P i is the probability value of the i-th category predicted by the model.

[0026] Preferably, the calculation expression of the feature value is: T = <x.x>, wherein, X is the second feature matrix of the node; and the calculation expression of the weight coefficient is: Wherein, T i is the feature value of the i-th node.

[0027] Preferably, the expression of the fully connected layer is: y = Wx + b; wherein, W is the weight matrix, x is the input matrix, and b is the bias vector; and the expression of the convergence layer is: y = a ht ·w h ·x h +α m ·w m ·x m +α ht ·w t ·x t ; wherein, a ht is the correction coefficient of the two end nodes, equal to 0.45, a m is the correction coefficient of the intermediate node, equal to 0.1, W h is the first node weight coefficient, W m is the intermediate node weight coefficient, W t is the tail node weight coefficient, x hx is a second feature matrix corresponding to the head node m x is a second feature matrix corresponding to the intermediate node t x is a second feature matrix corresponding to the tail node.

[0028] Preferably, the expression of the StepLR learning rate decay strategy is: Wherein, lr is an initial learning rate, gamma is a learning rate decay factor, e is an index of the training number, and s is a learning rate decay step. initial The expression of the Dropout regularization method is: Wherein, p is a dropout probability, m is a mask matrix, and takes 0 or 1, and is subject to Bernoulli distribution; x is an input matrix; the expression of the LayerNorm regularization method is: Wherein, gamma1 is a learnable parameter; mu is the mean of x, is the standard deviation containing a numerical stability term, and beta is a learnable offset coefficient.

[0029] Preferably, the visualization tool is neo4j.

[0030] The application also provides a production line design parameter correlation relationship analysis system, comprising:

[0031] A graph construction module is used to execute the S1 step in claim 1, and a correlation relationship graph of the production line design parameters is constructed.

[0032] A feature extraction module is used to execute the S2 step in claim 1, and a feature extraction model is constructed, comprising a multi-head graph attention self-feature extraction unit and a graph attention four-type feature extraction unit.

[0033] A correlation relationship reasoning module is used to execute the S3 step in claim 1, and correlation relationship reasoning of the production line design parameters is performed, comprising a third data preprocessing unit, a third model building unit, a third model training unit and a correlation relationship reasoning unit.

[0034] Compared with the related art, the production line design parameter correlation reasoning method and system provided by the application has the following beneficial effects: 1. Strong asymmetric relationship modeling capability: the graph attention mechanism and the bidirectional triangular graph neural network are used to break through the modeling limitations of traditional methods on asymmetric parameter relationships and accurately capture complex dependencies between parameters; 2. Complete causal relationship identification: combined with multi-head graph attention and four-type feature extraction, the system systematically integrates node itself and four-dimensional category information to avoid loss of causal relationships in analysis; 3. Self-verification and generalization capability: through reverse feature matrix processing and bidirectional input verification of reasoning data, the reliability of the model results is ensured; 4. Combined with regularization methods such as Dropout and LayerNorm, the ability of the model to learn from data and generalize to different scenarios is enhanced; 5. Efficient analysis and visualization: based on the neo4j visualization graph, the intuitive presentation and dynamic update of parameter relationships are realized, the complexity of production line design parameter correlation analysis is reduced, the design optimization efficiency is improved, and strong support is provided for intelligent manufacturing production line parameter collaborative design. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A flowchart of a production line design parameter correlation reasoning method is provided in the application.

[0036] Figure 2 A schematic diagram of the correlation relationship graph is provided in the application.

[0037] Figure 3 A multi-head graph attention mechanism self-feature extraction model structure diagram is provided in the application.

[0038] Figure 4 A graph attention mechanism four-type feature extraction model structure diagram is provided in the application.

[0039] Figure 5 A diagram of the head node, tail node and intermediate node is provided in the application.

[0040] Figure 6 A bidirectional triangular graph neural network reasoning model structure diagram is provided in the application.

[0041] Figure 7 A production line design parameter correlation analysis system structure diagram is provided in the application. DETAILED DESCRIPTION

[0042] The application provides a production line design parameter correlation reasoning method and system, aiming to solve the problems of poor asymmetric relationship modeling and causal relationship loss in existing production line design parameter correlation relationships.

[0043] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the scope of protection of the present application.

[0044] Please refer to the accompanying Figures 1-7 As shown in the drawings, the present application provides a production line design parameter correlation reasoning method, comprising: S1, constructing a correlation graph of production line design parameters; using a visualization tool to abstract the design parameters of four dimensions of configuration, motion, control and optimization of the production line into nodes with unique ids, labeling the classification attributes of each node according to the four types; and abstracting the mutual correlation between each design parameter into a labeled edge; based on the node id, four type classification attributes and labeled edge, a production line correlation graph containing node feature attributes and edge relationship labels is constructed.

[0045] In the present embodiment, the visualization tool uses neo4j graph database to complete the construction and visualization display of the production line correlation graph. Of course, the visualization tool of the present application is not limited to neo4j, and any graph database or visualization platform (such as TigerGraph, Graphviz, etc.) that can realize node attribute labeling, edge relationship labeling and graph dynamic updating is suitable for the technical solution of the present application.

[0046] S2, feature extraction model construction, including multi-head graph attention self-feature extraction and graph attention four-type feature extraction. It aims to convert the original design parameter data into a low-dimensional space mapping with stronger representation ability for subsequent correlation reasoning.

[0047] It should be noted that the purpose of multi-head graph attention self-feature extraction is to map the id data of the nodes in the graph to a low-dimensional space, thereby increasing the amount of information it contains; at the same time, considering that there is a correlation between the node and the adjacent node, in order to facilitate subsequent reasoning, part of the useful information of the adjacent node needs to be added to its own features. However, after completing the extraction of the node's own features, since the node's feature matrix does not contain the node's four-type information, and the correlation of the design parameters will be affected by the four-type, after completing the multi-head graph attention self-feature extraction, the node needs to be extracted for four-type features using the line graph attention four-type feature extraction method.

[0048] Specifically, the step of the multi-head graph attention mechanism self-feature extraction includes: first data preprocessing: arranging all nodes in ascending order according to node id and marking labels, and constructing a first adjacency matrix according to the association relationship between the head node and the tail node; first model building: adopting a multi-head attention network, converting the node id into a matrix through an embedding layer, and building a four-layer multi-head graph attention convolution layer as a first initial model; wherein the activation function of the first three convolution layers of the four-layer multi-head graph attention convolution layer is LeakyRelu, and the activation function of the fourth convolution layer of the four-layer multi-head graph attention convolution layer is Softmax; first model training: inputting the sorted node id and the first adjacency matrix into the first initial model, training the first initial model based on a cross-entropy loss function and an Adam optimizer, obtaining a multi-head graph attention mechanism self-feature extraction model, and the multi-head graph attention mechanism self-feature extraction model outputs a first feature matrix containing node self-feature and adjacent node association information.

[0049] Further, the step of the graph attention four-type feature extraction includes: second data preprocessing: marking labels according to the four-type categories to which the nodes belong, and constructing a second adjacency matrix; second model building: adopting a general graph attention mechanism, building a five-layer graph attention convolution layer as a second initial model; wherein the activation function of the first four convolution layers in the five-layer graph attention convolution layer is LeakyRelu, and the activation function of the fifth convolution layer in the five-layer graph attention convolution layer is Softmax; second model training: inputting the first feature matrix and the second adjacency matrix into the second initial model, training the second initial model based on a cross-entropy loss function and an Adam optimizer, obtaining a graph attention mechanism four-type feature extraction model, and the graph attention mechanism four-type feature extraction model outputs a second feature matrix containing four-type information.

[0050] In the embodiment, the expression of the activation function LeakyRelu is: LeakyReLU(x) = max(0, x) + a min(0, x); wherein a is the slope of the negative activation value, and x is the input matrix; the expression of the activation function Softmax is: wherein X i is the i-th element in the input matrix, X j is the j-th element in the input matrix, K is the column number of the input matrix, the Softmax function can convert an unnormalized matrix (usually the output of the model) into a probability distribution, so that each value in the output is between 0 and 1, and the sum of all outputs is 1; the expression of the cross-entropy loss function is: wherein N is the total number of categories, y i is the encoding of the true label, and P iThe probability value of the i-th class predicted by the model, that is, the output after Softmax normalization, is denoted as y. In addition, the Adam optimizer is used to accelerate the convergence speed of the loss value of the model.

[0051] S3, reasoning of the correlation relationship of the line design parameters. The role of this step is to realize the reasoning of the correlation relationship between parameters through a bidirectional triangle graph neural network, and to solve the problems of intermediate node influence and relationship directionality being ignored in traditional graph neural networks.

[0052] Further, the step of reasoning the association relationship of the production line design parameters specifically comprises: S31, third data preprocessing: rearranging the second feature matrix according to the ascending order of node id, and abstracting the first adjacency matrix and the second adjacency matrix into an undirected graph; traversing the undirected graph through a breadth-first search algorithm (BFS) to obtain the shortest path with a path length greater than 2 between any two nodes, determining the intermediate node based on the shortest path, extracting the second feature matrix corresponding to the first-middle-tail three-node pair and splicing it into a feature vector in order; taking the feature vector containing the association relationship between the first node and the tail node as learning data, and taking the rest as reasoning data; performing label marking, feature matrix exchange and reverse label calculation on the learning data, dividing the learning data into a training set, a validation set and a test set in a ratio of 8:1:1, and synchronously performing reverse feature matrix processing on the reasoning data; S32, third model building: extracting the second feature matrix corresponding to the first node, the tail node and the intermediate node from the learning data; calculating the feature values and weight coefficients of each node based on the extracted second feature matrix; providing a bidirectional triangular graph neural network model composed of seven fully connected layers and one convergence layer as a third initial model; inputting the second feature matrix of the first node, the tail node and the intermediate node into three independent fully connected layers respectively, and the activation function connected by the three fully connected layers is LeakyRelu; inputting the matrix output by the three fully connected layers into the convergence layer, introducing the end node correction coefficient and the intermediate node correction coefficient in the convergence layer to enhance the feature importance of the first node and the tail node, and generating a single matrix after convergence; sequentially passing the single matrix after convergence through four fully connected layers, wherein the activation function connected by the first three fully connected layers of the four fully connected layers is LeakyRelu, and the activation function connected by the fourth fully connected layer is Softmax; S33, third model training: inputting the training set into the third initial model, training the third initial model based on the cross-entropy loss function, the Adam optimizer, the StepLR learning rate decay strategy, the Dropout regularization method and the LayerNorm regularization method, and obtaining an association relationship reasoning model; S34, association relationship reasoning: inputting the reasoning data and the reasoning data processed by the reverse feature matrix into the association relationship reasoning model respectively, adding the results, and if the sum is N (N is the type of association relationship), then the two nodes have the corresponding label association relationship, otherwise not; reasoning the first node id, the tail node id and the relationship name through the association relationship reasoning model, and adding the function statements in the visualization tool into the graph.

[0053] In the embodiment, the expression of reverse label calculation is: y -= 2N1-1-y2, wherein: y- is a reverse label, N1 is a label category, and y2 is a label value; a calculation expression of the eigenvalue is: T = <x.x>, wherein X is the second feature matrix of the node; a calculation expression of the weight coefficient is: wherein Ti is the eigenvalue of the i th node. A calculation expression of the full connection layer is: y = Wx + b; wherein W is a weight matrix, x is an input matrix, and b is a bias vector; a calculation expression of the pooling layer is: y = α ht ·w h ·x h +α m ·w m ·x m +α ht ·w t ·x t ; wherein α ht is a correction coefficient of the two end nodes and is equal to 0.45, α m is a correction coefficient of the intermediate node and is equal to 0.1, W h is a weight coefficient of the first node, W m is a weight coefficient of the intermediate node, W t is a weight coefficient of the last node, x h is the second feature matrix corresponding to the first node, x m is the second feature matrix corresponding to the intermediate node, and x t is the second feature matrix corresponding to the last node.

[0054] It is worth mentioning that a calculation expression of the StepLR learning rate decay strategy is: wherein lr initial is an initial learning rate, γ is a learning rate decay factor, e is an index of the training number, and s is a learning rate decay step; a calculation expression of the Dropout regularization method is: wherein p is a dropout probability; m is a mask matrix, which is 0 or 1 and is subject to a Bernoulli distribution; and x is an input matrix; a calculation expression of the LayerNorm regularization method is: wherein γ1 is a learnable parameter; μ is a mean value of x, is a standard deviation with a numerical stability term, and β is a learnable offset coefficient.

[0055] The application further provides a production line design parameter correlation relationship analysis system 100, comprising:

[0056] a graph construction module 10 configured to perform the S1 step in claim 1 to construct a correlation graph of the production line design parameters;

[0057] The feature extraction module 20 is configured to perform the step S2 in claim 1, and constructs a feature extraction model, including a multi-head graph attention self-feature extraction unit 201 and a graph attention four-type feature extraction unit 202.

[0058] The correlation reasoning module 30 is configured to perform the step S3 in claim 1, and performs correlation reasoning of the production line design parameters, including a third data preprocessing unit 301, a third model building unit 302, a third model training unit 303 and a correlation reasoning unit 304.

[0059] Compared with the related art, the production line design parameter correlation reasoning method and system provided by the application has the following beneficial effects: 1. Strong asymmetric relationship modeling capability: the graph attention mechanism and the bidirectional triangular graph neural network are used to break through the modeling limitations of traditional methods on asymmetric parameter relationships, and to accurately capture the complex dependencies between parameters; 2. Complete causal relationship identification: combined with multi-head graph attention and four-type feature extraction, the node itself and four-dimensional category information are systematically integrated to avoid loss of causal relationships in analysis; 3. Self-verification and generalization ability are excellent: through reverse feature matrix processing and bidirectional input verification of reasoning data, the reliability of the model results is ensured; 4. Combined with Dropout, LayerNorm and other regularization methods, the ability of the model to learn from data and generalize to different scenarios is enhanced; 5. Efficient analysis and visualization: based on the neo4j visualization graph, the intuitive presentation and dynamic update of the parameter relationship are realized, the complexity of the production line design parameter correlation analysis is reduced, and the design optimization efficiency is improved, which provides strong support for intelligent manufacturing production line parameter collaborative design.

[0060] The above-mentioned embodiments should be understood as illustrative rather than limiting the scope of the application. The scope of the application is defined by the claims. Some non-essential improvements and adjustments made by those skilled in the art without departing from the spirit and scope of the application also belong to the protection scope of the application.

Claims

1. A production line design parameter correlation relationship inference method, characterized by, The method comprises: S1, constructing a correlation graph of production line design parameters; using a visualization tool to abstract the design parameters of the four dimensions of the configuration, motion, control and optimization of the production line into nodes with unique ids, labeling the classification attributes of each node according to the four type categories; and abstracting the mutual correlation between the design parameters into edges with labels; based on the node id, the four type category attributes and the edges with labels, a production line correlation graph is constructed, which includes node feature attributes and edge relationship labels; S2, feature extraction model construction, comprising: Multi-head graph attention feature extraction: First data preprocessing: arrange all nodes in ascending order of node id and mark the labels, and construct a first adjacency matrix according to the correlation between the head node and the tail node; First model building: adopt a multi-head attention network, convert the node id into a matrix through an embedding layer, and build a four-layer multi-head graph attention convolution layer as a first initial model; wherein the activation function of the first three convolution layers of the four-layer multi-head graph attention convolution layer is LeakyRelu, and the activation function of the fourth convolution layer of the four-layer multi-head graph attention convolution layer is Softmax; First model training: input the sorted node id and the first adjacency matrix into the first initial model, train the first initial model based on the cross-entropy loss function and the Adam optimizer, obtain a multi-head graph attention mechanism feature extraction model, and the multi-head graph attention mechanism feature extraction model outputs a first feature matrix containing node self-feature and adjacent node correlation information; Four-type feature extraction of graph attention: Second data preprocessing: label the labels according to the four-type categories to which the nodes belong, and construct a second adjacency matrix; Second model building: adopt a general graph attention mechanism, build a five-layer graph attention convolution layer as a second initial model; wherein the activation function of the first four convolution layers in the five-layer graph attention convolution layer is LeakyRelu, and the activation function of the fifth convolution layer in the five-layer graph attention convolution layer is Softmax; Second model training: input the first feature matrix and the second adjacency matrix into the second initial model, train the second initial model based on the cross-entropy loss function and the Adam optimizer, obtain a graph attention mechanism four-type feature extraction model, and the graph attention mechanism four-type feature extraction model outputs a second feature matrix containing four-type information; S3, reasoning of the correlation of production line design parameters: Third data preprocessing: rearrange the second feature matrix according to the ascending order of node id, abstract the first adjacency matrix and the second adjacency matrix into an undirected graph; traverse the undirected graph through the breadth-first search algorithm BFS, obtain the shortest path between any two nodes with a path length greater than 2, determine the intermediate node based on the shortest path, extract the second feature matrix corresponding to the first-intermediate-tail three nodes and splice it into a feature vector in order; The feature vector containing the association relationship of the head node and the tail node label is taken as learning data, and the rest is taken as reasoning data; label marking, feature matrix exchange and reverse label calculation are performed on the learning data, the learning data is divided into a training set, a validation set and a test set in a ratio of 8:1:1, and reverse feature matrix processing is simultaneously performed on the reasoning data; Third model building: a second feature matrix corresponding to the positions of the head node, the tail node and the intermediate node is extracted from the learning data; the feature values and the weight coefficients of each node are calculated based on the extracted second feature matrix; a bidirectional triangular graph neural network model composed of seven fully connected layers and one convergence layer is provided as a third initial model; the second feature matrices of the head node, the tail node and the intermediate node are respectively input into three independent fully connected layers, and the activation function connected by the three fully connected layers is LeakyRelu; the matrices output by the three fully connected layers are input into the convergence layer, and the end node correction coefficient and the intermediate node correction coefficient are introduced in the convergence layer to enhance the feature importance of the head node and the tail node, and a single matrix is generated after convergence; the single matrix after convergence is sequentially passed through four fully connected layers, wherein the activation function connected by the first three fully connected layers of the four fully connected layers is LeakyRelu, and the activation function connected by the fourth fully connected layer is Softmax; Third model training: the training set is input into the third initial model, and the third initial model is trained based on a cross-entropy loss function, an Adam optimizer, a StepLR learning rate decay strategy, a Dropout regularization method and a LayerNorm regularization method to obtain an association relationship reasoning model; Association relationship reasoning: the reasoning data and the reasoning data processed by the reverse feature matrix are respectively input into the association relationship reasoning model, the results are added, and if the sum is N1, N1 is the type of association relationship, then the two nodes have the corresponding label association relationship, otherwise not; the head node id, the tail node id and the relationship name of the existing relationship are inferred by the association relationship reasoning model, and the function statements in the visualization tool are added to the graph.

2. The method of claim 1, wherein, The expression of the activation function LeakyRelu is: ; wherein, a is the slope of the negative activation value control, x is the input matrix.

3. The method of claim 1, wherein, The expression of the activation function Softmax is: ; wherein, x i is the i-th element in the input matrix, x j is the j-th element in the input matrix, and K is the number of columns of the input matrix.

4. The method of claim 3, wherein, The expression of the cross-entropy loss function is: ; where, N 2 is the total number of classes, y i is the encoding of the true label, P i is the probability value of the i-th class predicted by the model.

5. The method of claim 4, wherein, The calculation expression of the characteristic value is: T = <x1.x1>, wherein x1 is the second characteristic matrix of the node; and the calculation expression of the weight coefficient is: , wherein T i is the characteristic value of the i th node.

6. The method of claim 5, wherein , the expression of the full connection layer is: ; wherein, W is a weight matrix, x is an input matrix, and b is a bias vector; the expression of the convergence layer is: ; wherein, α ht is a two-end node correction coefficient equal to 0.45, α m is a middle node correction coefficient equal to 0.1, W h is a head node weight coefficient, W m is a middle node weight coefficient, W t is a tail node weight coefficient, x h is a second feature matrix corresponding to the head node, x m is a second feature matrix corresponding to the middle node, and x t is a second feature matrix corresponding to the tail node.

7. The method of claim 6, wherein, The expression of the StepLR learning rate decay strategy is: ; wherein, lr initial is an initial learning rate, γ is a learning rate decay factor, e is an index of a training number, and s is a learning rate decay step; the expression of the Dropout regularization method is: ; wherein, p is a dropout probability; m is a mask matrix, and is 0 or 1, and is subject to a Bernoulli distribution; x is an input matrix; the expression of the LayerNorm regularization method is: ; wherein, γ1 is a learnable parameter; μ is a mean value of x, is a standard deviation containing a numerical stability term, and β is a learnable offset coefficient.

8. The method of claim 1, wherein, The visualization tool is neo4j.

9. A production line design parameter correlation relationship analysis system characterized by comprising: Comprise: A graph construction module for performing the S1 step in claim 1 to construct an association relationship graph of the production line design parameters; A feature extraction module for performing the S2 step in claim 1 to construct a feature extraction model, including a multi-head graph attention self-feature extraction unit and a graph attention four-type feature extraction unit; An association relationship reasoning module for performing the S3 step in claim 1 to perform association relationship reasoning of the production line design parameters, including a third data preprocessing unit, a third model building unit, a third model training unit and an association relationship reasoning unit.

Images