Industrial product surface defect tracing method, system and device

Abstract

The invention discloses an industrial product surface defect traceability method, system and device, and relates to the technical field of industrial product surface defect traceability, and the method comprises the steps: collecting actual defect time sequence data of a product and supplementary data generated by a digital twin system, and building a multi-view topological expression; extracting consistency characteristics of each view to obtain space-time integration characteristics of defect evolution; performing instance mining according to the multi-view topological expression and the space-time integration characteristics to establish a surface defect quality analysis knowledge graph; defining type vectors of head entities and tail entities for various relationships in the knowledge graph, and generating initial semantic representations of newly added entities; constructing a knowledge graph evolution process of dynamic graph updating to perform sub-graph level updating, and performing graph enhancement by generating soft facts and confidence thereof; a reinforcement learning recommendation framework is constructed, and product surface defect traceability is realized; according to the method, the reasoning capability of the model under the sparse atlas is improved, and the reliability and interpretability of a reasoning path are maintained.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial product quality analysis and defect tracing, and in particular to a method, system and device for tracing the surface defects of industrial products. Background Art

[0002] In modern manufacturing systems, the detection, tracking and quality assessment of surface defects are crucial to ensuring product performance.

[0003] The following problems are common in surface defect data in existing industrial scenarios: (1) Uneven distribution of collected data: Due to factors such as equipment layout and complexity of working conditions, defect data has problems such as sparse distribution, short time span, and poor availability; (2) Difficulty in correlation modeling: The relationship between defect status, process parameters, quality factors and final product performance is complex and has strong nonlinear and multi-stage influence characteristics; (3) Complex evolution process: Defects are affected by multiple factors such as temperature, load, and stress during the manufacturing cycle, and their status evolves dynamically over time, making it difficult to model and track.

[0004] However, traditional quality assessment methods, such as expert rule-based statistical analysis, regression prediction, or neural network models, can often only handle static, single scenarios and lack the ability to semantically express, trace the source, and dynamically update defect instances. Summary of the Invention

[0005] In view of the shortcomings of existing technologies that can only handle static and single scenarios and lack the ability to semantically express, trace and infer defect instances and dynamically update them, the present invention proposes a method, system and device for tracing the surface defects of industrial products, thereby solving the problems existing in the existing technologies.

[0006] A method for tracing the source of surface defects of industrial products comprises the following steps: Collect real-time defect time series data of industrial products and supplementary data generated by the digital twin system; and establish a multi-view topological representation including physical view, process parameter view and quality result view based on the real-time defect time series data and supplementary data; Extract the consistent features and private features between the structural features of each view, model the evolution process of the defect instance features of the consistent features and private features in the time dimension, and obtain the spatiotemporal integrated features; Based on multi-view topological representation and spatiotemporal integration features, case mining is performed to establish a knowledge graph for industrial product surface defect quality analysis. Type features of head and tail entities are set for each type of relationship between entities in the knowledge graph. The initial semantic representation of the newly added entities is generated by aggregating the type features of the adjacency relationships of the newly added entities. Based on the initial semantic representation of the newly added entities, a local subgraph is constructed for the newly added entities and their facts. The local subgraph is updated through a graph neural network to mine the path rules of the local subgraph. Based on the path rules, soft facts of the knowledge graph are generated. By combining the update process of local subgraphs and soft facts, a reinforcement learning recommendation framework is constructed; the surface defect traceability of industrial products is achieved through the reinforcement learning recommendation framework.

[0007] Furthermore, the extraction of consistent features and private features between the structural features of each view, modeling the defect instance feature evolution process of the consistent features and private features in the time dimension, and obtaining spatiotemporal integrated features specifically includes the following steps: Graph convolutional neural network (GCN) or attention network is used to extract structural features of physical view, process parameter view and quality result view respectively; The maximum cross-correlation strategy is used to align the structural features of each view and extract consistent features and private features; A recursive neural network is used to model the evolution of defect instance features in the time dimension for consistent features and private features, and obtain spatiotemporal integrated features.

[0008] Furthermore, a composite attention mechanism constructed by combining graph neural network with semantic similarity and rule confidence is used to aggregate the type features of the newly added entity adjacency relationships.

[0009] Furthermore, the updating of the local subgraph by the graph neural network specifically includes updating the embedding of entities and entity relationships, which are respectively expressed as: in 、 、 , e Represents the variable u is the embedding of the entity, r Represents the variable z is the embedding of the relation, l =0 indicates the embedding of the entity initially selected when performing a partial update of the knowledge graph. Represents the entity to be updated e connected entities; are three learnable weight matrices for entity embedding updates; is a learnable weight matrix for relation embedding update; the conversion function Accepts relation entities as input and aggregates information from adjacent links. Perform element-wise product, expressed as .

[0010] Furthermore, the generation of soft facts of the knowledge graph according to the path rules specifically includes the following steps: For the knowledge graph, successful paths are accumulated through reinforcement learning methods and formalized into path rules; the successful paths are paths that can infer the correct target entity; Utilize path rules to identify potential inferred facts in the knowledge graph and generate soft facts.

[0011] Furthermore, the subgraph update and soft facts are combined to construct a reinforcement learning recommendation framework, and the surface defect traceability of industrial products is achieved through the reinforcement learning recommendation framework, which specifically includes the following steps: The Markov decision process is represented by a quadruple; the quadruple includes state, action, transition and reward; specifically includes: and entities move The entity where the step is located Define each state , is the set of all states, r q For the target query relationship; for the state , defining its action space as entity The union of adjacent edges and self-loop edges: ,in, for The inverse relationship, Represents the entity to be updated e connected entities, r express e and The relationship between Move to the next state , define the transition between states; after reaching the maximum number of moves, stop at state The final entity Get rewarded ; A reinforcement learning strategy is constructed based on the state of the current entity, target relationship, historical path, and actions including edges and their confidence in the knowledge graph, as well as rewards consisting of target matching accuracy and path semantic consistency. Starting from the defect node through a reinforcement learning strategy, a multi-hop path search is performed in the knowledge graph to obtain an inference path including the quality factor prediction results and the path formed; Surface defect traceability of industrial products can be achieved based on inference paths.

[0012] The present invention also includes a system for tracing the surface defects of industrial products, comprising: The acquisition module is used to collect real-time defect time series data of industrial products and supplementary data generated by the digital twin system; and based on the real-time defect time series data and supplementary data, it establishes a multi-view topological representation including physical view, process parameter view and quality result view; The feature extraction module is used to extract the consistent features and private features between the structural features of each view, and to model the evolution process of defect instance features in the time dimension for the consistent features and private features to obtain spatiotemporal integrated features. The knowledge graph establishment module is used to conduct case mining based on multi-view topological representation and spatiotemporal integration features, establish a knowledge graph for industrial product surface defect quality analysis, and set type features for head and tail entities for each type of relationship between entities in the knowledge graph. The module generates an initial semantic representation of the newly added entities by aggregating the type features of the adjacency relationships of the newly added entities. Based on the initial semantic representation of the newly added entities, a local subgraph is constructed for the newly added entities and their facts. The local subgraph is updated through a graph neural network to mine the path rules of the local subgraph. Based on the path rules, soft facts of the knowledge graph are generated. The reasoning module is used to build a reinforcement learning recommendation framework by combining the update process of the local subgraph and soft facts; the reinforcement learning recommendation framework is used to realize the traceability of surface defects of industrial products.

[0013] The present invention also includes a computer device for tracing the surface defects of industrial products, including: a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the industrial product surface defect tracing method are implemented.

[0014] The present invention also includes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the industrial product surface defect tracing method.

[0015] The present invention provides a method for tracing the surface defects of industrial products, which has the following beneficial effects: The present invention enhances the quality analysis capability by integrating the supplementary data generated by the digital twin system; uses the graph structure and multi-view fusion features to express the complex, high-dimensional semantic relationship between defects and quality factors; realizes the dynamic evolution representation of defects by constructing a knowledge graph evolution process that supports dynamic graph updates, making up for the lack of semantic expression, traceability reasoning and dynamic update capabilities of defect instances in traditional methods; introduces soft facts into the knowledge graph by mining path rules, and uses soft facts to complete the implicit relationships in the graph, thereby improving the structural integrity and information density of the knowledge graph and effectively expanding the connectivity and reasoning space of the knowledge graph; this method not only improves the reasoning ability of the model under sparse graphs, but also maintains the reliability and interpretability of the reasoning path. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of a method for tracing the surface defects of industrial products based on a dynamically updated knowledge graph in an embodiment of the present invention; Figure 2 Schematic diagram of a knowledge graph feature extraction method supporting autonomous evolution in an embodiment of the present invention; Figure 3 This is a schematic diagram of a knowledge graph case composed of defect entities in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0018] This paper proposes a complex surface defect quality analysis method based on a dynamically updated knowledge graph, such as Figure 1 As shown, the specific steps include: S1. Multi-source data fusion and graph construction: Collect actual defect time series data and supplementary data generated by the digital twin system, fuse them to generate an initial set of triples containing defect instances, quality factors, and process parameters, and construct a multi-view topological representation of entities and their relationships through graph learning. Specifically, it includes:

[0019] Collect real time series data from surface defect detection systems and use digital twin systems to generate diverse simulated defect data.

[0020] Extract features such as defect physical morphology, process parameters, quality indicators, etc. to form a multi-view representation.

[0021] Construct multi-view topological representations of entities (defects, processes, quality) and relationships (such as "trigger," "affect," and "cause").

[0022] S2. Multi-view feature extraction and spatiotemporal fusion modeling: Extract the spatiotemporal integration features of multi-source data, design a multi-view neural network to extract the discriminative features under each view, and extract the consistent features through the maximum cross-correlation strategy. Further, a recursive neural network is used to encode the time series data to obtain a spatiotemporal fusion representation of the defect evolution. Specifically, it includes:

[0023] Graph convolutional neural networks (GCNs) or attention networks are used to extract structural features from each view. Multi-view feature extraction includes the defect physical view, process parameter view, and quality result view, corresponding to data in different feature domains. Feature encoding is performed using graph convolutional networks or self-attention networks.

[0024] A maximum cross-correlation strategy is designed to align the feature spaces of each view, extract shared features and retain private features.

[0025] Use recurrent neural networks (such as LSTM and GRU) to model the evolution of defect instance features in the time dimension to obtain a spatiotemporal integrated representation.

[0026] S3, type-constrained entity encoding mechanism. Relationship extraction is performed based on the multi-view topological expression constructed in S1, and instance mining is performed on the full-cycle spatiotemporal integration features extracted in S2 to establish a product surface defect quality analysis knowledge graph; a type-constrained encoding mechanism is introduced to define the type vectors of the head entity and the tail entity for each type of relationship in the knowledge graph, and the initial embedding representation of the new node is constructed based on its association relationship; a composite attention mechanism constructed by combining graph neural networks with semantic similarity and rule confidence is used to perform weighted aggregation of node neighbor information to improve the representation ability of entities and relationships in the graph, where the type vector is a learnable parameter that describes the entity types at the input and output ends of the relationship, respectively, and is used in the initial entity encoding to replace random initialization, thereby improving the accuracy of the semantic representation of new entities; when aggregating adjacent node information, the composite attention mechanism jointly calculates the attention weight based on the semantic similarity between the current query relationship and the adjacent relationship and its co-occurrence confidence in the mined path rules. Specifically including:

[0027] For each type of relationship, a learnable type embedding of the head entity and the tail entity is set to express the semantic role of the entity in the relationship.

[0028] For newly added entities (such as new defect instances), they are aggregated through the type features of their adjacency relationships to generate their initial semantic representations, avoiding random initialization.

[0029] In order to improve the accuracy of the initial vector of the newly added entity, the present invention adopts the following method for vectorization: is the entity vector initialized by bert instead of random initialization vector, The aggregated features are obtained by aggregating the features of entities connected to the entity. The two are concatenated and passed through an MLP module to obtain the final initialization vector.

[0030] S4. Dynamic graph evolution and enhancement mechanism: Build a knowledge graph evolution module that supports dynamic graph updates, perform subgraph-level updates on newly added defect instances and their associated information, generate soft facts through path rule mining, and integrate their confidence to enhance the graph. Specifically, this includes:

[0031] A local subgraph is constructed for the newly added entity and its facts, and the local area is quickly updated through the graph neural network. Facts are some descriptions about the newly added entity.

[0032] Mining high-confidence path rules in the graph (e.g., "defect A → stress concentration → performance degradation") generates soft facts. Path rule mining relies on a reinforcement learning agent to perform multi-hop path reasoning in the knowledge graph, inferring possible relationships or links between new entities.

[0033] Use soft facts to complete the implicit relationships in the graph and improve the structural integrity and information density of the knowledge graph.

[0034] Specifically, during the update of a local subgraph of the knowledge graph, the embedding update method of entities and relationships is as follows: Different from the existing relational graph neural network that randomly initializes the base embedding of each entity for training, this paper introduces an adaptive relationship aggregation layer to learn the base embedding of each entity by separately aggregating its link relationships.

[0035] for Entities in , generating its basic embedding as follows: in , . are two learnable weight matrices. It's a relationship The learnable embedding is initialized using the Xavier method and as an activation function.

[0036] in , , . are three learnable weight matrices for entity embedding updates, subscripted as e Indicates that the variable is the embedding of the entity, the subscript is r Indicates that the variable is the embedding of the relationship. The superscript 0 refers to the embedding of the entity selected when performing a local update of the knowledge graph, that is, the superscript l =0. Is a learnable weight matrix for relation embedding update. The conversion function It accepts relation entities as input and aggregates information from adjacent links. Perform element-wise product, expressed as . ; Represents the entity to be updated e connected entities, r express e and The relationship between u It refers to the embedded vector, which is the data that the network model can process. Neural networks cannot process direct text and must be represented as vectors.

[0037] Soft Fact Enhancement: In knowledge graphs, newly added entities often face the problem of sparse connections (i.e., few associated facts), making path reasoning difficult or ineffective. To alleviate this problem, this paper introduces soft fact enhancement. By mining path rules and generating candidate facts with confidence, it effectively expands the connectivity and reasoning space of the knowledge graph. Specifically, it includes:

[0038] (1) Path rule mining: During model training or inference, the present invention accumulates a large number of successful paths (successful paths refer to paths that lead to the correct target entity) through reinforcement learning methods. The present invention formalizes these paths into path rules, for example:

[0039] r q ⇐r1∧r2∧⋯∧r k Among them, r q is the target query relation, and the right side is the intermediate relations that appear in sequence in the reasoning path.

[0040] (2) Soft Fact Generation: Using the path rules mined in the previous step, the present invention can identify a series of potential inferred facts in the graph, called soft facts. Specifically, for a triple sequence that meets the following path conditions:

[0041] (e s ,r1,e1),(e1,r2,e2),…,(ek −1,r k ,e t ) If there is a corresponding rule r q ⇐r1∧⋯∧r k , then soft facts can be generated: (e s ,r q ,e t ).

[0042] (3) Confidence evaluation and screening: To control the quality of soft facts, this paper introduces a frequency-based confidence index to measure the reliability of a path rule. The confidence of a rule p is defined as follows:

[0043] Among them: Pos(p): the number of facts generated by rule p that are verified to be correct; neg(p): the number of facts generated by rule p that are judged to be incorrect.

[0044] For each soft fact (e s ,r q ,e t ), the present invention sets its confidence as the confidence of its generation rule prq, that is: Con((e s ,r q ,e t ))=Con(prq) During the reasoning process, the present invention introduces these soft facts as candidate actions and embeds their confidence into the action representation to assist the policy network in making selections. If the triple corresponding to an action comes from the soft fact set Ts, its action vector is recorded as:

[0045] at=(r i ,e i ,Con((e t ,r i ,e i ))) By introducing confidence-weighted soft facts, this module not only improves the model's reasoning ability under sparse graphs, but also maintains the reliability and interpretability of the reasoning path.

[0046] S5. Defect quality reasoning and traceability driven by reinforcement learning: Based on reinforcement learning, a defect reasoning agent is built, which defines the state, action, and reward mechanism. It performs multi-hop path search in the knowledge graph, generates explainable defect quality reasoning results, and implements defect traceability and analysis. The state of the reinforcement learning reasoning agent includes the current defect entity, query relationship, historical path, current hop count, etc. The action includes the next hop relationship in the graph and its confidence. The reward includes the score of hitting the correct target entity and the semantic relevance of the path. Specifically, it includes:

[0047] Adopting a reinforcement learning strategy based on the state-action-reward framework, starting from the defect node, searching for the reasoning path; The state includes the current entity, target relationship, historical path, etc., and the action includes the edge in the graph and its confidence; The reward consists of target matching accuracy and path semantic consistency; The final output is the prediction results of the quality outcomes caused by defects and the path to their formation, making the results interpretable and supporting decision-making. The reasoning results include the prediction results of quality factors and their formation paths, which are used to provide explainable defect traceability information to support comprehensive product quality assessment and optimization decision-making.

[0048] This paper formalizes inductive knowledge graph reasoning as a Markov decision process and trains a reinforcement learning agent to perform walk-based reasoning on the knowledge graph.

[0049] (a) The general Markov decision process consists of a quaternion Each element is defined as follows:

[0050] Status: Combined with known queries and move The entity where the step is located Define each state , is the set of all possible states.

[0051] Action: For status , the possible action space is an entity The union of adjacent edges and self-loop edges. for The self-loop edge of an entity keeps the agent in the current state.

[0052] Conversion: Conversion function Defines the transitions between states. Afterwards, the agent Move to the next state .

[0053] Reward: After reaching the maximum number of moves, the agent stops in state The final entity At this point, the agent receives a reward ,Right now

[0054] (b) Policy Network: To solve the Markov decision problem, the agent needs a policy to decide what action to take in each state. The policy is parameterized using a neural network. In the policy network, the input consists of three parts: The embedding vector of Embedding vector and path history of Encoding. LSTM is used to encode the path history as follows:

[0055] in For special starting relationship Embedded, is the encoding of the path history. Then the policy network is defined as:

[0056] in is the stacked input of all actions, is the softmax operator, and are the learnable weight matrices of the two fully connected layers.

[0057] (c) Optimizer: The goal of the policy network is to maximize the expected reward of all training queries: in 、 They are the action and status of the last step respectively.

[0058] Based on the same inventive concept, the present invention also proposes an industrial product surface defect tracing system, comprising: The acquisition module is used to collect real-time defect time series data of industrial products and supplementary data generated by the digital twin system; and based on the real-time defect time series data and supplementary data, establish a multi-view topological representation including physical view, process parameter view and quality result view.

[0059] The feature extraction module is used to extract the consistent features and private features between the structural features of each view, model the defect instance feature evolution process of the consistent features and private features in the time dimension, and obtain the spatiotemporal integrated features.

[0060] The knowledge graph establishment module is used to conduct instance mining based on multi-view topological representation and spatiotemporal integration features, establish a knowledge graph for surface defect quality analysis of industrial products, and set type features of head entities and tail entities for various relationships between entities in the knowledge graph; generate the initial semantic representation of the newly added entities by aggregating the type features of the adjacency relationships of the newly added entities; construct a local subgraph for the newly added entities and their facts based on the initial semantic representation of the newly added entities; update the local subgraph through the graph neural network and mine the path rules of the local subgraph; generate soft facts of the knowledge graph based on the path rules.

[0061] The reasoning module is used to build a reinforcement learning recommendation framework by combining the update process of the local subgraph and soft facts; the reinforcement learning recommendation framework is used to realize the traceability of surface defects of industrial products.

[0062] The present invention also proposes a computer device for tracing the surface defects of industrial products, comprising: a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the method for tracing the surface defects of industrial products are implemented.

[0063] The present invention also proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the industrial product surface defect tracing method.

[0064] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for tracing the source of surface defects of industrial products, characterized in that: The following steps are involved: Collect real-time defect time series data of industrial products and supplementary data generated by digital twin systems; Based on real-time defect timing data and supplementary data, a multi-view topology representation including physical view, process parameter view and quality result view is established; Extract the consistent features and private features between the structural features of each view, model the evolution process of the defect instance features of the consistent features and private features in the time dimension, and obtain the spatiotemporal integrated features; Based on multi-view topological representation and spatiotemporal integration features, case mining is performed to establish a knowledge graph for industrial product surface defect quality analysis. Type features of the head and tail entities are set for each type of relationship between entities in the knowledge graph. By aggregating the type features of the adjacency relationships of the newly added entities, an initial semantic representation of the newly added entities is generated. Based on the initial semantic representation of the newly added entities, a local subgraph of the newly added entities and their facts is constructed. The local subgraph is updated through a graph neural network to mine the path rules of the local subgraph. Based on the path rules, soft facts of the knowledge graph are generated. By combining the update process of local subgraphs and soft facts, a reinforcement learning recommendation framework is constructed; the surface defect traceability of industrial products is achieved through the reinforcement learning recommendation framework.

2. The method for tracing the surface defects of industrial products according to claim 1, characterized in that: The extraction of consistent features and private features between various view structure features, modeling the defect instance feature evolution process of the consistent features and private features in the time dimension, and obtaining spatiotemporal integrated features specifically includes the following steps: Graph convolutional neural network (GCN) or attention network is used to extract structural features of physical view, process parameter view and quality result view respectively; The maximum cross-correlation strategy is used to align the structural features of each view and extract consistent features and private features; A recursive neural network is used to model the defect instance feature evolution process of consistency features and private features in the time dimension to obtain spatiotemporal integrated features.

3. The method for tracing the surface defects of industrial products according to claim 1, characterized in that: A composite attention mechanism constructed by combining graph neural network with semantic similarity and rule confidence is used to aggregate the type features of the newly added entity adjacency relationships.

4. The method for tracing the surface defects of industrial products according to claim 1, characterized in that: The updating of the local subgraph by the graph neural network specifically includes the embedding update of entities and entity relationships, which are respectively expressed as: in 、 、 , e Represents the variable u is the embedding of the entity, r Represents the variable z is the embedding of the relation, l =0 indicates the embedding of the entity initially selected when performing a partial update of the knowledge graph. Represents the entity to be updated e connected entities; are three learnable weight matrices for entity embedding updates; is a learnable weight matrix for relation embedding update; the conversion function Accepts relation entities as input and aggregates information from adjacent links. Perform element-wise product, expressed as .

5. The method for tracing the surface defects of industrial products according to claim 1, characterized in that: Generating soft facts of the knowledge graph according to the path rules specifically includes the following steps: For the knowledge graph, successful paths are accumulated through reinforcement learning methods and formalized into path rules; the successful paths are paths that can infer the correct target entity; Utilize path rules to identify potential inferred facts in the knowledge graph and generate soft facts.

6. The method for tracing the surface defects of industrial products according to claim 1, characterized in that: The subgraph update and soft facts are combined to construct a reinforcement learning recommendation framework, and the surface defect traceability of industrial products is achieved through the reinforcement learning recommendation framework, which specifically includes the following steps: The Markov decision process is represented by a quadruple; the quadruple includes state, action, transition and reward; specifically includes: and entities move The entity where the step is located Define each state , is the set of all states, r q For the target query relationship; for the state , defining its action space as entity The union of adjacent edges and self-loop edges: ,in, for The inverse relationship, Represents the entity to be updated e connected entities, r express e and The relationship between Move to the next state , define the transition between states; after reaching the maximum number of moves, stop at state The final entity Get rewarded ; Construct a reinforcement learning strategy based on the state of the current entity, target relationship, historical path, and actions including edges in the knowledge graph and their confidence, as well as rewards consisting of target matching accuracy and path semantic consistency; Starting from the defect node through a reinforcement learning strategy, a multi-hop path search is performed in the knowledge graph to obtain an inference path including the quality factor prediction results and the path formed; Surface defect traceability of industrial products can be achieved based on inference paths.

7. An industrial product surface defect tracing system, characterized in that: include: The acquisition module is used to collect real-time defect time series data of industrial products and supplementary data generated by the digital twin system; Based on real-time defect timing data and supplementary data, a multi-view topology representation including physical view, process parameter view and quality result view is established; The feature extraction module is used to extract the consistent features and private features between the structural features of each view, model the evolution process of the defect instance features of the consistent features and private features in the time dimension, and obtain the spatiotemporal integrated features; The knowledge graph building module is used to conduct case mining based on multi-view topological representation and spatiotemporal integration features, build a knowledge graph for industrial product surface defect quality analysis, and set type features for head and tail entities for each type of relationship between entities in the knowledge graph; By aggregating the type features of the adjacency relationships of the newly added entities, an initial semantic representation of the newly added entities is generated. Based on the initial semantic representation of the newly added entities, a local subgraph of the newly added entities and their facts is constructed. The local subgraph is updated through a graph neural network to mine the path rules of the local subgraph. Based on the path rules, soft facts of the knowledge graph are generated. The reasoning module is used to build a reinforcement learning recommendation framework by combining the local subgraph update process and soft facts; Surface defect traceability of industrial products through a reinforcement learning recommendation framework.

8. A computer device for tracing the surface defects of industrial products, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein when the processor executes the computer program, the steps of the industrial product surface defect tracing method according to any one of claims 1 to 6 are implemented.

9. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the industrial product surface defect tracing method according to any one of claims 1 to 6.

Images