Automatic digital twinning generation method based on data driving and graph learning and product quality prediction method and device

Abstract

The invention provides an automatic digital twinning generation method based on data driving and graph learning and a product quality prediction method and device, and belongs to the technical field of digital twinning. Comprising the steps of collecting physical characteristics and constructing a physical diagram; associating the entity graph features with regression models on other related features; taking the entity graph as a node, and modeling the entity graph by using a graph-to-graph conversion model to construct a digital twin graph; fusing the digital twin map with the domain knowledge-based DT model of the entity in the map to construct a fused digital twin map; and predicting physical characteristics of an unobservable entity according to the obtained fused digital twin map, and storing the fused digital twin map in a database. According to the method, the general data structure related to the processing framework is established, the heterogeneous physical entities and the relation thereof are modeled, dependence on domain knowledge is eliminated, an existing domain-based model and a data-driven model are fused, and the fidelity of digital twinning is enhanced.

Description

Technical Field

[0001] The present invention relates to an automatic digital twin generation method and a product quality prediction method and device based on data drive and graph learning, belonging to the technical field of digital twins. Background Art

[0002] Current digital twin research is unable to model heterogeneous and complex relationships between different physical entities in a unified way. To address the problem that the rich information contained in sensor data is not fully utilized, and the current digital twin research is unable to model heterogeneous and complex relationships between different physical entities in a unified way, various models must be used to represent different relationships and rely heavily on domain expertise. Therefore, physical-to-virtual modeling of entities and their relationships is usually performed from scratch on an application-by-application basis. This is expensive and time-consuming, and makes it almost infeasible to build scalable digital twins with a large number of heterogeneous entities. Summary of the invention

[0003] The purpose of the present invention is to provide an automatic digital twin generation method and a product quality prediction method and device based on data-driven and graph learning. By establishing a common data structure related to the processing framework, heterogeneous physical entities and their relationships are modeled, thus getting rid of the dependence on domain knowledge.

[0004] In order to achieve the above object, the present invention is implemented through the following technical solutions: The physical features of the physical entity are collected and preprocessed, and the significance test of the correlation method with the highest confidence level produced by all the features is performed to construct the entity graph.

[0005] A feature in the entity graph feature is associated with a regression model on each other related feature determined by the entity graph structure. When the value of any feature in the entity graph changes, the regression model updates the value of the feature related to the changed feature.

[0006] The entity graph is used as a node and a graph-to-graph conversion model is used to model the entity graph to construct a digital twin graph (DTG), in which the directed arcs connecting the nodes in the digital twin graph represent the relationships between the entity graphs.

[0007] Obtain a domain knowledge-based digital twin model of each entity in the entity graph, fuse the digital twin graph with the domain knowledge-based digital twin model of the entity in the graph, and construct a fused digital twin graph.

[0008] The physical characteristics of unobservable entities are predicted based on the acquired fused digital twin graph, and the fused digital twin graph is stored in the database. When the entity graph changes, the database is queried whether the same object exists. If so, it is added to the newly constructed digital twin graph according to the entity graph relationship stored in the database. Only the relationship between the unstored entity graph and other entity graphs of the newly constructed digital twin graph is trained.

[0009] Preferably, the specific method of using the entity graph as a node and using the graph-to-graph conversion model to model the entity graph and construct a digital twin graph is as follows: The entity graph at the tail end of the directed arc of the digital twin graph is taken as the source entity graph, and the entity graph at the head end of the directed arc is taken as the target entity graph, and the source entity graph and the target entity graph are graphically adapted.

[0010] A graph convolutional neural network is used as the encoder to encode the feature vectors in the source entity graph.

[0011] After decoding by the dot product decoder, the reconstructed target entity graph is obtained.

[0012] The graph convolutional neural network is iteratively trained using the loss function to make the reconstructed target entity graph close to the target entity graph.

[0013] Preferably, the specific method of using the graph convolutional neural network as an encoder to encode the feature vector in the source entity graph is as follows: Construct an adjacency matrix from the source entity graph and the feature matrix , so that the adjacency matrix Capturing sufficiently correlated feature pairs, the adjacency matrix Each element in ,in, Representation Node To Node The correlation coefficient of Represents the identity matrix.

[0014] The adjacency matrix and the feature matrix Input graph convolutional neural network encoding and get the encoding result , the specific parameterized graph convolutional neural network output is as follows: , in, represents the output of the last layer of graph convolutional neural network, Represents the weight matrix of the previous layer of graph convolutional neural network, Represents the number of layers of the graph convolutional neural network.

[0015] The specific steps of decoding the dot product decoder to reconstruct the target entity graph are as follows: The encoding result Perform node conversion to obtain the potential feature matrix of the source entity graph .

[0016] The latent feature matrix Transpose to get the transposed feature matrix , and transform the latent feature matrix and the transposed feature matrix Input to the dot product decoder to obtain the adjacency matrix of the reconstructed target entity graph .

[0017] The specific calculation formula is as follows: , in, Represents the element in row i and column j of the adjacency matrix of the reconstructed target entity graph, represents the identity matrix, represents the logistic sigmoid function, Expressed as a range , The number of features that represent the entity.

[0018] The adjacency matrix of the target entity graph will be reconstructed Transformed into the reconstructed target entity graph.

[0019] The loss function is as follows: , , , in, Represents the adjacency matrix The sparsity of Represents the gTarget adjacency matrix i OK j Column elements, Represents the actual target entity graph node The characteristic vector of Represents the node in the target entity graph that is reconstructed The characteristic vector of is a hyperparameter whose value is greater than 0.

[0020] Preferably, the specific manner of graphic adaptation is as follows: If the number of features of the source entity graph is smaller than that of the target entity graph, dummy features and random arc weights are padded to the source entity graph, where each dummy feature has a 0 feature vector and has inward connections with other real features.

[0021] If the number of features in the source entity graph is greater than the number of features in the target entity graph, strongly correlated features in the source entity graph are merged into super features represented by cluster centroids through K-means clustering, where the value of K is equal to the number of features in the target entity graph.

[0022] Preferably, the specific method of integrating the digital twin graph with the DT model based on domain knowledge of the entities in the graph is as follows: Establish a comparative fusion integration model to obtain the model values ​​of the two entity graphs , if the model value , the connection relationship between the two entity graphs remains unchanged. If the model value , then delete the connection relationship between the two entity graphs; the comparison fusion integration model is as follows: , in, , Represents a threshold variable, and its value range is greater than 0. Represents the final loss function value of the graph convolutional neural network, Indicates whether there is a connection relationship between two entity graphs in the DT model based on domain knowledge. Its value is 0 if there is no connection relationship, and 1 if there is a connection relationship.

[0023] For each contrast fusion ensemble model, an equivalent generator network is created and trained to obtain a fused entity model.

[0024] The digital twin graph and the DT model based on domain knowledge of the entities in the graph are input into the fused entity model to obtain the fused digital twin graph.

[0025] Preferably, the process of training the equivalent generator network is as follows: Random Gaussian noise is input into the equivalent generator network and the integrated model respectively to predict all feature values ​​of the entity graph; Use the discriminator to distinguish the prediction results of the equivalent generator network from those of the integrated model, backpropagate the difference to the equivalent generator, and iterate until the discriminator cannot distinguish the prediction results of the equivalent generator network from those of the integrated model.

[0026] Preferably, the equivalent generator network includes a first GCN layer, a first Relu activation function, a second GCN layer, a second Relu activation function, a node conversion module, an embedding matrix, and a dot product decoder topology in sequence, and each module has an input and output relationship in sequence.

[0027] A product quality prediction method based on automatic digital twin, comprising: Establishing a fused digital twin graph using the automatic digital twin generation method based on data-driven and graph learning; Integrate the digital twin The entity graphs are respectively input into the dimensionality reduction feature extraction module; the dimensionality reduction feature extraction module is composed of a first graph neural network convolution layer, a first graph neural network pooling layer, a second graph neural network convolution layer, and a second graph neural network pooling layer in sequence, and each module has an input and output relationship in sequence.

[0028] Add the average pooled feature vectors of the first graph neural network pooling layer and the second graph neural network pooling layer to obtain global features of different scales , the pooled feature vectors obtained by passing all entity graphs through the second graph neural network pooling layer are combined into the reconstructed entity graph.

[0029] The reconstructed entity graph is processed by the dimension reduction feature extraction module to obtain the average pooling feature vector of the first graph neural network pooling layer, the average pooling feature vector of the second graph neural network pooling layer, and the pooling feature vector of the second graph neural network pooling layer. , add the average pooling feature vector of the first graph neural network pooling layer and the average pooling feature vector of the second graph neural network pooling layer to obtain global features of different scales .

[0030] All global features , And the pooled feature vector Input the neural network, and then pass the output of the neural network output layer through the ReLU activation function layer to obtain the product quality prediction result.

[0031] Preferably, the reconstructed entity graph combination method is as follows: an undirected line segment is generated between two entity graphs that have a relationship in the fused digital twin graph.

[0032] The advantages of the present invention are as follows: The present invention proposes a digital twin graph that does not require domain knowledge and is completely data-driven. The existing domain-based model is integrated with the data-driven model. This enables DTG to utilize domain knowledge at any time, thereby enhancing the fidelity of the digital twin. The present invention allows DTG to perform complex system simulations and update real-time changes in any entity state or system structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0034] Figure 1 Schematic diagram of the processing framework of the digital twin graph of the present invention.

[0035] Figure 2 It is a schematic diagram of the digital twin graph structure of the present invention and its related G2G transformation model.

[0036] Figure 3 This is a schematic diagram of the generator structure of the present invention.

[0037] Figure 4 This is a schematic diagram of real-time model fusion using CFIN in the present invention.

[0038] Figure 5 This is a flow chart of the prediction model algorithm of the present invention.

[0039] Figure 6 This is the automatic digital twin diagram DTG of the air-conditioning outdoor unit production line of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] Example 1 like Figure 1 As shown, an automatic digital twin generation method based on data-driven and graph learning includes: S1: Collect and preprocess the physical features of the physical entity, perform a significance test (such as a t-test) using the correlation method (such as the Pearson coefficient) that produces the highest confidence on average using all the features, and construct an entity graph.

[0042] This embodiment takes air conditioning production as an example, and the following physical characteristics need to be collected: Copper tube assembly production line: assembly pressure, assembly speed, ambient temperature, ambient humidity, and assembly gap.

[0043] U-shaped elbow production line: equipment speed, equipment temperature, vibration level, ambient humidity, and ambient temperature.

[0044] Welding production line: flame temperature, heating time, electrode dosage, equipment temperature, ambient humidity, and vibration level.

[0045] Helium leak detection production line: detection pressure, sensitivity, leakage rate, ambient temperature, ambient humidity, detection time, equipment temperature.

[0046] Vacuum pumping and refrigerant injection production line: pumping speed, pumping time, pumping pressure, refrigerant quantity, refrigerant pressure, refrigerant temperature, equipment temperature.

[0047] Shell stamping production line: equipment speed, equipment temperature, vibration level, ambient humidity, ambient temperature, and stamping pressure.

[0048] When purchasing a compressor: rated power, cooling capacity, noise level, refrigeration coefficient, and rated pressure.

[0049] Purchase fans and: diameter, speed, air volume, rated power, noise level.

[0050] S2: Associating a feature in the entity graph feature with a regression model on each other related feature determined by the entity graph structure. When the value of any feature in the entity graph changes, the regression model updates the value of the feature related to the changed feature.

[0051] S3: Use the entity graph as a node and use the graph-to-graph transformation model to model the entity graph to construct a digital twin graph (DTG), in which the directed arcs connecting the nodes in the digital twin graph represent the relationship between the entity graphs. Unlike traditional GNNs, where GNN node relationships are modeled as simple scalars (decay rates), DTG uses a G2G transformation model that can describe complex relationships between physical entities with high fidelity.

[0052] like Figure 2 As shown in Figure 2 As shown in Figure 1, it models the mapping from source entity graph i to target entity graph j. More specifically, the G2G model is used to automatically update the feature values ​​of the target graph whenever the feature values ​​of the source graph change. This update rule usually varies between different source-target pairs, so a data-driven approach is adopted to learn the rule based on historical data.

[0053] The specific method of constructing the digital twin graph is as follows: S301: The entity graph at the tail end of the directed arc of the digital twin graph is used as the source entity graph, and the entity graph at the head end of the directed arc is used as the target entity graph, and the source entity graph and the target entity graph are adapted. The specific method of the graphic adaptation is as follows: If the number of features in the source entity graph is less than the number of features in the target entity graph, pseudo features and random arc weights are added to the source entity graph, each pseudo feature has a 0 feature vector and has inward connections with other real features. These filled features will use the real feature vectors to compose their feature vectors during the encoding process, while not affecting the real features or the target entity features. Therefore, the final 0 feature vector will not contribute to the transformation of node features or corresponding topology, and the parameters of the final trained gDecode and GCN encoders will be Will not be affected by spurious features in the source solid graph.

[0054] If the number of features in the source entity graph is greater than the number of features in the target entity graph, strongly correlated features in the source entity graph are merged into super features represented by cluster centroids through K-means clustering, where the value of K is equal to the number of features in the target entity graph.

[0055] S302: Use the graph convolutional neural network as an encoder to encode the feature vector in the source entity graph. The specific method is as follows: Construct an adjacency matrix from the source entity graph and the feature matrix , so that the adjacency matrix Capturing sufficiently correlated feature pairs, the adjacency matrix Each element in ,in, Representation Node To Node The correlation coefficient of The dimension is , represents the number of features of the entity, Represents the length of the feature vector.

[0056] Parameterize the encoder as follows: , , in, Indicates The output of the layer, Indicates The weight matrix of the layer, , Represents the number of layers of the graph convolutional neural network.

[0057] The adjacency matrix and the feature matrix Input graph convolutional neural network encoding and get the encoding result , the specific parameterized graph convolutional neural network output is as follows: , in, represents the output of the last layer of graph convolutional neural network, Represents the weight matrix of the previous layer of graph convolutional neural network, Represents the number of layers of the graph convolutional neural network.

[0058] S303: After decoding by the dot product decoder, the reconstructed target entity graph is obtained. The specific steps are as follows: The encoding result Perform node conversion to obtain the potential feature matrix of the source entity graph , The dimension is .

[0059] The latent feature matrix Transpose to get the transposed feature matrix , and transform the latent feature matrix and the transposed feature matrix Input to the dot product decoder to obtain the adjacency matrix of the reconstructed target entity graph .

[0060] The specific calculation formula is as follows: , in, represents the identity matrix, represents the logistic sigmoid function, Expressed as a range , The number of features that represent the entity.

[0061] The adjacency matrix of the target entity graph will be reconstructed Transformed into the reconstructed target entity graph.

[0062] S304: Iteratively train the graph convolutional neural network using a loss function to make the reconstructed target entity graph close to the target entity graph. The loss function is as follows: , , , in, , , Represents the adjacency matrix The sparsity of Represents the gTarget adjacency matrix i OK j Column elements, Represents the actual target entity graph node The characteristic vector of Represents the node in the target entity graph that is reconstructed The characteristic vector of To adjust the hyperparameters of the priority between the two losses ( >0).

[0063] because The maximum possible number of arcs, therefore, adding S before the first term puts pressure on connected arcs when S is large (sparser graph), thus mitigating extreme (too sparse or too dense) graph topologies. The reason for avoiding extremes is that such a matrix tends to ignore under-represented classes (connected or unconnected arcs), making it difficult to learn class membership.

[0064] Used by the source and target entity graphs Training the entire G2G transformation model can learn the parameters of the GCN encoder .

[0065] like Figure 6 As shown, in this embodiment, the digital twin graph established by taking the air-conditioning production line as an example includes entities such as copper tube assembly production line, U-shaped elbow production line, welding production line, helium leak detection production line, vacuum refrigerant injection production line, purchased compressor, purchased fan and shell stamping production line. Arrows ①②③④⑤⑥⑦⑧⑨⑩ are G2G transformations between entity graphs, and the total number of entities is 8. Each entity graph contains the relationship between the characteristic vector of each parameter node and the node, and the arrows in the entity graph are the relationship between the nodes. The characteristic vector takes five dimensions, among which the node characteristic time vector parameters in the purchased compressor and purchased fan entity graphs remain unchanged, so only one-dimensional data is displayed. Some entity graphs in the air-conditioning production line are depicted as Figure 3 The corresponding parts of ⅠⅡ are block diagrams.

[0066] S4: Obtain the domain knowledge-based DT model of each entity in the entity graph, fuse the digital twin graph with the domain knowledge-based DT model of the entity in the graph, and construct a fused digital twin graph.

[0067] Since the connection relationship between each two entity graphs in the digital twin graph constructed in the above steps is quite different from the connection relationship of the existing DT model based on domain knowledge, it is necessary to construct a fused digital twin graph to solve this problem; the specific method is as follows: S401: Establish a comparative fusion integration model to obtain the model values ​​of the two entity graphs , if the model value , the connection relationship between the two entity graphs remains unchanged. If the model value , then delete the connection relationship between the two entity graphs; the comparison fusion integration model is as follows: , in, , represents the threshold variable ( , the specific value shall be adjusted according to the actual situation). Represents the final loss function value of the graph convolutional neural network, Indicates whether there is a connection relationship between two entity graphs in the DT model based on domain knowledge. Its value is 0 if there is no connection relationship, and 1 if there is a connection relationship.

[0068] The ensemble model may end up having a large number of parameters, increasing its risk of overfitting. Moreover, due to the potential multitude of entities in the physical system, the entire DTG may contain a large number of parameters due to the ensemble model. These may jeopardize the scalability of the comprehensive digital twin during construction and simulation. Therefore, an equivalent generator network is constructed.

[0069] S402: Create and train an equivalent generator network for each comparison fusion integration model to obtain a fusion entity model. The equivalent generator network includes a first GCN layer, a first Relu activation function, a second GCN layer, a second Relu activation function, a node conversion module, an embedding matrix, and a dot product decoder topology, and each module has an input and output relationship in turn.

[0070] like Figure 3 As shown, this is the structure diagram of the generator, in which: Relu: activation function , GCNLayer is the GCN layer.

[0071] Dot product decoder topology: Use weighted binary cross entropy loss as the topology reconstruction loss between the generator output graph structure adjacency matrix and the contrast fusion ensemble model output graph structure adjacency matrix.

[0072] Node conversion: refers to the process of aggregating and updating information of nodes in the graph. In the node conversion operation, each node aggregates and updates information based on its own characteristics and the characteristics of its neighboring nodes to obtain a new node feature representation.

[0073] Embedding matrix: The feature matrix after stage transformation multiplied by the transpose of this matrix.

[0074] The GCN layer is the core component of the GCN model. Its function is to learn the representation of nodes from the adjacency matrix and node features of the graph. The calculation process of the GCN layer is shown in the following formula: , in, represents the input of the lth layer, represents the adjacency matrix, represents the activation function, represents the weight matrix, , is the identity matrix, for The degree matrix of is a diagonal matrix whose diagonal elements represent the degree of each node. The node degree can be expressed as That is, the number of edges connected to a vertex). In a directed graph, a vertex The degree of the vertex The sum of the out-degree and in-degree of The number of directed edges outgoing and incoming vertices The number of directed edges of the nodes is arranged in order to form a diagonal matrix .

[0075] The process of training the equivalent generator network is as follows: Random Gaussian noise is input into the equivalent generator network and the integrated model respectively to predict all the feature values ​​of the entity graph.

[0076] Use the discriminator to distinguish the prediction results of the equivalent generator network from those of the integrated model, backpropagate the difference to the equivalent generator, and iterate until the discriminator cannot distinguish the prediction results of the equivalent generator network from those of the integrated model.

[0077] S403: Input the digital twin graph and the DT model of the entities in the graph based on domain knowledge into the fused entity model to obtain a fused digital twin graph.

[0078] S5: Predict the physical characteristics of the unobservable entities based on the acquired fused digital twin graph, and store the fused digital twin graph in the database. When the entity graph changes, query the database to see if the same object exists. If so, add it to the newly constructed digital twin graph based on the entity graph relationship stored in the database. Only train the relationship between the unstored entity graph and other entity graphs of the newly constructed digital twin graph.

[0079] It should be noted that the most typical scenario for system-wide simulation is to deal with the state updates of interconnected entities. Usually, only the states of some entities are observable (i.e., their IoT sensor data are timely fed into the digital twin), while other entities are not, but also need to be updated (due to the state changes of other entities) in order to reflect the latest state of the entire system. In another case, the physical system undergoes structural changes, or the virtual twin is simulating structural changes, for example, entities collapse or are replaced by different types, new entities are added, or connections between entities are disconnected / replaced / added. In this case, the originally trained G2G transformation models of those affected entities can no longer be used for reasoning. To handle the first case, the proposed G2G transformation model is used to propagate the state changes of observable entities to the entire system graph until a new stable state is reached (i.e., convergence). Note that the G2G model is fundamentally different from existing GNNs, where the propagation rule is a simple decay rate (scalar), while the rules expressed in G2G can capture more complex relationships between real entities.

[0080] To handle the second case, new G2G models and new entity graphs are needed to replace the affected ones. This can be achieved by repeating the first step, but only for the affected areas, instead of rebuilding the entire system graph from scratch. In fact, this process can be accelerated further by creating a DTG database that holds a copy of all entity graphs and G2G models once built or (pre-)trained (which can be done offline). Therefore, when a new entity graph or G2G is needed, the GTG database is first queried to retrieve the same object (if it already exists), otherwise just train a new one. Call it a lazy training strategy. After replacing the G2G model, the second scenario becomes the first one, and just run the same process to update the entire DTG.

[0081] Example 2 like Figure 5 As shown, a product quality prediction method based on automatic digital twins is shown. This embodiment takes air conditioning production as an example, and includes: S1: Use the automatic digital twin generation method based on data-driven and graph learning as described in any one of claims 1-7 to establish a fused digital twin graph for air-conditioning production.

[0082] S2: Integrate the digital twin image The entity graphs are respectively input into the dimensionality reduction feature extraction module; the dimensionality reduction feature extraction module is composed of a first graph neural network convolution layer, a first graph neural network pooling layer, a second graph neural network convolution layer, and a second graph neural network pooling layer in sequence, and each module has an input and output relationship in sequence.

[0083] S3: Add the average pooled feature vectors of the first graph neural network pooling layer and the second graph neural network pooling layer to obtain global features of different scales , the pooled feature vectors obtained by passing all entity graphs through the second graph neural network pooling layer are combined into the reconstructed entity graph.

[0084] S4: The reconstructed entity graph is processed by the dimension reduction feature extraction module to obtain the average pooling feature vector of the first graph neural network pooling layer, the average pooling feature vector of the second graph neural network pooling layer, and the pooling feature vector of the second graph neural network pooling layer. , add the average pooling feature vector of the first graph neural network pooling layer and the average pooling feature vector of the second graph neural network pooling layer to obtain global features of different scales .

[0085] S5: All global features , And the pooled feature vector Input the neural network, and pass the output of the neural network output layer through the ReLU activation function layer to obtain the product quality prediction result. The three output quantities are: air conditioner life, air conditioner cooling capacity, and air conditioner failure rate.

[0086] The reconstructed entity graph combination method is as follows: an undirected line segment is generated between two entity graphs that have a relationship in the fused digital twin graph.

[0087] In the present invention, this method is applied to the automatic digital twin modeling of the air-conditioning outdoor unit production line, collects data from important production line nodes of the production line, automatically builds the relationship between each node and updates the twin model in real time, and simulates and predicts the output, cooling capacity / heating capacity, air volume, noise level, and service life of the air-conditioning outdoor unit. The work of the production line can be better adjusted, greatly improving production quality and efficiency. At the same time, this method has great versatility and expansion capabilities. This method is not limited to the application of air-conditioning production lines and can be easily applied to multiple fields.

[0088] The embodiment of the present disclosure also provides a product quality prediction device based on automatic digital twins, including a processor and a memory. Optionally, the device may also include a communication interface and a bus. Among them, the processor, the communication interface, and the memory can communicate with each other through the bus. The communication interface can be used for information transmission. The processor can call the logic instructions in the memory to execute the product quality prediction method based on automatic digital twins of the above embodiment.

[0089] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0090] The memory, as a computer-readable storage medium, can be used to store software programs and computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the present disclosure. The processor executes the function application and data processing by running the program instructions / modules stored in the memory, that is, realizing the product quality prediction method based on automatic digital twin in the above embodiment.

[0091] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory may include a high-speed random access memory and may also include a non-volatile memory.

[0092] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An automatic digital twin generation method based on data-driven and graph learning, characterized in that: include: Collect and preprocess the physical features of the physical entity, use all the features to perform significance tests on the correlation method with the highest confidence on average, and construct the entity graph; Associating a feature in the entity graph feature with a regression model on each other related feature determined by the entity graph structure, when the value of any feature in the entity graph changes, the regression model updates the value of the feature related to the changed feature; The entity graph is used as a node and a graph-to-graph conversion model is used to model the entity graph to construct a digital twin graph, wherein the directed arcs connecting the nodes in the digital twin graph represent the relationship between the entity graphs; Obtain the DT model based on domain knowledge of each entity in the entity graph, fuse the digital twin graph with the DT model based on domain knowledge of the entity in the graph, and construct a fused digital twin graph; The physical characteristics of unobservable entities are predicted based on the acquired fused digital twin graph, and the fused digital twin graph is stored in the database. When the entity graph changes, the database is queried whether the same object exists. If so, it is added to the newly constructed digital twin graph according to the entity graph relationship stored in the database. Only the relationship between the unstored entity graph and other entity graphs of the newly constructed digital twin graph is trained.

2. The automatic digital twin generation method based on data-driven and graph learning according to claim 1 is characterized in that: The specific method of using the entity graph as a node and using the graph-to-graph conversion model to model the entity graph and construct a digital twin graph is as follows: The entity graph at the tail end of the directed arc of the digital twin graph is used as the source entity graph, and the entity graph at the head end of the directed arc is used as the target entity graph, and the source entity graph and the target entity graph are adapted; Use graph convolutional neural network as encoder to encode feature vectors in source entity graph; After decoding by the dot product decoder, the reconstructed target entity graph is obtained; The graph convolutional neural network is iteratively trained using the loss function to make the reconstructed target entity graph close to the target entity graph.

3. The automatic digital twin generation method based on data-driven and graph learning according to claim 2 is characterized in that: The specific method of using the graph convolutional neural network as an encoder to encode the feature vector in the source entity graph is as follows: Construct an adjacency matrix from the source entity graph and the feature matrix , so that the adjacency matrix Capturing sufficiently correlated feature pairs, the adjacency matrix Each element in ,in, Representation Node To Node The correlation coefficient of represents the identity matrix; The adjacency matrix and the feature matrix Input graph convolutional neural network encoding and get the encoding result , the specific parameterized graph convolutional neural network output is as follows: ， in, represents the output of the last layer of graph convolutional neural network, Represents the weight matrix of the previous layer of graph convolutional neural network, Represents the number of layers of the graph convolutional neural network; The specific steps of decoding the dot product decoder to reconstruct the target entity graph are as follows: The encoding result Perform node conversion to obtain the potential feature matrix of the source entity graph ; The latent feature matrix Transpose to get the transposed feature matrix , and transform the latent feature matrix and the transposed feature matrix Input to the dot product decoder to obtain the adjacency matrix of the reconstructed target entity graph ; The specific calculation formula is as follows: ， in, Represents the element in row i and column j of the adjacency matrix of the reconstructed target entity graph, represents the identity matrix, represents the logistic sigmoid function, Expressed as a range , The number of features that represent the entity; The adjacency matrix of the target entity graph will be reconstructed Transformed into a reconstructed target entity graph; The loss function is as follows: ， ， ， in, Represents the adjacency matrix The sparsity of Represents the gTarget adjacency matrix i OK j Column elements, Represents the actual target entity graph node The characteristic vector of Represents the node in the target entity graph that is reconstructed The characteristic vector of is a hyperparameter whose value is greater than 0.

4. The automatic digital twin generation method based on data-driven and graph learning according to claim 2 is characterized in that: The specific methods of graphics adaptation are as follows: If the number of features in the source entity graph is less than the number of features in the target entity graph, pseudo features and random arc weights are added to the source entity graph, where each pseudo feature has 0 feature vector and has inward connections with other real features; If the number of features in the source entity graph is greater than the number of features in the target entity graph, strongly correlated features in the source entity graph are merged into super features represented by cluster centroids through K-means clustering, where the value of K is equal to the number of features in the target entity graph.

5. The automatic digital twin generation method based on data-driven and graph learning according to claim 1 is characterized in that: The specific method of integrating the digital twin graph with the DT model based on domain knowledge of the entities in the graph is as follows: Establish a comparative fusion integration model to obtain the model values ​​of the two entity graphs , if the model value , the connection relationship between the two entity graphs remains unchanged. If the model value , then delete the connection relationship between the two entity graphs; the comparison fusion integration model is as follows: ， in, , Represents a threshold variable, and its value range is greater than 0. Represents the final loss function value of the graph convolutional neural network, Indicates whether there is a connection relationship between two entity graphs in the DT model based on domain knowledge. Its value is 0 if there is no connection relationship, and 1 if there is a connection relationship. Create and train an equivalent generator network for each contrast fusion ensemble model to obtain a fusion entity model; The digital twin graph and the DT model based on domain knowledge of the entities in the graph are input into the fused entity model to obtain the fused digital twin graph.

6. The automatic digital twin generation method based on data-driven and graph learning according to claim 5 is characterized in that: The process of training the equivalent generator network is as follows: Random Gaussian noise is input into the equivalent generator network and the integrated model respectively to predict all feature values ​​of the entity graph; Use the discriminator to distinguish the prediction results of the equivalent generator network from those of the integrated model, backpropagate the difference to the equivalent generator, and iterate until the discriminator cannot distinguish the prediction results of the equivalent generator network from those of the integrated model.

7. The automatic digital twin generation method based on data-driven and graph learning according to claim 6 is characterized in that: The equivalent generator network includes a first GCN layer, a first Relu activation function, a second GCN layer, a second Relu activation function, a node conversion module, an embedding matrix, and a dot product decoder topology, and each module has an input and output relationship in turn.

8. A product quality prediction method based on automatic digital twin, characterized in that: include: Using the automatic digital twin generation method based on data-driven and graph learning as described in any one of claims 1 to 7 to establish a fused digital twin graph; Integrate the digital twin The entity graphs are respectively input into the dimension reduction feature extraction module; the dimension reduction feature extraction module is composed of a first graph neural network convolution layer, a first graph neural network pooling layer, a second graph neural network convolution layer, and a second graph neural network pooling layer in sequence, and each module is in an input and output relationship in sequence; Add the average pooled feature vectors of the first graph neural network pooling layer and the second graph neural network pooling layer to obtain global features of different scales , the pooled feature vectors obtained by passing all entity graphs through the second graph neural network pooling layer are combined into a reconstructed entity graph; The reconstructed entity graph is processed by the dimension reduction feature extraction module to obtain the average pooling feature vector of the first graph neural network pooling layer, the average pooling feature vector of the second graph neural network pooling layer, and the pooling feature vector of the second graph neural network pooling layer. , add the average pooling feature vector of the first graph neural network pooling layer and the average pooling feature vector of the second graph neural network pooling layer to obtain global features of different scales ; All global features , And the pooled feature vector Input the neural network, and then pass the output of the neural network output layer through the ReLU activation function layer to obtain the product quality prediction result.

9. The product quality prediction method based on automatic digital twin according to claim 8 is characterized in that: The reconstructed entity graph combination method is as follows: an undirected line segment is generated between two entity graphs that have a relationship in the fused digital twin graph.

10. A product quality prediction device based on automatic digital twin, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the product quality prediction method based on automatic digital twin as described in any one of claims 8-9 when running the program instructions.

Images