Large-scale customization quality prediction method based on transfer learning

Abstract

The invention discloses a large-scale customization quality prediction method based on transfer learning. The invention designs a large-scale customization quality prediction method for solving the problems that structures among variables are difficult to excavate and the quality of products with few samples is difficult to predict due to modular production, customized assembly and inconsistent number of different product samples in large-scale customization production. The method comprises the following steps: firstly, mining a causal relationship among source domain variables by utilizing a PC causal discovery method aiming at source domain data with a richer sample size, carrying out structure migration on a part of fixed structures from a source domain to a target domain by utilizing mechanism knowledge in order to obtain the causal relationship among target domain variables, and carrying out target domain few-sample amplification through a generative adversarial network; a causal relationship is complemented in a target domain after variable amplification through a PC causal discovery method, a target domain variable causal graph structure is obtained, then a heterogeneous causal graph attention network is utilized to carry out source domain and target domain model training, and target domain quality prediction is carried out through an MAML transfer learning method with MMD distance weighting. Aiming at the characteristics of large-scale customized production of customized products, the transfer learning-based quality prediction method suitable for large-scale customized production is formed.

Description

Technical Field

[0001] The present invention relates to a mass customization quality prediction method based on transfer learning, and belongs to the field of soft measurement. Background Art

[0002] Mass customization, a production model that balances enterprise productivity with consumer customization needs, is gaining widespread adoption across increasingly competitive industries, including semiconductor manufacturing, apparel manufacturing, and automotive manufacturing. The core principle of mass customization is to reduce internal product diversity through standardized components and then increase external diversity through customized assembly of different modules to create a variety of products, thereby satisfying the customization needs of diverse consumers.

[0003] The automotive manufacturing process, a typical example of mass customization, meets consumer demand for diverse products through the selection and selection of different accessories. However, in this mass customization process, the number of different models varies significantly due to differences in market positioning and demand. Some product types have high inventory and relatively rich actual test data and quality data, while others have less inventory and lower demand and only limited actual test data. Therefore, predicting the quality characteristics of these products is a challenge.

[0004] Current quality prediction models can be categorized as mechanism-based, expert knowledge-based, and data-driven. Mechanism models rely on mathematical models of the reaction process mechanism, but the complexity of modern industrial processes makes it difficult to construct accurate mechanism models. Expert knowledge models, on the other hand, are not very versatile and are limited by the difficulty of accumulating expert knowledge, making them primarily used as auxiliary tools. Data-driven quality prediction methods are widely used due to their wide applicability and limited application limitations. However, they have certain requirements for the amount of training data. Excessive data can significantly impact model training results. Transfer learning, on the other hand, transfers knowledge learned in the source domain task to the target domain task, improving the prediction performance of the target domain task model. Therefore, it is widely used in situations where data is insufficient.

[0005] At the same time, the lack of sample size significantly impacts the learning of inter-variable relationships within products. Automotive parts, as characteristic variables, exhibit significant inter-variable relationships. If these relationships are not exploited, the predictive model's effectiveness and interpretability will decline. However, in the automotive manufacturing process, a typical mass customization process, there are certain connections between different product types. This not only lays the foundation for transfer learning in soft sensing models, but also allows for the inter-variable structure to be learned and transferred from a rich sample of product types.

[0006] In addition, due to the different variables and structures of different models, there are large differences between their data. Therefore, there are significant domain differences between the source network and the target network in transfer learning, which also poses a huge challenge to transfer learning.

[0007] In summary, to address the above issues, it is necessary to develop a large-scale customized quality prediction model. Summary of the Invention

[0008] In response to the shortcomings of existing technologies and the characteristics of large-scale customization, the present invention designs a quality prediction method based on transfer learning. Compared with commonly used prediction methods, transfer learning can transfer the knowledge of classes with richer sample sizes to classes with smaller sample sizes, thereby obtaining better prediction results in the target domain.

[0009] This paper first uses the PC causal discovery method to mine causal relationships between source domain variables, targeting source domain data with a richer sample size. To obtain causal relationships between target domain variables, the paper leverages mechanistic knowledge to transfer partially fixed structures from the source domain to the target domain. A generative adversarial network is then used to amplify the target domain with a small number of samples. In the amplified target domain, the PC causal discovery method is used to complete the causal relationships and obtain the causal graph structure of the target domain variables. The paper then uses the MAML transfer learning method with MMD distance weighting to predict target domain quality. In response to the characteristics of mass customization in the production of customized products, a transfer learning-based quality prediction framework suitable for mass customization has been developed and validated in a dataset of the automotive manufacturing assembly process, a typical mass customization process.

[0010] The technical solutions of the present invention are as follows:

[0011] First of all, based on the characteristics of the large-scale customized production model, this patent roughly divides the manufacturing process into three steps: modular design, customized assembly, and product production and testing. In this process, the form of collected data will be defined, including accessory quality data, customized assembly data, and product quality data.

[0012] Modular design is to improve the standardization of accessory design and form universal modules to meet the needs of diversified product adaptation. Then, through customized assembly, different types and quantities of accessories are selected to build a product series for customers to choose from. Finally, the selected products are manufactured and tested, and product quality parameters are used to ensure the quality of the final product. The core data collected in the above process include: Accessory quality data: key quality parameters (usually one is selected) are used to characterize the quality status of accessories as input data. Customized coding vectors are used to record the types and quantities of accessories used in product assembly. Key quality parameters of products are selected in the testing phase to reflect the quality level of the final product.

[0013] Here, the data collected during the mass customization manufacturing process can be defined as X:

[0014]

[0015] Where X represents the production and testing data of the product, including customized assembly data, accessory quality data, and product quality data; n represents the number of products actually produced; a represents the total number of accessory categories; q represents the total number of quality parameters of all accessories used in the product production process; l represents the total number of quality indicators obtained in the product testing process. The data mainly includes the source domain data X s and target domain data X t .

[0016] During mass customization, different product types are produced based on the different selection and assembly of accessories. Due to differences in market positioning and demand, the number of different car models varies greatly. Some product types have high inventory and relatively rich actual test data and quality data, while some product types with small inventory and low demand have only very little actual test data. Therefore, types with richer data are used as source domain data, and product types with relatively scarce data are used as target domain data. The purpose of designing a quality prediction model is to extract relevant knowledge from the source domain data and apply it to the target domain so that the target domain with relatively scarce data can also obtain accurate quality prediction results.

[0017] To achieve target domain quality prediction, it is necessary to leverage knowledge from the source domain. To this end, we first use the PC causal discovery method to mine causal relationships between source domain variables and determine the relationships between them. This method consists of three steps: constructing a fully connected skeleton graph, removing redundant edges, and determining causal directions to determine the final causal graph structure.

[0018] PC causal discovery requires computing the correlations and conditional independence between variables from the data. First, a fully connected skeleton graph (i.e., an undirected graph without any edges) is constructed. A conditional independence test is then used to assess the conditional independence relationships between variables. Based on the results of the conditional independence test, the PC algorithm removes conditionally independent edges between variables to reduce redundant connections in the graph.

[0019] Starting with an empty set of conditions, gradually increase the size of the set. For each pair of adjacent variables, test whether they are independent under the given set of conditions. If a set of conditions exists that makes the adjacent variables independent, remove the corresponding edge and record the corresponding separating set. Repeat until all possible sets of conditions have been tested. Here, we use the chi-squared test. The chi-squared test formula and degrees of freedom formula are as follows.

[0020]

[0021] Among them Oijk It is the joint observation frequency of variables X and Y under a specific combination of conditional values. ijk is the expected frequency under the assumption that variables X and Y are independent. |Z| is the number of categories of the conditional variable Z.

[0022] Meanwhile, during the edge removal process, the undirected graph is directed, resulting in the final causal graph structure. Meek's rule is used here to determine the direction. If A→B exists and BC is undirected, and A and C are not adjacent, then the direction is B→C. If A→B→C exists and AC is undirected, then the direction is A→C. This also avoids loops, ensuring that the final graph is a directed acyclic graph (DAG).

[0023] The PC causal algorithm is used to mine the causal structure between variables in source domain samples. Based on vehicle model information and knowledge of automobile manufacturing mechanisms, some fixed relationships between variables are transferred from the source domain to the target domain. This fixes some relatively stable causal structures in the target domain as priors, generating a priori causal structure graph. To minimize the sample size, a generative adversarial network is used to amplify the sample size, and the PC causal discovery method is then used to complete the entire causal graph structure.

[0024] Generative Adversarial Network (GAN) achieves data generation through adversarial training. Its core consists of two modules: Generator (G) and Discriminator (D). The generator is trained from the prior noise distribution p z (usually a standard Gaussian distribution) samples the latent variable z and maps it to the data space G(z; θ z ); The discriminator receives real data x~p data Or generate samples G(z), and output a scalar representing the probability that the input is real data. The two are optimized through the minimax game:

[0025]

[0026] After randomly initializing the generator and discriminator, the training process alternately updates D and G. First, the discriminator is updated, the generator is fixed, a batch of samples is sampled from the real data and a batch of noise is sampled from the noise distribution to generate fake data, and the gradient is maximized:

[0027]

[0028] Then update the generator, fix the discriminator, sample the noise, and minimize it through gradient descent:

[0029]

[0030] When the discriminator reaches the optimal solution When , the generator objective is equivalent to minimizing the Jensen-Shannon divergence. By alternating updates to the generator and discriminator, when the optimal solution is reached, the data generated by the generator cannot be identified as fake data by the discriminator. Through adversarial training, the generated distribution asymptotically approximates the true distribution, generating new data that is closest to the original data, thereby achieving data gain.

[0031] On this basis, the PC causal discovery algorithm is used to learn the relationship between variables in the target domain data, and based on the prior causal structure diagram, the causal graph structure of the entire target domain is completed.

[0032] Transfer learning is then used to make the final prediction. Transfer learning is used here because, while GANs can amplify data, a small target domain data size prevents the generative adversarial network from generating large amounts of data, otherwise it will experience model collapse and data instability. Therefore, transfer learning is still necessary to transfer knowledge from the source domain data to the target domain, reducing the difficulty of target domain model training and improving the target domain model's prediction accuracy.

[0033] Based on the obtained causal graph structure, the heterogeneous causal graph attention network HCGAT is selected for the mass customization quality prediction model. Because the same spare part may be used more than one in the manufacturing process of mass customization, this will lead to multi-scale problems in the manufacturing data, that is, the dimensions of different variables are different. In order to deal with this problem, the heterogeneous causal graph attention network is selected. HCGAT uses the weight parameter matrix of different nodes to map node variables of different dimensions to the same dimension, and then uses the variable features of the same dimension through the graph attention network for training.

[0034] In order to unify the feature space of heterogeneous nodes, a learnable weight matrix is ​​used to perform linear transformation on the original features:

[0035] h′ i =W i ·x i

[0036] where x i is the original feature of node i, W i It is a weight parameter matrix designed for this node type for mapping. This operation converts the heterogeneous node features x i Projected into a unified dimensional space to meet the input requirements of the homogeneous graph model.

[0037] In the projection feature h i 'At this point, the attention mechanism is constructed, and the node representation update formula is defined as:

[0038]

[0039] Where σ is a nonlinear activation function, α ij is the attention coefficient representation between node i and node j, and W is the attention weight matrix.

[0040] The HCGAT model is constructed in the source domain. To ensure that source domain knowledge can be effectively transferred to the target domain, the concept of MAML meta-learning is borrowed. MAML is very different from the traditional pre-training and fine-tuning parameter scheme. In transfer learning, the pre-training and fine-tuning parameter scheme uses large-scale source domain data to pre-train model parameters as a universal feature extractor, and then fine-tunes them on a small sample of target domain data. Its effectiveness depends on the similarity of the distribution of the source and target domains. MAML uses a meta-learning framework to learn quickly adaptable initialization. It needs to consider the distribution of both datasets simultaneously. Although the weights obtained by MAML after training are not globally optimal for both tasks, after training, they can quickly adapt to different tasks, ensuring that they converge to the global optimum.

[0041] To achieve this goal, we add MMD loss to the loss function. MMD loss (Maximum Mean Discrepancy Loss) is a kernel-based probability distribution difference measure widely used in generative models, domain adaptation, and other fields. Its core principle is to measure the difference between two distributions by the distance between the sample means in the Reproducing Kernel Hilbert Space (RKHS). Its formula is:

[0042]

[0043] The loss function is defined as the mean square error (MSE) on the supervision node plus the MMD loss. The formula for the mean square error loss is:

[0044]

[0045] Where n is the number of nodes, f(x) and y are the predicted value and the true value respectively.

[0046] The overall error formula is:

[0047] Loss=Loss MSE +α·Loss MMD

[0048] Where α is the weight between the two.

[0049] The Adam optimizer is used to iteratively update the network weights. The algorithm adjusts the learning rate through adaptive moment estimation and has better convergence than traditional stochastic gradient descent.

[0050] The corresponding source domain HCGAT model is obtained by training with sample data in the source domain:

[0051] y s =HCGAT(X s ,A s )

[0052] After the source domain model is trained, the network model parameters are fixed, and then passed to the quality prediction network designed for the target domain as initialization parameters. Then, the sample data amplified by GAN is used for training, and the entire network parameters are fine-tuned to obtain the final target domain quality prediction model.

[0053] The trained target domain quality prediction HCGAT model is used for quality prediction of this type of product. Its hyperparameters are transferred to the target domain model, and the expanded target domain samples are used for training to obtain the final target domain quality prediction model:

[0054] y t =HCGAT(X t ,A t )

[0055] The beneficial effects of the present invention are mainly reflected in:

[0056] The present invention analyzes the relevant characteristics and problems of the mass customization production model, especially the problem of sample size imbalance in mass customization. In order to extract the target domain causal graph structure, the mechanism knowledge is first used to fix part of the structure of the source domain as the prior causal graph structure, and then the generative adversarial network is used to expand the number of target domain samples and the target domain causal graph is completed based on the PC causal discovery algorithm to ensure the accuracy and interpretability of the target domain causal graph. Then, for the multi-scale problem, the heterogeneous causal graph attention network is used for prediction. In order to ensure the effect of transfer learning, the MAML meta-learning idea is used to introduce the MMD loss to ensure the training of the target domain, so as to obtain a better target domain quality prediction effect. The mass customization quality prediction method based on transfer learning proposed in the present invention not only solves the basic problems such as multi-scale in mass customization, but also solves the problem of sample size imbalance by transferring the structural knowledge in the product data with rich sample size to ensure the construction of the causal graph, and uses the sample data characteristics in the source domain, combined with the distribution difference of the source domain and target domain data, to effectively use the source domain knowledge to train the target domain model, so as to obtain a better target domain quality prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the manufacturing process for mass customization;

[0058] Figure 2 This is the quality prediction framework diagram;

[0059] Figure 3 It is a causal discovery method for PC;

[0060] Figure 4 Schematic diagram of generative adversarial network;

[0061] Figure 5 Flowchart for transfer learning DETAILED DESCRIPTION

[0062] Reference Figures 1 to 5 The present invention is further described.

[0063] according to Figure 2 The flowchart shown in the figure is combined with actual data processing and model construction to introduce each step of the method in detail.

[0064] This patent describes the core manufacturing process and data collection based on the characteristics of mass customization production. The process mainly includes: modular design, customized assembly, and product production and testing. The data collected in this process include: accessory quality data, customized assembly data and product quality data, such as Figure 1 shown.

[0065] Modular design is to improve the standardization of accessory design and form universal modules to meet the needs of diversified product adaptation. Then, through customized assembly, different types and quantities of accessories are selected to build a product series for customers to choose from. Finally, the selected products are manufactured and tested, and product quality parameters are used to ensure the quality of the final product. The core data collected in the above process include: Accessory quality data: key quality parameters (usually one is selected) are used to characterize the quality status of accessories as input data. Customized coding vectors are used to record the types and quantities of accessories used in product assembly. Key quality parameters of products are selected in the testing phase to reflect the quality level of the final product.

[0066] Here, the data collected during the mass customization manufacturing process can be defined as X:

[0067]

[0068] Where X represents the production and testing data of the product, including customized assembly data, accessory quality data, and product quality data; n represents the number of products actually produced; a represents the total number of accessory categories; q represents the total number of quality parameters of all accessories used in the product production process; l represents the total number of quality indicators obtained in the product testing process. The data mainly includes the source domain data X s and target domain data X t .

[0069] During mass customization, different product types will be produced based on the different selection and assembly of accessories. Due to differences in market positioning and demand, the number of different car models varies greatly. Some product types have high inventory and relatively rich actual test data and quality data, while some product types with less inventory and lower demand have only a small amount of actual test data. Therefore, the types with richer data are used as source domain data, and the product types with relatively scarce data are used as target domain data. The purpose of designing a quality prediction model is to extract relevant knowledge from the source domain data and apply it to the target domain so that the target domain with relatively scarce data can also obtain accurate quality prediction results. Therefore, the quality prediction of the present invention can be divided into two parts: source domain model training and target domain model testing. Figure 2 shown.

[0070] In order to achieve quality prediction in the target domain, it is necessary to utilize the knowledge of the source domain. To this end, the PC causal discovery method is first used to mine the causal relationship between source domain variables and determine the relationship between variables. The method can be divided into three steps: constructing a fully connected skeleton graph, deleting redundant edges, and determining the causal direction. The final causal graph structure is determined, such as Figure 3 shown.

[0071] PC causal discovery requires computing the correlations and conditional independence between variables from the data. First, a fully connected skeleton graph (i.e., an undirected graph without any edges) is constructed. A conditional independence test is then used to assess the conditional independence relationships between variables. Based on the results of the conditional independence test, the PC algorithm removes conditionally independent edges between variables to reduce redundant connections in the graph.

[0072] 1. Start with an empty set of conditions and gradually increase the size of the set. For each pair of adjacent variables, test whether they are independent under the given set of conditions. If a set of conditions exists that makes the adjacent variables independent, remove the corresponding edge and record the corresponding separating set. Repeat until all possible sets of conditions have been tested. Here, we use the chi-square test. The chi-square test formula and degrees of freedom formula are as follows.

[0073]

[0074] Among them O ijk It is the joint observation frequency of variables X and Y under a specific combination of conditional values. ijk is the expected frequency under the assumption that variables X and Y are independent. |Z| is the number of categories of the conditional variable Z.

[0075] Meanwhile, during the edge removal process, the undirected graph is directed, resulting in the final causal graph structure. Meek's rule is used here to determine the direction. If A→B exists and BC is undirected, and A and C are not adjacent, then the direction is B→C. If A→B→C exists and AC is undirected, then the direction is A→C. This also avoids loops, ensuring that the final graph is a directed acyclic graph (DAG).

[0076] The PC causal algorithm is used to mine the causal structure between variables in source domain samples. Based on vehicle model information and knowledge of automobile manufacturing mechanisms, some fixed relationships between variables are transferred from the source domain to the target domain. This fixes some relatively stable causal structures in the target domain as priors, generating a priori causal structure graph. To minimize the sample size, a generative adversarial network is used to amplify the sample size, and the PC causal discovery method is then used to complete the entire causal graph structure.

[0077] Generative Adversarial Network (GAN) achieves data generation through adversarial training. Its core consists of two modules: Generator (G) and Discriminator (D). Figure 4 As shown. The generator is derived from the prior noise distribution p z (usually a standard Gaussian distribution) samples the latent variable z and maps it to the data space G(z; θ z ); The discriminator receives the real data xp data Or generate samples G(z), and output a scalar representing the probability that the input is real data. The two are optimized through the minimax game:

[0078]

[0079] After randomly initializing the generator and discriminator, the training process alternately updates D and G. First, the discriminator is updated, the generator is fixed, a batch of samples is sampled from the real data and a batch of noise is sampled from the noise distribution to generate fake data, and the gradient is maximized:

[0080]

[0081] Then update the generator, fix the discriminator, sample the noise, and minimize it through gradient descent:

[0082]

[0083] When the discriminator reaches the optimal solution When , the generator objective is equivalent to minimizing the Jensen-Shannon divergence. By alternating updates to the generator and discriminator, when the optimal solution is reached, the data generated by the generator cannot be identified as fake data by the discriminator. Through adversarial training, the generated distribution asymptotically approximates the true distribution, generating new data that is closest to the original data, thereby achieving data gain.

[0084] On this basis, the PC causal discovery algorithm is used to learn the relationship between variables in the target domain data, and based on the prior causal structure diagram, the causal graph structure of the entire target domain is completed.

[0085] Then, transfer learning is used to make the final prediction, such as Figure 5 As shown in Figure 2, transfer learning is used here because, while GANs can amplify data, a small target domain data size prevents the generative adversarial network from generating large amounts of data, otherwise it will experience model collapse and data instability. Therefore, transfer learning is still necessary to transfer knowledge from source domain data to the target domain, reducing the difficulty of target domain model training and improving the target domain model's prediction accuracy.

[0086] Based on the obtained causal graph structure, the heterogeneous causal graph attention network HCGAT is selected for the mass customization quality prediction model. Because the same spare part may be used more than one in the manufacturing process of mass customization, this will lead to multi-scale problems in the manufacturing data, that is, the dimensions of different variables are different. In order to deal with this problem, the heterogeneous causal graph attention network is selected. HCGAT uses the weight parameter matrix of different nodes to map node variables of different dimensions to the same dimension, and then uses the variable features of the same dimension through the graph attention network for training.

[0087] In order to unify the feature space of heterogeneous nodes, a learnable weight matrix is ​​used to perform linear transformation on the original features:

[0088] h i ′=W i ·x i

[0089] where x i is the original feature of node i, W i It is a weight parameter matrix designed for this node type for mapping. This operation converts the heterogeneous node features x i Projected into a unified dimensional space to meet the input requirements of the homogeneous graph model.

[0090] In the projection feature h i 'At this point, the attention mechanism is constructed, and the node representation update formula is defined as:

[0091]

[0092] Where σ is a nonlinear activation function, α ij is the attention coefficient representation between node i and node j, and W is the attention weight matrix.

[0093] The HCGAT model is constructed in the source domain. To ensure that source domain knowledge can be effectively transferred to the target domain, the concept of MAML meta-learning is borrowed. MAML is very different from the traditional pre-training and fine-tuning parameter scheme. In transfer learning, the pre-training and fine-tuning parameter scheme uses large-scale source domain data to pre-train model parameters as a universal feature extractor, and then fine-tunes them on a small sample of target domain data. Its effectiveness depends on the similarity of the distribution of the source and target domains. MAML uses a meta-learning framework to learn quickly adaptable initialization. It needs to consider the distribution of both datasets simultaneously. Although the weights obtained by MAML after training are not globally optimal for both tasks, after training, they can quickly adapt to different tasks, ensuring that they converge to the global optimum.

[0094] To achieve this goal, we add MMD loss to the loss function. MMD loss (Maximum Mean Discrepancy Loss) is a kernel-based probability distribution difference measure widely used in generative models, domain adaptation, and other fields. Its core principle is to measure the difference between two distributions by the distance between the sample means in the Reproducing Kernel Hilbert Space (RKHS). Its formula is:

[0095]

[0096] The loss function is defined as the mean square error (MSE) on the supervision node plus the MMD loss. The formula for the mean square error loss is:

[0097]

[0098] Where n is the number of nodes, f(x) and y are the predicted value and the true value respectively.

[0099] The overall error formula is:

[0100] Loss=Loss MSE +α·Loss MMD

[0101] Where α is the weight between the two.

[0102] The Adam optimizer is used to iteratively update the network weights. The algorithm adjusts the learning rate through adaptive moment estimation and has better convergence than traditional stochastic gradient descent.

[0103] The corresponding source domain HCGAT model is obtained by training with sample data in the source domain:

[0104] y s =HCGAT(X s ,A s )

[0105] After the source domain model is trained, the network model parameters are fixed, and then passed to the quality prediction network designed for the target domain as initialization parameters. Then, the sample data amplified by GAN is used for training, and the entire network parameters are fine-tuned to obtain the final target domain quality prediction model.

[0106] The trained target domain quality prediction HCGAT model is used for quality prediction of this type of product. Its hyperparameters are transferred to the target domain model, and the expanded target domain samples are used for training to obtain the final target domain quality prediction model:

[0107] y t =HCGAT(X t ,A t )

Claims

1. A mass customization production quality prediction method based on transfer learning, characterized in that: The method comprises the following steps: Step 1: Based on the characteristics of the mass customization production model, define its manufacturing process and related data collection. The manufacturing process is divided into three steps: modular design, customized assembly, and product production and testing: 1.1 Modular design: Improve the standardization of accessory design and form universal modules to meet the adaptation needs of diverse products; 1.2 Customized assembly: By selecting different types and quantities of accessories, we can build a product series for customers to choose from; 1.3 Production and Testing: Execute the manufacturing and testing of selected products, and use product quality parameters to ensure the final product is qualified; The core data collected in the above process includes: 1) Accessory quality data: Key quality parameters (usually one) are used to characterize the quality of the accessories as input data for this solution; 2) Customized assembly data: Customized encoding vectors are used to record the types and quantities of components used in product assembly. 3) Product quality data: Select key quality parameters (usually one) during the testing phase to reflect the quality level of the final product; Step 2: In the mass customization process, such as the automobile manufacturing process, the number of different car models varies greatly due to differences in market positioning and demand. Some product types have high inventory and relatively rich actual test data and quality data, while some product types with low inventory and low demand have only limited actual test data. Therefore, product types with abundant samples are used as source domain data, and product types with relatively scarce samples are used as target domain data. The entire quality prediction method can be divided into source domain model training and target domain quality prediction. Step 3: To obtain the causal graph structure of the target domain, we use the source domain structure transfer, generative adversarial network sample amplification, and PC causal discovery method to ensure the integrity and reliability of the target domain causal graph structure. The steps are as follows: 3.1 Source Domain Causal Graph Construction: First, for source domain data with relatively rich samples, we use the PC causal discovery method to determine the final source domain causal graph structure through three steps: constructing a fully connected skeleton graph, deleting redundant edges, and determining the causal direction. 3.2 Target domain prior causal graph construction: Based on vehicle model information and knowledge of automobile manufacturing mechanisms, some fixed relationships between variables are transferred from the source domain causal graph structure to the target domain, so that some relatively stable causal structures are fixed in the target domain as priors, generating a prior causal structure graph; 3.3 Obtaining the complete target domain causal graph structure: First, to address the problem of small causal graph samples, we use generative adversarial networks to amplify them. Then, we use the PC causal discovery method to learn the target domain causal graph. Based on the target domain prior causal graph, we obtain the complete target domain causal graph structure. Step 4: Based on the obtained source domain causal graph, construct and train the heterogeneous causal graph neural network of the target domain, and design a weighted loss function with MMD loss to reduce the training requirements of the target domain and improve the training effect of the target domain. The steps are as follows: 4.1 To unify the feature space of heterogeneous nodes, a learnable weight matrix is ​​used to perform linear transformation on the original features: h′ i =W i ·x i where x i is the original feature of node i, W i It is a weight parameter matrix designed for this node type for mapping. This operation converts the heterogeneous node features x i Projecting to a unified dimensional space to meet the input requirements of the homogeneous graph model; 4.2 Projection feature h′ i At this point, the attention mechanism is constructed and trained. The node representation update formula is defined as: Where σ is a nonlinear activation function, α ij is the attention coefficient representation between node i and node j, W is the attention weight matrix; 4.3 To ensure that source domain knowledge can be effectively transferred to the target domain, we borrow the concept of MAML meta-learning. We add MMD loss to the loss function, and the overall error formula is: Loss=Loss MSE +α·Loss MMD Where α is the weight between the two; Step 5: In the target domain quality prediction phase, the parameters of the source domain model after training are transferred to the target domain model, and the parameters are fine-tuned through target domain training to obtain the final target domain quality prediction model for quality prediction. y t =HCGAT(X t ,A t ) 2. The method for predicting mass customization production quality based on transfer learning according to claim 1, characterized in that: The customized assembly data defined in 1 above is as follows: Here, the data collected during the mass customization manufacturing process can be defined as X: Where X represents the production and testing data of the product, including customized assembly data, accessory quality data, and product quality data; n represents the number of products actually produced; a represents the total number of accessory categories; q represents the total number of quality parameters of all accessories used in the product production process; l represents the total number of quality indicators obtained in the product testing process. The data mainly includes the source domain data X s and target domain data X t .

3. The method for predicting mass customization production quality based on transfer learning according to claim 2, characterized in that: The process of 4.3 is as follows: 3.1 Add MMD loss to the loss function to measure the difference between the two distributions. The formula is: 3.2 The loss function is defined as the mean square error (MSE) on the supervision node plus the MMD loss. The formula for the mean square error loss is: Where n is the number of nodes, f(x) and y are the predicted value and the true value respectively. 3.3 The overall error formula is: Loss=Loss MSE +α·Loss MMD Where α is the weight between the two, and the source domain model is trained using the MMD weighted loss function.

4. The method for predicting mass customization production quality based on transfer learning according to claim 3, characterized in that: The process of step 5 is as follows: 4.1 After the source domain model is trained, the network model parameters are fixed and then passed to the quality prediction network designed for the target domain as initialization parameters. The network model is then trained using the sample data amplified by the GAN, and the overall network parameters are fine-tuned to obtain the final target domain quality prediction model. 4.2 Use the trained target domain quality prediction HCGAT model to predict the quality of this type of product. y t =HCGAT(X t ,A t ) 5. The method for predicting mass customization production quality based on transfer learning according to claim 4, characterized in that: In response to the problems of sample size imbalance and domain differences in the mass customization production model, the scarcity of target domain samples is solved through causal structure migration and sample expansion, and the source domain data and knowledge are effectively utilized to overcome the data distribution differences between the source domain and the target domain, reducing the difficulty of target domain training. The resulting heterogeneous causal attention network for target domain quality prediction significantly improves the quality prediction effect of target domain products in mass customization scenarios.

Images