A product assembly process knowledge pushing method, system and device

Abstract

This invention discloses a method, system, and device for pushing product assembly process knowledge in the field of semantic recognition and knowledge graph construction. The method includes: constructing a knowledge graph based on acquired product assembly process information using a trained Bert-BiLSTM-CRF model and an LLM-KE model to obtain a first initial knowledge graph and a second initial knowledge graph; filtering the first and second initial knowledge graphs using a structural rationality judgment module (SRDM); evaluating the filtered first and second initial knowledge graphs using a trained LLM-KGE model; constructing a product assembly process knowledge graph (APKG-CP) based on the evaluation results; and pushing assembly process knowledge based on Bayesian network intelligent reasoning. This invention can complete knowledge pushing under different process design requirements, improving the efficiency of human-machine collaborative work.

Description

Technical Field

[0001] This invention relates to the field of semantic recognition and knowledge graph construction technology, and in particular to a method, system and device for pushing knowledge about product assembly processes. Background Technology

[0002] Complex products, such as satellites and aircraft, are characterized by complex structures, high manufacturing technology content, and complex development processes. In the development of complex products, assembly work accounts for 20% to 70% of the total workload, averaging 45%. This high proportion requires process design systems to strike a balance between automation efficiency and human ergonomics. With the advent of Industry 5.0, the goal of intelligent manufacturing has shifted from technology-driven production to knowledge-driven efficiency improvement through human-machine collaboration. In this paradigm, assembly, as the final stage of complex product development, not only focuses on technical implementation but also emphasizes the deep integration of expert knowledge, human experience, and artificial intelligence to improve operational efficiency. However, in the process of efficiently reusing knowledge from assembly process knowledge with complex semantic relationships, both the traditional Bidirectional Encoder Representation from Transformers (BERT) model and the more recent large language model (LLM) have their own shortcomings. How to combine LLM and BERT, utilizing their respective semantic recognition advantages for knowledge graph construction to solve the problem of knowledge incompleteness during knowledge delivery, is a very promising research direction. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a product assembly process knowledge push method, system and equipment that can complete knowledge push under different process design requirements with low hardware resources, output assembly process knowledge that is accurately matched with process requirements and can be efficiently reused, so as to improve the efficiency of human-machine collaborative operation.

[0004] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0005] In a first aspect, the present invention provides a method for pushing product assembly process knowledge, including:

[0006] Based on the obtained product assembly process information, a knowledge graph is constructed using the trained Bert-BiLSTM-CRF model to obtain the first initial knowledge graph, and a knowledge graph is constructed using the trained LLM-KE model to obtain the second initial knowledge graph.

[0007] The first and second initial knowledge graphs are filtered using the structural rationality judgment module SRDM to obtain the filtered first initial knowledge graph and the filtered second initial knowledge graph.

[0008] The trained LLM-KGE model is used to evaluate the first and second initial knowledge graphs after screening, and the product assembly process knowledge graph APKG-CP is constructed based on the evaluation results.

[0009] Based on the product assembly process knowledge graph APKG-CP, product assembly process knowledge is pushed out using Bayesian network intelligent reasoning.

[0010] Optionally, the training process for the Bert-BiLSTM-CRF model, LLM-KE model, and LLM-KGE model includes:

[0011] Based on the product assembly process information obtained, a training knowledge graph is constructed based on human experience;

[0012] The text in the product assembly process information is used as the input data for the model, and the training knowledge graph is used as the answer when evaluating the model to construct a training set;

[0013] Set the probability of each answer in the Bert-BiLSTM-CRF model and the distillation model LLM, input the training set into the Bert-BiLSTM-CRF model and the distillation model LLM, and train the model to obtain the trained Bert-BiLSTM-CRF model and LLM-KE model.

[0014] The training set is expanded by using explicit negative samples from a knowledge graph containing randomly generated false information, as well as erroneous samples generated by Bert-BiLSTM-CRF and LLM-KE during training, to obtain an expanded training set.

[0015] Set the response of the distillation model LLM to True or False, input the expanded training set into the distillation model LLM, train the model, and obtain the trained LLM-KGE model.

[0016] Optionally, the structural rationality judgment module SRDM is used to make the following preliminary error judgment on the structure of the knowledge graph:

[0017] ,

[0018] in, The answer represents the knowledge graph. This indicates an answer template. Used to determine the answer Does it conform to the answer template? , Indicates an answer Matching answer template , Indicates an answer Does not conform to the answer template; Represents an entity, Indicates a relationship. Represents text, Represents a set of relations. Used to determine entities in the answer Extracted from text and relationships Does it belong to a set of relations? , Represents the entities in the answer. Extracted from text And relationship Belongs to a set of relations ; Used to determine the answer Is it acceptable? Indicates an answer Acceptable. Indicates an answer Unacceptable.

[0019] Optionally, the step of using the trained LLM-KGE model to evaluate the first and second initial knowledge graphs after screening, and constructing a product assembly process knowledge graph APKG-CP based on the evaluation results, includes:

[0020] Calculate the confidence score of the Bert-BiLSTM-CRF model's response based on the probability distribution of each category output by the last Softmax layer of the Bert-BiLSTM-CRF model.

[0021] The confidence level of the response of the LLM-KE model was calculated using the multiple sampling method;

[0022] The confidence scores of the responses from the Bert-BiLSTM-CRF and LLM-KE models were corrected using the LLM-KGE model.

[0023] The answer confidence scores of the modified Bert-BiLSTM-CRF model and the LLM-KE model are weighted and fused to obtain the fused answer confidence score.

[0024] Based on the confidence level of the fused answers and a preset threshold, answers with confidence levels higher than the preset threshold are selected to construct a product assembly process knowledge graph APKG-CP.

[0025] Optionally, the calculation of the response confidence of the Bert-BiLSTM-CRF model is achieved using the following formula:

[0026] ,

[0027] in, This represents the confidence level of the response from the Bert-BiLSTM-CRF model. , Indicates the first The predicted probabilities of each category, Indicates the first The predicted probabilities of each category, Indicates the first The predicted probabilities of each category, Indicates the total number of categories;

[0028] The calculation of the response confidence of the LLM-KE model using the multisampling method is achieved through the following formula:

[0029] ,

[0030] in, This indicates the confidence level of the response from the LLM-KE model. This indicates the number of times the same result appears in the responses of the LLM-KE model. This represents the total number of responses in the LLM-KE model;

[0031] The correction of the response confidence of the Bert-BiLSTM-CRF model and the LLM-KE model using the LLM-KGE model is achieved by the following formula:

[0032] ,

[0033] in, This represents the confidence level of the response from the modified Bert-BiLSTM-CRF model. This represents the confidence correction coefficient of the Bert-BiLSTM-CRF model. This indicates the number of times the same result occurs in the response of the LLM-KGE model as in the response of the Bert-BiLSTM-CRF model. This represents the total number of responses from the Bert-BiLSTM-CRF model; This represents the confidence level of the response from the modified LLM-KE model. This represents the confidence correction coefficient for the LLM-KE model. This indicates the number of times the same result occurs in the LLM-KGE model's response as in the LLM-KE model's response. This represents the total number of responses in the LLM-KE model;

[0034] The weighted fusion of the response confidence scores of the modified Bert-BiLSTM-CRF model and the LLM-KE model is achieved through the following formula:

[0035] ,

[0036] in, This indicates the confidence level of the fused response. This represents the static weights of the Bert-BiLSTM-CRF model. This represents the static weights of the LLM-KE model.

[0037] Optionally, the product assembly process knowledge graph APKG-CP is a multigraph structure with node labels and directed edge labels, and its expression is as follows:

[0038] ,

[0039] in, Represents a set of entities. Represents a set of relations. Represents a collection of attributes. A set of triplets representing knowledge of product assembly processes.

[0040] Optionally, the method of pushing assembly process knowledge based on Bayesian network intelligent reasoning includes:

[0041] Based on the compilation requirements of the current process content in the assembly process information, extract the assembly elements involved in the current process.

[0042] Based on the assembly elements involved in the current process, the product assembly process knowledge graph APKG-CP is used for retrieval to obtain a knowledge graph retrieval network that meets the requirements, and a Bayesian network structure diagram is constructed based on this retrieval network.

[0043] Based on the Bayesian network structure diagram and the relevant data of each entity in the product assembly process knowledge graph APKG-CP, the Bayesian network parameters are learned through the EM algorithm to obtain the conditional probability table corresponding to the Bayesian network structure.

[0044] Based on the Bayesian network structure and its corresponding conditional probability table, the probability of assembly elements appearing under various assembly process requirements is calculated, and product assembly process knowledge is pushed out.

[0045] Optionally, the Bayesian network parameter learning via the EM algorithm includes:

[0046] Initialize Bayesian network parameters ;

[0047] Repeat the following steps until the Bayesian network parameters are obtained. convergence:

[0048] Using the first Bayesian network parameters in the next iteration Repair missing data in the product assembly process knowledge graph APKG-CP, and obtain the expected log-likelihood function based on the repaired data:

[0049] ,

[0050] in, This represents the total number of samples in the Product Assembly Process Knowledge Graph APKG-CP. Indicates the first The sample set in the next iteration Indicates the first In the nth iteration One sample, express The set of missing values ​​in the middle. This represents the expected log-likelihood function. Represents an entity, Indicates the parameters of the Bayesian network and the repaired sample Under the condition of missing variables Values The conditional probability, Indicates the parameters of the Bayesian network Under the conditions, the sample With missing variables Values The probability of them occurring simultaneously;

[0051] By optimizing the Bayesian network parameters by maximizing the expected log-likelihood function, the first... Bayesian network parameters in the next iteration :

[0052] ,

[0053] in, This represents the function that takes the maximum value.

[0054] Secondly, the present invention provides a product assembly process knowledge push system, comprising:

[0055] The initial knowledge graph construction module is used to: construct a knowledge graph using a trained Bert-BiLSTM-CRF model based on the obtained product assembly process information to obtain a first initial knowledge graph, and construct a knowledge graph using a trained LLM-KE model to obtain a second initial knowledge graph;

[0056] The filtering module is used to: filter the first initial knowledge graph and the second initial knowledge graph using the structural rationality judgment module SRDM to obtain the filtered first initial knowledge graph and the filtered second initial knowledge graph.

[0057] The product assembly process knowledge graph construction module is used to: evaluate the first and second initial knowledge graphs after screening using a trained LLM-KGE model, and construct the product assembly process knowledge graph APKG-CP based on the evaluation results;

[0058] The knowledge push module is used to push product assembly process knowledge based on the product assembly process knowledge graph APKG-CP and Bayesian network intelligent reasoning.

[0059] Thirdly, the present invention provides a computer device, comprising:

[0060] Memory, used to store computer instructions;

[0061] A processor for executing the computer instructions to implement the steps of the product assembly process knowledge push method described in any one of the first aspects.

[0062] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0063] 1. A dynamic weighted fusion algorithm based on LLM confidence evaluation using knowledge graphs can fully utilize the advantages of the Bert model and LLM: the Bert-BiLSTM-CRF model can achieve efficient task fine-tuning, accurate semantic similarity calculation, and short text classification under low resource conditions; LLM-KE has advantages such as generalized understanding and creative generation; by using the respective advantages of both to construct the knowledge graph to make up for their respective defects, a dynamically enhanced assembly process knowledge graph is obtained to solve the problem of knowledge incompleteness during knowledge push.

[0064] 2. The solution utilizes lightweight BERT models, distilled large language models, and Bayesian networks, reducing the demand for GPU performance. It can be deployed locally, minimizing the risk of privacy leaks when running in the cloud.

[0065] 3. Construct a Bayesian network model for assembly components, assembly features, and operations. Under different process design requirements, complete knowledge push and output assembly process knowledge that is accurately matched with process requirements and can be efficiently reused, thereby improving the efficiency of human-machine collaborative operation. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the overall process of pushing product assembly process knowledge according to an embodiment of the present invention.

[0067] Figure 2 A detailed flowchart of the product assembly process knowledge push method provided according to an embodiment of the present invention;

[0068] Figure 3 A flowchart illustrating the construction process of the product assembly process knowledge graph APKG-CP provided in an embodiment of the present invention;

[0069] Figure 4 A Bayesian network structure diagram provided according to an embodiment of the present invention;

[0070] Figure 5 This is a knowledge graph substructure diagram provided according to an embodiment of the present invention;

[0071] Figure 6 This refers to the directed acyclic graph corresponding to the knowledge graph substructure provided in the embodiments of the present invention.

[0072] Figure 7 This is a flowchart of the forward inference process of a Bayesian network provided according to an embodiment of the present invention;

[0073] Figure 8 A flowchart illustrating the reverse inference process of a Bayesian network according to an embodiment of the present invention;

[0074] Figure 9 A flowchart of bidirectional inference of a Bayesian network provided according to an embodiment of the present invention;

[0075] Figure 10 This is a substructure diagram of the knowledge graph of the upper compartment assembly of the payload compartment according to an embodiment of the present invention;

[0076] Figure 11 This is a Bayesian network structure diagram of the upper cargo compartment assembly according to an embodiment of the present invention. Detailed Implementation

[0077] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0078] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0079] Example 1:

[0080] This invention discloses a method for pushing product assembly process knowledge, with reference to... Figure 1 As shown, the specific steps include the following:

[0081] S1. Based on the obtained product assembly process information, a knowledge graph is constructed using the trained Bert-BiLSTM-CRF model to obtain the first initial knowledge graph, and a knowledge graph is constructed using the trained LLM-KE model to obtain the second initial knowledge graph.

[0082] S2, the first initial knowledge graph and the second initial knowledge graph are filtered using the structural rationality judgment module SRDM to obtain the filtered first initial knowledge graph and the filtered second initial knowledge graph;

[0083] S3, use the trained LLM-KGE model to evaluate the first and second initial knowledge graphs after screening, and construct the product assembly process knowledge graph APKG-CP based on the evaluation results;

[0084] S4. Based on the product assembly process knowledge graph APKG-CP, push product assembly process knowledge based on Bayesian network intelligent reasoning.

[0085] Specifically, in step S1, the specific textual information of the workshop assembly process is obtained, and a training knowledge graph is constructed based on human experience; the specific textual information of the workshop assembly process is used as the input data of the model, and the training knowledge graph is used as the answer when evaluating the model, thus constructing a training set, a validation set, and a test set;

[0086] The training set is input into Bert-BiLSTM-CRF (Bidirectional Encoder Representations from Transformers-Bidirectional Long Short-Term Memory-Conditional RandomField) for training, resulting in a trained Bert-BiLSTM-CRF. The training set is then input into the distillation model LLM for fine-tuning, resulting in the knowledge extraction LLM (LLM-KE model). The answers from the Bert-BiLSTM-CRF model and the distillation model LLM are the probabilities of each answer. Based on the trained Bert-BiLSTM-CRF and the fine-tuned LLM-KE model, a preliminary knowledge graph is constructed using complex product assembly process information that was not previously labeled.

[0087] The training set is expanded using explicit negative samples from a knowledge graph containing randomly generated false information, as well as erroneous samples generated during the training of Bert-BiLSTM-CRF and LLM-KE, to obtain an expanded training set. Another distilled model, LLM, is then fine-tuned using the expanded training set to obtain a knowledge graph evaluation model (LLM-KGE). The LLM-KGE model uses a simple binary classification method to set the answer result: True for correct answers and False for incorrect answers.

[0088] In step S2, the structural rationality judgment module SRDM is used to make the following preliminary error judgment on the structure of the knowledge graph:

[0089] ,

[0090] in, The answer represents the knowledge graph. This indicates an answer template. Used to determine the answer Does it conform to the answer template? , Indicates an answer Matching answer template , Indicates an answer Does not conform to the answer template; Represents an entity, Indicates a relationship. Represents text, Represents a set of relations. Used to determine entities in the answer Extracted from text and relationships Does it belong to a set of relations? , Represents the entities in the answer. Extracted from text And relationship Belongs to a set of relations ; Used to determine the answer Is it acceptable? Indicates an answer Acceptable. Indicates an answer Unacceptable.

[0091] In step S3, the trained LLM-KGE model is used to evaluate the first and second initial knowledge graphs after screening, and the product assembly process knowledge graph APKG-CP is constructed based on the evaluation results, including:

[0092] Calculate the confidence score of the Bert-BiLSTM-CRF model's response based on the probability distribution of each class output by the last Softmax layer:

[0093] ,

[0094] in, This represents the confidence level of the response from the Bert-BiLSTM-CRF model. , Indicates the first The predicted probabilities of each category, Indicates the first The predicted probabilities of each category, Indicates the first The predicted probabilities of each category, Indicates the total number of categories;

[0095] Calculate the confidence score of the LLM-KE model response using the multiple sampling method:

[0096] ,

[0097] in, This indicates the confidence level of the response from the LLM-KE model. This indicates the number of times the same result appears in the responses of the LLM-KE model. This represents the total number of responses in the LLM-KE model;

[0098] The confidence scores of the responses from the Bert-BiLSTM-CRF and LLM-KE models were corrected using the LLM-KGE model:

[0099] ,

[0100] in, This represents the confidence level of the response from the modified Bert-BiLSTM-CRF model. This represents the confidence correction coefficient of the Bert-BiLSTM-CRF model. This indicates the number of times the same result occurs in the response of the LLM-KGE model as in the response of the Bert-BiLSTM-CRF model. This represents the total number of responses from the Bert-BiLSTM-CRF model; This represents the confidence level of the response from the modified LLM-KE model. This represents the confidence correction coefficient for the LLM-KE model. This indicates the number of times the same result occurs in the LLM-KGE model's response as in the LLM-KE model's response. This represents the total number of responses in the LLM-KE model;

[0101] The confidence scores of the modified Bert-BiLSTM-CRF model and the LLM-KE model are weighted and fused to obtain the fused confidence scores:

[0102] ,

[0103] in, This indicates the confidence level of the fused response. This represents the static weights of the Bert-BiLSTM-CRF model. Represents the static weights of the LLM-KE model;

[0104] Based on the confidence level of the fused answers and a preset threshold, answers with a confidence level higher than the preset threshold are selected. A top-down approach, commonly used in domain knowledge graph construction, is employed. The assembly process card, the most core component of the assembly process document, is used as the assembly process data to construct APKG-CP. APKG-CP uses the assembly process elements in the assembly process card as nodes and the various semantic relationships between them as edges to build a structured semantic network, which facilitates the display, mining, and analysis of assembly process knowledge.

[0105] The product assembly process knowledge graph APKG-CP is a multigraph structure with node and directed edge labels, and its expression is as follows:

[0106] ,

[0107] in, Represents a set of entities. Represents a set of relations. Represents a collection of attributes. A set of triplets representing knowledge of product assembly processes.

[0108] In step S4, the process of pushing assembly process knowledge based on Bayesian network intelligent reasoning includes:

[0109] S4.1, Extraction of process planning requirements: When compiling different process contents, extract the assembly elements such as components, assembly features, and operations involved in this process based on the compilation requirements of the current process contents.

[0110] S4.2, Process Knowledge Graph Retrieval: Based on the assembly elements required for process compilation, a retrieval is performed in the complex product assembly process knowledge graph to obtain a knowledge graph retrieval network that meets the requirements, and a Bayesian network structure diagram is constructed based on this retrieval network.

[0111] S4.3, Bayesian network parameter learning: After obtaining the Bayesian network structure based on the knowledge graph, the parameters are learned through the EM algorithm based on the relevant data of each entity in the knowledge graph to obtain the conditional probability table corresponding to the Bayesian network structure.

[0112] S4.4, Bayesian Network Intelligent Reasoning: After obtaining the Bayesian network structure diagram and conditional probability table, the reasoning characteristics of the Bayesian network are used to calculate the probability of the assembly elements appearing under various process preparation requirements, thereby realizing the push of assembly process knowledge.

[0113] The Bayesian network model is constructed based on the assembly process knowledge graph, and the Bayesian network structure is designed based on the schema layer of the assembly process knowledge graph. The Bayesian network structure uses nodes from the knowledge graph as nodes, representing events where a certain entity appears; it uses edges from the knowledge graph as directed edges, representing relationships between two events. It is necessary to ensure that the direction of the directed edges conforms to the logical relationship between two assembly elements in the assembly process knowledge, and that there is exactly one directed line segment between two nodes. Based on the three assembly elements—assembly components, assembly features, and operations—that can represent assembly operations and the relationships between these elements in the schema layer of the knowledge graph, the structure of the knowledge graph (see reference) is... Figure 5 Transform into a directed acyclic graph (see reference) Figure 6 Thus, a Bayesian network structure based on the assembly process knowledge graph is obtained.

[0114] The Directed Acyclic Graph (DAG) and the Conditional Probability Table are the two components of a Bayesian network. The DAG can express the conditional independence and dependency relationships between nodes in the network, while the Conditional Probability Table is used to calculate the probability of each basic event. The Bayesian network parameter learning is based on the knowledge graph of complex product assembly processes, and the frequency of occurrence of each entity under different conditions is statistically analyzed as data for parameter learning.

[0115] In practice, the APKG-CP may suffer from some missing data. To address this, the Expectation Maximization (EM) algorithm is used to iterate through the training data, and then the maximum expectation is used to find an approximate maximum likelihood estimate. For training data D, the EM algorithm starts from... a certain initial value Start, initial value The result can be generated randomly; then, two steps are performed alternately: the E-step (Expectation Step) and the M-step (Maximization Step). The E-step is used to calculate the expectation, and the M-step is used to calculate the maximum likelihood estimate. Assume that t iterations have been performed, and the estimate is obtained... The (t+1)th iteration consists of the following E-steps and M-steps:

[0116] (1) Step E: Using the first Bayesian network parameters in the next iteration Repair missing data in the product assembly process knowledge graph APKG-CP, and obtain the expected log-likelihood function based on the repaired data:

[0117] ,

[0118] in, This represents the total number of samples in the Product Assembly Process Knowledge Graph APKG-CP. Indicates the first The sample set in the next iteration Indicates the first In the nth iteration One sample, express The set of missing values ​​in the middle. This represents the expected log-likelihood function. Represents an entity, Indicates the parameters of the Bayesian network and the repaired sample Under the condition of missing variables Values The conditional probability, Indicates the parameters of the Bayesian network Under the conditions, the sample With missing variables Values The probability of them occurring simultaneously;

[0119] (2) M-step: Optimize the Bayesian network parameters by maximizing the expected log-likelihood function to obtain the first step. Bayesian network parameters in the next iteration :

[0120] ,

[0121] in, This represents the function that takes the maximum value.

[0122] Repeat the above steps until the Bayesian network parameters are obtained. The convergence completes the parameter learning of the Bayesian network model.

[0123] Currently, most general recommendation systems are built based on user interest models and are widely used in e-commerce platforms, short video platforms, and other fields. These systems push similar items based on the user's current search information. In the process planning of complex products, the recommendation of specific procedures is context-based; that is, the recommendation system can push various types of assembly elements that are contextually related to the searched content, rather than similar assembly elements of the same type. This embodiment leverages the reasoning capabilities of Bayesian networks to implement knowledge push from an assembly process knowledge graph, enabling real-time delivery of assembly process guidance knowledge in the process planning scenario.

[0124] A Bayesian network's topology is a directed acyclic graph. Nodes represent variables or propositions (which can be observable variables, latent variables, unknown parameters, etc.), and directed edges between nodes describe causal relationships or conditional dependencies. Essentially, the existence of an arrow from one node to another implies a causal relationship or conditional independence between the corresponding random variables. (Reference) Figure 4 The diagram shows a simple Bayesian network structure. to In a Bayesian network model, the directed arrows between nodes indicate the relationships between them. For example, a node... Points to node This indicates that there is a dependency between the two, and the node With nodes There is no arrow connecting the two nodes to indicate that they are independent. It is obvious that there may be more than one directed edge between the nodes.

[0125] When two nodes are connected by a single arrow, it indicates that one node is the "cause" (parent) and the other is the "effect" (child). This connection generates a conditional probability value between the two nodes. The conditional probability reflects the strength of the connection between the two nodes. Figure 4 In this context, conditional probability is used to represent nodes. and The strength of connections between nodes. A node's state changes are heavily influenced by the state of its parent node. Therefore, when the parent node exhibits different state combinations, the child nodes will display different probability distributions, thus forming a conditional probability table for that node. Node The conditional probability table is shown in Table 1. Under normal circumstances, it is assumed that each node has only two states, "0" and "1". It is worth noting that, regardless of how many parent nodes a node has, given that the states of its parent nodes are fixed, the sum of the probabilities of its two states is always 1.

[0126] Table 1: Nodes The conditional probability table.

[0127]

[0128] When performing probability calculations, Bayes' theorem must be followed. Before the calculation, the prior probability and posterior probability must be determined. The formula for calculating the posterior probability is called the conditional probability formula, which leads to Bayes' theorem.

[0129] The reasoning pattern varies depending on the known and unknown quantities. (Refer to...) Figures 7-9 As shown, it can be divided into: forward reasoning, backward reasoning, and two-way reasoning; among them, , As the cause node, , It is both an intermediate node and a result node of the previous level. , , This is the result node. Assuming each node has only two states, "0" and "1", the reasoning process is as follows:

[0130] (1) Forward reasoning, also known as causal reasoning, is reasoning along the direction of a directed graph. Assume that in... In the event of occurrence, the reasoning node The probability of occurrence, such as Figure 7 As shown. The nodes can be obtained from the known conditions. , State probabilities, expressed as conditional probability values. , This is represented. Then, using the law of total probability, we further obtain the node... The probability of its occurrence is calculated using the following formula:

[0131] ,

[0132] in, and Indicates the node status. Represents a node Under the conditions, The probability, Represents a node Under the conditions, The probability, Represents a node Under the conditions, The probability, Represents a node Under the conditions, The probability of.

[0133] Taking this as an example, the conditional probability of each node can be calculated, thereby completing forward reasoning.

[0134] (2) Reverse reasoning, also known as evidence reasoning, is reasoning in reverse direction of the directed graph. Assume nodes... The occurrence of this event can be investigated by inverse reasoning using known prior probabilities and conditional probabilities to explore the reasons for the result, such as... Figure 8 As shown. By finding the node Under the conditions that it occurs, the cause node , The conditional probability of occurrence, i.e. , The values ​​are compared to generate a recommendation list, and then the appropriate recommendations are inferred. The cause of the occurrence. The calculation formula is as follows:

[0135] ,

[0136] in, Represents a node Under the condition, node The probability, Represents a node Under the condition, node The probability, Represents a node The probability, Represents a node The probability of;

[0137] In the formula, The prior probability is known. and Obtained from the following formula:

[0138] ,

[0139] ,

[0140] in, and Indicates the node status. express Under the conditions, The probability, Represents a node Under the conditions, The probability, express Under the conditions, The probability of;

[0141] At this point, we can find... The value of is used as an example to calculate . The value of is used to complete the reverse reasoning.

[0142] (3) Two-way reasoning is a combination of forward and backward reasoning, which can deduce the probability of intermediate nodes. Assume a node... and nodes Simultaneous occurrence, inferring the result of the intermediate node state based on known prior probabilities and conditional probabilities, such as Figure 9 As shown. Similar to forward and backward reasoning, the conditional probabilities of intermediate nodes are calculated using Bayes' theorem, and then a reasoning list is generated. Details are omitted here.

[0143] The three inference methods of the described Bayesian network can deduce the remaining required assembly elements based on different assembly elements in the requirements during process planning. Therefore, the advantages of Bayesian network inference can be combined with assembly process knowledge graphs and applied to process planning scenarios during assembly process design, providing process engineers with real-time guidance and suggestions on assembly processes that meet requirements, thereby improving the efficiency of process design.

[0144] Example 2:

[0145] This invention discloses a method for pushing product assembly process knowledge, with reference to... Figure 2 As shown, it includes the following steps:

[0146] Prepare process documents, train Bert-BiLSTM-CRF, fine-tune LLM-KE and LLM-KGE, and output the final knowledge graph APKG-CP based on dynamic weighted fusion of confidence, realizing dynamic incremental completion of the existing knowledge graph; in terms of knowledge push based on Bayesian network, first execute "process step preparation requirements" and "extract assembly elements: components, assembly features, operations" according to the requirements, and with the support of "knowledge graph retrieval" provided by APKG-CP, perform "constructing Bayesian network structure based on search assembly elements", adjust the "parameter learning" strategy, and select the optimal "execute parameter learning to obtain conditional probability table"; finally, realize "Bayesian network analysis" through "intelligent analysis" to obtain the "recommendation list".

[0147] like Figure 3 As shown, in this embodiment, the construction of the knowledge graph APKG-CP includes the following steps:

[0148] By leveraging expert experience, specific tags for relevant information in process documents and the relationships between those tags are identified, thus constructing a knowledge graph based on expert experience.

[0149] Based on expert experience knowledge graphs, Bert-BiLSTM-CRF is trained, and LLM-KE and LLM-KGE are fine-tuned;

[0150] The initial knowledge graph information output by Bert-BiLSTM-CRF and LLM-KE is filtered using the Knowledge Graph Structure Reasonableness Determination (SRDM) module, and the filtered information is then fed into LLM-KGE to generate confidence correction coefficients. and ;

[0151] Calculate the confidence level of each answer for Bert-BiLSTM-CRF and LLM-KE, and multiply it by the confidence level correction factor generated by LLM-KGE. and This yields a given answer and its corresponding confidence score.

[0152] Finally, by setting thresholds, results with high confidence are selected to construct a highly accurate knowledge graph for complex product assembly, APKG-CP.

[0153] like Figure 5 and Figure 6As shown in this embodiment, the Bayesian network structure construction process is illustrated, where different colors represent different types of entities. Nodes in the knowledge graph are used as nodes in the Bayesian network structure, representing events where a certain entity occurs. Edges in the knowledge graph are used as directed edges in the Bayesian network structure, representing the relationship between two events. It is important to note that the direction of the directed edges must conform to the logical relationship between two assembly elements in the assembly process knowledge, and it must be ensured that there is exactly one directed line segment between two nodes. Based on the three assembly elements in the knowledge graph pattern layer—assembly components, assembly features, and operations—that can represent assembly operations, and the relationships between these elements, the structure of the knowledge graph is transformed into a directed acyclic graph (DAG). Thus, the Bayesian network structure based on the assembly process knowledge graph is obtained.

[0154] like Figures 7-9 As shown, in this embodiment, the reasoning mode differs depending on the known and unknown quantities, and can be categorized into: forward reasoning, backward reasoning, and bidirectional reasoning. The figure shows a three-level Bayesian network. , As the cause node, , It is both an intermediate node and a result node of the previous level. , , This is the result node. Assuming each node has only two states, "0" and "1", the reasoning process is executed.

[0155] The present invention will be further described below with reference to specific embodiments.

[0156] This invention uses DeepSeek-R1-Distill-Qwen-1.5B as the basic LLM model. First, to demonstrate the effectiveness of the proposed APKG-CP construction method, preliminary experiments were conducted using the SemEval2010 Task 8 dataset. Then, a dataset was constructed using assembly process cards from an aerospace company in Shanghai (APC-AS) to construct the APKG-CP, verifying its practical feasibility. Finally, accurate delivery of process knowledge graphs was achieved based on a Bayesian network model. The dataset description is as follows:

[0157] SemEval2010 Task 8 defines 10 types of relationships between nouns, including: Cause-Effect, Instrument-Agency, Product-Producer, Content-Container, Entity-Origin, Entity-Destination, Component-Whole, Member-Collection, Message-Topic, and Other. Nouns not belonging to the first nine types of relationships are classified as "Other". The dataset contains 8000 training data entries and 2717 data entries for validation and testing. Each sample contains a complete sentence and identifies two entities whose semantic relationship within the sentence falls under one of the 10 relationship types.

[0158] The APC-AS dataset is constructed using assembly process cards as raw data. Through years of digitalization practice, it has accumulated massive amounts of assembly process data in the production of products with numerous models and diverse structures. By organizing the assembly process cards, extracting the names of processes or steps and recording detailed operation information, 1462 valid data entries were obtained, forming the original text set, totaling 98485 words.

[0159] (I) SemEval2010 Task 8 Experiment

[0160] Entity recognition and relation classification tasks were evaluated separately, and the F1 score was used to evaluate our model. Only completely correct predictions were counted; even if a word did not match the annotated entity, it was still classified as an error. Under strict metrics, BERT-BiLSTM-CRF, LLM-KE, and KGIC-LDM (A Knowledge Graph Intelligent Completion Framework Based on Lightweight Distillation Models) achieved F1 scores of 88.6%, 65.1%, and 91.4% for entity recognition, respectively. Table 2 compares several excellent algorithms for relation classification on SemEval2010 Task 8. It shows that KGIC-LDM achieves the best classification performance compared to traditional algorithms such as CR-CNN (Classifying Relations by Ranking with Convolutional Neural Networks), FCM (Factor-based Compositional Embedding Models), and R-BERT (Relation Classification BERT). While LLM-KE scores relatively low, KGIC-LDM improves performance by 3.7% and 9.4% compared to Bert-BiLSTM-CRF and LLM-KE, respectively. This demonstrates that the dynamic weighted fusion strategy of confidence effectively integrates the advantages of the latter two algorithms, proving the feasibility and effectiveness of the proposed dynamic incremental knowledge graph construction method.

[0161] Table 2 compares the relationship prediction results of different models on the SemEval2010 Task 8 dataset.

[0162]

[0163] (II) APC-AS Experiment

[0164] Experimental validation was obtained in SemEval2010 Task 8, and the same method was used to perform the dynamic incremental construction task of APKG-CP on the APC-AS dataset.

[0165] Step S1, build APKG-CP;

[0166] Step S2, Process Preparation Requirements Acquisition: To verify the feasibility of the Bayesian network construction method and the effectiveness of the intelligent push technology for assembly process knowledge graphs, a case study of the assembly process knowledge of a satellite's upper compartment is selected. First, a Bayesian model is constructed based on the knowledge graph of the satellite's upper compartment assembly process to verify the feasibility of the push technology process. Then, a random process in the upper compartment assembly is randomly selected to verify the accuracy of the pushed content.

[0167] During process preparation, process guidance is sometimes needed to develop new process content. This assembly process guidance is typically derived from existing assembly process cases and expressed as "perform XX operation on XX assembly feature of XX component." For example, "remeasure the parallelism accuracy between the upper surface of the payload compartment top plate and the reference platform," where the component is the "payload compartment top plate," the assembly feature is "parallelism," and the operation is "remeasure accuracy." These three assembly elements are entities in a knowledge graph. When proposing preparation requirements, one or two assembly elements are usually known. Assembly process guidance is formed by pushing the most likely third assembly element. Based on the known assembly elements, the most probable result is obtained using three inference modes of a Bayesian network, thus enabling the complete assembly process guidance to be pushed.

[0168] Step S3, Process Knowledge Graph Retrieval: In the actual process preparation, it is first necessary to determine the scope of the assembly process guidance. For example, when preparing process content related to the assembly of the upper cargo compartment of the payload bay, the scope of the process guidance should be selected as the upper cargo compartment assembly of the payload bay, and content related to the upper cargo compartment assembly of the payload bay should be retrieved from the assembly process knowledge graph. Then, component entities, assembly feature entities, and operation entities related to the upper cargo compartment assembly are selected from the assembly process knowledge graph, and a subgraph of the assembly process knowledge graph is generated for subsequent use, such as... Figure 10 As shown.

[0169] Step S4: Construct a Bayesian network model: Based on the structure of the assembly process knowledge graph subgraph, construct a Bayesian network structure, using component entities as parent nodes, assembly feature entities as first-level child nodes, and operation entities as second-level child nodes, constructing a directed acyclic graph with the relationships between entities as directed edges. Simultaneously, since the knowledge graph contains many entities with the same name, their essential operations and logic are identical, only belonging to different process contents. These entities are simplified when constructing the directed acyclic graph to reduce the computational load during parameter training and Bayesian inference. Furthermore, if the requirements can be further refined, for example, from the content related to "load compartment upper compartment assembly" to the content related to "load compartment upper bulkhead assembly," refining the retrieval requirements to the assembly element part, a more concise Bayesian network model with better inference performance will be obtained. The Bayesian network structure constructed based on the knowledge graph subgraph of "load compartment upper bulkhead assembly" as the process compilation requirement is as follows: Figure 11 As shown.

[0170] After constructing the directed acyclic graph (DAG), parameter learning is required to build a conditional probability table. An initial dataset is formed by combining the frequency of assembly element entities in the knowledge graph subgraph within the entire assembly process knowledge graph. Then, the EM algorithm is used to learn the parameters of the initial dataset, and the conditional probability table is obtained after the algorithm converges. Step S5, Bayesian network inference: After constructing the Bayesian network model, the state of unknown assembly elements under the current conditions is calculated based on real-time process planning requirements, and then sorted according to probability to form a recommendation list.

[0171] The recommended list under these conditions was calculated. Verified by process engineers, the relationship between the three assembly elements in actual assembly process planning is "C1, F1, O3". Operation O3, which has the highest probability in the recommended list, is the optimal assembly guidance scheme in actual process planning. Similarly, even after changing other inference modes, the Bayesian network can still push results that meet the actual process planning requirements, verifying the effectiveness of the intelligent push technology. After obtaining the optimal solution through Bayesian network inference, complete process information can be retrieved from the assembly process knowledge graph based on this assembly process guidance, i.e., using the assembly elements in this process guidance as search conditions, guiding process engineers in the process planning.

[0172] Therefore, by leveraging the similarity between the assembly process knowledge graph and the Bayesian network, and utilizing the reasoning pattern of the Bayesian network, the assembly process knowledge graph can push out the assembly process knowledge that best meets the needs of process planning for assembly process guidance, thereby increasing the utilization rate of assembly process knowledge and improving the quality and efficiency of process planning.

[0173] Example 3:

[0174] Based on the same inventive concept as Embodiment 1, this embodiment of the invention discloses a product assembly process knowledge push system, comprising:

[0175] The initial knowledge graph construction module is used to: construct a knowledge graph using a trained Bert-BiLSTM-CRF model based on the obtained product assembly process information to obtain a first initial knowledge graph, and construct a knowledge graph using a trained LLM-KE model to obtain a second initial knowledge graph;

[0176] The filtering module is used to: filter the first initial knowledge graph and the second initial knowledge graph using the structural rationality judgment module SRDM to obtain the filtered first initial knowledge graph and the filtered second initial knowledge graph.

[0177] The product assembly process knowledge graph construction module is used to: evaluate the first and second initial knowledge graphs after screening using a trained LLM-KGE model, and construct the product assembly process knowledge graph APKG-CP based on the evaluation results;

[0178] The knowledge push module is used to: push product assembly process knowledge based on the product assembly process knowledge graph APKG-CP and Bayesian network intelligent reasoning.

[0179] The specific functions of each module are described in the relevant content of Implementation Example 1, and will not be repeated here.

[0180] Example 4:

[0181] This embodiment provides a computer device, including:

[0182] Memory, used to store computer instructions;

[0183] A processor is used to execute the computer instructions to implement the steps of the product assembly process knowledge push method described in Embodiment 1.

[0184] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0185] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0186] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0187] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0188] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A product assembly process knowledge push method, characterized by, include: Based on the obtained product assembly process information, a knowledge graph is constructed using the trained Bert-BiLSTM-CRF model to obtain the first initial knowledge graph, and a knowledge graph is constructed using the trained LLM-KE model to obtain the second initial knowledge graph. The first and second initial knowledge graphs are filtered using the structural rationality judgment module SRDM to obtain the filtered first initial knowledge graph and the filtered second initial knowledge graph. The trained LLM-KGE model is used to evaluate the first and second initial knowledge graphs after screening, and the product assembly process knowledge graph APKG-CP is constructed based on the evaluation results. Based on the product assembly process knowledge graph APKG-CP, product assembly process knowledge is pushed out using Bayesian network intelligent reasoning. The training process for the Bert-BiLSTM-CRF model, LLM-KE model, and LLM-KGE model includes: Based on the product assembly process information obtained, a training knowledge graph is constructed based on human experience; The text in the product assembly process information is used as the input data for the model, and the training knowledge graph is used as the answer when evaluating the model to construct a training set; Set the probability of each answer in the Bert-BiLSTM-CRF model and the distillation model LLM, input the training set into the Bert-BiLSTM-CRF model and the distillation model LLM, and train the model to obtain the trained Bert-BiLSTM-CRF model and LLM-KE model. The training set is expanded by using explicit negative samples from a knowledge graph containing randomly generated false information, as well as erroneous samples generated by Bert-BiLSTM-CRF and LLM-KE during training, to obtain an expanded training set. Set the distillation model LLM to True or False, input the expanded training set into the distillation model LLM, train the model, and obtain the trained LLM-KGE model. The structure rationality judgment module SRDM is used to make the following preliminary error judgments on the structure of the knowledge graph: ， in, The answer represents the knowledge graph. This indicates an answer template. Used to determine the answer Does it conform to the answer template? , Indicates an answer Matching answer template , Indicates an answer Does not conform to the answer template; Represents an entity, Indicates a relationship. Represents text, Represents a set of relations. Used to determine entities in the answer Extracted from text and relationships Does it belong to a set of relations? , Represents the entities in the answer. Extracted from text And relationship Belongs to a set of relations ; Used to determine the answer Is it acceptable? Indicates an answer Acceptable. Indicates an answer Unacceptable.

2. The product assembly process knowledge push method according to claim 1, characterized in that, The process involves evaluating the first and second initial knowledge graphs after screening using a trained LLM-KGE model, and constructing a product assembly process knowledge graph APKG-CP based on the evaluation results, including: Calculate the confidence score of the Bert-BiLSTM-CRF model's response based on the probability distribution of each category output by the last Softmax layer of the Bert-BiLSTM-CRF model. The confidence level of the response of the LLM-KE model was calculated using the multiple sampling method; The confidence scores of the responses from the Bert-BiLSTM-CRF and LLM-KE models were corrected using the LLM-KGE model. The answer confidence scores of the modified Bert-BiLSTM-CRF model and the LLM-KE model are weighted and fused to obtain the fused answer confidence score. Based on the confidence level of the fused answers and a preset threshold, answers with confidence levels higher than the preset threshold are selected to construct a product assembly process knowledge graph APKG-CP.

3. The product assembly process knowledge push method according to claim 2, characterized in that, The confidence score of the Bert-BiLSTM-CRF model is calculated using the following formula: , in, This represents the confidence level of the response from the Bert-BiLSTM-CRF model. , Indicates the first The predicted probabilities of each category, Indicates the first The predicted probabilities of each category, Indicates the first The predicted probabilities of each category, Indicates the total number of categories; The calculation of the response confidence of the LLM-KE model using the multisampling method is achieved through the following formula: , in, This indicates the confidence level of the response from the LLM-KE model. This indicates the number of times the same result appears in the responses of the LLM-KE model. This represents the total number of responses in the LLM-KE model; The correction of the response confidence of the Bert-BiLSTM-CRF model and the LLM-KE model using the LLM-KGE model is achieved by the following formula: ， in, This represents the confidence level of the response from the modified Bert-BiLSTM-CRF model. This represents the confidence correction coefficient of the Bert-BiLSTM-CRF model. This indicates the number of times the same result occurs in the response of the LLM-KGE model as in the response of the Bert-BiLSTM-CRF model. This represents the total number of responses from the Bert-BiLSTM-CRF model; This represents the confidence level of the response from the modified LLM-KE model. This represents the confidence correction coefficient for the LLM-KE model. This indicates the number of times the same result occurs in the LLM-KGE model's response as in the LLM-KE model's response. This represents the total number of responses in the LLM-KE model; The weighted fusion of the response confidence scores of the modified Bert-BiLSTM-CRF model and the LLM-KE model is achieved through the following formula: ， in, This indicates the confidence level of the fused response. This represents the static weights of the Bert-BiLSTM-CRF model. This represents the static weights of the LLM-KE model.

4. The product assembly process knowledge dissemination method according to claim 1, characterized in that, The product assembly process knowledge graph APKG-CP is a multigraph structure with node and directed edge labels, and its expression is as follows: ， in, Represents a set of entities. Represents a set of relations. Represents a collection of attributes. A set of triplets representing knowledge of product assembly processes.

5. The product assembly process knowledge dissemination method according to claim 1, characterized in that, The method of pushing assembly process knowledge based on Bayesian network intelligent reasoning includes: Based on the compilation requirements of the current process content in the assembly process information, extract the assembly elements involved in the current process. Based on the assembly elements involved in the current process, the product assembly process knowledge graph APKG-CP is used for retrieval to obtain a knowledge graph retrieval network that meets the requirements, and a Bayesian network structure diagram is constructed based on this retrieval network. Based on the Bayesian network structure diagram and the relevant data of each entity in the product assembly process knowledge graph APKG-CP, the Bayesian network parameters are learned through the EM algorithm to obtain the conditional probability table corresponding to the Bayesian network structure. Based on the Bayesian network structure and its corresponding conditional probability table, the probability of assembly elements appearing under various assembly process requirements is calculated, and product assembly process knowledge is pushed out.

6. The product assembly process knowledge push method according to claim 5, characterized in that, The method of learning Bayesian network parameters using the EM algorithm includes: Initialize Bayesian network parameters ; Repeat the following steps until the Bayesian network parameters are obtained. convergence: Using the first Bayesian network parameters in the next iteration Repair missing data in the product assembly process knowledge graph APKG-CP, and obtain the expected log-likelihood function based on the repaired data: ， in, This represents the total number of samples in the Product Assembly Process Knowledge Graph APKG-CP. Indicates the first The sample set in the next iteration Indicates the first In the nth iteration One sample, express The set of missing values ​​in the middle. This represents the expected log-likelihood function. Represents an entity, Indicates the parameters of the Bayesian network and the repaired sample Under the condition of missing variables Values The conditional probability, Indicates the parameters of the Bayesian network Under the conditions, the sample With missing variables Values The probability of them occurring simultaneously; By optimizing the Bayesian network parameters by maximizing the expected log-likelihood function, the first... Bayesian network parameters in the next iteration : ， in, This represents the function that takes the maximum value.

7. A product assembly process knowledge push system, characterized in that, include: The initial knowledge graph construction module is used to: construct a knowledge graph using a trained Bert-BiLSTM-CRF model based on the obtained product assembly process information to obtain a first initial knowledge graph, and construct a knowledge graph using a trained LLM-KE model to obtain a second initial knowledge graph; The filtering module is used to: filter the first initial knowledge graph and the second initial knowledge graph using the structural rationality judgment module SRDM to obtain the filtered first initial knowledge graph and the filtered second initial knowledge graph. The product assembly process knowledge graph construction module is used to: evaluate the first and second initial knowledge graphs after screening using a trained LLM-KGE model, and construct the product assembly process knowledge graph APKG-CP based on the evaluation results; The knowledge push module is used to push product assembly process knowledge based on Bayesian network intelligent reasoning, according to the product assembly process knowledge graph APKG-CP. Based on the product assembly process knowledge graph APKG-CP, product assembly process knowledge is pushed out using Bayesian network intelligent reasoning. The training process for the Bert-BiLSTM-CRF model, LLM-KE model, and LLM-KGE model includes: Based on the product assembly process information obtained, a training knowledge graph is constructed based on human experience; The text in the product assembly process information is used as the input data for the model, and the training knowledge graph is used as the answer when evaluating the model to construct a training set; Set the probability of each answer in the Bert-BiLSTM-CRF model and the distillation model LLM, input the training set into the Bert-BiLSTM-CRF model and the distillation model LLM, and train the model to obtain the trained Bert-BiLSTM-CRF model and LLM-KE model. The training set is expanded by using explicit negative samples from a knowledge graph containing randomly generated false information, as well as erroneous samples generated by Bert-BiLSTM-CRF and LLM-KE during training, to obtain an expanded training set. Set the distillation model LLM to True or False, input the expanded training set into the distillation model LLM, train the model, and obtain the trained LLM-KGE model. The structure rationality judgment module SRDM is used to make the following preliminary error judgments on the structure of the knowledge graph: ， in, The answer represents the knowledge graph. This indicates an answer template. Used to determine the answer Does it conform to the answer template? , Indicates an answer Matching answer template , Indicates an answer Does not conform to the answer template; Represents an entity, Indicates a relationship. Represents text, Represents a set of relations. Used to determine entities in the answer Extracted from text and relationships Does it belong to a set of relations? , Represents the entities in the answer. Extracted from text And relationship Belongs to a set of relations ; Used to determine the answer Is it acceptable? Indicates an answer Acceptable. Indicates an answer Unacceptable.

8. A computer device, characterized in that, include: Memory, used to store computer instructions; A processor for executing the computer instructions to implement the steps of the product assembly process knowledge push method according to any one of claims 1-6.

Images