A method for constructing a knowledge base for the entire discrete manufacturing process with knowledge association and dynamic evolution

Abstract

The present invention discloses a method for constructing a knowledge base for the whole process of discrete manufacturing with knowledge association and dynamic evolution, including the following steps: collecting data of the whole process of discrete manufacturing and preprocessing; taking product quality control as the starting point, constructing an association map of initial stage parameters and final stage product quality, and learning the map; dividing a single knowledge map into different levels according to different stages of the whole process of discrete manufacturing, capturing the association information between different levels based on association features, and calculating the similarity or association strength between association features based on an association algorithm; and dynamically updating the knowledge base based on the evaluation of the contribution of graph nodes. The present invention takes product quality control as the starting point, introduces association features based on the association of knowledge graphs between layers to better capture the relationship between different levels; and realizes the dynamic update and evolution of the knowledge base according to the contribution evaluation of the nodes.

Description

Technical Field

[0001] The present invention relates to the fields of industrial big data and intelligent manufacturing, and in particular to a method for constructing a knowledge base for the entire discrete manufacturing process with knowledge association and dynamic evolution. Background Art

[0002] With the increasing complexity of the entire discrete manufacturing process and the sharp increase in the amount of information, the manufacturing industry is facing more complex knowledge management and quality control challenges. The traditional manufacturing knowledge base manages different process data such as process processing, part processing, component assembly and overall assembly separately, each forming an independent subsystem. However, the finished product quality of the product is not absolutely independent of its original processing technology documents, and there is a certain correlation between them. For example, in the production process, the quality parameters of the process processing stage may directly affect the process parameters of the part processing stage, and the quality data of the part processing stage may be fed back to the process processing stage, thereby affecting production decisions and process optimization. Therefore, the traditional manufacturing knowledge base often finds it difficult to meet the needs of the ever-changing manufacturing environment, and the knowledge association and update methods often have certain limitations. Summary of the invention

[0003] Purpose of the invention: The present invention aims to provide a knowledge system that is progressive in layers, and a method for constructing a knowledge base of discrete manufacturing full-process knowledge association and dynamic evolution that can be dynamically updated and evolved based on product quality control.

[0004] Technical solution: The method for constructing a knowledge base for the entire discrete manufacturing process knowledge association and dynamic evolution described in the present invention comprises the following steps:

[0005] (1) Collect and pre-process data from the entire discrete manufacturing process;

[0006] (2) Taking product quality control as the starting point, the features from the initial stage to the final stage of discrete manufacturing are associated and mapped based on the association relationship algorithm to obtain qualitative knowledge association; the complex relationship between the parameters in the initial stage of discrete manufacturing and the product quality in the final stage is modeled using a neural network, and a correlation map between the parameters in the initial stage and the product quality in the final stage is constructed, and the map is learned;

[0007] (3) Divide a single knowledge graph into different levels according to different stages of the entire discrete manufacturing process, capture the association information between different levels based on association features, and calculate the similarity or association strength between association features based on the association algorithm;

[0008] (4) Dynamically update the knowledge base based on graph node contribution evaluation.

[0009] Preferably, the method of mapping the features from the initial stage to the final stage of discrete manufacturing based on the association relationship algorithm to obtain qualitative knowledge association includes: using an association rule mining method to calculate the association strength of the features from the initial stage to the final stage, and the association strength is measured by support and confidence;

[0010] Support:

[0011] Confidence:

[0012] in, represents the feature matrix in the initial stage, Represents the final product quality indicator matrix, and count indicates that the data set also contains item sets and The number of samples, N represents the total number of samples; Support represents the item set and The frequency of simultaneous occurrence and the higher the support, the higher the frequency of the association rule in the data set and the stronger the correlation.

[0013] Preferably, a graph representation learning method is used, including a graph attention network GAT, a graph convolutional neural network GCN or its variants ChebNet or GraphSAGE to learn the graph.

[0014] Preferably, the associated features include:

[0015] For the level Nodes in , its associated features for:

[0016] in, It is The number of nodes in the level, It is Nodes in the hierarchy The characteristic vector of Indicates Hierarchy Node With Node The weight of the association between .

[0017] Preferably, the association algorithm comprises:

[0018] First, define the incidence matrix , where the elements Representation Node and The goal of the association matrix is ​​to maximize the sum of the similarities of associated features:

[0019]

[0020] in Represents the similarity calculation function;

[0021] The optimization problem is solved by the Hungarian algorithm, which is mathematically expressed as follows:

[0022]

[0023] in is a permutation, representing the matching of nodes; Indicates l +1 level and Tier k nodes that match the nodes.

[0024] By solving this optimization problem, we get the incidence matrix , which indicates the level and level The optimal connection between nodes.

[0025] Preferably, in step (4), the contribution evaluation is performed by analyzing each node in the knowledge base, quantitatively or qualitatively evaluating the importance of the node through the gradient of the node to the quality index and the statistical analysis of the node's historical data, and calculating the contribution score of each node to characterize its influence on product quality.

[0026] Preferably, in step (4), the dynamic update of the knowledge base includes knowledge addition, knowledge replacement and knowledge deletion.

[0027] Preferably, the knowledge enhancement includes:

[0028] Using autoencoders to extract new knowledge representations from data ; Add new knowledge representation to the knowledge graph and set node The original representation of , the update formula is:

[0029] in, is the learning rate, That is the new knowledge representation of the node.

[0030] Preferably, the knowledge replacement includes:

[0031] First, evaluate the node contribution: using the nodes in the current knowledge base, calculate the node contribution score as , the contribution score represents the node's contribution to the quality index The gradient of For Node The partial derivatives are calculated as: ,in Is a node Knowledge representation;

[0032] Next, set the replacement confidence threshold , if the contribution score of the current node , the current node representation is replaced by the new knowledge representation extracted using the autoencoder model.

[0033] Preferably, the knowledge deletion includes:

[0034] First, evaluate the node contribution: using the nodes in the current knowledge base, calculate the node contribution score as , the contribution score represents the node's contribution to the quality index The gradient of For Node The partial derivatives are calculated as: ,in Is a node Knowledge representation;

[0035] Next, set the deletion threshold , if the contribution score of the current node , then the node is considered to be deleted, and the nodes and their related edges that meet the deletion judgment conditions are deleted from the knowledge base.

[0036] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. The present invention includes establishing a single knowledge graph, multi-level knowledge graph association and knowledge update. In the method of establishing a single knowledge graph, an association relationship algorithm and a neural network are used to express the knowledge of each stage in the entire discrete manufacturing process through qualitative association relationships and quantitative neural networks; then, through the multi-level knowledge graph association method, the knowledge graph association between layers is defined, and association features are introduced to better capture the relationship between different levels; finally, according to the contribution score of the node and the deletion conditions, the dynamic evolution of the knowledge base is realized; 2. It is possible to establish relatively independent knowledge graphs for multiple stages of the entire discrete manufacturing process, so that the knowledge update of each level is independent of other levels, so as to better adapt to changes in the manufacturing environment; at the same time, a contribution evaluation and dynamic update mechanism are introduced to improve the adaptability and flexibility of the knowledge base; 3. A hierarchical knowledge system is established to cope with the complexity of the entire discrete manufacturing process. This method starts from the quality aspect, takes product quality control as the starting point, and combines qualitative and quantitative knowledge to construct a dynamically evolving knowledge base. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1It is a schematic diagram of the knowledge graph structure of the entire discrete manufacturing process of the present invention;

[0038] Figure 2 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0040] like Figure 2 As shown, the method for constructing a knowledge base for the entire discrete manufacturing process knowledge association and dynamic evolution described in the present invention includes the following steps:

[0041] (1) Data collection and feature extraction

[0042] Using technical means such as sensors and equipment interfaces, real-time data from the entire discrete manufacturing process is collected, including but not limited to key parameters such as the raw material procurement stage, process design stage, tooling design stage, production debugging stage, process processing stage, parts processing stage, component assembly stage, overall assembly stage, packaging and transportation stage. The collected data is constructed into a data set, and feature screening and extraction are performed.

[0043] (2) Establish a single knowledge graph: Establish a comprehensive knowledge graph covering the entire discrete manufacturing process. This knowledge graph can cover all stages from raw material procurement to final product delivery, providing comprehensive reference and guidance for the manufacturing process.

[0044] This embodiment takes the discrete manufacturing process including the process processing stage, parts processing stage, component assembly stage, and overall assembly stage as an example for detailed description. Figure 1 shown.

[0045] The specific steps include:

[0046] (2.1) Qualitative knowledge association: Taking product quality control as the starting point, the characteristics from the initial stage to the final stage of discrete manufacturing are associated and mapped based on the association relationship algorithm. In this embodiment, it refers to the association mapping of the process parameters in the process processing stage to the product quality characteristics in the overall assembly stage to establish a qualitative knowledge graph.

[0047] The association relationship algorithm uses the association rule mining method:

[0048] Assume that samples, each containing Process parameters and The process parameters are expressed as a matrix , the dimension is , where each row represents a sample (e.g., a workpiece or a process) and each column represents a feature (e.g., different dimensions of process parameters). Product quality indicators are represented as matrices , the dimension is , where each row represents a sample (e.g., a product or a production batch) and each column represents a product quality indicator (e.g., a different quality characteristic).

[0049] The correlation strength between process parameters and product quality characteristics is calculated, and the correlation strength is measured by support and confidence.

[0050] Support:

[0051] Confidence:

[0052] In association rule mining, count means that the item set is also included in the data set. and The number of samples, N represents the number of samples; Support represents the item set and The frequency of co-occurrence, that is, the item set is included in all transactions at the same time and The higher the support, the more frequently the association rule appears in the data set, and the stronger the association.

[0053] (2.2) Quantitative knowledge modeling: Model the complex relationship between process parameters and product quality through deep learning methods. This embodiment uses a recurrent neural network , to better capture temporal or spatial knowledge.

[0054] The process parameter sequence Input into the RNN model to get the output sequence .in Indicates The process parameters of the samples, Represents the corresponding product quality indicator. By building a complex relationship model between process parameters and product quality through deep learning methods, and then using the output and relationship information of these models to further build and optimize the knowledge graph, it helps to form a more complete, accurate and practical knowledge system, providing strong support for the optimization and quality control of the manufacturing process.

[0055] (2.3) Graph representation learning: Qualitative and quantitative knowledge is expressed in the form of a graph, with process parameters and quality indicators as nodes in the graph, and edges representing the relationship between them. Graph representation learning methods, including but not limited to graph attention networks (GAT), graph convolutional neural networks (GCN) and their variants ChebNet and GraphSAGE, are used to learn graphs in order to better utilize the relationship between nodes.

[0056] In this embodiment, the GCN method is used, and the update rules of GCN are as follows:

[0057]

[0058] in It is The representation matrix of the layer nodes, is the adjacency matrix of the graph, is the degree matrix, is the weight matrix, is the activation function.

[0059] Through multiple layers of GCN iteration, an updated representation of each node can be obtained. Through graph representation learning, the relationship between nodes can be better utilized, the representation of nodes can be learned, and the complex relationship between process parameters and product quality can be captured in the graph structure, so that a single knowledge graph can be established more effectively and valuable information can be extracted from it for production process optimization and quality control.

[0060] (3) Multi-level knowledge graph association method: decompose a single knowledge graph into multiple levels according to the different stages of the entire discrete manufacturing process, and establish associations between the levels.

[0061] (3.1) Hierarchical division: First, a single knowledge graph is divided into different hierarchies, each of which represents a stage in the entire discrete manufacturing process, such as process processing, parts processing, component assembly, and overall assembly. is the total number of levels, Indicates the index of each level.

[0062] (3.2) Definition of correlation features: In each level, correlation features are introduced to capture the correlation information between different nodes. Nodes in , its associated features for:

[0063] in, It is The number of nodes in the level, It is Nodes in the hierarchy The characteristic vector of Indicates Hierarchy Node With Node The weight of the association between .

[0064] (3.3) Association algorithm: In order to establish associations between different levels, an association algorithm is introduced to calculate the similarity or association strength between associated features. In this embodiment, a graph matching-based method is used to establish associations between levels by maximizing the similarity between associated features, specifically including:

[0065] First, define the incidence matrix , where the elements Representation Node and The goal of the association matrix is ​​to maximize the sum of the similarities of associated features:

[0066]

[0067] in Represents the similarity calculation function, you can choose cosine similarity or Euclidean distance;

[0068] The optimization problem is solved by the Hungarian algorithm, which is mathematically expressed as follows:

[0069]

[0070] in is a permutation, indicating the matching of nodes. l +1 layer nodes are reordered or allocated to determine l The nodes of the layer and l The best match between nodes in the +1 layer; Indicates Layer Tier k nodes that match the nodes.

[0071] By solving this optimization problem, we get the incidence matrix , which indicates the level and level The optimal connection between nodes.

[0072] By defining association features and association algorithms, the association of knowledge graphs at different levels is achieved, which helps to build a more complex and comprehensive knowledge system, thereby improving the accuracy and effectiveness of product quality control. Through this progressive knowledge system, various knowledge in the discrete manufacturing process can be better organized and managed, and comprehensive monitoring and refined management of the production process can be achieved, thereby improving production efficiency and product quality.

[0073] (4) Dynamically update the knowledge base based on contribution evaluation.

[0074] As manufacturing technology and demand continue to develop, the knowledge in the knowledge base also needs to be continuously updated and improved. The knowledge update method involves evaluating the timeliness and importance of knowledge, identifying outdated or no longer applicable knowledge, replacing or deleting it in a timely manner, and adding new knowledge in a timely manner. This step can maintain the practicality and accuracy of the knowledge base and ensure that it can always provide effective support for the manufacturing process. It mainly includes two parts: contribution evaluation and dynamic update mechanism:

[0075] Among them, contribution evaluation is to determine the contribution of each node in the knowledge base to the quality and efficiency of the entire discrete manufacturing process by analyzing them. It includes: node importance evaluation and contribution calculation, that is, quantitative or qualitative evaluation of the influence of nodes through the gradient of quality indicators and statistical analysis of historical data of nodes, and then calculate the contribution score of each node based on the importance evaluation of the node to characterize its influence on production quality and efficiency.

[0076] The dynamic update mechanism dynamically adjusts and updates the knowledge base based on the results of contribution evaluation to keep it synchronized with the manufacturing environment. It includes outdated node identification: identifying nodes with contribution scores below the threshold, and determining that they may be outdated or no longer applicable to the current production environment; deletion judgment: formulating a deletion strategy based on the node contribution score to determine which nodes should be deleted or replaced; knowledge base update: according to the deletion strategy, the knowledge base is updated to delete outdated or no longer needed nodes, and new nodes are added or existing nodes are updated to maintain the accuracy and practicality of the knowledge base.

[0077] Specifically, it can be divided into knowledge addition, knowledge replacement and knowledge deletion.

[0078] (4.1) Knowledge increase:

[0079] (4.1.1) Data collection and preprocessing: Collect experimental data, including process parameters , product quality indicators . Perform preprocessing steps such as cleaning and normalization on the data to ensure data quality.

[0080] (4.1.2) New knowledge extraction: Use machine learning models to extract potential features or information from preprocessed data, that is, new knowledge representation. In this embodiment, an autoencoder is used to extract potential features from the collected raw data instead of directly using the raw data itself. The extracted new knowledge can be used to update the existing knowledge base, add new associations between process parameters and quality indicators, and improve the practicality and accuracy of the knowledge base. Specifically, it includes:

[0081] (a) Preprocess the relevant process parameters and product quality data, including cleaning and normalization, to construct a training data set as input data;

[0082] (b) Construction of autoencoder: An autoencoder usually consists of two parts: an encoder and a decoder. The encoder is used to compress the input data into a low-dimensional representation (latent space) to capture the main features of the data; the decoder is used to reconstruct the low-dimensional representation into the original data to evaluate the accuracy of the reconstruction.

[0083] (c) Training the autoencoder: Use the reconstruction error as a loss function, such as the mean squared error (MSE), to measure the difference between the reconstructed data and the original input data; optimize the model parameters through the back-propagation algorithm to minimize the reconstruction error and train the autoencoder to learn the latent features of the data.

[0084] (d) New knowledge extraction: After training is completed, the encoder part can be used to extract new knowledge. By encoding new input data, new latent representations can be obtained, which capture the relationship between important process parameters and quality indicators.

[0085] The autoencoder outputs a new knowledge representation : .

[0086] (4.1.3) Add the new knowledge representation to the knowledge graph. Let the node The original representation of , the update formula is:

[0087] in, is the learning rate, That is the new knowledge representation of the node.

[0088] (4.2) Knowledge replacement

[0089] (4.2.1) Node contribution evaluation: Using the nodes in the current knowledge base, the node contribution score is calculated as , the contribution score represents the node's contribution to the quality index The gradient of For Node The partial derivatives are calculated as: ,in Is a node Knowledge representation;

[0090] (4.2.2) New knowledge extraction: For each node, use the same autoencoder model to extract new knowledge representation

[0091] (4.2.3) Replacement decision: Formulate a replacement strategy. Replace when ,otherwise , Represents the replacement confidence threshold.

[0092] The replacement confidence threshold is calculated by collecting historical usage data of all knowledge items, including confidence scores, and selecting an appropriate confidence level by analyzing the performance of knowledge in historical data. For example, a fixed threshold such as 0.80 or 80% may be set.

[0093] The new representation of knowledge after replacement is:

[0094]

[0095] like , then take , otherwise take .

[0096] (4.3) Knowledge deletion:

[0097] (4.3.1) Node contribution evaluation: Calculate each node Contribution score ;

[0098] (4.3.2) Deletion judgment: Set the deletion judgment threshold , if the node's contribution score , then the node is considered deleted and marked as 1, otherwise it is marked as 0.

[0099] The deletion judgment threshold is calculated by collecting the usage frequency, contribution score and historical effect data of all knowledge items in the knowledge base, and setting a suitable contribution threshold by calculating the contribution scores of all knowledge items. For example, the lower limit of 30% of the knowledge contribution scores in the historical data can be set as the deletion judgment threshold.

[0100] (4.3.3) Knowledge base update: According to the deletion judgment, the nodes and their related edges that meet the deletion judgment conditions are deleted from the knowledge base:

[0101]

[0102] in, , and They represent the nodes, edges and related features that need to be deleted respectively.

Claims

1. A method for constructing a knowledge base for the entire discrete manufacturing process knowledge association and dynamic evolution, characterized in that: The following steps are involved: (1) Collect and preprocess data from the entire discrete manufacturing process; (2) Taking product quality control as the starting point, the features from the initial stage to the final stage of discrete manufacturing are associated and mapped based on the association relationship algorithm to obtain qualitative knowledge association; the complex relationship between the parameters in the initial stage of discrete manufacturing and the product quality in the final stage is modeled using a neural network, and a correlation map between the parameters in the initial stage and the product quality in the final stage is constructed, and the map is learned; (3) Divide a single knowledge graph into different levels according to different stages of the entire discrete manufacturing process, capture the association information between different levels based on association features, and calculate the similarity or association strength between association features based on the association algorithm; (3.1) Hierarchical division: First, a single knowledge graph is divided into different levels. Each level represents a stage of the entire discrete manufacturing process. Let L be the total number of levels and l represent the index of each level. (3.2) Definition of correlation features: In each level, correlation features are introduced to capture the correlation information between different nodes; for node i in level l, its correlation feature F l,i for: Among them, N l is the number of nodes at level l, X l,j is the feature vector of node j in level l, α l,i,j represents the association weight between node i and node j at level l; (3.3) Association algorithm: In order to establish associations between different levels, an association algorithm is introduced to calculate the similarity or association strength between associated features, including: First, define the incidence matrix M l,l+1 , where the element M l,l+1 (i, j) represents the strength of association between nodes i and j; the goal of the association matrix is ​​to maximize the sum of the similarities of associated features: Where sim() represents the similarity calculation function; The optimization problem is solved by the Hungarian algorithm, which is mathematically expressed as follows: Where λ is a permutation, indicating the matching of nodes; λ(k) represents the node in the l+1th layer that matches the kth node in the lth layer; (4) Dynamically update the knowledge base based on graph node contribution evaluation.

2. The method for constructing a knowledge base for discrete manufacturing full-process knowledge association and dynamic evolution according to claim 1 is characterized in that: The method of mapping the features from the initial stage to the final stage of discrete manufacturing based on the association relationship algorithm to obtain qualitative knowledge association includes: using an association rule mining method to calculate the association strength of the features from the initial stage to the final stage, and the association strength is measured by support and confidence; Support: Confidence: Among them, X represents the feature matrix of the initial stage, Y represents the product quality indicator matrix of the final stage, count represents the number of samples that contain both item sets X and Y in the data set, and N represents the total number of samples; Support represents the frequency of item sets X and Y appearing at the same time. The higher the support, the higher the frequency of the association rule appearing in the data set and the stronger the correlation.

3. The method for constructing a knowledge base for discrete manufacturing full-process knowledge association and dynamic evolution according to claim 1 is characterized in that: Graph representation learning methods are used to learn graphs, including graph attention network GAT, graph convolutional neural network GCN or its variants ChebNet or GraphSAGE.

4. The method for constructing a knowledge base for discrete manufacturing full-process knowledge association and dynamic evolution according to claim 1 is characterized in that: In step (4), the contribution evaluation is performed by analyzing each node in the knowledge base, quantitatively or qualitatively evaluating the importance of the node through the gradient of the node to the quality indicator and the statistical analysis of the node's historical data, and calculating the contribution score of each node to characterize its influence on product quality.

5. The method for constructing a knowledge base for discrete manufacturing full-process knowledge association and dynamic evolution according to claim 1 is characterized in that: In step (4), the dynamic update of the knowledge base includes knowledge addition, knowledge replacement and knowledge deletion.

6. The method for constructing a knowledge base for discrete manufacturing full-process knowledge association and dynamic evolution according to claim 5 is characterized in that: The knowledge increase includes: Use autoencoders to extract new knowledge representation K from data i ; Add the new knowledge representation to the knowledge graph, and set the original representation of node i to The update formula is: Where η is the learning rate, That is the new knowledge representation of node i.

7. The method for constructing a knowledge base for discrete manufacturing full-process knowledge association and dynamic evolution according to claim 5 is characterized in that: The knowledge replacement includes: First, evaluate the node contribution: use the nodes in the current knowledge base to calculate the node contribution score C i , the contribution score represents the gradient of the node to the quality index Y, and is calculated by the partial derivative of the quality index Y with respect to the node i: Where V i is the representation of node i; Next, set the replacement confidence threshold T replace , if the contribution score of the current node is C i <T replace , the current node representation is replaced by the new knowledge representation extracted using the autoencoder model.

8. The method for constructing a knowledge base for discrete manufacturing full-process knowledge association and dynamic evolution according to claim 5 is characterized in that: The knowledge deletion includes: First, evaluate the node contribution: use the nodes in the current knowledge base to calculate the node contribution score C i , the contribution score represents the gradient of the node to the quality index Y, and is calculated by the partial derivative of the quality index Y with respect to the node i: Where V i is the representation of node i; Next, set the deletion judgment threshold T delete , if the contribution score of the current node is C i <T delete , then the node is considered to be deleted, and the nodes and their related edges that meet the deletion judgment conditions are deleted from the knowledge base.

Images