Large model driven digital delivery work digital human intelligent interaction method and system

Abstract

The application relates to the technical field of data processing, and provides a digital human intelligent interaction method and system for a large model driven digital delivery factory, wherein factory equipment data and user instructions are acquired in real time through a modal fusion module; when an abnormality is detected, a fault threshold of a corresponding equipment component in an industrial knowledge graph is matched to generate a working condition semantic vector, and then an operation instruction set is output, so that the pertinence and accuracy of the operation instruction are improved; after the intelligent interaction module receives the operation instruction set, the factory equipment data is loaded in a virtual environment for simulation, a predicted execution effect is recorded, and a risk index is comprehensively judged, so that the risk is predicted in advance through virtual simulation, and inappropriate operation instructions can be effectively avoided from being directly applied to the field; after the feedback optimization module is executed in the field, the digital human controls the equipment to acquire an actual execution effect, calculates an execution deviation, and further judges based on the execution deviation; and this mechanism enables the system to continuously adjust and optimize according to the actual execution.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for intelligent interaction of digital humans in a large-scale model-driven digital delivery factory. Background Technology

[0002] In existing technologies, digital delivery factories have gradually introduced digital humans for equipment interaction and control. They acquire equipment data in real time through industrial sensors and execute operation commands in combination with preset programs. At the same time, large models are beginning to be used for scenarios such as fault diagnosis and parameter optimization, while the construction of industrial knowledge graphs provides support for equipment association analysis and historical data reuse.

[0003] However, existing technologies still have shortcomings. First, the integration of device data and user commands is low, and multimodal information lacks deep semantic association, resulting in insufficient accuracy in fault location and operation generation. Second, the simulation verification mechanism for operation commands during digital human interaction is imperfect, relying solely on simple parameter comparison to judge execution risks, which is insufficient to cope with chain reactions under complex working conditions. Third, the feedback optimization mechanism is weak, the analysis dimension of the deviation between predicted execution effect and actual effect is singular, and the knowledge graph update and command regeneration lack dynamic association, resulting in limited self-iteration capability of the system.

[0004] Based on the shortcomings of the existing technology, the technical problems to be solved in this application include: how to generate instructions based on a large model after modal fusion of data in a digital delivery factory to achieve intelligent interaction of digital humans. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a method and system for intelligent interaction of digital humans in a large-scale model-driven digital delivery factory.

[0006] Firstly, this application provides a large-model-driven digital delivery factory digital human intelligent interaction system, which includes: a modality fusion module, an intelligent interaction module, and a feedback optimization module;

[0007] The modal fusion module is used to acquire factory equipment data and user instructions in real time. When an abnormality is detected in the factory equipment data, the module locates the equipment component through the equipment identifier and matches the fault threshold of the equipment component in the industrial knowledge graph to generate a working condition semantic vector. At the same time, the module matches the fault mode based on the working condition semantic vector to dynamically generate alternative operations. Verification parameters are dynamically bound to each alternative operation to output an operation instruction set.

[0008] The intelligent interaction module is used to receive the operation instruction set, load factory equipment data in the virtual environment and perform simulation to record the predicted execution effect, and calculate the achievement degree and risk index of the verification parameters during the simulation execution. When the risk index is less than the risk threshold, the on-site execution is triggered. When the risk index is greater than or equal to the risk threshold, the operation instruction set is re-output and the verification parameters are reset.

[0009] The feedback optimization module is used to drive the digital human to control factory equipment after triggering on-site execution, obtain the actual execution effect of the operation instruction set, calculate the execution deviation between the predicted execution effect and the actual execution effect, and determine whether to incrementally update the industrial knowledge graph or regenerate the operation instruction set based on the execution deviation.

[0010] As an optional implementation, the output logic of the operation instruction set includes:

[0011] For each alternative operation, verification parameters are dynamically bound based on the equipment features in the working condition semantic vector;

[0012] The sub-steps of the alternative operations are sequentially ordered, and a time window is allocated to each sub-step to form an operation flow that includes timing constraints.

[0013] The operation process, verification parameters, and timing constraints are encapsulated and output as an operation instruction set.

[0014] As an optional implementation, the sub-logic for generating the working condition semantic vector includes:

[0015] It can acquire factory equipment data and user commands in real time, and when an anomaly is detected in the factory equipment data, it can locate the equipment component through the equipment identifier in the factory equipment data;

[0016] Based on the located equipment components, the fault thresholds of the equipment components in the industrial knowledge graph are matched, and the abnormal data of the equipment components are matched with the fault thresholds to initially associate abnormal equipment states with fault modes.

[0017] Equipment features are extracted from real-time factory equipment data using wavelet transform, and user commands are segmented and entity recognized using natural language processing to obtain command features.

[0018] The preprocessed device features, instruction features, and abnormal device states are semantically correlated through a multi-head attention mechanism to generate a fusion feature matrix that includes spatiotemporal correlation information.

[0019] By combining the failure modes of equipment components in the industrial knowledge graph, the fused feature matrix is ​​enhanced with context to generate operating condition semantic vectors.

[0020] As an optional implementation, the sub-logic for generating the alternative operation includes:

[0021] Calculate the cosine similarity between the semantic vector of the operating condition and the fault modes in the industrial knowledge graph to select an initial candidate operation set;

[0022] The initial candidate operation set is input into the industrial big model, and the industrial big model adjusts the parameters of the initial candidate operation set based on the equipment characteristics;

[0023] Based on factory safety regulations and equipment physical limitations, constraints are constructed. Based on these constraints, the initial candidate operation set after parameter adjustment is validated according to rules, and alternative operations are generated.

[0024] As an optional implementation, the logic for resetting the verification parameters includes:

[0025] During the simulation execution, the degree of matching between the actual value of the verification parameter and the parameter threshold is calculated in real time to quantify the degree of achievement of the verification parameter;

[0026] The risk index is output by nonlinearly fitting the weighted average of the achievement of the verification parameters, the number of abnormal fluctuations, and the risk of related equipment through an industrial large model.

[0027] When the risk index is less than the preset risk threshold, on-site execution is triggered. When the risk index is greater than or equal to the risk threshold, the operation instruction set is re-output, and the risk points in the simulation execution process are located. The verification parameters are dynamically reset by querying historical parameters in the industrial knowledge graph and combining the number of abnormal fluctuations.

[0028] As an optional implementation, the recording sub-logic for predicting execution results includes:

[0029] After receiving the set of operation instructions, the factory equipment data is loaded into the virtual environment based on the digital twin and the simulation is executed. The operation sub-steps are executed one by one in the virtual environment according to the timing constraints, and the equipment status of each time window is recorded synchronously to form simulation trajectory data.

[0030] During the simulation execution, the simulation trajectory data is mapped to the predicted values ​​of the verification parameters, the predicted execution progress of the operation sub-steps, and the predicted linkage response of the associated devices, forming a simulation log;

[0031] The simulation logs are correlated with the expected goals of the operation instruction set to generate and record the predicted execution results.

[0032] As an optional implementation, the incremental update logic of the industrial knowledge graph includes:

[0033] When the execution deviation is less than or equal to the deviation threshold, it is determined to be an incremental update of the industrial knowledge graph, and the deviation features of the execution deviation are extracted.

[0034] Input the deviation features into the industrial big model to generate deviation explanation rules for the causes of deviations, and then associate the deviation explanation rules with equipment components and failure modes in the industrial knowledge graph.

[0035] Assign initial confidence levels to the deviation interpretation rules, and encapsulate the deviation characteristics, deviation causes, and deviation interpretation rules of the executed deviations into deviation cases to incrementally update the industrial knowledge graph.

[0036] As an optional implementation, the sub-logic for calculating the execution deviation includes:

[0037] The system acquires the predicted execution effect from the records and drives the digital human to control factory equipment after triggering on-site execution in order to obtain the actual execution effect of the operation instruction set.

[0038] The absolute difference between the predicted and actual values ​​of the verification parameters is calculated to obtain the numerical deviation. The deviation of the time window between the predicted and actual execution progress of the operation sub-steps is determined to obtain the timing deviation. The matching degree between the predicted linkage response and the actual linkage response of the associated equipment is judged to obtain the correlation deviation.

[0039] The bias influence weights are determined based on industrial knowledge graphs. The bias influence weights are then weighted and summed with numerical bias, temporal bias, and correlation bias to calculate the execution bias between the predicted execution effect and the actual execution effect.

[0040] As an optional implementation, the logic for regenerating the operation instruction set includes:

[0041] When the execution deviation exceeds the deviation threshold, it is determined that the operation instruction set should be regenerated, and the cause of the execution deviation is extracted.

[0042] New constraints are generated based on the causes of deviations, and differentiated operation instruction sets are generated based on the new constraints. The differentiated operation instruction sets are then simulated and verified in a virtual environment.

[0043] Based on the predicted execution results verified by simulation, the operation instruction set is selected and regenerated, and synchronously recorded in the industrial knowledge graph.

[0044] Secondly, this application provides a large model-driven digital delivery factory digital human intelligent interaction method, which includes: real-time acquisition of factory equipment data and user instructions; when abnormalities are detected in the factory equipment data, locating the equipment component through the equipment identifier and matching the fault threshold of the equipment component in the industrial knowledge graph to generate a working condition semantic vector.

[0045] Fault modes are matched based on working condition semantic vectors to dynamically generate alternative operations, and verification parameters are dynamically bound to each alternative operation to output an operation instruction set.

[0046] After receiving the set of operation instructions, the factory equipment data is loaded in the virtual environment and the simulation is executed to record the predicted execution effect;

[0047] During the simulation execution, the achievement degree and risk index of the verification parameters are calculated. When the risk index is less than the risk threshold, on-site execution is triggered. When the risk index is greater than or equal to the risk threshold, the operation instruction set is re-output and the verification parameters are reset.

[0048] After triggering on-site execution, the digital human controls factory equipment, obtains the actual execution effect of the operation instruction set, calculates the execution deviation between the predicted execution effect and the actual execution effect, and determines whether to incrementally update the industrial knowledge graph or regenerate the operation instruction set based on the execution deviation.

[0049] Compared with the prior art, the beneficial effects of this application are: by acquiring factory equipment data and user instructions in real time through the modal fusion module, when abnormal equipment data is detected, the equipment component is located by means of equipment identifier, the fault threshold of the corresponding equipment component in the industrial knowledge graph is matched to generate the working condition semantic vector, and then alternative operations are dynamically generated and the verification parameters are bound to output the operation instruction set. This process closely combines the actual state of the equipment and user needs, and improves the pertinence and accuracy of the operation instructions.

[0050] After receiving the operation instruction set, the intelligent interaction module loads the equipment data in the virtual environment for simulation, records the predicted execution effect, and calculates the achievement degree and risk index of the verification parameters. Based on the risk index, it determines whether to trigger on-site execution or re-output the operation instruction set and reset the verification parameters. By predicting risks in advance through virtual simulation, it can effectively avoid applying inappropriate operation instructions directly to the field, ensure the safe operation of equipment, and improve the effectiveness of operation execution.

[0051] After the feedback optimization module is executed on-site, it drives the digital human control device to obtain the actual execution effect, calculates the execution deviation between the predicted execution effect and the actual execution effect, and determines whether to incrementally update the industrial knowledge graph or regenerate the operation instruction set based on the execution deviation. This mechanism enables the system to continuously adjust and optimize according to the actual execution situation. Whether it is updating the industrial knowledge graph or regenerating the operation instruction set, it helps to improve the system's adaptability and intelligence level, ensuring that the system can better cope with the complex operating conditions of the factory. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0053] Figure 1 The system flowchart of the large-model-driven digital delivery factory digital human intelligent interaction system provided in the embodiments of this application is shown below.

[0054] Figure 2 A logic diagram for resetting verification parameters of the large-model-driven digital delivery factory digital human intelligent interaction system provided in the embodiments of this application;

[0055] Figure 3 The calculation sub-logic diagram of the execution deviation of the large-model-driven digital delivery factory digital human intelligent interaction system provided in the embodiments of this application;

[0056] Figure 4 This is a flowchart illustrating the method for intelligent interaction between digital humans and digital delivery factories driven by a large model, as provided in an embodiment of this application. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0058] Example 1:

[0059] like Figure 1 The diagram shown illustrates a system flowchart for a large-model-driven digital delivery factory digital human intelligent interaction system, which includes a modal fusion module, an intelligent interaction module, and a feedback optimization module, according to an embodiment of this application.

[0060] The modal fusion module is used to acquire factory equipment data and user commands in real time. When an anomaly is detected in the factory equipment data, the module locates the equipment component through the equipment identifier and matches the fault threshold of the equipment component in the industrial knowledge graph to generate a working condition semantic vector. At the same time, it matches the fault mode based on the working condition semantic vector to dynamically generate alternative operations. It dynamically binds verification parameters to each alternative operation to output an operation instruction set.

[0061] Furthermore, the sub-logic for generating the working condition semantic vector includes:

[0062] It can acquire factory equipment data and user commands in real time, and when an anomaly is detected in the factory equipment data, it can locate the equipment component through the equipment identifier in the factory equipment data;

[0063] Based on the located equipment components, the fault thresholds of the equipment components in the industrial knowledge graph are matched, and the abnormal data of the equipment components are matched with the fault thresholds to initially associate abnormal equipment states with fault modes.

[0064] Equipment features are extracted from real-time factory equipment data using wavelet transform, and user commands are segmented and entity recognized using natural language processing to obtain command features.

[0065] The preprocessed device features, instruction features, and abnormal device states are semantically correlated through a multi-head attention mechanism to generate a fusion feature matrix that includes spatiotemporal correlation information.

[0066] By combining the failure modes of equipment components in the industrial knowledge graph, the fused feature matrix is ​​enhanced with context to generate operating condition semantic vectors.

[0067] It needs to be explained that the construction of the industrial knowledge graph is based on the entire lifecycle data of factory equipment. It is achieved through multi-source information integration and structural association. First, the basic attribute data of the factory equipment is acquired, including equipment model, component composition, installation location, and technical parameters. Among them, technical parameters include fault thresholds and operating ranges. The basic attribute data serves as the basic nodes of the industrial knowledge graph, such as equipment nodes and component nodes. Second, historical operating data, fault records, and maintenance cases of the equipment are collected to extract fault modes, the relationship between faults and components, and the correspondence between maintenance operations and fault resolution, which serve as the edges between nodes. In the structuring process, nodes and relationships are stored through a graph database. Node types and relationship types are defined and standardized through ontology. Node types include equipment, components, and fault modes, while relationship types include inclusion, fault association, and operation applicability. At the same time, nodes for deviation cases are reserved in the industrial knowledge graph to store information such as execution deviation characteristics and deviation interpretation rules, so as to realize the dynamic accumulation of knowledge.

[0068] In the real-time flow of factory equipment data, it is necessary to quickly identify abnormal equipment states and pinpoint the source of problems to provide precise targets for subsequent fault analysis. A sliding window mechanism is used to monitor factory equipment data in real time, with the window size set to a certain duration and the data updated at fixed intervals, based on statistical data. The criteria define a normal operating range. When multiple consecutive points of factory equipment data exceed this normal operating range within the window, it indicates that the factory equipment data is abnormal. Relying on the unique equipment identifier built into the factory equipment data, which includes the equipment code and component serial number, when an abnormality is detected in the factory equipment data, the specific abnormal equipment component is located according to the industrial knowledge graph. This avoids the spread of abnormal states, reduces the blindness of fault diagnosis, and ensures that subsequent analysis focuses on the accurate equipment object.

[0069] Different equipment components have different fault judgment criteria. It is necessary to combine the corresponding fault thresholds to clarify the nature of the anomaly and establish a preliminary connection between the abnormal equipment state and the fault mode. Based on the located equipment component, the corresponding fault threshold is retrieved from the industrial knowledge graph. Each equipment component in the industrial knowledge graph contains multiple sets of fault thresholds. For example, bearing components include multi-dimensional fault thresholds such as vibration frequency, temperature, and amplitude. Real-time abnormal data is compared with these fault thresholds item by item, and the degree of anomaly is judged. When a certain piece of factory equipment data exceeds the corresponding fault threshold and the degree of anomaly reaches a preset level, it is associated with the fault mode recorded in the industrial knowledge graph. This quantifies the severity of the abnormal state, provides fault background information for subsequent semantic fusion, and improves the accuracy of fault analysis.

[0070] Factory equipment data and user commands come in various raw forms and need to be transformed into processable feature information for semantic-level association and fusion. Vibration signals from factory equipment data undergo multi-level wavelet transform using wavelet basis functions to decompose them into approximation coefficients and detail coefficients. The energy entropy and peak factor of the detail coefficients are then extracted as vibration features. Temperature signals are filtered by moving average to remove noise, and the mean temperature and rate of temperature change are extracted as temperature features. For user commands, if they are voice commands, features are first extracted using Mel-frequency cepstral coefficients and then converted into text. For text commands, word segmentation is performed, followed by part-of-speech tagging and entity recognition to extract command features such as the operation object, action requirements, and time constraints. This transforms unstructured data into structured features, reduces data dimensionality, highlights key information, and lays the foundation for cross-modal semantic association.

[0071] Equipment features and command features belong to different modalities, requiring the establishment of semantic-level associations to fuse equipment status and user operational intentions. Preprocessed equipment features, command features, and abnormal equipment states are concatenated to obtain an initial feature vector. This initial feature vector is then input into a multi-head attention mechanism containing multiple attention heads, each with a specific hidden layer dimension. The multi-head attention mechanism captures the semantic association between equipment status and user operational intentions by calculating the attention weights between equipment features, command features, and abnormal equipment states. For example, the association between a user's command to reduce load and equipment current overload. After multi-layer multi-head attention calculations, a fused feature matrix including spatiotemporal association information is generated. Each element in the fused feature matrix contains the association information between the equipment status and user operational intentions at the corresponding time step. This breaks down the modal barriers between factory equipment data and user commands, achieving deep fusion of cross-modal information. The fused features better reflect the overall working conditions, providing a foundation for contextual enhancement of the fused feature matrix and enabling the generated working condition semantic vector to contain richer semantic layers.

[0072] The fusion feature matrix needs to incorporate historical failure modes of equipment components to enrich semantic information and more comprehensively represent the current operating condition. Information such as recent historical failure modes and maintenance records of the equipment components are retrieved from the industrial knowledge graph. This information is then linked to the fusion feature matrix via residual connections using a graph neural network. The input is a multi-layered fully connected network, ultimately generating an operating condition semantic vector. Some dimensions of this vector correspond to the real-time fusion feature matrix, while others correspond to historical context information. This allows the operating condition semantic vector to not only contain real-time status information but also incorporate historical experience and failure patterns, enhancing its ability to represent the operating condition. This provides a comprehensive description of the operating condition for generating candidate operations, making the selection of the initial candidate operation set more closely aligned with the current actual scenario.

[0073] Furthermore, the sub-logic for generating alternative operations includes:

[0074] Calculate the cosine similarity between the semantic vector of the operating condition and the fault modes in the industrial knowledge graph to select an initial candidate operation set;

[0075] The initial candidate operation set is input into the industrial big model, and the industrial big model adjusts the parameters of the initial candidate operation set based on the equipment characteristics;

[0076] Based on factory safety regulations and equipment physical limitations, constraints are constructed. Based on these constraints, the initial candidate operation set after parameter adjustment is validated according to rules, and alternative operations are generated.

[0077] It needs to be explained that the industrial big data model is built upon the industrial domain adaptation and fine-tuning of a general big data model. Its core is incorporating the operational patterns and logic of factory equipment. First, a general big data model with strong semantic understanding and reasoning capabilities is selected as the foundation. Then, historical factory data is used for domain pre-training. Training data includes textual data such as equipment operation logs, operation command sets, fault diagnosis reports, and maintenance manuals, as well as time-series data acquired by equipment sensors, which is converted into textual descriptions, such as a temperature rise from 30℃ to 50℃ over 10 minutes. Task scenarios are designed according to different needs to fine-tune the industrial big data model. First, the industrial big data model is trained to understand the semantic relationship between factory equipment data and user commands, such as associating a pressure reduction command with a valve opening adjustment operation. Second, the industrial big data model is trained to calculate a risk index based on simulation data, such as combining the achievement rate of verification parameters and the number of abnormal fluctuations to determine the risk index. Finally, the industrial big data model is trained to analyze the causes of deviations and generate deviation explanation rules, such as inferring from numerical deviations and equipment characteristics that pressure deviations stem from unreasonable threshold settings. Finally, during the fine-tuning process, a small amount of labeled industrial scenario data is used for supervised learning to ensure that the industrial big data model's output conforms to industrial operating standards.

[0078] Based on the current operating conditions, historically validated and effective operation cases are quickly identified to provide a reference basis for subsequent operation generation. Multiple recent fault modes related to the equipment component are retrieved from the case library of the industrial knowledge graph. The cases corresponding to each fault mode are transformed into case vectors. The cosine similarity between the current operating condition semantic vector and these case vectors is calculated, and a similarity threshold is set to filter out case vectors with high similarity as the initial candidate operation set. Each case includes information such as operation, execution parameters, and applicable scenarios. This leverages historical experience to quickly narrow down the operation scope, reduce the generation of invalid operations, improve the efficiency of operation generation, and provide initial samples for parameter adjustment of the large industrial model, making the adjusted operations more practically grounded.

[0079] Historical failure mode cases may not perfectly match the detailed differences in the current operating conditions. Therefore, it is necessary to dynamically optimize the parameters of the initial candidate operation set based on real-time equipment characteristics. The initial candidate operation set is input into an industrial big data model, which uses a large amount of industrial operation case data during pre-training. The input of the industrial big data model includes the current operating condition semantic vector and the initial candidate operation set. The industrial big data model adjusts the parameters of the initial candidate operation set based on the current real-time equipment characteristics. For example, for the valve opening in cases corresponding to historical failure modes, the industrial big data model dynamically adjusts it according to the proportional relationship between the current pressure value and the pressure value in cases corresponding to historical failure modes. For the runtime, it adjusts it based on the degree of current temperature anomaly to ensure sufficient cooling. During the parameter adjustment process, the industrial big data model uses an internal attention mechanism to focus on the correlation strength between equipment characteristics and operating parameters, ensuring the rationality of the adjustment and making the operation more adaptable to the current scenario. This enhances the adaptability of the operation to the real-time operating conditions, avoids poor operation results due to parameter rigidity, provides more realistic operation content for rule verification, and makes the verified candidate operations more feasible.

[0080] Operations must comply with factory safety regulations and equipment physical limitations to ensure their safety and feasibility. Constraints, comprising multiple rules, are constructed based on factory safety procedures and equipment physical limitations. Factory safety procedures include the requirement to disconnect power and display warning signs before operating high-voltage equipment, and the need for a hot work permit for any hot work. Equipment physical limitations include maximum motor speed limits and maximum valve opening limits. The initial candidate operation set, after parameter adjustments, undergoes rule verification, matching each operation with its constraints. For example, if an operation involves cutting on a high-pressure pipeline without mentioning a power-off procedure, it violates safety regulations and is eliminated. If the motor speed setting exceeds equipment physical limitations, it violates equipment limitations and is corrected and retained. Finally, multiple operations that fully comply with the constraints are retained as candidate operations. This process filters out unsafe and infeasible operations, ensuring factory production safety, improving the actual feasibility of operations, and providing qualified operational materials for the output of the operation instruction set, ensuring a safe and reliable foundation for the final output operation instruction set.

[0081] Specifically, the output logic of the operation instruction set includes:

[0082] For each alternative operation, verification parameters are dynamically bound based on the equipment features in the working condition semantic vector;

[0083] The sub-steps of the alternative operations are sequentially ordered, and a time window is allocated to each sub-step to form an operation flow that includes timing constraints.

[0084] The operation process, verification parameters, and timing constraints are encapsulated and output as an operation instruction set.

[0085] Measurable verification criteria need to be set for each candidate operation to evaluate its performance and ensure that it achieves the expected goals. Based on the equipment features in the operating condition semantic vector, verification parameters are bound to each candidate operation through a rule engine. The rule engine contains multiple binding rules. For example, for operations with abnormal vibration, the vibration frequency stability range and amplitude range are bound; for operations with abnormal temperature, the temperature drop rate and final stable temperature are bound. For example, for the candidate operation of replacing bearings, verification parameters such as vibration frequency stability range, amplitude range, and temperature range are bound; while for the operation of adjusting valve opening, parameters such as pressure stability range and flow fluctuation range are bound. The thresholds of the verification parameters are determined based on the normal operating range of the equipment and historical good operating results. This provides a clear basis for evaluating the operation effect and facilitates the comparison between subsequent simulation verification and actual execution results.

[0086] The execution order and timing of sub-steps affect the overall operational effectiveness and require careful planning to ensure smooth execution. Analyzing the logical dependencies of each sub-step in alternative operations and prioritizing their timing is crucial. For example, pressure testing must precede valve adjustment, and startup testing must follow equipment component replacement. Based on operational complexity and equipment response time, time windows are allocated to each sub-step. For instance, the shutdown step is assigned a time window based on historical data, while the more complex bearing disassembly step is assigned a longer time window. A certain amount of redundancy is included in the time window settings to handle unforeseen circumstances. This results in an operational flow with timing constraints, including the sub-step number, operation content, and time window. This prevents operational failures caused by disordered execution order of sub-steps, ensuring planned and orderly operation and improving operational efficiency.

[0087] The operation process and verification parameters need to be integrated into a standardized format for easy reception and processing by the intelligent interaction module. The operation process, verification parameters, and timing constraints should be encapsulated in a unified format to form an operation instruction set. This encapsulated instruction set includes machine-readable instructions, generation time, verification parameters, and natural language descriptions. After encapsulation, it is output to the intelligent interaction module, with verification methods used during transmission to ensure data integrity. This achieves standardized transmission of the operation instruction set, reduces errors during information transmission, ensures the intelligent interaction module accurately understands the operation requirements, provides clear instruction guidelines for subsequent simulation execution, and enables the simulation process to accurately simulate the operation effect.

[0088] The intelligent interaction module receives the operation instruction set, loads factory equipment data in the virtual environment and performs simulation to record the predicted execution effect. During the simulation execution, it calculates the achievement degree and risk index of the verification parameters. When the risk index is less than the risk threshold, it triggers on-site execution. When the risk index is greater than or equal to the risk threshold, it re-outputs the operation instruction set and resets the verification parameters.

[0089] Furthermore, the recording sub-logic for predicting execution results includes:

[0090] After receiving the set of operation instructions, the factory equipment data is loaded into the virtual environment based on the digital twin and the simulation is executed. The operation sub-steps are executed one by one in the virtual environment according to the timing constraints, and the equipment status of each time window is recorded synchronously to form simulation trajectory data.

[0091] During the simulation execution, the simulation trajectory data is mapped to the predicted values ​​of the verification parameters, the predicted execution progress of the operation sub-steps, and the predicted linkage response of the associated devices, forming a simulation log;

[0092] The simulation logs are correlated with the expected goals of the operation instruction set to generate and record the predicted execution results.

[0093] The execution process of the operation instruction set needs to be simulated in a virtual environment to predict the execution effect in advance and avoid the risks of direct execution on physical equipment. After receiving the operation instruction set, a virtual environment consistent with the physical environment of the factory is constructed based on digital twin technology, thus forming a digital delivery factory. The physical environment of the factory includes details such as equipment size, installation location, and pipeline routing. When loading factory equipment data into the virtual environment, not only basic information such as equipment identification, connection relationship, and equipment status are imported, but also the physical characteristic parameters and historical performance degradation curves of the equipment are included. The physical characteristic parameters include the thermal conductivity of the material and the mechanical transmission efficiency, and the historical performance degradation curve includes the relationship between the motor's operating years and output power. Then, the operation instruction set is executed accordingly. The timing constraints in the simulation drive the virtual equipment's actions synchronously as each sub-step of the operation is executed in the virtual environment. For example, when adjusting the valve opening, the virtual valve gradually changes its opening according to the actual mechanical response speed. At the same time, the simulation calculates the chain reaction of pressure loss and flow rate changes in the pipeline in real time. When recording the equipment status at each time window, in addition to conventional parameters such as pressure, speed, and temperature, it also includes in-depth data such as equipment vibration spectrum, energy consumption changes, and stress values ​​of key components, forming simulation trajectory data containing multi-dimensional information. The high consistency between the virtual environment and the physical equipment ensures the authenticity of the simulation results. The multi-dimensional status records can comprehensively reflect the impact of the operation on the equipment, reduce trial and error costs, and discover potential hidden problems in advance.

[0094] The simulation trajectory data is relatively scattered and needs to be organized and mapped to form structured information in order to clearly present the expected results of the operation. During the simulation execution, the acquired simulation trajectory data is processed in real time, mapping the simulation trajectory data into predicted values ​​of verification parameters, predicted execution progress of operation sub-steps, and predicted linkage responses of related equipment. The predicted values ​​of verification parameters are obtained by filtering signals related to the verification parameters from the simulation trajectory data, such as pipeline pressure sensor data related to pressure parameters, which are then filtered and denoised to be converted into numerical ranges that conform to verification standards. The predicted execution progress of operation sub-steps is calculated by comparing the ratio of the real-time execution time of the operation sub-step to the planned time window. The predicted linkage response of associated equipment is obtained by tracking the state changes of associated equipment after the target equipment is operated. For example, after adjusting the main pump flow, the complete response path of secondary pipeline pressure change, heat exchanger temperature change, and cooling system load change is recorded. The response delay time of each associated equipment is recorded and integrated into the simulation log according to the dual dimensions of time axis and operation sub-step. Each operation sub-step includes the predicted value of the corresponding verification parameter, the predicted execution progress, and the predicted linkage response of associated equipment. The multi-dimensional data mapping and structured integration enable the simulation log to clearly present both the direct effect of the operation and the indirect impact, which facilitates the rapid location of key information and potential problems in the operation process.

[0095] It is necessary to compare the simulation logs with the expected goals of the operation instruction set to determine whether the operation can achieve the preset effect, providing a basis for judgment in subsequent field execution or instruction re-output; to perform correlation analysis between the information in the simulation logs and the expected goals of the operation instruction set, to determine whether the predicted values ​​of various verification parameters meet expectations, and to analyze the stability and convergence speed of the verification parameters, such as fluctuation amplitude, duration, and time to recover from outliers to the normal range; to determine whether the progress of the operation sub-steps is reasonable, such as checking whether the operation sub-steps are completed within the planned time window, and the execution quality of the operation sub-steps, i.e., whether the valve adjustment is accurate and in place, and the detection. Does the operation sub-step cover all key items? Is the linkage response of related equipment normal? This includes whether it conforms to the expected coordination pattern and whether the response level is within a reasonable range. For example, when the load on the main equipment increases, does the auxiliary equipment adjust accordingly in a timely manner? Based on these correlation analysis results, generate and record the predicted execution effect. Multi-dimensional comparative analysis makes the predicted execution effect more comprehensive and objective. It not only clarifies whether the operation meets the standards but also points out problems and directions for improvement, providing sufficient basis for subsequent decisions. It also provides detailed effect evaluation data for verifying the achievement of parameters and calculating the risk index, enabling risk judgment to be more accurately combined with the actual performance of the operation.

[0096] Specifically, such as Figure 2 As shown, the logic for resetting the verification parameters includes:

[0097] During the simulation execution, the degree of matching between the actual value of the verification parameter and the parameter threshold is calculated in real time to quantify the degree of achievement of the verification parameter;

[0098] The risk index is output by nonlinearly fitting the weighted average of the achievement of the verification parameters, the number of abnormal fluctuations, and the risk of related equipment through an industrial large model.

[0099] When the risk index is less than the preset risk threshold, on-site execution is triggered. When the risk index is greater than or equal to the risk threshold, the operation instruction set is re-output, and the risk points in the simulation execution process are located. The verification parameters are dynamically reset by querying historical parameters in the industrial knowledge graph and combining the number of abnormal fluctuations.

[0100] It is necessary to measure the degree to which the operation instruction set meets the verification parameters during the simulation process to determine the effectiveness and reliability of the operation. During the simulation execution, the matching degree between the actual value of the verification parameter and the parameter threshold is calculated in real time. The matching degree includes whether the actual value of the verification parameter is within the range of the parameter threshold, and the deviation of the actual value of the verification parameter from the center value of the parameter threshold, such as the percentage deviation between the actual value and the theoretical value of temperature, to quantify the achievement of the verification parameter and reflect the control effect of the operation on the verification parameter. The multi-dimensional quantification method makes the evaluation of the achievement of the verification parameter more accurate and detailed, which can fully reflect the differences in the control effect of the operation on different parameters and provide more accurate basic data for risk assessment.

[0101] Achieving a single validation parameter is insufficient to fully reflect operational risk; multiple factors must be considered to comprehensively assess the risk level of operational execution. An industrial big data model is used to perform nonlinear fitting on the weighted average of validation parameter achievement, the number of abnormal fluctuations, and the risk of associated equipment. The weighted average of validation parameter achievement reflects the overall control effectiveness of the validation parameters; the number of abnormal fluctuations refers to the number of times factory equipment data is abnormal; and the risk of associated equipment includes overload caused by operations. The industrial big data model uses deep learning algorithms to perform nonlinear fitting on this information and incorporates corrections based on similar scenarios from historical cases. For example, if fluctuations in a certain type of factory equipment data have historically led to equipment failure, the industrial big data model will correspondingly increase the risk index for this scenario. The final output risk index reflects the overall risk of operational execution. This comprehensive consideration of multiple factors and correction based on historical cases allows the risk index to more comprehensively and accurately reflect the actual risk level of the operation, avoiding the limitations of a single indicator.

[0102] Depending on the magnitude of the risk index, different handling methods are required to ensure the safety and effectiveness of the operation. The calculated risk index is compared with a preset risk threshold. If the risk index is less than the threshold, the operational risk is within an acceptable range, triggering on-site execution. If the risk index is greater than or equal to the threshold, the operation carries a high risk, requiring a re-output of the operation instruction set. Simultaneously, risk points in the simulation process are identified. For example, is the frequent exceeding of a verification parameter due to unreasonable sub-steps or inappropriate threshold settings? The cause of the risk points is determined, such as timing conflicts in sub-steps leading to fluctuations in factory equipment data or changes in equipment properties, rendering the original verification parameter thresholds unusable. Then, by querying historical cases of similar scenarios in the industrial knowledge graph, not only are similar risk scenarios matched, but the adjustment effects of these cases are also referenced. In a certain case, the degree of risk reduction after narrowing the normal operating range of factory equipment data was demonstrated. Combined with the number of abnormal fluctuations in this simulation, verification parameters were dynamically reset. For example, for factory equipment data with frequent abnormal fluctuations, the normal operating range was narrowed and the fluctuation amplitude was restricted. For factory equipment data anomalies caused by equipment aging, the normal operating range was adjusted to adapt to the current performance of the equipment. This makes risk response more precise. The accurate location of risk points and reference to historical cases ensure the rationality and effectiveness of the verification parameter reset, guaranteeing the safety and applicability of subsequent operation instruction sets. When on-site execution is triggered, detailed predicted execution effects and risk warnings are provided to the feedback optimization module, facilitating comparison of actual effects and problem tracing. When the operation instruction set is re-output, targeted reset verification parameters are provided to the modal fusion module, enabling the newly generated operation instruction set to more accurately avoid risks.

[0103] The feedback optimization module is used to drive the digital human to control factory equipment after triggering on-site execution, obtain the actual execution effect of the operation instruction set, calculate the execution deviation between the predicted execution effect and the actual execution effect, and determine whether to incrementally update the industrial knowledge graph or regenerate the operation instruction set based on the execution deviation.

[0104] It needs to be explained that the construction of the digital human focuses on equipment control capabilities and interactive adaptability, and is divided into two parts: virtual image modeling and control logic embedding. Virtual image modeling is done through 3D modeling tools to design the appearance according to the needs of the factory scene, such as tooling shape and interactive interface integration, and to preset basic operation actions through motion capture technology, such as the limb movement sequence for pressing a button and adjusting a valve. Among them, control logic embedding is the core. Through API interface, the digital human is associated with the factory equipment control system and operation instruction set. After receiving instructions from the feedback optimization module, the digital human parses the operation sub-steps and time windows in the operation instruction set and converts them into corresponding control signals. At the same time, it obtains the equipment status in real time through sensors and adjusts the virtual image's action display accordingly. For example, when adjusting a valve, the digital human's virtual arm makes the corresponding action. In addition, the digital human also integrates simple interactive response logic. For example, during the operation, when the equipment malfunctions, the virtual image provides feedback on the status through preset expressions or prompts, realizing visual interaction with the on-site environment.

[0105] Furthermore, such as Figure 3 As shown, the sub-logic for calculating the deviation includes:

[0106] The system acquires the predicted execution effect from the records and drives the digital human to control factory equipment after triggering on-site execution in order to obtain the actual execution effect of the operation instruction set.

[0107] The absolute difference between the predicted and actual values ​​of the verification parameters is calculated to obtain the numerical deviation. The deviation of the time window between the predicted and actual execution progress of the operation sub-steps is determined to obtain the timing deviation. The matching degree between the predicted linkage response and the actual linkage response of the associated equipment is judged to obtain the correlation deviation.

[0108] The bias influence weights are determined based on industrial knowledge graphs. The bias influence weights are then weighted and summed with numerical bias, temporal bias, and correlation bias to calculate the execution bias between the predicted execution effect and the actual execution effect.

[0109] Deviation analysis needs to be based on the predicted execution effect and the actual execution effect. These two are the original basis for calculating the execution deviation and must be accurately obtained and ensured to be corresponding; retrieve the recorded predicted execution effect from the intelligent interaction module. This information includes the predicted value of the verification parameter, the predicted execution progress of the operation sub-steps, and the predicted linkage response of the associated devices. Ensure that it corresponds to the operation instruction set executed on-site this time. When executing on-site, the digital human controls the operation of the factory equipment, and the factory equipment data is obtained in real-time through the sensors on the factory equipment. At the same time, record the completion status of each operation of the digital human and summarize it to form the actual execution effect, which includes the actual value of the verification parameter, the actual execution progress of the operation sub-steps, and the actual linkage response of the associated devices; thus ensuring that the basic data used for deviation calculation is complete and corresponding, providing reliable original materials for accurately calculating various deviations later.

[0110] Calculate the execution deviation from three dimensions: numerical value, time sequence, and association, and obtain these three types of deviations respectively, which can comprehensively reflect the difference between the predicted execution effect and the actual execution effect; when calculating the numerical deviation, compare the predicted value and the actual value of the verification parameter to determine the absolute difference between the two. When determining the time sequence deviation, view the time window corresponding to the predicted execution progress of the operation sub-steps and compare it with the time window of the actual execution progress to determine the deviation degree between the time windows; when judging the association deviation, analyze the predicted linkage response and the actual linkage response of the associated devices, and comprehensively judge the matching degree of the two from the timeliness and amplitude of the linkage response. The lower the matching degree, the greater the association deviation. At the same time, record the specific performance of the associated devices not responding as expected; capture deviations from three different dimensions respectively to comprehensively present the differences in all aspects during the operation execution process and avoid the one-sidedness of single-dimensional analysis.

[0111] Judge the subsequent operations based on the comprehensive execution deviation index. Since the importance of various deviations is different, it is necessary to perform comprehensive calculations in combination with the deviation impact weights; according to the preset deviation impact weights in the industrial knowledge graph, these deviation impact weights are determined based on the impact degree of different deviations on equipment operation and production. First, standardize the numerical deviation, time sequence deviation, and association deviation, then multiply the numerical deviation, time sequence deviation, and association deviation by their corresponding deviation impact weights respectively, and then add these products to obtain the execution deviation, so as to overall reflect the difference degree between the predicted execution effect and the actual execution effect; thus comprehensively considering the impacts of various deviations, the obtained execution deviation can more comprehensively and objectively reflect the overall deviation of the operation execution, providing a key quantitative basis for judging whether to incrementally update the industrial knowledge graph or regenerate the operation instruction set, making the decision more in line with the actual situation.

[0112] Specifically, the incremental update logic of the industrial knowledge graph includes:

[0113] When the execution deviation is less than or equal to the deviation threshold, it is determined to be an incremental update of the industrial knowledge graph, and the deviation features of the execution deviation are extracted.

[0114] Input the deviation features into the industrial big model to generate deviation explanation rules for the causes of deviations, and then associate the deviation explanation rules with equipment components and failure modes in the industrial knowledge graph.

[0115] Assign initial confidence levels to the deviation interpretation rules, and encapsulate the deviation characteristics, deviation causes, and deviation interpretation rules of the executed deviations into deviation cases to incrementally update the industrial knowledge graph.

[0116] When the execution deviation is less than or equal to the deviation threshold, an incremental update of the industrial knowledge graph is performed. This step requires first making a judgment and extracting deviation features to provide a basis for subsequent updates. The calculated execution deviation is compared with the preset deviation threshold. If the execution deviation is less than or equal to the deviation threshold, an incremental update of the industrial knowledge graph is determined. Then, deviation features are extracted from the execution deviation. These deviation features include the verification parameters involved in the deviation, the operation sub-steps in which the deviation occurred, and related equipment information, etc. This accurately determines the scenarios suitable for incremental updates, avoids unnecessary reconstruction of the industrial knowledge graph, and at the same time, the extracted deviation features provide specific content for the subsequent generation of deviation interpretation rules.

[0117] Integrating deviation features into an industrial knowledge graph, generating deviation interpretation rules, and associating them with relevant equipment components enables the industrial knowledge graph to better record and reflect deviation situations. The extracted deviation features are input into a large-scale industrial model, which, combined with existing knowledge and data, analyzes the causes of deviations and forms deviation interpretation rules. These rules clearly define the deviation conditions and causes. These rules are then associated with corresponding equipment components and failure modes in the industrial knowledge graph, establishing a link between deviations and equipment / failures. This provides a clear explanation of the causes of deviations and, through association with the industrial knowledge graph, enables structured storage of deviation knowledge, facilitating subsequent queries and applications. It also provides core rule content for encapsulating deviation cases, ensuring that incrementally updated industrial knowledge graphs contain complete deviation-related information.

[0118] Deviation-related information is incrementally updated into the industrial knowledge graph in the form of cases to enrich its content and provide reference for subsequent operations. An initial confidence level is assigned to the generated deviation interpretation rules, determined based on the historical processing results of similar deviations. Information such as the deviation characteristics, causes, and interpretation rules of the executed deviation are integrated and encapsulated to form a complete deviation case. This deviation case is then added to the industrial knowledge graph, completing the incremental update. This process supplements the content of the industrial knowledge graph, increasing the recording and analysis of deviations in actual operations, providing more reference for the subsequent generation of operation instruction sets. The updated industrial knowledge graph provides richer knowledge support for the modal fusion module to generate operation instruction sets, helping to improve the accuracy of the operation instruction sets.

[0119] Specifically, the logic for regenerating the operation instruction set includes:

[0120] When the execution deviation exceeds the deviation threshold, it is determined that the operation instruction set should be regenerated, and the cause of the execution deviation is extracted.

[0121] New constraints are generated based on the causes of deviations, and differentiated operation instruction sets are generated based on the new constraints. The differentiated operation instruction sets are then simulated and verified in a virtual environment.

[0122] Based on the predicted execution results verified by simulation, the operation instruction set is selected and regenerated, and synchronously recorded in the industrial knowledge graph.

[0123] When the execution deviation exceeds the deviation threshold, the operation instruction set needs to be regenerated. The cause of the deviation is key to the regeneration, making the newly generated operation instruction set more targeted. By comparing the overall execution deviation with the deviation threshold, if the execution deviation exceeds the deviation threshold, it is determined that the operation instruction set needs to be regenerated. By analyzing various deviations between the predicted execution effect and the actual execution effect, combined with the equipment's operating status and operation process, the specific causes of the deviation are identified, such as unreasonable verification parameter settings or problems with the timing arrangement of operation sub-steps. This accurately determines the scenarios in which the operation instruction set needs to be regenerated. At the same time, the clear causes of deviation provide a clear direction for improvement in subsequent regeneration work, avoiding blind adjustments and providing specific reasons for the generation of new constraints, ensuring that the constraints can effectively solve the problems that cause the deviation.

[0124] The operation instruction set is regenerated by generating new constraints based on the causes of deviations, and the effectiveness of the new operation instruction set needs to be verified through simulation. Based on the extracted causes of deviations, new constraints are formulated, such as adjusting the reasonable range of verification parameters and optimizing the timing of operation sub-steps. Based on these new constraints, the modal fusion module generates multiple different operation instruction sets, which differ in parameter settings and step arrangements. For example, the first operation instruction set only adjusts the pressure threshold, the second only extends the interval time, and the third adjusts both the pressure threshold and the interval time. Then, in a virtual environment, these differentiated operation instruction sets are simulated and executed under the same initial conditions as in-situ execution, and their predicted execution effects are observed and recorded. The new constraints ensure that the generated operation instruction set can be improved to address the causes of deviations, while simulation verification checks the effectiveness of the new operation instruction set in advance, reducing the risk of in-situ execution.

[0125] The most suitable set of operation instructions is selected from multiple differentiated sets and recorded in the industrial knowledge graph for future reference in similar scenarios. Based on the predicted execution results obtained from simulation verification, each differentiated set of operation instructions is evaluated to assess its effectiveness in resolving deviations and the feasibility of the operation. For example, the evaluation assesses whether the pressure deviation of the first set of operation instructions is reduced, whether the delay of the operation sub-steps in the second set is improved, and whether the third set of operation instructions simultaneously solves the problems of pressure deviation and operation sub-step delays. The set of operation instructions with the best evaluation results is selected as the regenerated set of operation instructions. This set, along with its generation process and simulation results, is recorded in the industrial knowledge graph for future reference and retrieval. This ensures that the regenerated set of operation instructions effectively solves deviation problems, improves the accuracy of operation execution, and, by recording it in the industrial knowledge graph, accumulates and reuses experience. It provides reliable instruction content for the regenerated set of operation instructions to be executed on-site and also provides a reference case for the feedback optimization module to handle similar situations in the future.

[0126] Example 2:

[0127] like Figure 4 The diagram shows a flowchart illustrating a large-model-driven intelligent interaction method for digital human-based delivery factories, as provided in this application embodiment. The method includes:

[0128] Real-time acquisition of factory equipment data and user commands; when an anomaly is detected in the factory equipment data, the equipment component is located by the equipment identifier, and the fault threshold of the equipment component in the industrial knowledge graph is matched to generate a working condition semantic vector.

[0129] Fault modes are matched based on working condition semantic vectors to dynamically generate alternative operations, and verification parameters are dynamically bound to each alternative operation to output an operation instruction set.

[0130] After receiving the set of operation instructions, the factory equipment data is loaded in the virtual environment and the simulation is executed to record the predicted execution effect;

[0131] During the simulation execution, the achievement degree and risk index of the verification parameters are calculated. When the risk index is less than the risk threshold, on-site execution is triggered. When the risk index is greater than or equal to the risk threshold, the operation instruction set is re-output and the verification parameters are reset.

[0132] After triggering on-site execution, the digital human controls factory equipment, obtains the actual execution effect of the operation instruction set, calculates the execution deviation between the predicted execution effect and the actual execution effect, and determines whether to incrementally update the industrial knowledge graph or regenerate the operation instruction set based on the execution deviation.

[0133] Since the principle of the method in this application embodiment is similar to that of the system described in this application embodiment, the implementation of the method is the same as that of the system, and the repeated parts will not be described again.

Claims

1. A digital delivery work digital human intelligent interaction system driven by a large model, characterized in that, Comprise: A modal fusion module, an intelligent interaction module, and a feedback optimization module; The modal fusion module is used to acquire factory equipment data and user instructions in real time, and when an abnormality in the factory equipment data is detected, it is positioned to the equipment component through the equipment identifier, and the fault threshold of the equipment component in the industrial knowledge graph is matched to generate a working condition semantic vector, and a fault mode is matched based on the working condition semantic vector to dynamically generate alternative operations, and verification parameters are dynamically bound for each alternative operation to output an operation instruction set; The intelligent interaction module is used to receive the operation instruction set, load the factory equipment data in the virtual environment and execute simulation to record the predicted execution effect, and calculate the achievement degree and risk index of the verification parameters during the simulation execution, and when the risk index is less than the risk threshold, the scene execution is triggered, and when the risk index is greater than or equal to the risk threshold, the operation instruction set is re-output and the verification parameters are reset; The feedback optimization module is used to drive the digital person to control the factory equipment after triggering the scene execution, acquire the actual execution effect of the operation instruction set, calculate the execution deviation between the predicted execution effect and the actual execution effect, and judge whether to incrementally update the industrial knowledge graph or re-generate the operation instruction set based on the execution deviation.

2. The large model-driven digital delivery work digital human intelligent interaction system of claim 1, wherein, The output logic of the operation instruction set comprises: For each alternative operation, dynamically bind the verification parameters based on the equipment features in the working condition semantic vector; The operation sub-steps of the alternative operation are time-sequentially sorted, and each operation sub-step is assigned a time window to form an operation process including time constraints; The operation process, verification parameters, and time constraints are encapsulated and output as the operation instruction set.

3. The large model-driven digital delivery work digital human intelligent interaction system of claim 2, wherein, The generation sub-logic of the working condition semantic vector comprises: Acquire factory equipment data and user instructions in real time, and when an abnormality in the factory equipment data is detected, position to the equipment component through the equipment identifier in the factory equipment data; Based on the positioned equipment component, match the fault threshold of the equipment component in the industrial knowledge graph, match the abnormal data of the equipment component with the fault threshold, and preliminarily associate the abnormal equipment state with the fault mode; Extract the equipment features of the real-time acquired factory equipment data through wavelet transform, and perform word segmentation and entity recognition on the user instructions through natural language processing to obtain instruction features; Pretreated equipment features, instruction features, and abnormal equipment states are semantically associated through a multi-head attention mechanism to generate a fusion feature matrix including spatio-temporal correlation information; Combine the fault modes of the equipment components in the industrial knowledge graph to contextually enhance the fusion feature matrix to generate a working condition semantic vector.

4. The large model-driven digital delivery work digital human intelligent interaction system of claim 3, wherein, The generation sub-logic of the alternative operation comprises: Calculate the cosine similarity between the working condition semantic vector and the fault modes in the industrial knowledge graph to filter out an initial candidate operation set; Input the initial candidate operation set into an industrial large model, and the industrial large model adjusts the parameters of the initial candidate operation set based on the equipment features; Construct constraint conditions according to the factory safety regulations and equipment physical limitations, and perform rule verification on the initial candidate operation set after parameter adjustment based on the constraint conditions to generate alternative operations.

5. The large model-driven digital delivery work digital human intelligent interaction system of claim 4, wherein, The reset logic of the verification parameters comprises: The matching degree between the actual value of the verification parameter and the parameter threshold is calculated in real time during the simulation execution process to quantify the achievement degree of the verification parameter; The risk index is output by nonlinear fitting of the weighted average value of the achievement degree of the verification parameter, the number of abnormal fluctuations, and the risk of the associated equipment through the industrial large model; When the risk index is less than the preset risk threshold, the scene execution is triggered, and when the risk index is greater than or equal to the risk threshold, the operation instruction set is re-output, and the risk point in the simulation execution process is located, the historical parameter adjustment in the industrial knowledge graph is queried, and the verification parameter is dynamically reset according to the number of abnormal fluctuations.

6. The large model-driven digital delivery work digital human intelligent interaction system of claim 5, wherein, The record sub-logic of the predicted execution effect includes: After receiving the operation instruction set, the factory equipment data is loaded and simulation is performed in the virtual environment based on the digital twin, the operation sub-steps are sequentially executed in the virtual environment according to the time sequence constraint, and the equipment state of each time window is recorded synchronously to form simulation trajectory data; During the simulation execution process, the simulation trajectory data is mapped into the predicted value of the verification parameter, the predicted execution progress of the operation sub-steps, and the predicted linkage response of the associated equipment to form a simulation log; The simulation log is associated with the expected target of the operation instruction set to generate and record the predicted execution effect.

7. The large model-driven digital delivery work digital human intelligent interaction system of claim 6, wherein, The incremental update logic of the industrial knowledge graph includes: When the execution deviation is less than or equal to the deviation threshold, it is judged that the industrial knowledge graph is incrementally updated, and the deviation feature of the execution deviation is extracted; The deviation feature is input into the industrial large model to generate a deviation explanation rule of the deviation cause, and the deviation explanation rule is associated with the equipment components and fault modes in the industrial knowledge graph; The deviation explanation rule is assigned an initial confidence, and the deviation feature, the deviation cause, and the deviation explanation rule of the execution deviation are encapsulated as a deviation case to incrementally update the industrial knowledge graph.

8. The large model-driven digital delivery work digital human intelligent interaction system of claim 7, wherein, The calculation sub-logic of the execution deviation includes: The recorded predicted execution effect is obtained, and the digital human controls the factory equipment after triggering the scene execution to obtain the actual execution effect of the operation instruction set; The absolute difference between the predicted value and the actual value of the verification parameter is calculated to obtain a numerical deviation, the deviation degree of the time window between the predicted execution progress and the actual execution progress of the operation sub-steps is determined to obtain a timing deviation, and the matching degree between the predicted linkage response and the actual linkage response of the associated equipment is judged to obtain an association deviation; The deviation influence weight is determined based on the industrial knowledge graph, and the deviation influence weight, the numerical deviation, the timing deviation, and the association deviation are weighted and summed to calculate the execution deviation between the predicted execution effect and the actual execution effect.

9. The large model-driven digital delivery work digital human intelligent interaction system of claim 8, wherein, The re-generation logic of the operation instruction set includes: When the execution deviation is greater than the deviation threshold, it is judged that the operation instruction set is regenerated, and the deviation cause of the execution deviation is extracted; A new constraint condition is generated based on the deviation cause, a differentiated operation instruction set is generated based on the new constraint condition, and simulation verification is performed on the differentiated operation instruction set in the virtual environment; The operation instruction set is selected and regenerated based on the predicted execution effect after the simulation verification, and is recorded to the industrial knowledge graph synchronously.

10. A large-model-driven digital delivery factory digital human intelligent interaction method, implemented based on any one of claims 1-9, characterized in that, It includes: Real-time acquisition of plant equipment data and user instructions, when detecting abnormal plant equipment data, positioning to the equipment component through equipment identification, and matching the fault threshold of the equipment component in the industrial knowledge graph to generate a working condition semantic vector; Based on the working condition semantic vector, match the fault mode to dynamically generate the alternative operation, dynamically bind the verification parameter for each alternative operation to output the operation instruction set; After receiving the operation instruction set, load the plant equipment data in the virtual environment and execute the simulation to record the predicted execution effect; During the simulation execution process, calculate the achievement degree and risk index of the verification parameter, when the risk index is less than the risk threshold, trigger the scene execution, when the risk index is greater than or equal to the risk threshold, re-output the operation instruction set and reset the verification parameter; After triggering the scene execution, drive the digital person to control the plant equipment, acquire the actual execution effect of the operation instruction set, calculate the execution deviation between the predicted execution effect and the actual execution effect, and judge whether to incrementally update the industrial knowledge graph or regenerate the operation instruction set based on the execution deviation.

Images