A modular method for deconstructing and reorganizing complex products based on graph neural network

Abstract

The present invention discloses a complex product deconstruction and reorganization modularization method based on graph neural network, including: collecting raw data; optimizing raw data into valid data; selecting R nodes with higher degrees from degree matrix d in descending order as important nodes, and determining the number k of product functional modules by calculating the cosine similarity of functional features between important nodes; deconstructing and reorganizing complex product connection relationships; analyzing each functional module, reallocating those that are mistakenly divided into other functional modules, and finding change propagation connection parts between functional modules. The present invention can consider the modularization processing of complex products from the dual characteristics of the explicit features of complex product parts themselves and the invisible features of adjacent parts implied by structural information, so as to achieve more efficient and accurate description of complex products.

Description

Technical Field

[0001] The present invention relates to the technical field of complex product development, and in particular to a complex product deconstruction and reorganization modularization method based on graph neural network. Background Art

[0002] In the process of complex product development, due to the influence of user needs and other uncertain factors, it is inevitable to encounter the problem of design changes. However, complex products have the characteristics of high coefficient dimensions, large number of units, and strong structural coupling between parts. When designers change one or more of the parts, it will inevitably cause changes to the related parts. However, due to the characteristics of complex products, it is difficult for designers to draw a clear change path, which may cause repeated changes, or incomplete changes will cause service performance to fail to meet customer requirements, greatly increasing time consumption and operational complexity. However, the ability of complex product changes to quickly adapt to the market plays a vital role in the core competitiveness of enterprises. Therefore, how to help designers respond to changes quickly and clearly depict the change path is of great significance to the development of enterprises.

[0003] At present, the redescription of complex products is mostly modularized. For example, the invention with patent number CN108427852A proposes a crane grab module decoupling method, which first determines the various parts of the crane grab, establishes the grab design structure matrix, performs clustering and reorganization, and obtains a module division scheme; according to the obtained module division scheme, the dependency between the parts is considered, the degree of mutual influence between them is determined, the grab module association matrix is ​​established, and the association dependency between the two modules is calculated; the grab modules are decoupled by reducing the association dependency between the modules. However, similar technologies only consider explicit functions or structural information when modularizing, and rarely consider the modularization of complex products from the dual characteristics of the explicit characteristics of the complex product parts themselves and the invisible characteristics of the adjacent parts implied by the structural information. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a modular method for deconstructing and reconstructing complex products based on graph neural networks, which can consider the modular processing of complex products from the dual characteristics of the explicit characteristics of complex product components themselves and the invisible characteristics of adjacent components implied by structural information, thereby achieving more efficient and accurate description of complex products.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The embodiment of the present invention proposes a modular method for deconstructing and reconstructing complex products based on a graph neural network, and the modular method includes the following steps:

[0007] S1: Collect original data, including service performance, component names, coupling relationships between components, and functional characteristics of components of complex products;

[0008] S2: Preprocess the raw data and optimize the preprocessed raw data into valid data:

[0009] Number the n nodes of the complex product, construct the complex network design structure matrix model A to describe the connection relationship of each component, design the function feature matrix X of each component to describe the functional characteristics of each component, and design the degree matrix d of each component to express how many connection relationships each component contains;

[0010] S3: From the degree matrix d, select R nodes with higher degrees in descending order as important nodes, and determine the number of product functional modules k by calculating the cosine similarity of the functional characteristics between the important nodes;

[0011] S4: Deconstruct and reorganize complex product connections:

[0012] The graph autoencoder is used to aggregate the features between the component and the surrounding parts as the implicit feature Z of the component itself; for the implicit feature Z finally obtained, the FCM algorithm is used to perform fuzzy clustering on it according to the number of modules k obtained in step S3, and the membership matrix of each component to each module is output:

[0013]

[0014] Select each component M i，j The maximum value is considered to be M i，j Corresponding modules;

[0015] S5, analyze each functional module, reallocate those that are mistakenly divided into other functional modules, and find the connection components that propagate changes between functional modules.

[0016] Furthermore, in step S2, the process of preprocessing the original data includes the following steps:

[0017] Perform integrity analysis on the original data, delete erroneous data, and repair or supplement abnormal data.

[0018] Furthermore, in step S2, the process of optimizing the preprocessed raw data into valid data includes the following steps:

[0019] S21: Number the n nodes of the complex product as A1, A2, A3...A n ;

[0020] S22: Construct a complex network design structure matrix model A for describing the connection relationship between components:

[0021]

[0022] In the formula, a i，j =1 means that part i and part j are connected, a i，j =0 means that there is no connection between part i and part j;

[0023] S23: Design a matrix of functional characteristics of each component to describe the functional characteristics of each component:

[0024]

[0025] In the formula, the element x in the matrix X is α，β represents the βth attribute value of the αth component, α=1, 2, ..., n, β=1, 2, ..., m;

[0026] S24: Design the degree matrix of each component To express how many connections each component has, the elements in the degree matrix d Representatives The component is connected to the αth component.

[0027] Furthermore, in step S3, the process of determining the number k of product function modules includes the following steps:

[0028] Select R nodes with higher degrees from the degree matrix d and calculate the cosine similarity of the functional features between each important node:

[0029]

[0030] In the formula, A (R)d represents a subset of components with higher degrees selected from the degree matrix, and Represents the features of the components in the selected higher-degree subset;

[0031] The important nodes with high similarity are regarded as a functional module class, and the number of functional modules k is finally determined.

[0032] Furthermore, in step S4, the process of using the graph autoencoder to aggregate the features between the component and the surrounding parts as the implicit features Z of the component itself includes the following steps:

[0033] Input the adjacency matrix A and the feature matrix X into the encoder GAT network in GAE to assign component weights:

[0034]

[0035] In the formula, and represents the implicit features of the components output after the feature matrix X in the above text is extracted by the graph neural network. Leaky ReLU is the neural network activation function. W is related to is the neural network parameter, exp represents the sotmax activation function;

[0036] And the initial form of the implicit features of the parts is extracted through the graph neural network:

[0037]

[0038] The final implicit features are obtained by jointly optimizing the decoder and encoder in GAE:

[0039]

[0040] In the formula, and x α，m Corresponding.

[0041] Furthermore, in step S5, the process of reallocating the wrongly divided into other functional modules includes the following steps:

[0042] Select the parts in each functional module that are not connected to other parts, calculate the cosine similarity between this part and the center of other module classes, and multiply it by the membership degree of this functional module:

[0043]

[0044] The module with the highest score is the one to which this part belongs.

[0045] Furthermore, in step S5, the process of finding the change propagation connection components between the functional modules includes the following steps:

[0046] S51: Select a node with a higher degree in any module, calculate the cosine similarity between this node and other modules, and if the cosine similarity with this module is the highest, use it as the inter-module change propagation node;

[0047] S52: Repeat step S52 to finally determine the changed connection nodes between all modules.

[0048] The beneficial effects of the present invention are:

[0049] The present invention can consider the modular processing of complex products from the dual characteristics of the explicit features of complex product parts themselves and the invisible features of adjacent parts implied by structural information, thereby achieving more efficient and accurate description of complex products. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of a modular method for deconstructing and reorganizing complex products based on graph neural networks according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The present invention will now be described in further detail with reference to the accompanying drawings.

[0052] It should be noted that the terms such as "upper", "lower", "left", "right", "front", "back", etc. cited in the invention are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments in their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0053] Figure 1 This is a flow chart of a complex product deconstruction and reorganization modularization method based on a graph neural network according to an embodiment of the present invention. Figure 1 , the method comprises the following steps:

[0054] Step S1: Collecting raw data

[0055] The original data includes information such as the service performance, component names, coupling relationships between components, and functional characteristics of components of complex products.

[0056] Step S2: Analyze and process the original data

[0057] During the data collection and transmission process, abnormalities, errors or losses may occur, resulting in reduced availability of the original data. If the original data is directly used to decouple and reconstruct complex products, the accuracy of complex product characterization will be reduced. The analysis and processing process of the original data is: integrity analysis of the original data, deletion of erroneous data, repair or supplementation of abnormal data, and finally optimization of the original data into valid data, thereby improving the availability of the data.

[0058] An important method for decoupling and reconstructing complex products is to modularize them. Each functional module in a complex product is centered on one or more important components, supplemented by other parts. Based on the above characteristics of complex products, we first number the n nodes of the complex product A = [A1, A2, A3...A n ], and then construct a complex network design structure matrix model To describe the connection relationship between each component, we will design the functional feature matrix of each component. Describe the functional characteristics of each component, and finally design the degree matrix of each component To express how many connections each component has.

[0059] Step S3: Determine the number of product function modules

[0060] We first select R nodes with higher degrees from the degree matrix d, and then calculate the cosine similarity of the functional features between each important node:

[0061]

[0062] The important nodes with high similarity are regarded as a functional module class, and the number of functional modules k is finally determined.

[0063] Step S4: Deconstruct and reconstruct complex product connection relationships

[0064] S41: Get implicit feature Z

[0065] The features of complex product parts are not only related to themselves, but also to the associated parts. In addition, the importance of other parts around the parts varies, and different weight coefficients need to be assigned. At this time, we use GAE (graph autoencoder) to aggregate the features between the parts and the surrounding parts as the implicit features of the parts themselves. We first input the adjacency matrix A and the feature matrix X into the encoder GAT network in GAE to assign weights to the parts:

[0066]

[0067] And initially extract the hidden features Then use the decoder and encoder in GAE to jointly optimize and obtain the final implicit features

[0068] S42: Use implicit feature Z to divide functional modules

[0069] For the final implicit feature Z, the number of modules K is obtained according to step S3. We use the FCM algorithm to perform fuzzy clustering and output the membership matrix of each component to each module. We select each component M i，j The maximum value is considered to be M i，j The corresponding module.

[0070] S43: Eliminate the misclassification of components into other functional modules

[0071] To prevent the problem of incorrectly dividing components into modules, we first select the parts in each functional module that are not connected to other parts, calculate the cosine similarity between this part and the center of the class of other modules, and multiply it by the membership degree of this functional module:

[0072]

[0073] The module with the highest score is the one to which this part belongs.

[0074] S44: Finding components that connect changes between functional modules

[0075] We select a point with higher degree in the module and calculate the cosine similarity between this node and other modules. If the cosine similarity with this module is the highest, it will be used as the change propagation node between modules. Repeat this step to finally determine the change connection nodes between all modules.

[0076] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A modular method for deconstructing and reorganizing complex products based on graph neural networks, characterized in that: The modular method comprises the following steps: S1: Collect original data, including service performance, component names, coupling relationships between components, and functional characteristics of components of complex products; S2: Preprocess the raw data and optimize the preprocessed raw data into valid data: Number the n nodes of the complex product, construct the complex network design structure matrix model A to describe the connection relationship of each component, design the function feature matrix X of each component to describe the functional characteristics of each component, and design the degree matrix d of each component to express how many connection relationships each component contains; S3: From the degree matrix d, select R nodes with higher degrees in descending order as important nodes, and determine the number of product functional modules k by calculating the cosine similarity of the functional characteristics between the important nodes; S4: Deconstruct and reorganize complex product connections: The graph autoencoder is used to aggregate the features between the component and the surrounding parts as the implicit feature Z of the component itself; for the implicit feature Z finally obtained, the FCM algorithm is used to perform fuzzy clustering on it according to the number of modules k obtained in step S3, and the membership matrix of each component to each module is output: Select each component M i，j The maximum value is considered to be M i，j Corresponding modules; S5, analyze each functional module, reallocate those that are mistakenly divided into other functional modules, and find the connection components that propagate changes between functional modules.

2. The complex product deconstruction and reorganization modularization method based on graph neural network according to claim 1 is characterized in that: In step S2, the process of preprocessing the original data includes the following steps: Perform integrity analysis on the original data, delete erroneous data, and repair or supplement abnormal data.

3. The complex product deconstruction and reorganization modularization method based on graph neural network according to claim 1 is characterized in that: In step S2, the process of optimizing the preprocessed raw data into valid data includes the following steps: S21: Number the n nodes of the complex product as A1, A2, A3...A n ; S22: Construct a complex network design structure matrix model A for describing the connection relationship between components: In the formula, a i，j =1 means that part i and part j are connected, a i，j =0 means that there is no connection between part i and part j; S23: Design a matrix of functional characteristics of each component to describe the functional characteristics of each component: In the formula, the element x in the matrix X is α，β represents the βth attribute value of the αth component, α=1, 2, ..., n, β=1, 2, ..., m; S24: Design the degree matrix of each component To express how many connections each component has, the elements in the degree matrix d Representatives The component is connected to the αth component.

4. The complex product deconstruction and reorganization modularization method based on graph neural network according to claim 3 is characterized in that: In step S3, the process of determining the number k of product function modules includes the following steps: Select R nodes with higher degrees from the degree matrix d and calculate the cosine similarity of the functional features between each important node: In the formula, A (R)d represents a subset of components with higher degrees selected from the degree matrix, and Represents the features of the components in the selected higher-degree subset; The important nodes with high similarity are regarded as a functional module class, and the number of functional modules k is finally determined.

5. The complex product deconstruction and reorganization modularization method based on graph neural network according to claim 3 is characterized in that: In step S4, the process of using the graph autoencoder to aggregate the features between the component and the surrounding parts as the implicit features Z of the component itself includes the following steps: Input the adjacency matrix A and the feature matrix X into the encoder GAT network in GAE to assign component weights: In the formula, and represents the implicit features of the components output after the feature matrix X in the above text is extracted by the graph neural network. Leaky ReLU is the neural network activation function. W is related to is the neural network parameter, exp represents the softmax activation function; And the initial form of the implicit features of the parts is extracted through the graph neural network: The final implicit features are obtained by jointly optimizing the decoder and encoder in GAE: In the formula, and x a，m Corresponding.

6. The complex product deconstruction and reorganization modularization method based on graph neural network according to claim 4 is characterized in that: In step S5, the process of reallocating the wrongly divided into other functional modules includes the following steps: Select the parts in each functional module that are not connected to other parts, calculate the cosine similarity between this part and the center of other module classes, and multiply it by the membership degree of this functional module: The module with the highest score is the one to which this part belongs.

7. The complex product deconstruction and reorganization modularization method based on graph neural network according to claim 3 is characterized in that: In step S5, the process of finding the change propagation connection components between functional modules includes the following steps: S51: Select a node with a higher degree in any module, calculate the cosine similarity between this node and other modules, and if the cosine similarity with this module is the highest, use it as the inter-module change propagation node; S52: Repeat step S52 to finally determine the changed connection nodes between all modules.

Images