A body-in-white welding plant management and control system

Abstract

The present application relates to a kind of white body welding workshop management and control system, it relates to the technical field of automobile manufacturing, the system includes: semantic data fusion layer, is deployed with welding workshop digital twin knowledge graph;Large model cognitive reasoning layer is configured to generate the first decision instruction carrying fault root cause and corresponding repair information, or generate the second decision instruction carrying equipment configuration parameter set;Interactive execution layer is configured to push the fault root cause and repair information carried in the first decision instruction to operation terminal, or the equipment configuration parameter set carried in the second decision instruction is encapsulated as equipment-level control instruction and is issued to corresponding workshop physical equipment, the quick traceability of fault root and the reconfigurable management and control of system are realized.

Description

Technical Field

[0001] This invention relates to the field of automobile manufacturing technology, and in particular to a control system for a body-in-white welding workshop. Background Technology

[0002] Body-in-white welding is a core process in automobile manufacturing. Its production site involves various physical devices such as welding robots, resistance welding equipment, automated tooling, and material handling carts. These devices have complex semantic relationships involving tooling manipulation, workpiece clamping, and material flow. With the development of intelligent manufacturing, welding workshops have widely deployed information systems such as manufacturing execution systems and data acquisition and monitoring systems to record equipment status, process parameters, and production progress.

[0003] However, in existing welding workshop management systems, equipment operation data, process parameters, material status, and tooling information are scattered across multiple heterogeneous systems. These systems operate independently and lack a unified model to explicitly express the semantic relationships between physical entities such as equipment, tooling, materials, and process parameters within the workshop. When equipment failures or quality anomalies occur, engineers need to manually search across systems and rely on experience to deduce the cause of the failure, making it impossible to quickly locate the root cause entity based on the inherent logical relationships between data. When the production line needs to switch models or adjust processes, the system cannot automatically adapt to the changes, requiring manual modification of the parameter configurations and logic codes of the relevant equipment one by one. This results in system adjustments being heavily reliant on manual experience and inefficient. Therefore, existing technologies lack a unified management and control solution that can express the various entity instances in the welding workshop and their semantic relationships, and achieve intelligent reasoning and system reconfiguration based on this unified expression. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a white body welding workshop control system, which solves the technical problems of difficult fault tracing and difficult system reconstruction in the prior art.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] This invention provides a white body welding workshop control system, comprising:

[0009] The semantic data fusion layer is deployed with a digital twin knowledge graph of the welding workshop. The digital twin knowledge graph of the welding workshop includes multiple entity instances and their semantic relationships. The semantic data fusion layer is configured to collect physical equipment data in the workshop in real time and assign it to the attributes of the corresponding entity instances in the digital twin knowledge graph of the welding workshop, so that the knowledge graph is dynamically updated with the status of the physical equipment.

[0010] The large model cognitive reasoning layer is configured to receive natural language commands or system alarm events. When the input is a system alarm event, it determines the abnormal equipment based on the system alarm event, and uses the entity instance corresponding to the abnormal equipment as the starting point to perform multi-hop causal reasoning through semantic association to identify the entity instance of the root cause of the fault, and generates a first decision command carrying the root cause of the fault and corresponding repair information. When the input is a natural language command representing the intention to change the production line configuration, it extracts the attributes or semantic associations of the entity instance to be modified from the natural language command, modifies the corresponding attributes or semantic associations in the digital twin knowledge graph of the welding workshop, and generates a second decision command carrying the equipment configuration parameter set based on the modified knowledge graph.

[0011] The interactive execution layer has its input end coupled to the large model cognitive reasoning layer to receive the first decision instruction or the second decision instruction, and its output end is communicatively connected to the workshop physical equipment and the operation terminal. The interactive execution layer is configured to push the root cause of the fault and repair information carried in the first decision instruction to the operation terminal, or to encapsulate the set of equipment configuration parameters carried in the second decision instruction into equipment-level control instructions and send them to the corresponding workshop physical equipment.

[0012] Optionally, the semantic association relationship includes at least one or more of the following relationships: equipment manipulation relationship, tooling clamping relationship, material flow relationship, and spatial location association relationship between workstation and equipment.

[0013] Optionally, the semantic data fusion layer is also configured to generate an abnormal status label and associate it with the corresponding entity instance when the attribute value of an entity instance in the knowledge graph deviates from a preset threshold. The abnormal status label is used to trigger a system alarm event.

[0014] Optionally, when the large model cognitive reasoning layer performs multi-hop causal reasoning through semantic association, it takes the entity instance corresponding to the abnormal equipment as the starting point and traces back to the upstream related entity instance layer by layer along the semantic association. By comparing the deviation of the attribute values ​​of each entity instance along the way in the welding workshop digital twin knowledge graph from the corresponding preset threshold, the root cause entity instance of the fault is located.

[0015] Optionally, the large model cognitive reasoning layer is specifically configured to, when the input is a natural language instruction representing the intention to change the production line configuration, extract the equipment identifier to be replaced and the new equipment specification parameters from the natural language instruction, determine the attribute or semantic relationship of the entity instance to be modified based on the equipment identifier to be replaced and the new equipment specification parameters, modify the corresponding attribute or semantic relationship in the digital twin knowledge graph of the welding workshop, and generate a second decision instruction carrying a set of equipment configuration parameters based on the modified knowledge graph; wherein, the modification includes: removing the spatial location association between the entity instance corresponding to the equipment to be replaced and the entity instance corresponding to the workstation, creating a new entity instance and establishing a spatial location association between it and the entity instance corresponding to the workstation, and adjusting the associated process parameter attribute values ​​according to the new equipment specification parameters.

[0016] Optionally, the interactive execution layer is specifically configured to push the root cause of the fault and repair information carried in the first decision instruction to the operation terminal, and also locate and highlight the physical location corresponding to the root cause entity instance in the three-dimensional scene of the operation terminal according to the spatial location information of the root cause entity instance in the knowledge graph.

[0017] Optionally, the large model cognitive reasoning layer is also configured to calculate compensation parameter values ​​and generate a third decision instruction carrying compensation parameters when the deviation of the attribute value of the fault root cause entity instance is within a preset compensable range, based on the attribute values ​​of one-hop adjacent entity instances in the welding workshop digital twin knowledge graph that have a direct semantic relationship with the fault root cause entity instance; the interactive execution layer is also configured to encapsulate the compensation parameters carried in the third decision instruction into a closed-loop control instruction and send it to the physical device.

[0018] Optionally, the large model cognitive reasoning layer is also configured to calculate the degree of deviation between the attribute value of each candidate root cause entity instance and a preset threshold when the input is a system alarm event and multiple candidate root cause entity instances are identified after multi-hop causal reasoning, and to determine the candidate root cause entity instance with the largest deviation as the fault root cause entity instance.

[0019] Optionally, the semantic data fusion layer is also configured to automatically update the association state attributes of one-hop adjacent entity instances that have a direct semantic relationship with the entity instance when the attribute value of an entity instance in the knowledge graph is updated, based on the semantic relationship.

[0020] Optionally, when the large model cognitive reasoning layer performs multi-hop causal reasoning, it generates a structured reasoning evidence chain according to the order of each hop of the reasoning path. The reasoning evidence chain records the initial entity instance identifier, the semantic relationship type traversed, the intermediate entity instance identifier reached, and the degree of deviation of the attribute value of the intermediate entity instance for each hop. The first decision instruction contains the reasoning evidence chain. When the interactive execution layer pushes the first decision instruction to the operation terminal, it displays the reasoning evidence chain in the form of a visual directed graph in the three-dimensional scene of the operation terminal. The nodes in the visual directed graph are marked with different colors according to the degree of deviation of the attribute value, and the edges in the visual directed graph are labeled with their corresponding semantic relationship types.

[0021] (III) Beneficial Effects

[0022] The beneficial effects of this invention are:

[0023] This application provides a white body welding workshop management and control system. By constructing a digital twin knowledge graph of the welding workshop containing entity instances and their semantic relationships, it unifies the data of scattered physical equipment, tooling, materials, and process parameters within the workshop into a single semantic model. When a system alarm event is received, starting with the entity instance corresponding to the abnormal equipment, multi-hop causal reasoning is performed through the semantic relationships in the knowledge graph to identify the root cause entity instance of the fault and generate a first decision instruction carrying the root cause of the fault and repair information, which is then pushed to the operation terminal, thereby quickly locating the root cause of the fault. When a natural language instruction representing a production line configuration change intention is received, the system extracts the attributes or semantic relationships to be modified from the natural language instruction and modifies the knowledge graph. Then, based on the modified knowledge graph, a second decision instruction carrying a set of equipment configuration parameters is generated and sent to the corresponding physical equipment in the workshop, thereby achieving automatic reconstruction of the system at the semantic level. In this way, through the collaboration of the knowledge graph and the large model, rapid tracing of the root cause of the fault and reconfigurable control of the system are achieved. Attached Figure Description

[0024] Figure 1 This illustration shows a schematic diagram of a white body welding workshop control system provided in an embodiment of this application;

[0025] Figure 2 This document illustrates a flowchart of knowledge graph modification in a production line configuration change scenario, as provided in an embodiment of this application.

[0026] Figure 3 A schematic diagram of a white body welding workshop control system provided in an embodiment of this application is shown. Detailed Implementation

[0027] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] To address the challenges of fault tracing and system reconfiguration in existing technologies, this application provides a white body welding workshop management system. By constructing a digital twin knowledge graph of the welding workshop containing entity instances and their semantic relationships, it unifies data on dispersed physical equipment, tooling, materials, and process parameters within the workshop into a single semantic model. When a system alarm event is received, starting with the entity instance corresponding to the abnormal equipment, multi-hop causal reasoning is performed through the semantic relationships in the knowledge graph to identify the root cause entity instance of the fault and generate a first decision instruction carrying the root cause and repair information, which is then pushed to the operation terminal, thereby quickly locating the root cause of the fault. When a natural language instruction representing a production line configuration change is received, the system extracts the attributes or semantic relationships to be modified from the natural language instruction and modifies the knowledge graph. Then, based on the modified knowledge graph, a second decision instruction carrying a set of equipment configuration parameters is generated and sent to the corresponding physical equipment in the workshop, thereby achieving automatic system reconfiguration at the semantic level. In this way, through the collaboration of the knowledge graph and the large model, rapid fault root cause tracing and reconfigurable system management are achieved.

[0029] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0030] Please see Figure 1 , Figure 1 A schematic diagram of a white body welding workshop control system provided in an embodiment of this application is shown. Figure 1 As shown, the system includes a semantic data fusion layer, a large-scale model cognitive reasoning layer, and an interactive execution layer. The semantic data fusion layer provides the knowledge graph query and retrieval foundation for the large-scale model cognitive reasoning layer. The output of the large-scale model cognitive reasoning layer is coupled to the input of the interactive execution layer to transmit decision instructions. The output of the interactive execution layer is communicatively connected to the workshop physical equipment and the operating terminal, respectively.

[0031] The semantic data fusion layer is equipped with a digital twin knowledge graph of the welding workshop, which includes multiple entity instances and their semantic relationships. The semantic data fusion layer is configured to collect physical equipment data in the workshop in real time and assign it to the attributes of the corresponding entity instances in the digital twin knowledge graph of the welding workshop, so that the knowledge graph is dynamically updated with the status of the physical equipment.

[0032] The large model cognitive reasoning layer is configured to receive natural language commands or system alarm events. When the input is a system alarm event, it determines the abnormal equipment based on the system alarm event, and uses the entity instance corresponding to the abnormal equipment as the starting point to perform multi-hop causal reasoning through semantic association to identify the entity instance of the root cause of the fault, and generates a first decision command carrying the root cause of the fault and corresponding repair information. When the input is a natural language command representing the intention to change the production line configuration, it extracts the attributes or semantic associations of the entity instance to be modified from the natural language command, modifies the corresponding attributes or semantic associations in the digital twin knowledge graph of the welding workshop, and generates a second decision command carrying the equipment configuration parameter set based on the modified knowledge graph.

[0033] The interactive execution layer has its input end coupled to the large model cognitive reasoning layer to receive the first decision instruction or the second decision instruction, and its output end is communicatively connected to the workshop physical equipment and the operation terminal. The interactive execution layer is configured to push the root cause of the fault and repair information carried in the first decision instruction to the operation terminal, or to encapsulate the set of equipment configuration parameters carried in the second decision instruction into equipment-level control instructions and send them to the corresponding workshop physical equipment.

[0034] It should be understood that the specific execution processes of the semantic data fusion layer, the large model cognitive reasoning layer, and the interactive execution layer can all be set according to actual needs, and the embodiments of this application are not limited thereto.

[0035] Optionally, the semantic data fusion layer deploys a welding workshop digital twin knowledge graph (hereinafter referred to as the knowledge graph). This knowledge graph is constructed based on the welding workshop domain ontology and contains multiple entity instances and semantic relationships between them. Entity instances are digital mappings of physical entities in the workshop, each with several attributes representing the entity's state, parameters, or location. For example, a welding robot is mapped to an entity instance, which may contain attributes such as current current value, operating temperature, and coordinate position. Semantic relationships express the physical or logical connections between different entity instances. Furthermore, these semantic relationships include at least one or more of the following:

[0036] Equipment manipulation relationships refer to the associations formed when equipment performs operations on tooling or workpieces, and are used to characterize the control logic of "which equipment operates which tooling". For example, the equipment manipulation relationship between a welding robot entity instance and an automated tooling entity instance indicates that the welding robot performs welding operations on the tooling.

[0037] Tooling clamping relationship refers to the association formed when a tooling positions and clamps a workpiece, used to characterize the clamping logic of "which tooling clamps which workpiece". For example, the tooling clamping relationship between a side panel tooling entity instance and a side panel assembly entity instance indicates that the tooling is currently clamping the side panel assembly.

[0038] Material flow relationships refer to the associations formed when workpieces or assemblies are transferred between different workstations according to the process sequence. They are used to characterize the logistics logic of "which workpiece flows from which workstation to which workstation". For example, when a floor assembly entity instance flows from the floor welding workstation to the final assembly workstation, this flow process is recorded as a material flow relationship.

[0039] The spatial relationship between a workstation and equipment refers to the spatial attribution between a workstation and the equipment fixedly installed on it, representing the installation positioning logic of "which equipment is fixed to which workstation." This relationship is not considered immutable due to equipment replacement, but rather means that the equipment's current installation position on the workstation is a definite spatial attribution. For example, a welding robot is fixed to an automated welding workstation, and the two are linked by a spatial relationship. When the equipment is replaced, this relationship is updated accordingly to reflect the association between the new equipment and the workstation.

[0040] Furthermore, this semantic data fusion layer is configured to collect real-time operational data of the physical equipment in the workshop and assign the collected data to the attributes of the corresponding entity instances in the knowledge graph. This allows the knowledge graph to be dynamically updated as the physical equipment status changes, thereby achieving synchronous mapping of the physical workshop in the digital space. For example, when the output current value of a resistance welding device changes, the data fusion layer assigns this current value to the "current value" attribute of the entity instance corresponding to the resistance welding device; when the clamping cylinder positioning signal of a tooling changes abruptly, the signal value is assigned to the "clamping status" attribute of the entity instance corresponding to the tooling.

[0041] Furthermore, the semantic data fusion layer maintains data consistency within the knowledge graph. Specifically, when the attribute value of an entity instance in the knowledge graph is updated, the semantic data fusion layer automatically updates the association status attributes of one-hop adjacent entity instances that have a direct semantic relationship with that entity instance, based on semantic association relationships. A one-hop adjacent entity instance refers to an entity instance directly connected to the knowledge graph through a semantic association relationship. Association status attributes refer to the status information of an entity instance related to adjacent entities, such as the attribute representing the clamping position in a tooling entity instance, the attribute representing the online or offline status in an equipment entity instance, and the attribute representing the location of the workstation in a material entity instance. For example, when the clamping cylinder position signal of an entity instance corresponding to a tooling changes from "clamped" to "released," the semantic data fusion layer not only updates the clamping status attribute of the tooling entity instance but also automatically updates the association status attributes of the workpiece entity instance clamped by the tooling based on the tooling clamping relationship (e.g., updating the workpiece status from "positioned" to "not positioned"), ensuring that the status information of associated entity instances in the knowledge graph remains synchronized and consistent. Through this mechanism, the state change of any entity instance in the knowledge graph can be automatically propagated to adjacent entity instances along semantic relationships, thereby ensuring that the state information of related entities in the entire knowledge graph always maintains semantic consistency, providing an accurate data foundation for causal reasoning in the subsequent cognitive reasoning layer of the large model.

[0042] Optionally, the large-scale cognitive reasoning layer integrates a large language model. This large language model is a deep learning model with massive parameters, pre-trained on a wide range of text data through self-supervised learning. It is capable of understanding and generating natural language and possesses contextual understanding, semantic parsing, and logical reasoning abilities. This large-scale cognitive reasoning layer receives external input, including natural language instructions and system alarm events. The large language model interacts with the digital twin knowledge graph of the welding workshop through Retrieval Enhanced Generation (RAG). Specifically, RAG is a technical architecture that combines an external knowledge base (i.e., the knowledge graph in this embodiment) with the large language model: when the large language model receives input, it first retrieves matching entity instances and their relationships in the knowledge graph based on the semantic keywords in the input. Then, it injects the retrieved structured knowledge as contextual information into the prompts of the large language model, enabling the large language model to reason and generate based on the retrieved real knowledge rather than the general knowledge in its own pre-trained parameters. In this way, the large language model can acquire specific knowledge of the welding workshop directly related to the current input (such as the real-time status of specific equipment, the location information of tooling, the current value of process parameters, etc.), thereby overcoming the problem of insufficient knowledge in industrial professional fields of general large language models and avoiding the generation of content that does not match the actual situation. The large language model can be implemented using various open-source or commercial models. For example, it can use DeepSeek series models (such as DeepSeek-V3 or DeepSeek-R1), Tongyi Qianwen series models (such as Qwen-7B or Qwen-32B), Llama series models, or ChatGLM series models, etc. All of the above models are large language models known in the art. Those skilled in the art can choose a suitable model as the large language model for the cognitive reasoning layer of the large model according to the actual deployment environment, computing resources, and performance requirements; this application does not limit this choice. During the process of real-time collection of workshop physical equipment data and assignment of attributes to corresponding entity instances in the knowledge graph, the semantic data fusion layer continuously monitors whether the attribute values ​​of each entity instance deviate from the preset threshold. The preset threshold refers to the normal value range set for each attribute of each entity instance. This threshold is predetermined based on process specifications or equipment nameplate parameters. When the attribute value of an entity instance deviates from the corresponding preset threshold, the semantic data fusion layer generates a status anomaly label and associates this label with the corresponding entity instance. This status anomaly label triggers a system alarm event. For example, when the current value of a welding robot exceeds the preset normal operating current range, the semantic data fusion layer generates a current anomaly label and associates it with the entity instance corresponding to the welding robot, simultaneously triggering a system alarm event. The large model cognitive reasoning layer performs fault root cause reasoning or production line configuration change processing depending on the input type.

[0043] When the input is a system alarm event, the large-scale model cognitive reasoning layer first determines the physical device that malfunctioned based on the alarm event and maps this physical device to a corresponding abnormal device entity instance in the knowledge graph. Based on this, the large-scale model cognitive reasoning layer uses this abnormal device entity instance as a starting point and performs multi-hop causal reasoning upstream along semantic relationships. Multi-hop causal reasoning refers to starting from the starting entity instance, tracing along a semantic relationship to the directly related parent entity instance, then tracing along the next layer of semantic relationships, and so on, until the root cause entity instance causing the fault is located. A hop refers to the entity instance reached after traversing a semantic relationship. During this process, the large-scale model cognitive reasoning layer compares the attribute values ​​of each entity instance along the way with corresponding preset thresholds, and locates the root cause entity instance based on the degree of deviation. For example, if a resistance welding machine at a certain workstation experiences an abnormal welding current alarm, the large model's cognitive inference layer starts with the entity instance corresponding to the resistance welding machine, traces along spatial relationships to the workstation, and then traces along tooling clamping relationships to the tooling at that workstation. It compares the deviation of the attribute values ​​of each entity instance along the way from the corresponding preset thresholds. If it finds that the clamping cylinder pressure value of the tooling deviates from the preset threshold, the tooling is determined to be the root cause entity instance, because the tooling not being clamped properly leads to a larger gap in the plate material, thus causing the abnormal welding current, rather than a malfunction of the welding machine itself. When the large model's cognitive inference layer identifies multiple candidate root cause entity instances after multi-hop causal reasoning, it calculates the deviation of the attribute values ​​of each candidate root cause entity instance from the corresponding preset thresholds, and determines the candidate root cause entity instance with the largest deviation as the root cause entity instance. For example, if the candidate root cause entity instances include the tooling (clamping pressure deviating from the threshold by 15%) and the welding machine (current deviating from the threshold by 5%), then the tooling with the larger deviation is determined to be the root cause entity instance.

[0044] During the reasoning process, the large model's cognitive reasoning layer generates a structured chain of reasoning evidence sequentially for each hop of the reasoning path, rather than simply outputting the final judgment result. Each hop of the reasoning evidence chain records the initial entity instance identifier, the type of semantic relationship traversed, the intermediate entity instance identifier reached, and the degree of deviation of the attribute values ​​of that intermediate entity instance. For example, starting from "Welding Machine_03," it reaches "Assembly Station_02" through spatial location relationships, recording the degree of deviation of the station's state attribute values; then, through tooling clamping relationships, it reaches "Side Panel Tooling_05," recording the degree of deviation of the clamping pressure of that tooling. The reasoning evidence chain is transmitted to the interactive execution layer along with the first decision instruction for subsequent visualization. After identifying the root cause entity instance of the fault, the large model's cognitive reasoning layer generates the first decision instruction. The first decision instruction carries the root cause of the fault and the corresponding repair information. The repair information includes at least the name and location identifier of the root cause entity instance in the knowledge graph, a description of the fault cause, and a list of related entity instances affected by the fault, derived from the semantic relationships in the knowledge graph, so that maintenance personnel can fully understand the scope of the fault's impact.

[0045] When the input is a natural language instruction representing an intention to change the production line configuration, the large-scale model's cognitive reasoning layer extracts the attributes or semantic relationships of the entity instance to be modified from the natural language instruction and modifies the corresponding attributes or semantic relationships in the knowledge graph. For example, a natural language instruction could be "Replace welding robot No. 2 at assembly station B with a new robot of model XX" or "Adjust the clamping force parameter of side panel fixture_05 to 5000N." The large-scale model's cognitive reasoning layer performs semantic parsing on the natural language instruction, extracting the target object to be modified (such as equipment identification), the type of modification operation (such as equipment replacement, parameter adjustment), and the new parameter values ​​or specifications. As a typical scenario, a production line configuration change specifically involves equipment replacement. In this case, the large-scale model's cognitive reasoning layer extracts the equipment identification to be replaced and the new equipment specification parameters from the natural language instruction, thereby determining the knowledge graph content that needs to be modified. Figure 2As shown, the specific modification operation includes three steps: Step S210, remove the spatial location association between the entity instance corresponding to the equipment to be replaced and the entity instance corresponding to the workstation, that is, remove the installation affiliation association between the old equipment and the workstation in the knowledge graph; Step S220, create a new entity instance corresponding to the specifications of the new equipment, and establish the spatial location association between the new entity instance and the entity instance corresponding to the workstation, that is, establish the installation affiliation association between the new equipment and the workstation in the knowledge graph; Step S230, adjust the associated process parameter attribute values ​​according to the specifications of the new equipment. For example, if the rated current range, rated voltage range, working pressure range, etc. of the new welding robot are different from those of the old equipment, then the current setting value, voltage setting value, pressure setting value, etc. in the process parameter attributes will be updated accordingly. In addition, production line configuration changes may also involve other types of modifications besides equipment replacement, such as adjusting the position coordinate parameters of a tooling, or modifying the welding sequence parameters of a welding point in a certain process. These modification operations can all be achieved by directly updating the attribute values ​​of the corresponding entity instances in the knowledge graph. The principle is the same as that of the aforementioned equipment replacement scenario, and will not be elaborated here. After the above modifications are completed, the large model's cognitive reasoning layer generates a second decision instruction based on the modified knowledge graph. The second decision instruction carries a set of device configuration parameters, which contains all the parameter configuration information required for the normal operation of the new device (including but not limited to current setpoints, voltage setpoints, operating timing, safety protection parameters, etc.), for subsequent distribution to the physical device.

[0046] For the fault root cause entity instance identified in the system alarm event handling, the large model cognitive reasoning layer further determines whether the deviation of the attribute value of the fault root cause entity instance is within a preset compensable range. The compensable range refers to the upper limit of the deviation of the attribute value that the physical equipment can correct through parameter adjustment. This upper limit is predetermined based on the equipment's design redundancy and allowable process deviations. If the deviation does not exceed this upper limit, it is within the compensable range; if the deviation exceeds this upper limit, it cannot be repaired through parameter adjustment and requires manual intervention. When the deviation of the attribute value of the fault root cause entity instance is within the preset compensable range, the large model cognitive reasoning layer calculates the compensation parameter value based on the attribute values ​​of one-hop adjacent entity instances in the knowledge graph that have a direct semantic relationship with the fault root cause entity instance. The calculation logic of the compensation parameter value is determined based on the difference between the attribute value of the one-hop adjacent entity instance and the standard value, and the impact of this difference on the fault root cause entity instance. For example, if the root cause of the fault is that the welding current at a certain workstation deviates from the standard value, and this deviation is caused by air pressure fluctuations at an adjacent workstation, the large model cognitive inference layer obtains the current air pressure value of the air pressure regulating valve (i.e., a hop adjacent entity instance) at that adjacent workstation, and calculates the required current compensation value based on the preset mapping relationship between the air pressure deviation value and the current compensation value. The mapping relationship can be an empirical formula fitted through historical data, or it can be an analytical expression based on the physical model. The specific form depends on the actual process type, and this application does not limit it. After calculating the compensation parameter value, the large model cognitive inference layer generates a third decision instruction carrying the compensation parameter. The third decision instruction is sent by the interactive execution layer to the physical device to complete automatic repair without manual intervention. If the deviation exceeds the preset compensable range, the large model cognitive inference layer does not generate a third decision instruction, and the first decision instruction only carries the root cause of the fault and repair information, which is then prompted by the operation terminal for manual intervention by maintenance personnel.

[0047] Furthermore, after the chain of reasoning evidence is transmitted to the interactive execution layer along with the first decision command, the interactive execution layer displays the chain of reasoning evidence as a visualized directed graph in the 3D scene of the operating terminal. Each node in the visualized directed graph corresponds to an entity instance recorded in the chain of reasoning evidence, and nodes are marked with different colors according to the degree of deviation of the attribute values ​​of that entity instance. For example, nodes with small deviations in attribute values ​​are displayed in green, those with medium deviations in yellow, those with large deviations in red, and the entity instance of the root cause of the fault with the largest deviation in dark red with a flashing border, so as to intuitively distinguish the importance of each node in the fault causal chain. Each edge in the visualized directed graph corresponds to a semantic relationship recorded in the chain of reasoning evidence, and each edge is labeled with its corresponding semantic relationship type, such as spatial location relationship, tooling clamping relationship, or material flow relationship, etc. Through the above visualized directed graph, the user of the operating terminal can intuitively view the complete reasoning path from the abnormal equipment to the root cause of the fault, as well as the state deviation of each entity instance along the way, so as to quickly understand the causal relationship of the fault and assist in decision-making.

[0048] After each multi-hop causal reasoning is completed and verified, such as Figure 3 As shown, the system can also be configured with a knowledge evolution layer to extract reusable reasoning experience from the reasoning evidence chain, continuously optimizing the efficiency of subsequent reasoning. The knowledge evolution layer is coupled to the large model cognitive reasoning layer, and is used to extract reusable reasoning experience in real time from the reasoning evidence chain generated by each multi-hop causal reasoning, transforming the extracted experience into real-time constraint rules, which are directly injected into the reasoning process of the large model cognitive reasoning layer. It should be noted that the knowledge evolution layer does not store the complete reasoning evidence chain for future queries; instead, it extracts two types of information from the evidence chain in real time and transforms them into real-time constraint rules for the reasoning engine.

[0049] The first category is the real-time pruning rule for invalid paths. Once a multi-hop causal reasoning is completed and verified as valid—that is, the fault root cause entity instance located along the reasoning evidence chain is indeed repaired and the fault is resolved—the knowledge evolution layer extracts the branch paths that were traversed but ultimately not confirmed as the fault root cause from the reasoning evidence chain of that reasoning. Here, a branch path refers to the path chosen by the large model cognitive reasoning layer during multi-hop causal reasoning when faced with multiple optional semantic relationship branches at a certain entity instance node, but ultimately that branch does not reach the fault root cause entity instance. For example, in a certain reasoning, the large model cognitive reasoning layer starts from the entity instance corresponding to the abnormal welding machine, facing two optional branches: tracing back to the workstation along the spatial location relationship and tracing back to the manipulated tooling along the equipment manipulation relationship. The reasoning engine first attempts to trace back to the tooling along the equipment manipulation relationship, finds that the tooling's attribute values ​​are normal, then backtracks and continues reasoning along the spatial location relationship, ultimately locating the fault root cause at the workstation level. In this process, the branch from the equipment manipulation relationship to the tooling is the branch path that has been traversed but ultimately not identified as the root cause of the fault. The knowledge evolution layer transforms this branch path into a real-time pruning rule and directly injects it into the inference engine of the large model's cognitive inference layer. This real-time pruning rule takes effect immediately at the start of the next inference: when the inference engine faces the same semantic relationship branch selection again, the branch marked as a historical invalid path is constrained to be skipped in the first step of inference, and the inference engine does not need to traverse and verify along the branch before backtracking. This is fundamentally different from the method of storing the inference evidence chain for subsequent queries. In the storage query method, the inference engine still needs to complete the traversal of invalid branches first, and then query the history database to confirm that the branch is invalid, and the inference process itself is not changed; while the real-time pruning rule directly blocks invalid branches at the source of inference, the inference process itself is changed, and invalid branches are eliminated in the first step, avoiding the waste of computing resources.

[0050] The second category is real-time priority guidance for valid paths. When a reasoning path is repeatedly verified as valid in multiple multi-hop causal reasoning, that is, when the fault root cause entity instances located along the reasoning path are indeed repaired and the fault is resolved, the knowledge evolution layer extracts the complete jump sequence of the reasoning path from the reasoning evidence chain of these multiple reasonings and transforms the complete jump sequence into a real-time priority guidance rule. The complete jump sequence refers to the complete path starting from the initial abnormal equipment entity instance, passing through each hop of semantic association type and reaching the intermediate entity instance identifier, until finally reaching the fault root cause entity instance. For example, if the evidence chain record of a certain reasoning is "welding machine_03—spatial location association → assembly station_02—tooling clamping relationship → side panel tooling_05", then this complete jump sequence is extracted as a valid path. In subsequent reasoning, when the large model cognitive reasoning layer faces multiple optional semantic association branches, the branch that matches the valid path already stored in the knowledge evolution layer is given a higher exploration priority. The reasoning engine prioritizes reasoning along this direction, rather than indiscriminately traversing all optional branches. The matching degree between the inference path and the effective path is determined based on the similarity of the initial entity instance type, the abnormal attribute type, and the semantic association type of the first hop. For example, when an abnormal current alarm occurs again at the same workstation, the inference engine first proceeds along the branch with the highest matching degree with the historical effective path, i.e., the direction of the spatial location association, to quickly verify whether the path is still applicable under the current working conditions. If applicable, the root cause of the fault is directly located. If the deviation of the attribute values ​​of the entity instances along the path under the current working conditions exceeds the preset range of the deviation distribution recorded in the historical effective path, it indicates that the working conditions have changed and the historical effective path is no longer fully applicable. In this case, the inference engine still performs the inference again according to the complete multi-hop causal inference process and submits the new inference results to the knowledge evolution layer for a new round of experience extraction and rule updates. Through the above mechanism, the knowledge evolution layer realizes the real-time extraction and immediate reuse of inference experience, and the inference efficiency under common fault scenarios continues to improve with the increase of system running time.

[0051] It should be noted that the real-time pruning rules for invalid paths and the real-time priority guidance for valid paths are two independently operating mechanisms that can be configured individually or simultaneously. In actual deployment, the system can choose to enable one or both mechanisms based on the size of the knowledge graph and available computing resources. Furthermore, the pruning rules and priority guidance rules stored in the knowledge evolution layer are time-sensitive. When the knowledge graph is reconstructed in the large model's cognitive reasoning layer due to process changes or equipment replacement, the knowledge evolution layer simultaneously clears historical rules incompatible with the new knowledge graph structure to prevent outdated rules from misleading current reasoning.

[0052] Optionally, the input of this interactive execution layer is coupled to the large model's cognitive reasoning layer to receive a first decision instruction, a second decision instruction, or a third decision instruction. The output of the interactive execution layer is communicatively connected to the physical equipment in the workshop and the operating terminal. The interactive execution layer executes different output actions according to the type of decision instruction received.

[0053] Upon receiving the first decision command, the interaction execution layer pushes the root cause and repair information carried in the command to the operating terminal. The operating terminal can be a handheld tablet, an industrial touchscreen, or a workshop control panel display, allowing maintenance personnel to view fault information. Simultaneously with pushing the root cause and repair information, the interaction execution layer locates and highlights the physical location of the root cause entity instance in the 3D scene of the operating terminal based on its spatial location information in the knowledge graph. Spatial location information refers to the coordinate data of the entity instance, pre-stored in the knowledge graph and updated as the equipment's installation location changes. This coordinate data corresponds to the actual physical location in the workshop coordinate system. Highlighting can be achieved by marking the physical location in the 3D scene with a prominent marker (such as a flashing circle, a highlighted border, or a pointing arrow), enabling users of the operating terminal to quickly locate the specific location of the fault within the workshop layout without manually searching through numerous devices.

[0054] Furthermore, when the first decision instruction includes a chain of reasoning evidence, the interactive execution layer also displays this chain as a visualized directed graph in the 3D scene of the operating terminal. Each node in the visualized directed graph corresponds to an entity instance recorded in the chain of reasoning evidence, and nodes are marked with different colors according to the degree of deviation of the entity instance's attribute values. For example, nodes with small attribute value deviations are displayed in green, those with medium deviations in yellow, those with large deviations in red, and the entity instance of the root cause of the fault with the largest deviation is displayed in dark red with a flashing border, to intuitively distinguish the importance of each node in the fault causal chain. Each edge in the visualized directed graph corresponds to a semantic relationship recorded in the chain of reasoning evidence, and each edge is labeled with its corresponding semantic relationship type, such as spatial location relationship, tooling clamping relationship, or material flow relationship. Through the above visualized directed graph, the user of the operating terminal can intuitively view the complete reasoning path from the abnormal equipment to the root cause of the fault, as well as the state deviation of each entity instance along the way, thereby quickly understanding the causal relationship of the fault and assisting in maintenance decisions.

[0055] Upon receiving the second decision instruction, the interactive execution layer encapsulates the set of device configuration parameters carried in the second decision instruction into a device-level control instruction and sends it to the corresponding physical equipment in the workshop. The set of device configuration parameters contains all the parameter configuration information required for the normal operation of the new equipment, such as current setpoints, voltage setpoints, operating timing, and safety protection parameters. The interactive execution layer encapsulates the above parameters into a device-level control instruction according to a communication protocol recognizable by the corresponding physical equipment (such as industrial Ethernet protocols such as Profinet, EtherCAT, or OPC UA), and sends it to the corresponding physical equipment through the industrial network, completing the automatic configuration of the equipment parameters without the need for manual input of parameters on the equipment control panel.

[0056] Upon receiving a third decision instruction, the interactive execution layer encapsulates the compensation parameters carried in the third decision instruction into a closed-loop control instruction and sends it to the corresponding physical device. The closed-loop control instruction is an adjustment instruction formed based on the current device status feedback. After its issuance, the device performs the adjustment operation according to the compensation parameters. The interactive execution layer continuously monitors the device's status feedback after execution until it confirms that the attribute value has returned to the preset threshold range, thus completing the closed loop and achieving automatic repair. For example, when receiving a third decision instruction containing a current compensation value, the interactive execution layer encapsulates this compensation value into a current adjustment instruction recognizable by the welding robot and sends it to the corresponding welding robot controller. The welding robot continues to work according to the adjusted current value. The interactive execution layer continuously monitors the robot's current value feedback until it confirms that the current value has returned to the preset threshold range, thus completing the closed loop.

[0057] When the interactive execution layer receives multiple decision commands simultaneously, it queues and schedules them according to the execution priority carried in each command, prioritizing commands related to production line safety risks or downtime risks to ensure that critical faults and configuration changes can be responded to with the shortest possible delay. Simultaneously, the interactive execution layer provides real-time feedback on the execution status of each command (e.g., issued, executing, completed, or failed) to the operation terminal, allowing managers to understand the command execution progress.

[0058] In summary, by employing the above technical solution, this application has the following advantages:

[0059] First, this application constructs a digital twin knowledge graph of the welding workshop, containing entity instances and their semantic relationships, through a semantic data fusion layer. This unifies the data of physical equipment, tooling, materials, and process parameters, which were originally scattered across different systems in the workshop, into a single semantic model, eliminating data silos and enabling the computer system to explicitly express and utilize the inherent relationships between entity instances. Based on this, the large-model cognitive reasoning layer utilizes the semantic relationships in the knowledge graph for multi-hop causal reasoning. Starting from abnormal equipment, it traces upstream related entity instances layer by layer along semantic relationships. By comparing the deviation of the attribute values ​​of each entity instance along the way from preset thresholds, it accurately locates the root cause entity instance of the fault and pushes the root cause and repair information to the operation terminal through the interactive execution layer. This significantly improves the efficiency and accuracy of fault diagnosis, overcoming the inefficient method of engineers manually searching information across systems and inferring faults based on experience in traditional approaches.

[0060] Secondly, the large model cognitive reasoning layer can also respond to natural language instructions that represent the intention to change the production line configuration. It can extract the attributes or semantic relationships of the entity instances to be modified from the natural language instructions, automatically modify the knowledge graph and generate a second decision instruction carrying the set of equipment configuration parameters. This instruction is then sent to the corresponding physical equipment in the workshop through the interactive execution layer. This enables the system to be quickly reconstructed at the semantic level in the scenario of production line configuration change, without the need for manual modification of PLC code and equipment parameters one by one, which greatly reduces the downtime and labor costs of production line adjustment.

[0061] Furthermore, the semantic data fusion layer maintains data consistency within the knowledge graph. When the attribute value of an entity instance is updated, the association state attributes of its one-hop adjacent entity instances are automatically updated based on semantic relationships. This ensures that the state information of associated entity instances in the knowledge graph remains synchronized and consistent, providing an accurate data foundation for causal reasoning in the large model's cognitive reasoning layer. Additionally, when the large model's cognitive reasoning layer identifies multiple candidate root source entity instances through multi-hop causal reasoning, it calculates the deviation of each candidate root source entity instance's attribute value from a preset threshold. The candidate root source entity instance with the largest deviation is then identified as the fault root source entity instance, further improving the accuracy and certainty of root source localization.

[0062] Building upon this foundation, the large-scale model cognitive reasoning layer, when the attribute value deviation of the root cause entity instance is within a preset compensable range, calculates compensation parameters based on the attribute values ​​of its one-hop adjacent entity instances and generates a third decision instruction. This instruction is then issued to the physical device by the interactive execution layer in a closed-loop control manner, achieving automatic fault repair without manual intervention. Simultaneously, during the reasoning process, the large-scale model cognitive reasoning layer generates a structured chain of reasoning evidence according to the sequence of each hop in the reasoning path. The interactive execution layer displays this chain of reasoning evidence as a visualized directed graph in the 3D scene of the operating terminal, allowing users to intuitively view the complete reasoning path from the faulty device to the root cause and the state deviation of each intermediate entity instance, significantly enhancing the interpretability and credibility of the reasoning results. When pushing fault root cause and repair information, the interactive execution layer also locates and highlights the physical location of the root cause entity instance in the 3D scene of the operating terminal based on its spatial location information in the knowledge graph, enabling maintenance personnel to quickly pinpoint the specific location of the fault within the workshop layout.

[0063] In addition, the system can optionally be configured with a knowledge evolution layer, which extracts invalid branch paths from the reasoning evidence chain of each multi-hop causal reasoning in real time and transforms them into pruning rules, and extracts valid paths and transforms them into priority guidance rules, and directly injects them into the reasoning process of the cognitive reasoning layer of the large model. This enables the real-time reuse of reasoning experience, and makes the reasoning efficiency under common failure scenarios continuously improve as the system runs longer.

[0064] It should be understood that the above-described white body welding workshop control system is merely exemplary, and those skilled in the art can make various modifications based on the above method, and such modified solutions also fall within the protection scope of this application.

[0065] 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.

[0066] 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, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0067] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0068] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0069] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0070] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A control system for a body-in-white welding workshop, characterized in that, include: The semantic data fusion layer is deployed with a digital twin knowledge graph of the welding workshop, which includes multiple entity instances and their semantic relationships. The semantic data fusion layer is configured to collect physical equipment data in the workshop in real time and assign it to the attributes of the corresponding entity instances in the digital twin knowledge graph of the welding workshop, so that the knowledge graph is dynamically updated with the status of the physical equipment. The large model cognitive reasoning layer is configured to receive natural language instructions or system alarm events. When the input is the system alarm event, it determines the abnormal device based on the system alarm event, and takes the entity instance corresponding to the abnormal device as the starting point, performs multi-hop causal reasoning through the semantic association relationship to identify the entity instance of the root cause of the fault, and generates a first decision instruction carrying the root cause of the fault and corresponding repair information. When the input is a natural language instruction representing the intention to change the production line configuration, the attributes or semantic relationships of the entity instance to be modified are extracted from the natural language instruction, and the corresponding attributes or semantic relationships in the welding workshop digital twin knowledge graph are modified. A second decision instruction carrying a set of equipment configuration parameters is generated based on the modified knowledge graph. The interactive execution layer has its input end coupled to the large model cognitive reasoning layer to receive the first decision instruction or the second decision instruction, and its output end is communicatively connected to the workshop physical equipment and the operation terminal; the interactive execution layer is configured to push the fault root cause and repair information carried in the first decision instruction to the operation terminal, or to encapsulate the set of equipment configuration parameters carried in the second decision instruction into equipment-level control instructions and send them to the corresponding workshop physical equipment.

2. The system according to claim 1, characterized in that, The semantic associations include at least one or more of the following relationships: equipment manipulation relationship, tooling clamping relationship, material flow relationship, and spatial location association between workstation and equipment.

3. The system according to claim 1, characterized in that, The semantic data fusion layer is further configured to generate an abnormal status label and associate it with the corresponding entity instance when the attribute value of an entity instance in the knowledge graph deviates from a preset threshold. The abnormal status label is used to trigger the system alarm event.

4. The system according to claim 1, characterized in that, When the large model cognitive reasoning layer performs multi-hop causal reasoning through the semantic association, it takes the entity instance corresponding to the abnormal device as the starting point and traces back to the upstream associated entity instance layer by layer along the semantic association. By comparing the deviation of the attribute values ​​of each entity instance along the way in the welding workshop digital twin knowledge graph from the corresponding preset threshold, the root cause entity instance of the fault is located.

5. The system according to claim 2, characterized in that, The large model cognitive reasoning layer is specifically configured to, when the input is a natural language instruction representing the intention to change the production line configuration, extract the equipment identifier to be replaced and the new equipment specification parameters from the natural language instruction, determine the attribute or semantic association of the entity instance to be modified based on the equipment identifier to be replaced and the new equipment specification parameters, modify the corresponding attribute or semantic association in the welding workshop digital twin knowledge graph, and generate a second decision instruction carrying a set of equipment configuration parameters based on the modified knowledge graph; wherein, the modification includes: removing the spatial location association between the entity instance corresponding to the equipment to be replaced and the entity instance corresponding to the workstation, creating a new entity instance and establishing a spatial location association between it and the entity instance corresponding to the workstation, and adjusting the associated process parameter attribute values ​​according to the new equipment specification parameters.

6. The system according to claim 1, characterized in that, The interactive execution layer is specifically configured to push the root cause of the fault and repair information carried in the first decision instruction to the operation terminal, and also to locate and highlight the physical location corresponding to the root cause entity instance in the three-dimensional scene of the operation terminal according to the spatial location information of the root cause entity instance in the knowledge graph.

7. The system according to claim 6, characterized in that, The large model cognitive reasoning layer is further configured to calculate compensation parameter values ​​and generate a third decision instruction carrying the compensation parameters when the attribute value deviation of the fault root cause entity instance is within a preset compensable range, based on the attribute values ​​of one-hop adjacent entity instances in the welding workshop digital twin knowledge graph that have a direct semantic relationship with the fault root cause entity instance; the interactive execution layer is further configured to encapsulate the compensation parameters carried in the third decision instruction into a closed-loop control instruction and send it to the physical device.

8. The system according to claim 1, characterized in that, The large model cognitive reasoning layer is also configured to calculate the degree of deviation between the attribute value of each candidate root cause entity instance and a preset threshold when the input is the system alarm event and multiple candidate root cause entity instances are identified after multi-hop causal reasoning, and to determine the candidate root cause entity instance with the largest deviation as the fault root cause entity instance.

9. The system according to claim 1, characterized in that, The semantic data fusion layer is also configured to automatically update the association status attributes of one-hop adjacent entity instances that have a direct semantic relationship with the entity instance when the attribute value of an entity instance in the knowledge graph is updated, based on the semantic association relationship.

10. The system according to claim 1, characterized in that, When performing multi-hop causal reasoning, the large model cognitive reasoning layer generates a structured chain of reasoning evidence according to the order of each hop in the reasoning path. Each hop of the chain of reasoning evidence records the identifier of the starting entity instance, the type of semantic relationship traversed, the identifier of the intermediate entity instance reached, and the degree of deviation of the attribute value of the intermediate entity instance. The first decision instruction includes the chain of reasoning evidence. When the interactive execution layer pushes the first decision instruction to the operating terminal, it displays the chain of reasoning evidence in the form of a visual directed graph in the three-dimensional scene of the operating terminal. The nodes in the visual directed graph are marked with different colors according to the degree of deviation of the attribute value, and the edges in the visual directed graph are labeled with their corresponding semantic relationship types.

Images