Large-scale customized quality prediction method based on heterogeneous causal graph attention network

Abstract

The invention discloses a large-scale customized quality prediction method based on a heterogeneous causal graph attention network. The quality prediction method is designed for the characteristics of modular production, customized assembly and the like of a large-scale customized production mode and the problems of multiple scales and data mixing in data. The method comprises the following steps: firstly, providing a PCAM causal discovery algorithm, constructing a causal skeleton by utilizing mechanism and data combination, and obtaining a reliable causal graph structure based on an additive noise model and by introducing a scoring mechanism; then designing a heterogeneous causal graph attention network HCGAT for a multi-scale problem based on a causal graph structure, mapping node features of different dimensions to the same dimension by using different weight parameter matrixes, and training through the graph attention network; and finally, aiming at the characteristics of large-scale customized production of customized products, a quality prediction framework suitable for large-scale customized production is provided, customized codes are designed through customized assembly information, and a filter and a judger are constructed based on the customized codes to classify and process data. And performing quality prediction on different types of product data by using HCGAT as a sub-model.

Description

Technical Field

[0001] The invention relates to a mass customization quality prediction method based on a heterogeneous causal graph attention network, and belongs to the field of soft measurement. Background Art

[0002] With the development of productivity, market competition is becoming more and more fierce. In order to meet the growing personalized needs of consumers, more and more companies are beginning to adopt or consider adopting the mass customization production model (MC). As a comprehensive production model that integrates application technology, management, standards, etc., mass customization enables companies to provide high-quality and efficient diversified and customized products to the market at low cost of mass production. The key idea of ​​MC is to standardize similar product family parts and structures by using methods such as modularization to reduce the internal diversity of products, and then use strategies such as customized assembly on the basis of common parts to increase the external diversity of products.

[0003] Mass customization (MC) has been applied to many production fields such as automobiles, clothing, furniture, and electronic products because of its advantages of taking into account both consumer demand and corporate benefits. Therefore, it has attracted the attention of many scholars. However, since mass customization is a comprehensive production model that includes production, assembly, management, and transportation, most of the previous research aims to ensure product diversity and cost control.

[0004] Quality control, as one of the important issues in the manufacturing industry, has been less studied in mass customization production. At present, the quality control method based on prediction is more accurate and reliable, and this method is heavily dependent on the accuracy of quality prediction. Therefore, this paper mainly focuses on the quality prediction problem in mass customization production.

[0005] The current quality prediction methods can generally be divided into methods based on mechanism models, methods based on expert knowledge, and methods based on data-driven. Among them, the performance of the mechanism model depends on whether the mathematical model of the reaction process mechanism can be accurately established. For complex and large-scale modern industries, mechanism modeling has encountered great obstacles. In addition, due to the difficulty in accumulating prior knowledge, the expert knowledge model also lacks universality and is mostly used as an auxiliary. Data-driven quality prediction methods are widely used in industrial processes due to their advantages such as low difficulty of use, wide range of use, and high prediction accuracy. Multivariate statistical methods such as PCA, machine learning methods such as SVM, and deep learning methods such as LSTM have all been used in quality prediction. Deep learning methods have better potential due to their better prediction effects and the advantage of being able to cope with high-dimensional data in complex industrial systems. However, the existing deep learning methods will face the following problems in the application of quality prediction in large-scale customized production.

[0006] First, the relationship between parts has an important impact on the prediction of final product quality. Graph neural networks are generally used to utilize the correlation between variables in the system, but the general graph structure considers the correlation between variables. Due to the functional dependence or collaboration between parts in mass customization production, there are causal relationships between parts that need to be mined. If the causal relationship between variables is not explored and utilized, the final prediction effect will be affected.

[0007] At the same time, the multi-scale problem is also a difficult problem faced in the quality prediction of large-scale customized products. In the production of customized products, the types of module accessories in each product are different, and the number of accessories is not exactly the same, which leads to the inconsistency of the variable dimensions of the collected data. The traditional prediction model must ensure the consistency of the variable dimensions due to mapping requirements.

[0008] In summary, to address the above issues, it is necessary to develop a quality prediction model suitable for mass customization. Summary of the invention

[0009] In view of the shortcomings of the existing technology and the characteristics of large-scale customization, the present invention designs a quality prediction method based on the heterogeneous causal graph attention network. Compared with the commonly used prediction methods, the heterogeneous causal graph attention network can not only utilize the correlation between variables through the causal graph, but also utilize multi-scale information to ensure the final prediction effect.

[0010] This paper first proposes a PCAM causal discovery algorithm, which uses the mechanism and data to jointly construct a causal skeleton, and introduces a scoring mechanism based on the additive noise model to obtain a reliable causal graph structure; then based on the causal graph structure, a heterogeneous causal graph attention network HCGAT is designed for multi-scale problems, and different weight parameter matrices are used to map node features of different dimensions to the same dimension, and then trained through the graph attention network. Finally, in view of the characteristics of mass customization of customized products, a quality prediction framework suitable for mass customization is proposed, and customized coding is designed through customized assembly information, and based on this, filters and judges are constructed to classify and process data, and then HCGAT is used as a sub-model to predict the quality of different types of product data.

[0011] The technical solution of the present invention is as follows:

[0012] First of all, this patent defines the manufacturing process and collected data of the mass customized production model based on the characteristics of the mass customized production model. The manufacturing process can be roughly divided into three steps: modular design, customized assembly, and product production and testing. In this process, the corresponding collected data is defined, including accessory quality data, customized assembly data, and product quality data.

[0013] Among them, in view of the characteristics of customized production and flexible co-production in mass customization, that is, different types and quantities of accessories are assembled into different types of products, and multiple products may be produced on the same production line, which leads to the possibility that the final collected data may contain multiple product information. Therefore, based on the customized assembly information of products in mass customization, a customized assembly data is designed to represent the types and quantities of accessories used to assemble products, encoded as an n-dimensional vector α = (a1, a2, ...a n ), where n represents the total number of accessory categories used in all products. If a=0, it means that no accessories of this type are used in the assembly of the product, and a=N means that N accessories of this type are used in the assembly of the product.

[0014] The quality prediction of the present invention can be divided into two parts: offline training and online testing. In offline training, the filter is first constructed by the designed customized assembly to determine the product category in the data. A customized coded data header is added to each piece of data, and the consistency of the coding is compared to determine whether the data belongs to the same product. The data of the same product is classified together, and the customized coding of the product is recorded for constructing a judge in the subsequent test data.

[0015] Then, causal discovery is performed separately in the data of different product types to construct the corresponding causal graph structure. Because the selection of different accessories may affect the causal relationship between accessories, building a separate causal graph can ensure better interpretability and accuracy.

[0016] A PCAM causal discovery method is designed. The process can be divided into three parts: the first part is the construction of the causal skeleton, the second part is the determination of the causal direction, and the third part is the overall structure optimization. Through these three steps, the final causal graph is determined for subsequent training and prediction.

[0017] First, we build a causal skeleton by combining prior knowledge and data. We first determine the accurate prior knowledge, including whether there is a causal relationship or even the causal direction information. Then, we use the data to perform HSIC independence tests on the remaining related variables to determine the causal relationship and build a skeleton diagram. This can improve the accuracy of the causal skeleton. When there is no prior knowledge, we can also build a skeleton diagram only through data.

[0018]

[0019] Where n is the number of samples, Tr is the trace of the matrix, and K X and K Y It is the kernel function calculation, where the Gaussian kernel function is used, J = I-1 / n, and is the n-order unit matrix.

[0020]

[0021] Then, on the determined causal skeleton, for each pair of variables with causal relationships, the additive noise model ANM method is used to determine the causal direction. In the production process, external factors such as workers' operating proficiency and assembly environment will interfere with the assembly process. In this case, the relationship between systems can be regarded as approximately consistent with the additive noise model, and the external noise is independent of the quality parameters of the accessories themselves. Therefore, the additive noise model can be used to construct forward and reverse functions, and its asymmetry can be used to determine the causal direction.

[0022] The judgment principle is that as an independent additive noise, it will not affect the cause variable, and since the result variable is obtained by the joint action of the cause variable and the noise variable, the result variable and the noise variable are not independent. Therefore, we construct a forward function f and a reverse function g to fit the relationship between the variable X and the variable Y, and judge the independence relationship between the noise NX and NY and the input variable.

[0023]

[0024]

[0025] If the function satisfies formula 1 but not formula 2, it means that variable X is the cause variable of variable Y.

[0026] The present invention performs forward fitting on the relationship between two variables through MLP, uses two hidden layers and a nonlinear activation function LeakyReLU for nonlinear fitting. As a deep learning algorithm, MLP can obtain better fitting effects for all nonlinear functions by controlling the Loss and the number of cycles.

[0027] Leaky ReLU=max(0,z)+α·min(0,r),0<α<<1

[0028] If NY and X are independent, that is, the direction is consistent with the data generation mechanism, then the forward hypothesis is correct, that is, variable X is the cause of variable Y. If not, then reject this hypothesis. Similarly, make assumptions and fits in the opposite direction, that is, X = g(Y) + NX, and use the same method to judge independence. If both directions are independent, then choose the side with a higher independence score. If both directions are not independent, then this method fails.

[0029] However, this method also has its defects, that is, ANM is generally used for one-to-one causal discovery. When there is a many-to-one causal relationship, its judgment effect will deteriorate. Therefore, the present invention introduces a scoring mechanism to further optimize the problem of inaccurate causal skeleton construction and inaccurate many-to-one causal effect.

[0030] Here, the Bayesian criterion is introduced as a scoring function. When there is ambiguity in independence judgment and ANM failure, that is, when the difference between the forward and reverse directions of ANM construction is small and both sides are independent and the causal direction cannot be determined, the final overall score is judged by adding or deleting new edges and changing the causal direction, thereby determining the final causal graph result.

[0031] BIC = k·ln(n)-2ln(L)

[0032] Based on the obtained causal graph, the heterogeneous causal graph attention network HCGAT is used to extract multi-scale information and use it for quality prediction. HCGAT sets the weight parameter matrix of different nodes, maps node features of different dimensions to the same dimension, and then performs quality prediction through the graph attention network.

[0033] First, different weight parameter matrices are used to map and embed features. The specific formula is as follows:

[0034] h ii =W i ·h i

[0035] Among them, h i are the characteristics of different nodes, W i It is a weight parameter matrix constructed for the feature dimension of the node and the feature dimension to be mapped. Through different W, all node features h are mapped to the same dimension. At this time, the isomorphic graph model can be used to process and train it.

[0036] At this point, the attention mechanism can be used to update the node. The specific formula is as follows:

[0037]

[0038] Where σ is the activation function, α ij It is the attention coefficient representation between nodes, W is the weight matrix in the attention mechanism, and W here is in the isomorphic graph, targeting the node features after being mapped to the same dimension as above.

[0039] The Adam algorithm is selected as the optimization algorithm. Adam is a first-order optimization algorithm that can replace the traditional stochastic gradient descent process. It can iteratively update the weights of the neural network based on the training data. The mean square error on all label nodes is selected as the loss function.

[0040]

[0041] Use each product data to train the corresponding sub-model

[0042] After the training is completed, the test phase begins. For the test data, the classification information in the judger is used to classify the data according to the customized coding, and the test data is divided into different product categories. The corresponding trained sub-model is selected according to the product category to make quality predictions.

[0043] The final output formula can be expressed as:

[0044] y i =HCGAT(X i ,A i )

[0045] The beneficial effects of the present invention are mainly reflected in:

[0046] The present invention analyzes the relevant characteristics and problems of the mass customization production model; in view of the characteristics of its data mixing, the filter and judge are constructed by designing customized coding to classify the data. The causal graph is jointly constructed by mechanism and data, and the overall structure of the causal graph is improved by the scoring mechanism, ensuring the accuracy and interpretability of the final causal graph. Prediction is performed using a heterogeneous causal graph attention network. The mass customization quality prediction method based on a heterogeneous causal graph attention network proposed in the present invention can solve the problems of multi-scale, data mixing, complex causal relationships, etc. in mass customization, and utilizes the correlation between variables and extracts information from multi-scale data to obtain better prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the manufacturing process for mass customization;

[0048] Figure 2 This is the quality prediction framework diagram;

[0049] Figure 3 It is the PCAM causal discovery method;

[0050] Figure 4 Updated procedures for HCGAT;

[0051] Figure 5 Training process for HCGAT;

[0052] Figure 6 Experimental results for quality prediction DETAILED DESCRIPTION

[0053] Reference Figures 1 to 6 The present invention is further described.

[0054] according to Figure 2 The flowchart shown in the figure is combined with actual data processing and model building to introduce each step of the method in detail.

[0055] First, this patent defines the manufacturing process and collected data of the mass customization production model based on the characteristics of the mass customization production model. The manufacturing process can be roughly divided into three steps: modular design, customized assembly, and product production and testing. In this process, the corresponding collected data is defined, including the quality data of accessories, customized assembly data, and product quality data. Figure 1 shown.

[0056] In order to reduce the internal diversity of products and increase the external diversity of products, mass customization production has put forward higher requirements for the accessory design and assembly design of products. First of all, the accessories of the products need to be redesigned. In order to adapt to more products, the accessory design tends to be more standardized to form universal accessories, and then enter the customized product production process. After the combination and assembly of different module accessories, a product series is formed. Then, according to the customer's choice, the production and testing links are carried out to obtain relevant product quality parameters to judge the quality of the produced products.

[0057] In the above process, three types of data can be obtained for data-driven quality prediction applications. In addition to the accessory quality data representing the input data and the product data representing the prediction target, a customized assembly data is defined to represent the customized assembly information.

[0058] It targets the characteristics of customized production and flexible co-production in mass customization, that is, different types and quantities of accessories are assembled into different types of products. At the same time, multiple products may be produced on the same production line, resulting in the possibility that the final collected data may contain multiple product information. Therefore, a customized encoding is designed based on the customized assembly information of products in mass customization, which represents the types and quantities of accessories used to assemble the product, and is encoded as an n-dimensional vector α = (a1, a2, ...a n ), where n represents the total number of accessory categories used in all products. If a=0, it means that no accessories of this type are used in the assembly of the product, and a=N means that N accessories of this type are used in the assembly of the product.

[0059] Therefore, the data that can be collected in the mass customization manufacturing process is defined as X:

[0060]

[0061] Where X pi represents the production and testing data of the i-th product, including customized assembly data, accessory quality data, and product quality data, p represents the total number of product categories; n pirepresents the number of products of category i actually produced; a represents the total number of accessory categories; qpi represents the total number of quality parameters of all accessories used in the production process of category i products; lpi represents the total number of quality indicators obtained in the testing process of category i products;

[0062] After the data definition is completed, the quality prediction of the present invention can be divided into two parts: offline training and online testing. Figure 2 shown.

[0063] In offline training, in order to classify and utilize mixed data, we first build a filter through the designed customized coding to determine the product category in the data. Add a customized coded data header to each piece of data, compare whether the coding is consistent to determine whether the data belongs to the same product, classify the data of the same product together, and record the customized coding of the product for building a judge in subsequent test data.

[0064] Then, due to the synergy and functional dependence between accessories, there is a causal relationship between accessories. At the same time, the selection of different accessories may affect the causal relationship between accessories. Therefore, in the data of different product types, causal discovery is performed separately and the corresponding causal graph structure is constructed. The separate construction of the causal graph structure can ensure better interpretability and accuracy.

[0065] A PCAM causal discovery method is designed. The process of PCAM causal discovery method is as follows Figure 3 As shown in the figure, the process can be divided into three parts: the first part is the construction of the causal skeleton, the second part is the determination of the causal direction, and the third part is the overall structure optimization. Through these three steps, the final causal graph is determined for subsequent training and prediction.

[0066] First, we build a causal skeleton by combining prior knowledge and data. We first determine accurate prior knowledge, including whether there is a causal relationship or even causal direction information, and use it as reliable knowledge. Then, we use data to perform HSIC independence tests on the remaining related variables to determine whether there is a causal relationship and build a skeleton diagram. This can improve the accuracy of the causal skeleton.

[0067] When there is no prior knowledge, it is also possible to construct a skeleton graph only from data.

[0068]

[0069] Where n is the number of samples, Tr is the trace of the matrix, and K X and K Y It is the kernel function calculation. Here, Gaussian kernel function is used, J=I-1 / n, which is the n-order unit matrix.

[0070]

[0071] Independence between variables means that there is no causal relationship, and non-independence between variables means that there is a causal relationship. The constructed causal skeleton is an undirected graph that may contain a small number of directed arrows (determined by accurate mechanism knowledge), which represents whether there is a causal relationship between the variables.

[0072] Then, on the determined causal skeleton, for each pair of variables with causal relationships, the ANMs method is used to determine the causal direction. In the production process, external factors such as workers' operating proficiency and assembly environment will interfere with the assembly process. In this case, the relationship between systems can be regarded as approximately consistent with the additive noise model, and the external noise is independent of the quality parameters of the accessories themselves. Therefore, the additive noise model can be used to construct forward and reverse functions, and its asymmetry can be used to determine the causal direction.

[0073] The judgment principle is that as an independent additive noise, it will not affect the cause variable, and since the result variable is obtained by the joint action of the cause variable and the noise variable, the result variable and the noise variable are not independent. Therefore, we construct a forward function f and a reverse function g to fit the relationship between the variable X and the variable Y, and judge the independence relationship between the noise NX and NY and the input variable.

[0074]

[0075] If the function satisfies formula 1 but not formula 2, it means that variable X is the cause variable of variable Y.

[0076] The present invention uses MLP to perform forward nonlinear fitting on the relationship between two variables, and uses two hidden layers and a nonlinear activation function LeakyReLU for nonlinear fitting. MLP, as a deep learning algorithm, can obtain better fitting effects for all nonlinear functions by controlling the Loss and the number of cycles.

[0077] Leaky ReLU=max(0,x)+α·min(0,r),0<α<<1

[0078] If NY and X are independent, that is, the direction is consistent with the data generation mechanism, then the positive hypothesis is correct, that is, variable X is the cause of variable Y. If not, then reject this hypothesis. Similarly, make an assumption in the opposite direction, that is, X = g(Y) + NX, and use the same method to judge independence. If both directions are independent, then choose the side with a higher independence score. If both directions are not independent, then this method fails.

[0079] However, this method also has its defects, that is, ANM is generally used for one-to-one causal discovery. When there is a many-to-one causal relationship, its judgment effect will deteriorate. Therefore, this paper introduces a scoring mechanism to further optimize the problem of inaccurate causal skeleton construction and inaccurate many-to-one causal effect.

[0080] Here we introduce the Bayesian criterion as the scoring function. When there is ambiguity in independence judgment and ANM failure, that is, when the difference between the forward and reverse directions of ANM construction is small and both sides are independent and the causal direction cannot be determined, we can determine whether the final overall score is improved by adding or deleting new edges and changing the causal direction, thereby determining the final causal graph result.

[0081] BIC = k·ln(n)-2ln(L)

[0082] Based on this, the heterogeneous causal graph attention network HCGAT is used to extract multi-scale information and use it for quality prediction. First, the number of accessories used in the product is determined through the customized assembly data of the product, and the data of different accessories are used as the feature input of the node. Then, HCGAT selects the corresponding weight parameter matrix for mapping according to the different feature dimensions of different nodes, and maps the variable features to the same dimension.

[0083] The update process of HCGAT is as follows Figure 3 As shown. First, different weight parameter matrices are used to map and embed features. The specific formula is as follows:

[0084] h ii =W i ·h i

[0085] Among them, h i are the characteristics of different nodes, W i It is a weight parameter matrix constructed for the feature dimension of the node and the feature dimension to be mapped. Through different W, all node features h are mapped to the same dimension. At this time, the isomorphic graph model can be used to process and train it.

[0086] At this point, the attention mechanism can be used to update the node. The specific formula is as follows:

[0087]

[0088] Where σ is the activation function, α ij It is the attention coefficient representation between nodes, W is the weight matrix in the attention mechanism, and W here is in the isomorphic graph, targeting the node features after being mapped to the same dimension as above.

[0089] Then HCGAT is trained. The training process is as follows Figure 4As shown. The Adam algorithm is selected as the optimization algorithm. Adam is a first-order optimization algorithm that can replace the traditional stochastic gradient descent process. It can iteratively update the weights of the neural network based on the training data. The mean square error on all label nodes is selected as the loss function.

[0090]

[0091] Use each product data to train the corresponding sub-model.

[0092] After the training is completed, the online testing phase begins.

[0093] First, for the test data, the classification information in the judge is used to classify the data according to the customized coding, and the test data is divided into different product categories. According to the product category, the corresponding trained HCGAT sub-model is selected to perform quality prediction.

[0094] The final output formula can be expressed as:

[0095] y i =HCGAT(X i ,A i )

[0096] The quality prediction effect is compared and verified by the automobile manufacturing simulation data set, which is generated by Simulink simulation based on the actual automobile production process.

[0097] The data set consists of two series of products, and the two products use a total of 12 accessories, of which product one uses 10 accessories, and product two uses a full set of 12 accessories. The present invention designs a representative parameter for each accessory, and at the same time, the number of each type of accessories is not exactly the same, and designs the quality parameters of the final product. The parameters of each type of accessories will affect the final product quality parameters. Therefore, product one has a total of 18 accessory quality parameters and a quality test parameter of the final automobile product, and product two has a total of 20 accessory parameters and a final automobile product test parameter. The input data is the quality parameter of each accessory in the automobile product, and the prediction target is the quality parameter of the automobile.

[0098] In order to verify the effect of the model, three methods were selected for comparison: MSPCA and using neural network for prediction, using PCAM causal composition method and using graph convolutional neural network GCN for prediction (HCGCN), and building causal graph through PC algorithm and using HCGAT for quality prediction. The relevant models used Adam optimizer to optimize the network, and the learning rate of all models was set to 0.001.

[0099] R2, MAE, RMSE and MAPE are used as prediction effect indicators, and the formulas are shown below.

[0100]

[0101] The experimental results are shown in the following table.

[0102]

[0103] At the same time, the prediction error is plotted into a graph, and the final result is as follows Figure 5 As shown in the figure, it can be found that by comparing the HCGAT and MSPCA methods, it is proved that processing multi-scale data by dimensionality reduction methods will lose some information, resulting in poor prediction results; by comparing HCGAT and HCGCN, it is proved that the advantage of introducing the attention mechanism; by comparing HCGAT and PC+HCGAT, it is proved that the quality of causal graph structure construction also has a great influence on the final prediction effect.

[0104] Therefore, the mass customization quality prediction based on HCGAT proposed in the present invention can better complete the mass customization quality prediction task, and the experiment proves the effectiveness and superiority of this method.

Claims

1. A mass customization production quality prediction method based on heterogeneous causal graph attention network, characterized in that: The method comprises the following steps: Step 1: Based on the characteristics of the mass customization production model, the manufacturing process and collected data of the mass customization production model are defined. The manufacturing process can be roughly divided into three steps: modular design, customized assembly, and product production and testing: 1.1 Higher requirements are placed on the design of product accessories. In order to adapt to more products, the design of accessories tends to be more standardized to form a universal accessory module; 1.2 During the customized assembly process, different types and quantities of accessories are selected to form a product series consisting of multiple products for customers to choose from; 1.3 Enter the production and testing phase, produce and test the selected products, and use product quality parameters to ensure the final product quality is qualified; In this process, the corresponding collected data is defined, including: 1) Accessory quality data: For each accessory, define its key quality parameter to represent the quality of the accessory. Usually, one key quality parameter is selected as a representative. This data is the input data of this solution. 2) Customized assembly data: define a customized encoding vector to represent the customized assembly information of the product, which is used to indicate the type and quantity of accessories used in the assembly of the product; 3) Product quality data: For the final product, during the testing phase, select its key quality parameters to represent the quality of the final product. Usually, one key quality parameter is selected as a representative, and this data is the target data of this solution. The data that can be collected in the mass customization manufacturing process is ultimately defined as X: Where X pi represents the production and testing data of the i-th product, including customized assembly data, accessory quality data, and product quality data, p represents the total number of product categories; n pi represents the number of products of category i actually produced; a represents the total number of accessory categories; qpi represents the total number of quality parameters of all accessories used in the production process of category i products; lpi represents the total number of quality indicators obtained in the testing process of category i products; Step 2: The overall quality prediction can be divided into two parts: offline training and online testing. In offline training, due to the characteristics of customized assembly and flexible co-production, the training data contains mixed data of multiple products. The customized coding vector in the collected data is used to build a filter for filtering and classifying large-scale customized data to obtain data of different product types. Step 3: A PCAM causal discovery method is designed to explore the relationship between variables and construct a causal graph structure for the data of different products classified by the filter. The steps are as follows: 3.1 Causal skeleton construction: First, use the mechanism knowledge and expert experience that are easily obtained in production to determine whether there is a causal relationship. Then use the independence test method between variables to determine whether there is a causal relationship between the remaining variables, and construct a causal skeleton diagram. 3.2 Determination of causal direction: Based on the causal skeleton diagram, the additive noise model is used to determine the causal direction between variables. The multi-layer perceptron MLP is used to perform forward and reverse nonlinear fitting, and the causal direction is determined by the independence between the cause variable and the noise. 3.3 Overall structure optimization: Using the Bayesian criterion as the scoring function, the additive noise model is judged to be ambiguous and invalid. By adding or deleting new edges and changing the causal direction, it is judged whether the final overall score is improved, so as to determine the final causal graph structure. Step 4: Based on the obtained causal graph, construct and train the heterogeneous causal graph neural network HCGAT. For each product data, train the corresponding sub-model. The steps are as follows: 4.1 Drawing on the idea of ​​heterogeneous graphs to deal with heterogeneous data, this invention uses different weight parameter matrices to deal with multi-scale data. For different variable dimensions, different weight parameter matrices are used to map them to the same dimension: h ii =W i ·h i Among them, h i are the characteristics of different nodes, W i It is a weight parameter matrix constructed for the feature dimension of the node and the feature dimension to be mapped. Through different W, all node features h are mapped to variables h of the same dimension. ii superior; 4.2 After obtaining variable data of the same dimension, the graph attention network is used for training, and the update function is as follows: Where σ is the activation function, α ij is the attention coefficient representation between nodes, W is the weight matrix in the attention mechanism, where W is in the isomorphic graph, targeting the node features after being mapped to the same dimension as above; Step 5: In the online testing phase, the general steps are as follows: 5.1 Obtain accessory quality data and customized assembly data as test data, extract the classification information in the filter during the training process to design a judge, and compare it with the encoding of the classification information in the filter to determine the product category to which the test data belongs; 5.2 According to the classification situation, select the corresponding trained HCGAT sub-model for quality prediction of this type of product.

2. According to claim 1, a method for predicting mass customization production quality based on heterogeneous causal graph attention network is characterized in that: The customized assembly data defined in 1 is as follows: According to the customized assembly information of products in mass customization, a customized assembly data is designed to represent the types and quantities of accessories used in assembling products. Its form is an n-dimensional vector, α = (a1, a2, ...a n ), where n represents the total number of accessory categories used in all products. If a=0, it means that this type of accessory is not used in the assembly of the product. If a=N, it means that N accessories of this type are used in the assembly of the product. The filter and judge are constructed through the designed customized assembly data for data classification. A customized coded data header is added to the data. The codes are compared to determine whether the data belongs to the same product.

3. According to claim 2, a method for predicting mass customization production quality based on heterogeneous causal graph attention network is characterized in that: The process of 3.1 is as follows: 3.1.1 First, obtain the accurate mechanism relationship and expert knowledge in the process of manufacturing as a priori, and determine whether there is a causal relationship between some variables and even the causal direction of variables. This part of information is relatively accurate and reliable; 3.1.2 Then use the data to perform HSIC independence test on the remaining related variables to determine the causal relationship. The threshold is designed to determine the independence relationship between variables. If it is less than the threshold, it means that the variables are independent and there is no causal relationship; if it is greater than the threshold, it means that the variables are not independent and there is a causal relationship: 3.1.3 When there is no prior knowledge, a skeleton diagram can be constructed only through data, which ensures the accuracy of the causal skeleton. At the same time, the relationship between variables is pre-screened through the causal skeleton diagram, reducing the computational complexity of causal discovery.

4. According to claim 3, a method for predicting mass customization production quality based on heterogeneous causal graph attention network is characterized in that: The process of 3.2 is as follows: 3.2.1 Use the additive noise model ANM method to determine the causal direction. First, assume that there is a positive causal relationship between the two variables. Use MLP to make a positive fit to the relationship between the two variables. Use two hidden layers and the nonlinear activation function LeakyReLU to ensure the nonlinear fitting effect. Use the error obtained after training as noise. 3.2.2 Additive noise is independent of the cause variable and not independent of the result variable. The independence relationship between noise and input variables and output variables is used to determine whether the hypothesis is met. If not, the hypothesis is rejected. 3.2.3 Make a similar judgment on the reverse relationship. If the function satisfies formula 1 but not formula 2, then it is determined that variable X is the cause variable of variable Y. If both directions are independent, the side with a higher independence score is selected. If both directions are not independent, this method fails. 3.2.4 Based on the determined causal skeleton, for each pair of variables with a causal relationship, the additive noise model (ANM) method is used to determine the causal direction.

5. According to claim 4, a method for predicting mass customization production quality based on heterogeneous causal graph attention network is characterized in that: The 4.1 process is as follows: 4.1.1 Through the customized assembly data of the product, determine the usage quantity of each accessory in the product, and construct vectors with the data of different accessories as the feature input of the node; 4.1.2 For different feature dimensions of different nodes, select the corresponding weight parameter matrix for mapping and map the variable features to the same dimension.

6. The method for predicting mass customization production quality based on heterogeneous causal graph attention network according to claim 5 is characterized by: Based on the obtained causal graph, the heterogeneous causal graph attention network HCGAT is used to extract multi-scale information and use it for quality prediction. This method utilizes the causal relationship between variables and extracts information from multi-scale data at the same time, which can achieve better prediction results.

Images