Cross-linking mixed process step product quality prediction method based on parallel collaborative graph network

Abstract

The invention discloses a cross-linking mixed process step product quality prediction method based on a parallel collaborative graph network, and relates to the technical field of product quality prediction, and the method comprises the steps: obtaining the input features of each process step of a to-be-detected product in the production and processing process, forming an input feature data set, carrying out the preprocessing of the input feature data set, and carrying out the prediction of the quality of the product according to the actual processing process; constructing a matching graph, an adjacent matrix and a mask graph of the to-be-tested product, and representing a connection relation and a transmission relation between the working steps by a graph structure; and finally, inputting the data into a trained product quality prediction model to obtain a product quality prediction result of the to-be-tested product. According to the product quality prediction model adopted by the invention, the influence conditions between the working step characteristics and the working procedures in the production and processing process are fully considered, and the cross-linking and mixing relationship between the working steps is further extracted through the advantages of the graph structure, so that the actual conditions in the production and processing field are better met, and higher prediction precision can be obtained.

Description

Technical Field

[0001] This application relates to the technical field of product quality prediction, and particularly to a method for predicting the quality of cross-linked and mixed process products based on a parallel collaborative graph network. Background Art

[0002] With the rapid development of industrial technology and the intensification of market competition, product quality has become the key to the survival and development of enterprises. However, traditional quality control methods mainly rely on post-inspection, which is not only costly but also easily causes irreversible damage to workpieces. Therefore, the current technology has put forward the demand for quality prediction that is synchronized with product processing, aiming to quickly and accurately predict the production quality status of products. This helps enterprises optimize processing plans, reduce manual inspection costs, and ultimately improve the production efficiency of high-quality products.

[0003] However, traditional quality prediction methods usually only focus on solving problems in a single production line. In actual manufacturing, due to the high customization of products and the increase in production speed, there are differences in the types of station platforms that different products pass through. In addition, there are complex situations where multiple stations process the same product in parallel and across process steps, which makes the product production line process flexible and variable, and the dependency relationships between process steps are complex and highly coupled, greatly increasing the difficulty of quality prediction during the product processing process. Specifically, current quality prediction methods are mainly divided into two categories: quality prediction based on physical information and quality prediction based on data-driven. Due to the complex cross-linking of each process step and the uncertainty of the production line, it is difficult to establish an accurate model for high-precision quality prediction using quality prediction methods based on physical information. Although data-driven quality prediction can make full use of the data mining capabilities of modern computers and get rid of the limitations of physical modeling, it also faces problems such as low prediction accuracy and poor reusability because it cannot consider the mutual influence of process parameters during the assembly process.

[0004] Some subsequent studies have added a parallel process and cross-process information integration module to data-driven quality prediction. For example, Chinese Patent Application Publication No. CN109711714A discloses a "method for predicting the quality of manufacturing and assembly products based on a parallel long short-term memory network", which can process parallel processes and cross-process production information through the introduction of an information integration module, thereby improving the accuracy of product quality prediction. However, in the face of the changing situations of different product production lines, the information integration module needs to process the corresponding information one by one, which significantly reduces the operation efficiency and parallelization degree of the model. At the same time, the product quality prediction model after adding the information integration module is also difficult to cope with the production and processing environment with more complex process connections.

[0005] In view of the above challenges, there is an urgent need to develop a product quality prediction method that can process complex multi-step couplings at high speed and with high accuracy and is flexible for production lines with variable product quality. Summary of the Invention

[0006] The purpose of this application is to provide a cross-linked hybrid process product quality prediction method based on a parallel collaborative graph network, which can flexibly handle the cross-linked hybrid process conditions existing in the production process and has higher prediction accuracy.

[0007] To achieve the above purpose, this application provides the following solutions:

[0008] A cross-linked hybrid process product quality prediction method based on a parallel collaborative graph network, comprising the following steps:

[0009] Obtain the input feature dataset of the product to be tested during the production process; the input feature dataset includes the input features of each process step of the product to be tested; the input feature is any process parameter and / or any measurement parameter of the product to be tested during the production process.

[0010] Preprocess the input feature dataset to obtain the preprocessed input feature dataset; the preprocessing includes outlier processing, missing value filling, grouping by process step, and feature dimensionality reduction; in the preprocessed input feature dataset, there are several input feature groups corresponding to each process step respectively.

[0011] According to the actual processing procedure of the product to be tested, construct the matching graph, adjacency matrix, and mask graph of the product to be tested; the matching graph includes several real nodes corresponding to each process step and a virtual node; the adjacency matrix is used to represent the connection relationship and transmission relationship between the real nodes in the matching graph; the mask graph is used to represent the process steps actually passed and the process steps not passed by the actual processing procedure of the product to be tested.

[0012] Input the preprocessed input feature dataset, the matching graph, adjacency matrix, and mask graph of the product to be tested into the trained product quality prediction model to obtain the product quality prediction result of the product to be tested; the product quality prediction model is a neural network model based on the parallel collaborative graph network structure, and the parallel collaborative graph network structure uses a bidirectional graph attention network to aggregate the cross-linked and hybrid information between each process step to obtain the process step aggregation feature, and performs regression according to the process step aggregation feature through a multi-layer feedforward neural network to obtain the product quality prediction result of the product to be tested.

[0013] Optionally, the product quality prediction model includes several process step feature encoding modules corresponding to each process step respectively, a bidirectional graph attention network, and a final processing quality prediction module; the bidirectional graph attention network includes a forward graph attention network and a reverse graph attention network.

[0014] Input the preprocessed input feature dataset, the matching graph of the product to be tested, the adjacency matrix, and the mask graph into the trained product quality prediction model to obtain the product quality prediction result of the product to be tested. The specific steps are as follows:

[0015] Input the input feature groups of each process step in the preprocessed input feature dataset into the process step feature encoding module corresponding to each process step for encoding to obtain the encoded and normalized quality parameter features of each process step.

[0016] According to the matching graph and the adjacency matrix of the product to be tested, construct graph structures in the forward graph attention network and the reverse graph attention network respectively; the effective nodes in the graph structures of the forward graph attention network and the reverse graph attention network all point to the virtual node; the effective node is the real node corresponding to the process step actually passed by the actual processing process of the product to be tested; the virtual node is used to extract the implicit quality parameters of each effective node in the graph structure.

[0017] Input the encoded and normalized quality parameter features of each process step into the corresponding real nodes in the graph structures of the forward graph attention network and the reverse graph attention network respectively, and use the bidirectional graph attention network to aggregate the cross-linked and mixed information between each process step to obtain the process step aggregation feature; when aggregating the cross-linked and mixed information between each process step, use the mask graph to mask the encoded and normalized quality parameter features of the process steps corresponding to the invalid nodes; the invalid node is the real node corresponding to the process step not passed by the actual processing process of the product to be tested. The process step aggregation feature is the feature of the virtual node information of each sample that aggregates all the effective node information through the attention mechanism.

[0018] In the final processing quality prediction module, use a multi-layer feedforward neural network to perform regression on the process step aggregation feature to obtain the product quality prediction result of the product to be tested.

[0019] According to the specific embodiments provided in this application, the following technical effects are disclosed in this application:

[0020] The present application provides a method for predicting the quality of cross-linked and mixed process products based on a parallel collaborative graph network. In this method, first, the input features of each process during the production and processing of the product to be tested are obtained, and an input feature data set is formed. Then, preprocessing such as outlier processing, missing value filling, grouping by process, and feature dimensionality reduction is performed on the input feature data set. Subsequently, according to the actual processing procedures of the product to be tested, a matching graph, an adjacency matrix, and a mask graph of the product to be tested are constructed to represent the connection relationship and transmission relationship between each process in the form of a graph structure. Finally, the preprocessed input feature data set, the matching graph, the adjacency matrix, and the mask graph of the product to be tested are all input into the trained product quality prediction model to obtain the product quality prediction result of the product to be tested. The product quality prediction model used is a neural network model based on a parallel collaborative graph network structure, in which a bidirectional graph attention network is used to aggregate the cross-linked and mixed information between each process to obtain a process aggregation feature, and a multi-layer feedforward neural network is used to perform regression based on the process aggregation feature to obtain the product quality prediction result of the product to be tested.

[0021] The above solution of the present application fully considers the influence situation between each station and process in the manufacturing and processing process, and further extracts the cross-linked and mixed relationship between processes through the advantages of the graph structure, which is more in line with the actual situation in the production and processing field and can achieve higher prediction accuracy. In addition, in the face of the complex association information between the cross-linked and mixed processes of different products, only by establishing the mapping relationship from the input features of different processes to the corresponding nodes of the graph structure according to the processing procedure information of product manufacturing, the product quality prediction model based on the parallel collaborative graph network can fuse the association information between the input features of each process and perform product quality prediction, without being restricted by the process transmission method and the product production and processing form, and can flexibly handle the situation of cross-linked and mixed processes existing in the production and processing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is a flowchart of a method for predicting the quality of cross-linked and mixed process products based on a parallel collaborative graph network provided by an embodiment of the present application.

[0024] Figure 2 It is a flowchart of step A2 in a method for predicting the quality of cross-linked and mixed process products based on a parallel collaborative graph network provided by an embodiment of the present application.

[0025] Figure 3It is a flowchart of step A4 in a method for predicting the quality of cross-linked hybrid process products based on a parallel collaborative graph network provided by an embodiment of the present application.

[0026] Figure 4 It is a schematic diagram of a quality feature encoding module in a method for predicting the quality of cross-linked hybrid process products based on a parallel collaborative graph network provided by an embodiment of the present application.

[0027] Figure 5 It is a schematic diagram of a bidirectional graph attention module in a method for predicting the quality of cross-linked hybrid process products based on a parallel collaborative graph network provided by an embodiment of the present application.

[0028] Figure 6 It is a schematic diagram of a final processing prediction module in a method for predicting the quality of cross-linked hybrid process products based on a parallel collaborative graph network provided by an embodiment of the present application.

[0029] Figure 7 It is a schematic diagram of the functional modules of a system for predicting the quality of cross-linked hybrid process products based on a parallel collaborative graph network provided by an embodiment of the present application.

[0030] Figure 8 It is a schematic diagram of the structure of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0031] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0032] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0033] In an exemplary embodiment, as Figure 1 shown, a method for predicting the quality of cross-linked hybrid process products based on a parallel collaborative graph network is provided, including the following steps:

[0034] A1. Obtain an input feature data set of the product to be tested during the production and processing process; the input feature data set includes the input features of each process step of the product to be tested during the production and processing process; the input features are any process parameters and / or any measurement parameters of the product to be tested during the production and processing process.

[0035] As an example, the input features can be cutting speed, feed rate, tool rotation speed, tool path, tool selection, machining path, workpiece coordinates, running time, surface roughness of the part, environmental temperature, humidity, lighting, dust reduction amount, etc. The process parameters and measurement parameters corresponding to the input features of each process step are not the same. The computer-aided process planning system (CAPP) provides a common data interface, and the distributed database stores the finally obtained input feature data set.

[0036] A2. Preprocess the input feature data set to obtain a preprocessed input feature data set; the preprocessing includes outlier processing, missing value filling, grouping by process step, and feature dimensionality reduction; in the preprocessed input feature data set, there are several input feature groups corresponding to each process step respectively.

[0037] In this embodiment, as Figure 2 shown, step A2 specifically includes the following steps:

[0038] A21. For each input feature in the input feature data set, compare the input feature with the input feature threshold, screen and eliminate the outliers in each input feature to obtain an input feature data set after outlier processing.

[0039] A22. For each input feature in the input feature data set after outlier processing, group them according to the process step when each input feature is recorded to obtain a grouped input feature data set; the grouped input feature data set includes several input feature groups corresponding to each process step.

[0040] A23. For any process step, according to the values of each input feature in the input feature group corresponding to the process step, fill the missing values in the input feature group to obtain an input feature data set after missing value filling.

[0041] Specifically, by reading the processing time sequence information of each product, the actual process steps passed by each product can be obtained. Thus, the situation where the input feature is a missing value can be divided into two groups: if the input feature contained in the actual process steps passed by the product is a missing value, it is a valid missing value, otherwise it is an invalid missing value. Subsequently, the following judgment is made on all the input features contained in each process step: if the number of products containing valid feature missing values is less than 30% of the total number of products actually passing through this process step, for the missing values in all products containing feature missing values, use the mean value of this feature for filling; if the number of products containing valid feature missing values is not less than 30% of the total number of products actually passing through this process step, for the missing values in all products containing feature missing values, construct machine learning regression analysis or classification models such as ridge regression, decision tree, support vector machine, etc., use the remaining features of the process step where the missing value is located as the input, and fill after obtaining the predicted value through repeated training and optimization of the model.

[0042] A24. Dimension reduction is performed on each input feature in the input feature dataset after filling in the missing values through an autoencoder, obtaining the preprocessed input feature dataset. Dimension reduction is performed on each high-dimensional input feature to obtain its non-linear mapping relationship from high dimension to low dimension, significantly improving the model performance, reducing the computational complexity and overfitting risk while retaining the main features of the data.

[0043] In this embodiment, in step A2, outlier processing, missing value filling, grouping by working steps, and feature dimension reduction are performed on the data in the input feature dataset. In some other embodiments, other preprocessing operations may also be included.

[0044] A3. According to the actual processing procedures of the product to be tested, a matching graph, an adjacency matrix, and a mask graph of the product to be tested are constructed; the matching graph includes several real nodes corresponding to each working step and a virtual node; the adjacency matrix is used to represent the connection relationship and transmission relationship between the real nodes in the matching graph; the mask graph is used to represent the working steps actually passed and the working steps not passed by the actual processing procedures of the product to be tested. In this embodiment, step A3 includes the following steps:

[0045] A31. Each working step in the production and processing process is regarded as a graph node, obtaining the initial matching graph of the product to be tested; the number of graph nodes in the initial matching graphs of all products is exactly the same. Each working step is regarded as a graph node, and the transmission between different working steps is regarded as the edge of the graph, creating a matching graph G i =(V i ,E i )(where i is the product number), V i =[v 1 ,v 2 ,…,v s ,v virtual (where s is the total number of working steps), enabling the subsequent graph network to identify and integrate the product information after graph structuring. The last node is a newly added virtual node for aggregating the information of the remaining nodes, and the numbers of the remaining nodes correspond one by one to the platform numbers. In this way, a mapping relationship from the input feature data of different working steps to the corresponding nodes of the graph structure can be established, facilitating subsequent parallel processing. The feature dimensions and orders of the graph nodes corresponding to each working step are the same, and the total number of graph nodes is the same, enabling subsequent parallel processing.

[0046] A32. Generate the adjacency matrix of the product to be tested according to the actual processing procedure of the product to be tested, and improve the connection relationship and transfer relationship of each graph node in the initial matching graph of the product to be tested based on the adjacency matrix to obtain the matching graph of the product to be tested. The adjacency matrix represents the connection relationship between the process steps of the product. If the product is transferred to process step y after processing at process step x, the matrix element value at coordinate (x, y) in the adjacency matrix is 1, otherwise it is 0. This enables the efficient expression of the graph structure transfer relationship in subsequent matrix calculations, where there are paths pointing to virtual nodes for all process stations passed by the product.

[0047] A33. Generate the mask graph of the product to be tested according to the actual processing procedure of the product to be tested; the mask graph is used to mask the input features in the preprocessed input feature dataset. Create a corresponding mask graph for each product, where the process stations actually passed by the product are 1 and those not passed are 0. Through the mask graph, all the input feature data corresponding to the process stations not passed by the product and the virtual nodes are set to 0, so that they will not affect the subsequent calculation and parameter optimization of the neural network, and finally complete the construction of the matching graph corresponding to the complex processing procedure of the product. Facing the pipeline structure with a small number of nodes but complex and variable connection methods, this graph-structured data can significantly improve the subsequent calculation parallelization degree and algorithm operation efficiency while keeping the occupied calculation space controllable.

[0048] A4. Input the preprocessed input feature dataset, the matching graph of the product to be tested, the adjacency matrix, and the mask graph into the trained product quality prediction model to obtain the product quality prediction result of the product to be tested; the product quality prediction model is a neural network model based on a parallel collaborative graph network structure. The parallel collaborative graph network structure uses a bidirectional graph attention network to aggregate the cross-mixed information between each process step to obtain the process step aggregation feature, and performs regression according to the process step aggregation feature through a multi-layer feedforward neural network to obtain the product quality prediction result of the product to be tested.

[0049] Specifically, in this embodiment, the product quality prediction model includes several process step feature encoding modules corresponding to each process step respectively, a bidirectional graph attention network, and a final processing quality prediction module; the bidirectional graph attention network includes a forward graph attention network and a reverse graph attention network.

[0050] Furthermore, as Figure 3 shown, step A4 specifically includes the following steps:

[0051] A41. Input the input feature groups of each process step in the preprocessed input feature dataset into the process step feature encoding module corresponding to the process step for encoding, and obtain the encoded and normalized quality parameter features of each process step. A corresponding quality feature encoding module is constructed for each process step. The quality feature encoding module consists of a quality prediction sub-model and a post-processing sub-model, and the quality prediction sub-model and the post-processing sub-model are connected in sequence. As an example, the present invention uses a multi-layer feedforward neural network to build the quality feature encoding module, enabling it to have the ability to extract implicit quality parameters and the ability to standardize dimensions at the same time. The input is the feature values of all products at this process step (m is the total number of products, d s is the initial feature length of this station), and the output is the encoded and normalized quality parameter feature of the process step (d emb is the encoded feature length. For the convenience of subsequent parallel computing, the encoded feature lengths of different stations are the same). The encoded and normalized quality parameter features of different stations will next be used as the input of the corresponding graph nodes of the bidirectional graph attention network, where the initial input of the virtual node is a zero vector

[0052] A42. According to the matching graph and adjacency matrix of the product to be tested, construct graph structures in the forward graph attention network and the reverse graph attention network respectively; the effective nodes in the graph structures of the forward graph attention network and the reverse graph attention network all point to the virtual node; the effective node is the real node corresponding to the process step actually passed by the actual processing process of the product to be tested; the virtual node is used to extract the implicit quality parameters of each effective node in the graph structure.

[0053] Each graph node except the virtual node in the above graph structure receives the encoded and normalized quality parameter feature of the corresponding process step. The virtual node is initially a zero vector. The effective nodes (the process steps actually passed by the product processing) in the forward graph attention network and the reverse graph attention network all point to the virtual node, and all products share the parameters of the bidirectional graph attention network.

[0054] The bidirectional graph attention network has a total of L layers. The input of the l-th layer is the updated node information of the l-1-th layer (where d h is the node feature length of the bidirectional graph attention network. When l = 1, H (l) =[e 1 ,e 2 ,…,e s ,e virtual ) and the adjacency matrix A of the matching graph of each product. Subsequently, the bidirectional graph attention network introduces the encoded structure information of the product matching graph, which is specifically expressed as follows:

[0055]

[0056] where are learnable parameters, LeakyReLU(·) is an activation function, and b∈N(a) represents that node b belongs to the set of nodes pointing to node a, that is, the set of row nodes corresponding to the column value of 1 where node a is located in the adjacency matrix. For the correlation value After softmax processing, α is obtained. To fully extract the association information between nodes, a multi-head attention mechanism is applied to optimize different feature parts of each process step site, thereby balancing the deviation that may be generated by the same attention mechanism;

[0057] At this time represents the forward multi-head graph attention network with the l-th layer adjacency matrix A. Then the forward propagation of the forward graph attention network can be expressed as:

[0058]

[0059] where (m is the number of products, s is the total number of process step sites, and d h is the node feature length of the bi-directional graph attention network) is the output of the (l - 1)-th layer of the bi-directional graph attention network, is the output of the l-th layer of the forward graph attention network. The input of the first layer is

[0060] Similarly, the forward propagation of the reverse graph attention network is:

[0061]

[0062] where A' is obtained by transposing the data of the adjacency matrix A except for the rows and columns where the virtual nodes are located. Taking node a as an example again, at this time N(a) represents the set of nodes pointed to by node a. Finally, by aggregating the forward, reverse, and current graph network node information, the following formula can be obtained:

[0063]

[0064] where are learnable parameters. At this time is the final output of the bi-directional graph attention network, At this time, the node feature information in the final output has undergone multiple aggregations of adjacent nodes, and can effectively extract the key information affecting the final quality of products in complex processing methods such as cross-process steps and parallel process steps;

[0065] The present invention uses matching graphs with similar structural formats for different products, aggregates the cross-linked and mixed information between sites through a graph attention network for different sites, and finally extracts the most effective implicit quality parameters from the virtual nodes in the matching graph, enabling parallel processing while well-targeting product quality prediction in complex manufacturing processes.

[0066] A43. Input the encoded and normalized quality parameter features of each process step into the corresponding real nodes in the graph structures of the forward graph attention network and the reverse graph attention network respectively. Use a bidirectional graph attention network to aggregate the cross-linked and mixed information between each process step to obtain a process step aggregation feature; the process step aggregation feature is the feature of the virtual node information of each sample that aggregates all effective node information through the attention mechanism. When aggregating the cross-linked and mixed information between each process step, use a masked graph to mask the encoded and normalized quality parameter features of the process steps corresponding to the invalid nodes; the invalid nodes are the real nodes corresponding to the process steps not passed by the actual processing procedure of the product to be tested.

[0067] A44. In the final processing quality prediction module, use a multi-layer feedforward neural network to perform regression on the process step aggregation feature to obtain the product quality prediction result of the product to be tested.

[0068] As an exemplary embodiment, in the bidirectional graph attention network, apply the multi-head attention mechanism to optimize different feature parts of each process step, and balance the deviation that may be generated by the same attention mechanism. In the final processing quality prediction module, use the sigmoid function or the linear activation function as the activation function of the last layer. Specifically, if the final problem is a classification problem, use the sigmoid function as the activation function of the last layer, and if the final problem is a regression problem, use the linear activation function as the activation function of the last layer.

[0069] In another embodiment, the cross-linked and mixed process step product quality prediction method based on the parallel collaborative graph network further includes the following steps:

[0070] B1. Obtain the input features of several products with known quality results during the production and processing process, and form several historical input feature data sets corresponding to each product.

[0071] B2. For any product with a known quality result, preprocess the historical input feature data set to obtain the preprocessed historical input feature data set of the product; the preprocessing includes outlier processing, missing value filling, grouping by process step, balanced random resampling, and feature dimensionality reduction; in the preprocessed historical input feature data set, there are several historical input feature groups corresponding to each process step respectively.

[0072] B3. For any product with a known quality result, construct the matching graph, adjacency matrix, and masked graph of the product according to the actual processing procedure of the product.

[0073] B4. Construct a product quality prediction model, and use the preprocessed historical input feature datasets, matching graphs, adjacency matrices, and mask graphs of different products as the inputs of the product quality prediction model respectively, and use the known quality results of the corresponding products as labels to train the product quality prediction model to obtain a trained product quality prediction model.

[0074] In this embodiment, step B2 specifically includes the following steps:

[0075] B21. For each input feature in the historical input feature dataset, compare the input feature with the input feature threshold, screen and remove the outliers in each input feature to obtain a historical input feature dataset after outlier processing. In the process of obtaining input feature data, some abnormal data will be generated due to instrument failures or human errors. These abnormal data should not be input into the subsequent model for training and prediction. Therefore, a threshold benchmark needs to be set, compare the input feature data of each process step of the product with the threshold benchmark value, and judge whether there are outliers or abnormal points in all input features. If there are outliers or abnormal points, they need to be deleted.

[0076] B22. For each input feature in the historical input feature dataset after outlier processing, group them according to the process step at which each input feature is recorded to obtain a grouped historical input feature dataset; the grouped historical input feature dataset includes several input feature groups corresponding to each process step.

[0077] B23. For any process step, fill the missing values in the input feature group according to the values of each input feature in the input feature group corresponding to the process step to obtain a historical input feature dataset after filling the missing values.

[0078] B24. Use the multi-label random resampling algorithm for process steps to perform balanced random resampling on the known quality results of the historical input feature dataset after filling the missing values to obtain a historical input feature dataset after label resampling. In the process of product production and manufacturing, the number of qualified products is often much larger than the number of unqualified products, which makes the dataset class imbalanced and affects the final prediction effect. Therefore, the multi-label sample balancing random resampling algorithm for process steps or the corresponding improved optimization algorithm is used to balance the positive and negative example samples of the dataset.

[0079] B25. Dimension reduction is performed on each input feature in the historical input feature dataset after label resampling through an autoencoder to obtain the preprocessed historical input feature dataset. The information redundancy of the process parameters and measurement parameters obtained in different process steps is often relatively high. If directly used, it will lead to a long model training time and high complexity, and it is also extremely easy to cause model overfitting. Therefore, an autoencoder needs to be constructed to perform dimension reduction on the historical input feature data of the product, mapping it from high-dimensional non-linear to low-dimensional.

[0080] During the process of constructing the model, the selection of many model hyperparameters and overfitting judgment are involved. The selection of appropriate hyperparameters and model complexity will be of great help to the model performance. Therefore, a validation set is divided from the dataset, and the cross-validation method is used to evaluate the performance of the current model on unknown data under different hyperparameter combinations. Specifically, after step B25, the cross-linked hybrid process step product quality prediction method based on the parallel collaborative graph network further includes the following steps:

[0081] B26. The preprocessed historical input feature dataset is divided into a historical input feature training dataset and a historical input feature validation dataset; the historical input feature training dataset is used to train the product quality prediction model, and the historical input feature validation dataset is used to evaluate the performance of the product quality prediction model on unknown data under a certain hyperparameter or complexity.

[0082] Specifically, in this embodiment, the process of training the product quality prediction model is as follows: The cross-entropy loss function or the mean squared error loss function is used as the loss function. Specifically, if the final problem is a classification problem, the cross-entropy loss function is used as the loss function; if the final problem is a regression problem, the mean squared error loss function is used as the loss function. According to the data in the historical input feature training dataset, the backpropagation algorithm is used between the multi-layer feedforward neural networks, so as to realize the update of the model parameters of the process step feature encoding module, the bidirectional graph attention network and the final processing quality prediction module in each process step; both the process step feature encoding module and the final processing quality prediction module are structured by using multi-layer feedforward neural networks.

[0083] Next, taking the actual dataset of the Bosch assembly line as an example, the specific implementation process of the above method of this application will be described. The Bosch assembly line dataset is a public dataset updated by Robert Bosch GmbH in Germany in the 2016 Kaggle competition. This dataset has a total of 4 production lines and 52 process steps. All process steps together contain 968 process parameters and measurement parameters, that is, the initial input feature data. In this dataset, the process steps passed by each product are complexly associated and have different temporal relationships. There is a parallel processing situation in this dataset, that is, a product may pass through multiple process steps at a certain moment. The initial input feature data of each product consists of continuous numerical features, discrete categorical features, and process step time point features. Part of the dataset content is shown in Table 1:

[0084] Table 1 Bosch assembly line dataset

[0085] Id L0_S0_F0 L0_S0_F2 … L0_S1_F28 L0_S2_F32 L0_S2_F36 … Response 879603 -0.023 0.011 … -0.026 -0.013 0.536 … 0 1042754 0.016 0.033 … 0.052 NaN NaN … 0 1507689 0.121 0.138 … -0.11 NaN NaN … 0

[0086] In this dataset, Id represents the part number, and "Lx_Sx_Fx" represents the part feature number. For example, "L0_S17_F431" represents the initial input feature data with the number 431, which is located at the 17th station of the 0th production line; Response represents the output quality feature of this part. If the value of Response is 1, it means that the part quality inspection is unqualified. If the value of Response is 0, it means that the part quality inspection is qualified.

[0087] After obtaining the above dataset, the following preprocessing steps of outlier processing, feature classification, missing value processing, balanced random resampling, feature dimensionality reduction, and dataset partitioning need to be carried out in sequence.

[0088] Calculate the upper bound (UB) = 75% quantile + (75% quantile - 25% quantile) * 1.5 and the lower bound (LB) = 25% quantile - (75% quantile - 25% quantile) * 1.5 for each feature. If a certain feature of a dataset is outside the boundary value, it is judged as an outlier and deleted. According to the process step number when the initial input feature data of the product is recorded, the input features are grouped for subsequent processing. In the Bosch dataset, the number after the product ID feature number "S" is the process step where this initial input feature is located.

[0089] After statistics, the number of products with valid feature missing values in each process step of this dataset is less than 30% of the total number of products actually passing through this process step. Therefore, for all missing input features, it is sufficient to fill them with the mean value of this feature. In the above dataset, the proportion of positive example samples with Response value of 0 and negative example samples with Response value of 1 is unbalanced. Therefore, the positive example samples and negative example samples of the dataset are respectively subjected to balanced random resampling using the multi-label random resampling algorithm for process steps or the corresponding improved algorithm. An autoencoder is constructed to reduce the dimension of the initial input feature data of Bosch assembly line products, mapping from high-dimensional non-linearity to low-dimensionality.

[0090] Finally, to facilitate the use of the cross-validation method to evaluate the performance of the current model on unknown data under a certain hyperparameter or complexity, this dataset is divided into a training set and a validation set in a ratio of 7:3. Subsequently, the cross-entropy loss function is used as the loss function, and based on the data in the training set, the backpropagation algorithm is used between the layers of the feedforward neural network to update the model parameters of the quality feature encoding module, the bi-directional graph attention network, and the final processing prediction module respectively.

[0091] In the Bosch assembly line dataset, there is a table of actual processing procedures for products, which records the process steps experienced by each product at different time sequences. If there are multiple process step information in one time sequence, it means that the product is being processed by these workstations simultaneously at this moment. For example, the actual processing procedure table of the product with ID 1606065 is shown in Table 2:

[0092] Table 2 Actual Processing Procedure Table for Products

[0093]

[0094] Regarding each process step as a graph node and the transfer between different process steps as the edge of the graph structure, a matching graph G i =(V i , E i )(where i is the product number). Since this dataset contains a total of 52 process steps, so V i =[v 1 , v 2 , …, v 52 , v virtual , enabling the subsequent graph network to recognize and integrate the product information after graph structuring. The last node is a virtual node, and the numbers of the remaining nodes correspond one-to-one with the platform numbers. In this way, a mapping relationship from different process step feature values and process parameters to the graph structure can be established, facilitating subsequent parallel processing by the graph network.

[0095] According to the actual processing procedures of each product, construct the corresponding adjacency matrix of the matching graph nodes, where there are paths pointing to virtual nodes for all the workstep stations passed by the product. Subsequently, create a corresponding mask graph for each product, with the stations actually passed by the product being 1 and the stations not passed being 0. Taking the product with id 1606065 as an example, from the initial time sequence, it can be determined that the values of the 0th, 1st, and 3rd rows and 5th column and the last column of the adjacency matrix are 1. In the one-dimensional mask matrix initialized to all 0s, the values at positions 0, 1, and 3 are set to 1. Similarly, for the connection relationships of the remaining time sequence work positions.

[0096] All the eigenvalue corresponding to the workstep stations not passed by the product and the virtual nodes are set to 0 through the mask graph, so that they will not affect the subsequent calculation and parameter optimization of the neural network. Finally, complete the construction of the matching graph corresponding to the complex processing procedures of the product. Facing the work position products with a small number of nodes but complex connection methods, this kind of graph-structured data can greatly improve the subsequent calculation parallelization degree and algorithm operation efficiency while keeping the occupied computing space controllable.

[0097] For this data set, the present invention uses the specific structure of the quality feature encoding module as shown in Figure 4 Since this data set contains a total of 52 worksteps, there are a total of 52 corresponding models in the quality feature encoding module. Similarly, each product also has 52 corresponding feature data groups. Due to the existence of the mask matrix, the input feature data groups corresponding to the stations not passed by the product are all 0 vectors. Parallelly input the corresponding feature data groups of the product according to the worksteps into the corresponding quality feature encoding module. The multi-layer feedforward neural network in the quality feature encoding module uses the vector (m is the total number of products, d s is the initial feature length of this station) as the input vector, where s = 1,..., 52. After being mapped by the multi-layer feedforward neural network, obtain (d emb is the encoded feature length).

[0098] The bidirectional graph attention network established based on the processed Bosch assembly line data set is composed of a forward graph attention network and a reverse graph attention network. Set the number of graph network layers to 3 layers. There are a total of 52 station graph nodes and 1 virtual node. Each graph node corresponds to the encoded feature of the corresponding platform. The virtual node is initially a vector of all 0s. The effective nodes (stations actually passed by the product processing) in the forward and reverse networks both point to the virtual node. All products share the parameters of the bidirectional graph attention network. In this embodiment, the bidirectional graph attention network uses the structure as shown in Figure 5 After the multi-layer feedforward neural network is mapped, the obtained e sInput the corresponding graph node according to the station number, perform multi-head attention mechanism calculation according to the matching graph adjacency matrix A of each product, and finally aggregate to obtain the final output of the bidirectional graph attention network. in

[0099] Virtual node information in the final output All valid node information, i.e., cross-linked parallel information of each product process step, is aggregated through the attention mechanism. Figure 6 The structure shown is used as the input of the final processing prediction module to obtain the category feature y of whether the final predicted product quality is qualified. In this embodiment, the prediction problem here is a classification problem, so the sigmoid function is selected as the last layer activation function of the final processing model, and the cross entropy loss function is used as the loss function during model training.

[0100] For classification problems, it is necessary to use the cross-validation method before prediction to obtain the best classification threshold by comparing and analyzing the accuracy of classification or using other evaluation indicators. Here, the classification threshold is selected as 0.50, that is, when the final output of the model is greater than or equal to 0.50, the model output is 1, indicating that the quality inspection result of the product is unqualified; conversely, the final output of the model is 0, indicating that the quality inspection result of the product is qualified. For the products shown in Table 1, the input feature data to be tested in the validation set is input into the quality prediction model for prediction, so as to obtain the quality features corresponding to the input feature data to be tested. After testing, the model output is 0.7982, which is greater than 0.50, so the predicted Response value of the product is 1, and the quality inspection result of the product is unqualified, which is consistent with the Response value of the original data.

[0101] The above solution of this application is different from most traditional quality prediction methods that attempt to convert complex working conditions into a production line with a single sequential and independent working condition. The parallel collaborative graph network fully considers the influence between each station process in the manufacturing and processing process, and further extracts the cross-linked and mixed relationships between the working steps through the advantages of the graph structure. Finally, the influence of different stations on the final quality of the product is fused through virtual nodes and the attention mechanism. Therefore, the parallel collaborative graph network is more in line with the actual situation in the production and processing field and can achieve higher prediction accuracy. Another key feature is that all products are stored with the same graph feature information, where the graph nodes, arrangement order, and node feature lengths are the same, while the adjacency matrix and mask matrix information record different graph feature transmission methods according to the actual station paths passed by the products. This enables the graph network module to receive the product graph feature information and the matching adjacency matrix in parallel, and then adopt different update methods for the graph nodes of different products according to the connection information of each station in the adjacency matrix. While sharing the internal parameters of the network unit, it reduces the model complexity and improves the parallelization degree of the algorithm, enabling this application to quickly and accurately predict the quality of products with mixed and variable production lines and cross-linked and mixed working steps.

[0102] Based on the same inventive concept, the embodiment of this application also provides a system for implementing the cross-linked and mixed working step product quality prediction method based on the parallel collaborative graph network involved above. The solution provided by this system to solve the problem is similar to the solution described in the above method. In an exemplary embodiment, as Figure 7 shown, a cross-linked and mixed working step product quality prediction system based on a parallel collaborative graph network is provided, including the following modules: a data acquisition module, a data preprocessing module, a matching graph generation module, a parameter optimization module, a quality prediction model construction module, a quality feature prediction module, and a distributed database.

[0103] The data acquisition module is connected to the CAPP system to obtain information such as input features or historical input features of the product during the production and processing process; and after performing preprocessing operations in the data preprocessing module, it is stored in the distributed database; subsequently, the matching graph generation module is used to generate a matching graph, an adjacency matrix, and a mask graph according to the actual processing procedures of the product, and store them in the distributed database; based on the quality prediction model construction module, a quality prediction model including a working step feature encoding module, a bidirectional graph attention network, and a final processing quality prediction module is constructed, and the parameter optimization module is called to optimize the parameters of the quality prediction model using the training set, and the obtained network parameters are stored in the distributed database; then, when actual quality feature prediction is required, the quality feature prediction module obtains the input features of each working step of the product to be tested and performs preprocessing. After generating the matching graph, the adjacency matrix, and the mask graph, the network parameters stored in the distributed database are used to call the quality prediction model to predict the quality of the product to be tested.

[0104] In the above solution provided by the present application, each part has been modularized. When facing a new production and processing task, only a matching graph, an adjacency matrix, and a mask matrix corresponding to the new product need to be created. The parameters inside each module of the parallel collaborative graph network do not need to be changed and will not affect a single module. Facing the complex correlation information between different product processing steps, only by establishing the mapping relationship from the feature values and process parameters of different processing steps to the corresponding nodes of the graph structure according to the processing procedure information of product manufacturing, the parallel collaborative graph network can fuse the correlation information between features for quality prediction, and will not be restricted by the process transfer method and the product production and processing form. The model proposed in the present application can flexibly handle the situation of cross-linked and mixed processing steps existing in the production and processing process.

[0105] Of course, Figure 7 the architecture shown is only exemplary. When implementing different functions, one or at least two components in the Figure 7 shown system can be omitted according to actual needs.

[0106] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 8 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the cross-linked and mixed processing step product quality prediction method based on the parallel collaborative graph network provided in the above embodiment can be implemented.

[0107] Those skilled in the art can understand that Figure 8 the structure shown in

[0108] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0109] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0110] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0111] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0112] Those of ordinary skill in the art can understand that all or part of the processes of implementing the above method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0113] In each of the embodiments provided in this application, the database involved may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., and is not limited thereto. In each of the embodiments provided in this application, the processor may be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., and is not limited thereto.

[0114] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0115] In this article, specific examples are used to elaborate on the principles and implementation manners of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A method for predicting the quality of cross-linked hybrid process products based on a parallel collaborative graph network, characterized in that: include: Acquire an input feature data set of the product to be tested during the production and processing process; the input feature data set includes input features of each process step in the production and processing process of the product to be tested; the input feature is any process parameter and / or any measurement parameter in the production and processing process of the product to be tested; Preprocessing the input feature data set to obtain a preprocessed input feature data set; the preprocessing includes outlier processing, vacancy filling, grouping by process step and feature dimension reduction; The preprocessed input feature data set includes a number of input feature groups corresponding to each process step; According to the actual processing procedures of the product to be tested, a matching graph, an adjacency matrix and a mask graph of the product to be tested are constructed; the matching graph includes a number of real nodes corresponding to each process step and a virtual node; the adjacency matrix is ​​used to characterize the connection relationship and the transfer relationship between the real nodes in the matching graph; the mask graph is used to characterize the process steps actually passed and the process steps not passed in the actual processing procedure of the product to be tested; The preprocessed input feature data set, the matching graph, the adjacency matrix and the mask graph of the product to be tested are input into the trained product quality prediction model to obtain the product quality prediction result of the product to be tested; the product quality prediction model is a neural network model based on a parallel collaborative graph network structure, and the parallel collaborative graph network structure adopts a bidirectional graph attention network to aggregate the cross-linked and mixed information between each process step to obtain the process step aggregation feature, and regresses according to the process step aggregation feature through a multi-layer feedforward neural network to obtain the product quality prediction result of the product to be tested.

2. The method for predicting the quality of cross-linked mixed process products based on a parallel collaborative graph network according to claim 1 is characterized in that: The product quality prediction model includes a plurality of work step feature encoding modules corresponding to each work step, a bidirectional graph attention network and a final processing quality prediction module; the bidirectional graph attention network includes a forward graph attention network and a reverse graph attention network; The preprocessed input feature data set, the matching graph, the adjacency matrix and the mask graph of the product to be tested are input into the trained product quality prediction model to obtain the product quality prediction result of the product to be tested, which specifically includes: Inputting the input feature groups of each process step in the preprocessed input feature data set into the process step feature encoding module of the corresponding process step for encoding, thereby obtaining the encoded normalized quality parameter features of each process step; According to the matching graph and adjacency matrix of the product to be tested, a graph structure is constructed in the forward graph attention network and the reverse graph attention network respectively; the valid nodes in the graph structure of the forward graph attention network and the reverse graph attention network all point to virtual nodes; the valid nodes are real nodes corresponding to the actual steps of the actual processing process of the product to be tested; the virtual nodes are used to extract implicit quality parameters of each valid node in the graph structure; The encoded normalized quality parameter features of each process step are respectively input into the corresponding real nodes in the graph structures of the forward graph attention network and the reverse graph attention network, and the cross-linked and mixed information between the process steps is aggregated using the bidirectional graph attention network to obtain the process step aggregation features; when aggregating the cross-linked and mixed information between the process steps, the encoded normalized quality parameter features of the process steps corresponding to the invalid nodes are shielded using the mask graph; the invalid nodes are the real nodes corresponding to the process steps that have not been passed in the actual processing procedures of the product to be tested; the process step aggregation features are the features of the virtual node information of each sample that aggregates all the valid node information through the attention mechanism; In the final processing quality prediction module, a multi-layer feedforward neural network is used to regress the process step aggregation features to obtain the product quality prediction result of the product to be tested.

3. The method for predicting the quality of cross-linked mixed process products based on a parallel collaborative graph network according to claim 1 is characterized in that: Preprocessing the input feature data set to obtain a preprocessed input feature data set specifically includes: For each input feature in the input feature data set, compare the input feature with an input feature threshold, screen and remove outliers in each input feature, and obtain an input feature data set after outlier processing; For each input feature in the input feature data set after outlier processing, grouping is performed according to the process step at which each input feature is recorded, so as to obtain a grouped input feature data set; the grouped input feature data set includes a plurality of input feature groups corresponding to each process step; For any process step, according to the value of each input feature in the input feature group corresponding to the process step, the missing values ​​in the input feature group are filled to obtain an input feature data set after the missing values ​​are filled; The autoencoder is used to reduce the dimension of each input feature in the input feature dataset after the missing values ​​are filled, so as to obtain the preprocessed input feature dataset.

4. The method for predicting the quality of cross-linked mixed process products based on a parallel collaborative graph network according to claim 1 is characterized in that: The method for predicting the quality of cross-linked mixed process step products based on the parallel collaborative graph network also includes: Obtain input features of several products with known quality results during the production process to form several historical input feature data sets corresponding to each product; For any product with known quality results, the historical input feature data set is preprocessed to obtain a preprocessed historical input feature data set of the product; the preprocessing includes outlier processing, vacancy filling, grouping by process step, balanced random resampling and feature dimension reduction; the preprocessed historical input feature data set includes a plurality of historical input feature groups corresponding to each process step; For any product with known quality results, a matching graph, an adjacency matrix and a mask graph of the product are constructed according to the actual processing procedures of the product; A product quality prediction model is constructed, and the preprocessed historical input feature data set, matching graph, adjacency matrix and mask graph of different products are used as inputs of the product quality prediction model respectively. The known quality results of the corresponding products are used as labels to train the product quality prediction model to obtain a trained product quality prediction model.

5. The method for predicting the quality of cross-linked mixed process products based on a parallel collaborative graph network according to claim 4 is characterized in that: Preprocessing the historical input feature data set to obtain the preprocessed historical input feature data set of the product specifically includes: For each input feature in the historical input feature data set, compare the input feature with an input feature threshold, screen and remove outliers in each input feature, and obtain a historical input feature data set after outlier processing; For each input feature in the historical input feature data set after outlier processing, grouping is performed according to the process step at which each input feature is recorded, so as to obtain a grouped historical input feature data set; the grouped historical input feature data set includes a plurality of input feature groups corresponding to each process step; For any process step, according to the value of each input feature in the input feature group corresponding to the process step, the vacant values ​​in the input feature group are filled to obtain a historical input feature data set after the vacant values ​​are filled; The multi-label random resampling algorithm is used to perform balanced random resampling on the known quality results of the historical input feature data set after the vacancy value is filled, and the historical input feature data set after label resampling is obtained; The autoencoder is used to reduce the dimension of each input feature in the historical input feature dataset after label resampling to obtain the preprocessed historical input feature dataset.

6. The method for predicting the quality of cross-linked mixed process products based on a parallel collaborative graph network according to claim 5 is characterized in that: After reducing the dimension of each input feature in the historical input feature data set after label resampling by the autoencoder to obtain the preprocessed historical input feature data set, the cross-linking mixed process step product quality prediction method based on the parallel collaborative graph network also includes: The preprocessed historical input feature data set is divided into a historical input feature training data set and a historical input feature verification data set; the historical input feature training data set is used to train the product quality prediction model, and the historical input feature verification data set is used to evaluate the performance of the product quality prediction model for unknown data under any hyperparameters or complexity.

7. The method for predicting the quality of cross-linked hybrid process products based on a parallel collaborative graph network according to claim 6 is characterized in that: The product quality prediction model is trained, specifically: using a cross entropy loss function or a mean square error loss function as a loss function, using a back propagation algorithm between multi-layer feedforward neural networks according to the data in the historical input feature training data set, so as to realize the update of model parameters of the work step feature encoding module, the bidirectional graph attention network and the final processing quality prediction module of each work step; the work step feature encoding module and the final processing quality prediction module are structures constructed using a multi-layer feedforward neural network.

8. The method for predicting the quality of cross-linked mixed process products based on a parallel collaborative graph network according to claim 1 is characterized in that: According to the actual processing procedures of the product to be tested, a matching graph, an adjacency matrix and a mask graph of the product to be tested are constructed, specifically including: Each process step in the production process is regarded as a graph node to obtain the initial matching graph of the product to be tested. The number of graph nodes in the initial matching graph of all products is exactly the same. Generate an adjacency matrix of the product to be tested according to the actual processing procedures of the product to be tested, and improve the connection relationship and transfer relationship of each graph node in the initial matching graph of the product to be tested based on the adjacency matrix to obtain the matching graph of the product to be tested; A mask image of the product to be tested is generated according to the actual processing steps of the product to be tested; the mask image is used to shield the input features in the preprocessed input feature data set.

9. The method for predicting the quality of cross-linked mixed process products based on a parallel collaborative graph network according to claim 1, characterized in that: In the bidirectional graph attention network, a multi-head attention mechanism is applied to optimize the different feature parts of each step and balance the deviations that may be caused by the same attention mechanism.

10. The method for predicting the quality of cross-linked mixed process products based on a parallel collaborative graph network according to claim 1, characterized in that: In the final processing quality prediction module, a sigmoid function or a linear activation function is used as the activation function of the last layer.

Images