Industrial equipment digital twin model automatic construction and cognition system and method
By constructing a multimodal large model and a four-layer hierarchical knowledge graph system, the problems of isolated models, single dimensions, and lagging maintenance in the digital modeling of industrial equipment are solved. This enables rapid and intelligent management of equipment and efficient digital twin simulation, supports natural language interaction and model self-evolution, and improves equipment interoperability and model accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGSHU INSTITUTE OF TECHNOLOGY
- Filing Date
- 2026-04-02
- Publication Date
- 2026-06-19
Smart Images
Figure CN122244394A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial internet technology, specifically to an automatic construction and cognition system and method for digital twin models of industrial equipment. Background Technology
[0002] With the deepening of Industry 4.0 and intelligent manufacturing, the digital integration of industrial equipment has become the foundation for achieving intelligent management, optimizing production efficiency, and reducing operation and maintenance costs. However, the current process of digital modeling for industrial equipment still faces many technical challenges, which seriously restrict the progress of industrial digital transformation, as follows:
[0003] 1. Isolated models and lack of interoperability: Different equipment manufacturers and industrial platforms use their own independent data model standards (such as EDDL, FDT, OPC UA, Modbus, BACnet, etc.), creating the "Rosetta Stone dilemma." The same model of equipment needs to be modeled repeatedly on different platforms, and the model data of different platforms cannot be shared or reused, forming new data silos and increasing the cost and difficulty of industrial system integration.
[0004] 2. Limited Model Dimensions: Existing digital models mainly focus on describing the static parameters of equipment (such as register addresses, rated power, and dimensions), lacking the characterization of dynamic behavior (such as efficiency curves, thermal characteristics, and response characteristics) and fault modes (such as fault codes, fault causes, and repair methods). This makes them unable to support advanced application scenarios such as deep digital twin simulation and predictive maintenance.
[0005] 3. Lagging Model Maintenance: Industrial equipment undergoes changes such as firmware upgrades, parameter adjustments, and component replacements during its life cycle. However, existing digital models cannot automatically synchronize these changes, requiring manual modification of model parameters. This leads to a gradual disconnect between the digital twin and the physical entity, compromising the accuracy and timeliness of the model and consequently affecting the reliability of subsequent applications.
[0006] 4. Low level of intelligence: Existing digital models are only used for accessing and displaying equipment data. They do not have the ability to perform intelligent diagnosis, fault prediction, performance analysis and natural language interaction based on the model. They fail to give full play to the potential value of equipment data and cannot meet the high-level needs of intelligent industrial management.
[0007] Therefore, there is an urgent need in this field for a system and method that can automatically, accurately, and multidimensionally construct digital twin models of industrial equipment and realize intelligent cognition and continuous evolution of the equipment, in order to solve the above-mentioned technical problems and promote the in-depth development of industrial digital transformation. Summary of the Invention
[0008] The purpose of this invention is to provide an automatic construction and cognition system for digital twin models of industrial equipment based on multimodal large models, so as to realize the paradigm leap of industrial equipment from "accessible" to "cognizable, simulable, predictable, and evolvable", reduce modeling costs, and improve the level of intelligent management of industrial equipment.
[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0010] An automated construction and cognition system for digital twin models of industrial equipment includes:
[0011] A multimodal information acquisition and parsing engine is used to automatically extract static attributes, dynamic behaviors, and communication protocol information of industrial equipment from various heterogeneous data sources, and generate structured data with confidence labels and traceability information.
[0012] The four-layer hierarchical knowledge graph construction module is connected to the multimodal information acquisition and parsing engine and is used to instantiate structured data into knowledge graph nodes and relationships to realize standardized organization and reasoning reuse of knowledge.
[0013] A multidimensional digital twin model generator, connected to the four-layer hierarchical knowledge graph construction module, is used to transform information in the knowledge graph into an executable multidimensional digital twin model. The model includes at least a communication model, a behavior model, and a fault model, forming a unified model bundle.
[0014] The cognitive interaction and self-evolution engine, connected to the multidimensional digital twin model generator, is used to provide a natural language interaction interface based on a large language model, and to trigger automatic calibration and version updates of the model based on the deviation between real-time data and the behavioral model.
[0015] Preferably, the multimodal information acquisition and parsing engine includes:
[0016] A multimodal data acquisition device for parallel acquisition of unstructured documents, semi-structured data, image data, video data, and structured data;
[0017] The hybrid parsing strategy module is used to perform semantic understanding of unstructured data using an AI multimodal large model parser and to accurately extract structured data using a rule engine parser.
[0018] The confidence scoring and source tracing module is used to attach a confidence score, information source anchor and parsing timestamp to each extracted information field. When the confidence of a key parameter is lower than the preset threshold, a human-machine collaborative review work order is automatically triggered, and the review result is used as a reinforcement learning signal to optimize the parsing engine.
[0019] The conflict resolution module is used to automatically resolve conflicts when the analysis results of the large model are inconsistent with those of the rule engine, based on confidence scores and rule priorities.
[0020] Preferably, the four-layer hierarchical knowledge graph construction module includes:
[0021] The four-layer knowledge graph architecture includes a basic parameter layer, a device type layer, a brand manufacturer layer, and a specific model layer; the specific model layer nodes automatically inherit all the basic parameters of the device type layer nodes.
[0022] A knowledge graph builder is used to construct a knowledge graph from the parsed results through inheritance mechanisms, relation extraction, and alignment fusion methods.
[0023] A knowledge reasoning engine to support type reasoning, compatibility reasoning, and default value reasoning.
[0024] Preferably, the specific process of the knowledge graph builder includes:
[0025] A. Entity Recognition and Hierarchical Classification: This involves identifying entity types and determining hierarchical classification from the structured data with confidence labels output by the multimodal information acquisition and parsing engine.
[0026] Identify the device type entity (such as "photovoltaic inverter") and classify it into the L2 device type layer;
[0027] Identify brand manufacturers (such as "Huawei" and "Sungrow Power") and classify them into the L3 brand manufacturer layer;
[0028] Identify specific model entities (such as "Huawei SUN2000-100KTL") and classify them into the L4 specific model layer;
[0029] Identify basic parameter entities (such as "voltage", "current", "efficiency") and classify them into the L1 basic parameter layer;
[0030] B. Inheritance mechanism implementation: Based on the hierarchical relationship of the four-level hierarchical architecture, inheritance links between entities are automatically established.
[0031] When a specific model layer node (such as "Huawei SUN2000-100KTL") is created, the system automatically finds the corresponding device type node in the L2 layer based on its "device type" attribute (such as "photovoltaic inverter") and establishes an "inherited" relationship edge.
[0032] All attributes (including basic parameters, common behavioral characteristics, etc.) of the inherited node (L2 layer) are automatically copied or referenced to the L4 layer node, avoiding the redefinition of parameters and realizing knowledge reuse;
[0033] At the same time, establish a "production" relationship edge between L4 nodes and L3 brand manufacturer nodes (e.g., "Huawei" produces "SUN2000-100KTL") to form a complete brand-model association;
[0034] C. Relation extraction and linking: Identify semantic relationships between entities from the parsing results and construct relation edges for the knowledge graph:
[0035] Parameter attribution relationship: Identify the "device has parameters" relationship, establish "has_parameter" relationship edges between L4 layer device model nodes and specific parameter nodes (such as "maximum efficiency" and "rated power"), and attach attributes such as parameter value (such as "98.6%)", unit, confidence level, etc.
[0036] Protocol support relationship: Identify the "device supports communication protocols" relationship and establish a "supports_protocol" relationship edge between the device model node and the protocol node (such as "Modbus TCP" or "OPC UA").
[0037] Functional composition relationship: Identify the "device contains components" relationship and establish composition relationship edges between the device and its sub-components (such as "MPPT module" and "cooling fan");
[0038] Fault Association: Identify the "equipment has fault codes" relationship, establish a "has_fault_code" relationship edge between the equipment model node and the fault code node (such as "F001" and "F021"), and associate information such as fault description, severity level, and repair suggestions;
[0039] Type membership: Establish a "belongs_to" relationship edge between L4 nodes and L2 nodes to clarify the equipment category to which the model belongs;
[0040] D. Alignment and fusion processing: merging and resolving conflicts of the same entity or relationship from different data sources.
[0041] Entity alignment: Based on entity name, attribute features, and context information, similarity calculation and rule matching are used to identify multiple descriptions pointing to the same entity (such as "Huawei SUN2000-100KTL" may be written as "SUN2000-100KTL" or "Huawei 100kW inverter" in different documents) and merge them into a single entity node.
[0042] Attribute fusion: When the same entity has attribute values from multiple sources, the optimal value is selected based on the confidence score, or multiple confidence values are fused (such as taking the average value or range value), and the source information of each source is retained;
[0043] Relation merging: Remove duplicate relation edges and merge relation attributes of the same type to form a unified knowledge representation;
[0044] E. Knowledge Graph Storage and Indexing: The completed knowledge graph is stored in a graph database, and a multi-dimensional index is established.
[0045] Indexes are built by level, supporting fast retrieval of entities at specific levels (such as all device types in L2).
[0046] Indexes are built based on entity type, enabling quick location of knowledge nodes of specific types;
[0047] Indexes are created based on relation types to support efficient relational queries and inference;
[0048] Each entity and relationship is accompanied by version information, confidence score, source anchor, and timestamp, enabling full lifecycle traceability of knowledge.
[0049] Preferably, the multidimensional digital twin model generator transforms information from the knowledge graph into an executable multidimensional digital twin model, specifically including:
[0050] The model mapper traverses the knowledge graph to find the L4 layer nodes corresponding to the target device model, their inherited L2 layer nodes, and associated L3 layer nodes, performing structured mapping according to a predefined model template.
[0051] Communication model mapping: Extract the node information connected by the "communication parameter" relationship edge associated with the device model in the graph, including the communication protocol type (such as Modbus, OPC UA), register address mapping table, data type definition, unit, scaling factor and access permission, and populate it into the JSON Schema template of the communication model to generate a machine-readable protocol configuration file;
[0052] Behavioral model mapping: Extract the knowledge nodes connected by relational edges such as "having performance curves", "having thermal characteristics", and "having aging patterns" associated with the device model in the graph, and convert the function expressions (such as efficiency curve η=f(load), thermal resistance coefficient), parameter value range, and fitting formula into executable code snippets, and encapsulate them into Python classes or function libraries of behavioral models.
[0053] Fault Model Mapping: Extract the knowledge nodes connected by relational edges such as "having fault codes", "having fault modes", "having diagnostic rules", and "having repair suggestions" associated with the equipment model in the graph, and convert the fault code dictionary, data change patterns (such as "temperature rises >15°C within 30 seconds"), diagnostic decision tree logic, and maintenance operation description into a structured fault model JSON file or a diagnostic script executable by the rule engine.
[0054] The model compiler compiles the intermediate representation files (JSON configuration, Python code snippets, diagnostic rule scripts) generated by the above mapping into a unified digital twin model package that can be directly loaded and called by the industrial IoT platform. The model package contains model metadata (version number, applicable device model, generation time), a list of dependent libraries, and runtime interface definitions.
[0055] The model validator verifies the accuracy of the newly generated model bundle using historical running data or simulation test data after the model is generated. It calculates the deviation between the model's predicted values and the actual values, detects internal logical conflicts in the model (such as duplicate communication addresses or contradictory fault rules), and marks the model as "verified" after verification and stores it in the model repository for runtime use.
[0056] Preferably, the multidimensional digital twin model in the multidimensional digital twin model generator includes:
[0057] The communication model includes a protocol mapping table, data type definitions, and access permission information.
[0058] Behavioral models, including efficiency curve functions, heat loss models, response characteristic models, and aging models, are used to describe the dynamic characteristics and physical behavior of equipment.
[0059] The fault model includes a fault code dictionary, a fault mode library, a diagnostic decision tree, and repair suggestions.
[0060] Preferably, the cognitive interaction and self-evolution engine includes:
[0061] The natural language interaction module is used to realize intent recognition, semantic parsing, multi-turn dialogue and result generation functions based on a large language model;
[0062] The model self-evolution module includes a deviation detection mechanism, a model calibration process, a performance degradation prediction mechanism, and a version management function. The deviation detection mechanism is used to continuously monitor the deviation between the real-time data of the device and the predicted values of the behavior model. When the deviation exceeds a preset threshold and exhibits batch characteristics, the model update process is triggered.
[0063] Preferably, the model self-evolution module includes:
[0064] (a) The deviation detection mechanism continuously monitors the deviation between real-time equipment data and behavioral model predictions, and uses a sliding window statistical method to calculate the mean, standard deviation, and trend of the deviation:
[0065] Set real-time monitoring value The model predicts the value. Then the deviation rate is defined. The system calculates the average deviation rate using a fixed sliding window. ,when If the value exceeds a preset threshold and the duration exceeds a set time, it is marked as abnormal; at the same time, a time-series trend analysis algorithm is used to calculate the slope of the deviation change. When k>0 and continues to rise, the early warning model may drift.
[0066] When multiple devices of the same model exhibit similar deviation patterns, a clustering analysis algorithm is used to calculate the similarity of the deviation sequences of the multiple devices: the distance between the deviation curves is calculated using the dynamic time warping method. When the deviation curves of multiple devices are clustered into the same cluster and the average distance within the cluster is less than the threshold, it is determined to be a batch feature, triggering the model calibration process.
[0067] (b) Model calibration process, including:
[0068] When the deviation occurs only in a single device, the system calculates the outlier index of that device's deviation compared to the average deviation of similar devices. ;in, This represents the outlier value of a single device. This represents the arithmetic mean of the deviations of similar equipment. This indicates other experimental measurements excluding outliers;
[0069] Then when the outlier index When the value exceeds the threshold, the individual device is determined to be abnormal, an alarm is triggered, and diagnostic suggestions are pushed.
[0070] When multiple devices of the same model exhibit the same deviation pattern, the system automatically collects the operating data of all devices of that model within a certain period, refits the behavioral model parameters using the least squares method or neural network regression, generates a new version of the model and marks it as "to be verified", and calculates the goodness of fit R2R2 and root mean square error RMSE between the new model and historical data as audit reference indicators.
[0071] Incremental learning mechanism: For slight deviations (such as an average deviation rate of 3%-5%), the system uses an online learning algorithm to fine-tune the model parameters without changing the main structure of the model, generate a model patch version, and record the amount of change in patch parameters and the effective time.
[0072] (c) Performance degradation prediction mechanism: Based on long-term equipment operation data, the exponential smoothing method or LSTM time series prediction model is used to predict the trend of key performance indicators (such as efficiency, temperature, vibration amplitude) and calculate the performance degradation rate. ;
[0073] in, This represents the key performance indicator characteristics at time point t+n. This represents the key performance indicator characteristics at time point t; at this point, based on the performance degradation rate... Predict the device performance at various time points within a certain future period, and generate preventative maintenance recommendations in advance when alarm thresholds are exceeded;
[0074] (d) The version management function maintains the complete version history of the model, including major versions (major structural updates), minor versions (parameter refitting), and patch versions (incremental fine-tuning). Each version records the update time, update reason, review record, version effective scope (such as the list of applicable devices), and version rollback dependencies; it supports one-click rollback to any historical version and automatically detects compatibility with the currently running data during rollback.
[0075] Preferably, it also includes a model version repository and runtime management module, used to store all historical version model bundles and manage the deployment strategies of different versions in the runtime environment:
[0076] Gray-scale deployment mechanism: After the new version model is generated, the system first selects 5%-10% of the same model devices for gray-scale deployment, continuously monitors the deviation changes for 24 hours. If the average deviation rate of the new version is reduced by more than the threshold compared with the old version and there are no abnormal alarms, the deployment will be automatically expanded to all devices.
[0077] Version Comparison Analysis: The system automatically generates a performance comparison report between the old and new versions of the model, including the prediction error distribution, anomaly detection accuracy, and computing resource consumption indicators for each load range, to assist experts in reviewing and making decisions;
[0078] Model lineage tracing: Records the derivation relationships of each version (e.g., v2.0 is generated by refitting the parameters of v1.0), constructs a lineage map of model versions, and supports tracing queries.
[0079] An automatic construction and cognition method for digital twin models of industrial equipment includes the following steps:
[0080] Step S1: Through the multimodal information acquisition and parsing engine, the structured information of industrial equipment is automatically extracted from various heterogeneous data sources, and structured data with confidence labels and traceability information is generated.
[0081] Step S2: Using a four-layer hierarchical knowledge graph construction module, the structured data is constructed into a knowledge graph according to the hierarchical structure of "basic parameters-equipment type-brand-model";
[0082] Step S3: Using a multidimensional digital twin model generator, transform the information in the knowledge graph into an executable multidimensional digital twin model, including at least a communication model, a behavior model, and a fault model;
[0083] Step S4: Through cognitive interaction and self-evolution engine, provide natural language interaction interface, and trigger automatic model calibration and version update based on the deviation between real-time data and behavioral model.
[0084] Compared with the prior art, the beneficial effects achieved by the present invention are:
[0085] This invention reduces the modeling time for a single device from 30-60 minutes to 1-2 minutes, improving modeling efficiency by over 90%, greatly reducing the implementation cost and manpower investment of industrial equipment digitization projects, and breaking through the bottleneck of the "first mile of digital transformation".
[0086] This invention employs a hybrid parsing strategy of "AI large model + rule engine", combined with confidence scoring and human-machine collaborative review mechanism, to ensure zero error in key parameters (such as communication address and rated parameters) and avoid on-site failures and economic losses caused by parameter configuration errors.
[0087] This invention is the first to integrate communication model (static layer), behavior model (dynamic layer), and fault model (experience layer) into a unified "model bundle", enabling digital twins not only to achieve device access, but also to support advanced applications such as simulation, prediction, and diagnosis, thus upgrading devices from "accessible" to "understandable, simulable, and predictable".
[0088] This invention achieves standardization and unification of heterogeneous data through a four-layer hierarchical knowledge graph, breaks down model barriers between different manufacturers and platforms, enables the reuse of device models and data interoperability, and eliminates data silos.
[0089] This invention utilizes a model self-evolution mechanism to automatically trigger model calibration and updates based on real-time operational data from physical devices, ensuring that the digital twin and the physical entity remain synchronized and guaranteeing the model's accuracy and timeliness.
[0090] This invention deeply integrates technologies such as multimodal large models, knowledge graphs, digital twins, and reinforcement learning to construct a complete and difficult-to-replicate technological innovation system. It has extremely high commercial value and strategic significance, and can provide core support for the digital transformation of industry. Attached Figure Description
[0091] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0092] Figure 1 This is an architecture diagram of an automatic construction and cognition system for digital twin models of industrial equipment according to the present invention;
[0093] Figure 2 This is a schematic diagram of the structure of the multimodal information acquisition and analysis engine of the present invention;
[0094] Figure 3This is a schematic diagram of the structure of the four-layer hierarchical knowledge graph of this invention;
[0095] Figure 4 This is a schematic diagram of the composition of the multidimensional digital twin model bundle of the present invention;
[0096] Figure 5 This is a flowchart illustrating the self-evolution mechanism of the model in this invention. Detailed Implementation
[0097] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0098] Please see Figures 1-5 The present invention provides the following technical solution:
[0099] Example 1: An automatic construction and cognition system for digital twin models of industrial equipment, comprising:
[0100] A multimodal information acquisition and parsing engine is used to automatically extract static attributes, dynamic behaviors, and communication protocol information of industrial equipment from various heterogeneous data sources, and generate structured data with confidence labels and traceability information.
[0101] Preferably, the multimodal information acquisition and parsing engine includes:
[0102] 1) The system receives the user-inputted device model "Huawei SUN2000-100KTL". The multimodal data acquisition unit automatically searches for and downloads the following relevant information for this model, and collects multi-source data in parallel, including:
[0103] Official product specifications (PDF format, 32 pages), including equipment rated parameters, performance indicators, dimensions, etc.
[0104] Modbus communication protocol document (PDF format, 15 pages), including register mapping table, communication parameters, etc.
[0105] The manufacturer's official website technical parameters page includes the latest firmware version of the device, upgrade instructions, etc.
[0106] The user manual includes wiring diagrams and fault code tables, which contain information on device wiring methods and fault code descriptions.
[0107] 2) The multimodal large model parser performs parallel parsing on the above data to extract key information:
[0108] Extract the following static parameters from the datasheet: rated power 100kW, maximum efficiency 98.6%, number of MPPTs 6, DC input voltage range 500-1100V, etc.
[0109] Extract the following from the communication protocol document: Modbus register mapping table, which contains 128 parameters including communication address, data type, and unit;
[0110] Identify the relationship between efficiency and load rate from the performance curve graph, and fit the efficiency function. ;
[0111] Extract the following from the fault code table: 56 fault codes and their corresponding descriptions, severity levels, and causes.
[0112] 3) The rule engine parser specifically handles the register mapping table to ensure accurate extraction of communication addresses and avoid communication failures due to address errors. For recurring parameters (such as DC voltage appearing in both the datasheet and the communication protocol document), the conflict resolution mechanism selects the parameter information from the communication protocol document with higher confidence based on the confidence scores of the two (datasheet parsing confidence score 0.95, communication protocol parsing confidence score 0.98).
[0113] 4) After parsing is complete, the system generates a parsing report, displaying the confidence distribution of each type of parameter:
[0114] Communication address confidence level: 0.98 (high), no manual verification required;
[0115] Static parameter confidence level: 0.95 (high), no manual review required;
[0116] Fault code confidence level: 0.88 (medium). Among them, the confidence levels of 3 fault codes (F015, F037, F049) are lower than 0.85, and an audit work order is automatically generated and pushed to the manual audit queue.
[0117] 5) After manual review is completed, the review results are used as reinforcement learning signals to optimize the parameters of the multimodal large model parser and improve the accuracy of subsequent fault code parsing.
[0118] The four-layer hierarchical knowledge graph construction module is connected to the multimodal information acquisition and parsing engine and is used to instantiate structured data into knowledge graph nodes and relationships to realize standardized organization and reasoning reuse of knowledge.
[0119] Preferably, the four-layer hierarchical knowledge graph construction module includes:
[0120] The parsed results are instantiated as knowledge graph nodes, and the knowledge graph is constructed according to a four-layer hierarchical architecture, as follows:
[0121] L1 Basic Parameter Layer: Common basic parameters such as voltage, current, power, frequency, temperature, and efficiency are predefined;
[0122] L2 Device Type Layer: The "PV Inverter" node inherits all the basic parameters of the L1 layer and adds PV inverter-specific parameters (such as MPPT voltage range, maximum DC input power, grid connection voltage, etc.).
[0123] L3 Brand Manufacturer Layer: Information about manufacturers associated with the "Huawei" node (headquarters in Shenzhen, official website link, product lines, etc.);
[0124] L4 specific model layer: The "Huawei SUN2000-100KTL" node inherits all attributes of the L2 layer (photovoltaic inverter) and the L3 layer (Huawei), and adds the parsed unique information (such as rated power of 100kW, maximum efficiency of 98.6%, support for Modbus TCP protocol, etc.).
[0125] At the same time, the following relationship edges are constructed to realize the association between entities:
[0126] "Huawei" → Production → "SUN2000-100KTL";
[0127] “SUN2000-100KTL” belongs to the category of “Photovoltaic Inverter”;
[0128] “SUN2000-100KTL” → has parameters → “Maximum efficiency: 98.6%”;
[0129] “SUN2000-100KTL” → Supported Protocols → “Modbus TCP”.
[0130] A multidimensional digital twin model generator, connected to the four-layer hierarchical knowledge graph construction module, is used to transform information in the knowledge graph into an executable multidimensional digital twin model. The model includes at least a communication model, a behavior model, and a fault model, forming a unified model bundle.
[0131] Preferably, the multidimensional digital twin model generator transforms information from the knowledge graph into an executable multidimensional digital twin model, specifically including:
[0132] The model mapper traverses the knowledge graph to find the L4 layer nodes corresponding to the target device model, their inherited L2 layer nodes, and associated L3 layer nodes, performing structured mapping according to a predefined model template.
[0133] Communication model mapping: Extract the node information connected by the "communication parameter" relationship edge associated with the device model in the graph, including the communication protocol type (such as Modbus, OPC UA), register address mapping table, data type definition, unit, scaling factor and access permission, and populate it into the JSON Schema template of the communication model to generate a machine-readable protocol configuration file;
[0134] Behavioral model mapping: Extract the knowledge nodes connected by relational edges such as "having performance curves", "having thermal characteristics", and "having aging patterns" associated with the device model in the graph, and convert the function expressions (such as efficiency curve η=f(load), thermal resistance coefficient), parameter value range, and fitting formula into executable code snippets, and encapsulate them into Python classes or function libraries of behavioral models.
[0135] Fault Model Mapping: Extract the knowledge nodes connected by relational edges such as "having fault codes", "having fault modes", "having diagnostic rules", and "having repair suggestions" associated with the equipment model in the graph, and convert the fault code dictionary, data change patterns (such as "temperature rises >15°C within 30 seconds"), diagnostic decision tree logic, and maintenance operation description into a structured fault model JSON file or a diagnostic script executable by the rule engine.
[0136] The model compiler compiles the intermediate representation files (JSON configuration, Python code snippets, diagnostic rule scripts) generated by the above mapping into a unified digital twin model package that can be directly loaded and called by the industrial IoT platform. The model package contains model metadata (version number, applicable device model, generation time), a list of dependent libraries, and runtime interface definitions.
[0137] The model validator verifies the accuracy of the newly generated model bundle using historical running data or simulation test data after the model is generated. It calculates the deviation between the model's predicted values and the actual values, detects internal logical conflicts in the model (such as duplicate communication addresses or contradictory fault rules), and marks the model as "verified" after verification and stores it in the model repository for runtime use.
[0138] The cognitive interaction and self-evolution engine, connected to the multidimensional digital twin model generator, is used to provide a natural language interaction interface based on a large language model, and to trigger automatic calibration and version updates of the model based on the deviation between real-time data and the behavioral model.
[0139] Preferably, the cognitive interaction and self-evolution engine includes:
[0140] After system deployment, the natural language interaction module allows maintenance personnel to interact with the digital twin model of the equipment using natural language, enabling functions such as equipment status query and fault diagnosis.
[0141] System: Intent recognition → Query today's real-time operating data of inverter #3 (load rate range 50%-80%) → Call the behavior model to calculate the theoretical efficiency at the corresponding load rate → Generate efficiency comparison report → Natural language feedback: "Inverter #3's average efficiency today is 96.5%, today's average load rate is 70%, corresponding to a theoretical efficiency of 97.1%. The real-time efficiency is 0.6 percentage points lower than the theoretical value, which is a slight deviation. It is recommended to check the cleanliness of the components or slight shading."
[0142] After running for one month, the model self-evolution module detected similar efficiency deviations in 10 inverters of the same model (real-time efficiency was 0.5-0.8 percentage points lower than the theoretical value), triggering the model update process.
[0143] Preferably, the model self-evolution module includes:
[0144] (a) The deviation detection mechanism continuously monitors the deviation between real-time equipment data and behavioral model predictions, and uses a sliding window statistical method to calculate the mean, standard deviation, and trend of the deviation:
[0145] Set real-time monitoring value The model predicts the value. Then the deviation rate is defined. The system calculates the average deviation rate using a fixed sliding window. ,when If the value exceeds a preset threshold and the duration exceeds a set time, it is marked as abnormal; at the same time, a time-series trend analysis algorithm is used to calculate the slope of the deviation change. When k>0 and continues to rise, the early warning model may drift.
[0146] When multiple devices of the same model exhibit similar deviation patterns, a clustering analysis algorithm is used to calculate the similarity of the deviation sequences of the multiple devices: the distance between the deviation curves is calculated using the dynamic time warping method. When the deviation curves of multiple devices are clustered into the same cluster and the average distance within the cluster is less than the threshold, it is determined to be a batch feature, triggering the model calibration process.
[0147] (b) Model calibration process, including:
[0148] When the deviation occurs only in a single device, the system calculates the outlier index of that device's deviation compared to the average deviation of similar devices. ;in, This represents the outlier value of a single device. This represents the arithmetic mean of the deviations of similar equipment. This indicates other experimental measurements excluding outliers;
[0149] Then when the outlier index When the value exceeds the threshold, the individual device is determined to be abnormal, an alarm is triggered, and diagnostic suggestions are pushed.
[0150] When multiple devices of the same model exhibit the same deviation pattern, the system automatically collects the operating data of all devices of that model within a certain period, refits the behavioral model parameters using the least squares method or neural network regression, generates a new version of the model and marks it as "to be verified", and calculates the goodness of fit R2R2 and root mean square error RMSE between the new model and historical data as audit reference indicators.
[0151] Incremental learning mechanism: For slight deviations (such as an average deviation rate of 3%-5%), the system uses an online learning algorithm to fine-tune the model parameters without changing the main structure of the model, generate a model patch version, and record the amount of change in patch parameters and the effective time.
[0152] (c) Performance degradation prediction mechanism: Based on long-term equipment operation data, the exponential smoothing method or LSTM time series prediction model is used to predict the trend of key performance indicators (such as efficiency, temperature, vibration amplitude) and calculate the performance degradation rate. ;
[0153] in, This represents the key performance indicator characteristics at time point t+n. This represents the key performance indicator characteristics at time point t; at this point, based on the performance degradation rate... Predict the device performance at various time points within a certain future period, and generate preventative maintenance recommendations in advance when alarm thresholds are exceeded;
[0154] (d) The version management function maintains the complete version history of the model, including major versions (major structural updates), minor versions (parameter refitting), and patch versions (incremental fine-tuning). Each version records the update time, update reason, review record, version effective scope (such as the list of applicable devices), and version rollback dependencies; it supports one-click rollback to any historical version and automatically detects compatibility with the currently running data during rollback.
[0155] Example 2: An automatic construction and cognition method for digital twin models of industrial equipment, comprising the following steps:
[0156] Step S1: Through the multimodal information acquisition and parsing engine, the structured information of industrial equipment is automatically extracted from various heterogeneous data sources, and structured data with confidence labels and traceability information is generated.
[0157] Step S2: Using a four-layer hierarchical knowledge graph construction module, the structured data is constructed into a knowledge graph according to the hierarchical structure of "basic parameters-equipment type-brand-model";
[0158] Step S3: Using a multidimensional digital twin model generator, transform the information in the knowledge graph into an executable multidimensional digital twin model, including at least a communication model, a behavior model, and a fault model;
[0159] Step S4: Through cognitive interaction and self-evolution engine, provide natural language interaction interface, and trigger automatic model calibration and version update based on the deviation between real-time data and behavioral model.
[0160] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An automatic construction and cognitive system for digital twin models of industrial equipment, characterized in that, include: A multimodal information acquisition and parsing engine is used to automatically extract static attributes, dynamic behaviors, and communication protocol information of industrial equipment from various heterogeneous data sources, and generate structured data with confidence labels and traceability information. The four-layer hierarchical knowledge graph construction module is connected to the multimodal information acquisition and parsing engine and is used to instantiate structured data into knowledge graph nodes and relationships to realize standardized organization and reasoning reuse of knowledge. A multidimensional digital twin model generator, connected to the four-layer hierarchical knowledge graph construction module, is used to transform information in the knowledge graph into an executable multidimensional digital twin model. The model includes at least a communication model, a behavior model, and a fault model, forming a unified model bundle. The cognitive interaction and self-evolution engine, connected to the multidimensional digital twin model generator, is used to provide a natural language interaction interface based on a large language model, and to trigger automatic calibration and version updates of the model based on the deviation between real-time data and the behavioral model.
2. The automatic construction and cognition system for digital twin models of industrial equipment as described in claim 1, characterized in that, The multimodal information acquisition and parsing engine includes: A multimodal data acquisition device for parallel acquisition of unstructured documents, semi-structured data, image data, video data, and structured data; The hybrid parsing strategy module is used to perform semantic understanding of unstructured data using an AI multimodal large model parser and to accurately extract structured data using a rule engine parser. The confidence scoring and source tracing module is used to attach a confidence score, information source anchor and parsing timestamp to each extracted information field. When the confidence of a key parameter is lower than the preset threshold, a human-machine collaborative review work order is automatically triggered, and the review result is used as a reinforcement learning signal to optimize the parsing engine. The conflict resolution module is used to automatically resolve conflicts when the analysis results of the large model are inconsistent with those of the rule engine, based on confidence scores and rule priorities.
3. The automatic construction and cognition system for digital twin models of industrial equipment as described in claim 1, characterized in that, The four-layer hierarchical knowledge graph construction module includes: The four-layer knowledge graph architecture includes a basic parameter layer, a device type layer, a brand manufacturer layer, and a specific model layer; the specific model layer nodes automatically inherit all the basic parameters of the device type layer nodes. A knowledge graph builder is used to construct a knowledge graph from the parsed results through inheritance mechanisms, relation extraction, and alignment fusion methods. A knowledge reasoning engine to support type reasoning, compatibility reasoning, and default value reasoning.
4. The automatic construction and cognition system for digital twin models of industrial equipment as described in claim 1, characterized in that, The multidimensional digital twin model in the multidimensional digital twin model generator includes: The communication model includes a protocol mapping table, data type definitions, and access permission information. Behavioral models, including efficiency curve functions, heat loss models, response characteristic models, and aging models, are used to describe the dynamic characteristics and physical behavior of equipment. The fault model includes a fault code dictionary, a fault mode library, a diagnostic decision tree, and repair suggestions.
5. The automatic construction and cognition system for digital twin models of industrial equipment as described in claim 1, characterized in that, The cognitive interaction and self-evolution engine includes: The natural language interaction module is used to realize intent recognition, semantic parsing, multi-turn dialogue and result generation functions based on a large language model; The model self-evolution module includes a deviation detection mechanism, a model calibration process, a performance degradation prediction mechanism, and a version management function. The deviation detection mechanism is used to continuously monitor the deviation between the real-time data of the device and the predicted values of the behavior model. When the deviation exceeds a preset threshold and exhibits batch characteristics, the model update process is triggered.
6. The automatic construction and cognition system for digital twin models of industrial equipment as described in claim 5, characterized in that, The model self-evolution module includes: (a) The deviation detection mechanism continuously monitors the deviation between real-time equipment data and behavioral model predictions, and uses a sliding window statistical method to calculate the mean, standard deviation, and trend of the deviation: Set real-time monitoring value The model predicts the value. Then the deviation rate is defined. The system calculates the average deviation rate using a fixed sliding window. ,when If the value exceeds a preset threshold and the duration exceeds a set time, it is marked as abnormal; at the same time, a time-series trend analysis algorithm is used to calculate the slope of the deviation change. When k>0 and continues to rise, the early warning model may drift. When multiple devices of the same model exhibit similar deviation patterns, a clustering analysis algorithm is used to calculate the similarity of the deviation sequences of the multiple devices: the distance between the deviation curves is calculated using the dynamic time warping method. When the deviation curves of multiple devices are clustered into the same cluster and the average distance within the cluster is less than the threshold, it is determined to be a batch feature, triggering the model calibration process. (b) Model calibration process, including: When the deviation occurs only in a single device, the system calculates the outlier index of that device's deviation compared to the average deviation of similar devices. ;in, This represents the outlier value of a single device. This represents the arithmetic mean of the deviations of similar equipment. This indicates other experimental measurements excluding outliers; Then when the outlier index When the value exceeds the threshold, the individual device is determined to be abnormal, an alarm is triggered, and diagnostic suggestions are pushed. When multiple devices of the same model exhibit the same deviation pattern, the system automatically collects the operating data of all devices of that model within a certain period, refits the behavioral model parameters using the least squares method or neural network regression, generates a new version of the model and marks it as "to be verified", and calculates the goodness of fit R2R2 and root mean square error RMSE between the new model and historical data as reference indicators for review. Incremental learning mechanism: For slight deviations, the system uses an online learning algorithm to fine-tune the model parameters without changing the main structure of the model, generate a model patch version, and record the amount of change in patch parameters and the effective time. (c) Performance degradation prediction mechanism: Based on long-term equipment operation data, the exponential smoothing method or LSTM time series prediction model is used to predict the trend of key performance indicators and calculate the performance degradation rate. ; in, This represents the key performance indicator characteristics at time point t+n. This represents the key performance indicator characteristics at time point t; at this point, based on the performance degradation rate... Predict the device performance at various time points within a certain future period, and generate preventative maintenance recommendations in advance when alarm thresholds are exceeded; (d) The version management function maintains the complete version history of the model, including major version, minor version and patch version. Each version records the update time, update reason, review record, version effective scope and version rollback dependency relationship; it supports one-click rollback to any historical version and automatically detects compatibility with the current running data during rollback.
7. The automatic construction and cognition system for digital twin models of industrial equipment as described in claim 1, characterized in that, It also includes a model version repository and runtime management module, used to store all historical version bundles of models and manage the deployment strategies of different versions in the runtime environment: Gray-scale deployment mechanism: After the new version model is generated, the system first selects 5%-10% of the same model devices for gray-scale deployment, continuously monitors the deviation changes for 24 hours. If the average deviation rate of the new version is reduced by more than the threshold compared with the old version and there are no abnormal alarms, the deployment will be automatically expanded to all devices. Version Comparison Analysis: The system automatically generates a performance comparison report between the old and new versions of the model, including the prediction error distribution, anomaly detection accuracy, and computing resource consumption indicators for each load range, to assist experts in reviewing and making decisions; Model lineage tracing: Records the derivative relationships of each version, constructs a lineage graph of model versions, and supports source tracing queries.
8. A method for automatically constructing and recognizing a digital twin model of industrial equipment according to any one of claims 1-7, characterized in that, Includes the following steps: Step S1: Through the multimodal information acquisition and parsing engine, the structured information of industrial equipment is automatically extracted from various heterogeneous data sources, and structured data with confidence labels and traceability information is generated. Step S2: Using a four-layer hierarchical knowledge graph construction module, the structured data is constructed into a knowledge graph according to the hierarchical structure of "basic parameters-equipment type-brand-model"; Step S3: Using a multidimensional digital twin model generator, transform the information in the knowledge graph into an executable multidimensional digital twin model, including at least a communication model, a behavior model, and a fault model; Step S4: Through cognitive interaction and self-evolution engine, provide natural language interaction interface, and trigger automatic model calibration and version update based on the deviation between real-time data and behavioral model.