A global fusion large model driven industrial agent interaction system and implementation method

Abstract

This invention discloses a fully integrated large-scale model-driven industrial intelligent agent interaction system and its implementation method, relating to the field of industrial internet technology. It includes: a fully integrated large-scale model training and inference module for modeling, training, lightweight output, and low-latency inference; a super memory storage module for storing industrial scenario data; an intelligent retrieval engine module for outputting optimized adjustment schemes; four core functional pillar modules: a carbon energy intelligent control optimization submodule, a holographic digital twin collaboration submodule, a hyperdimensional knowledge insight and decision-making submodule, and a process AI autonomous evolution submodule; a multimodal interaction portal module supporting multimodal interaction; and a data acquisition and execution control module for acquiring data, issuing commands, and providing feedback. This invention provides a fully integrated large-scale model-driven industrial intelligent agent interaction system and its implementation method, solving the problems of shortcomings in traditional large-scale models, isolated modules in industrial systems, biases in digital twins, and the lack of autonomous system evolution.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, and in particular to an industrial intelligent agent interaction system and implementation method driven by a large-scale, fully integrated model. Background Technology

[0002] As industrial intelligent systems develop towards full-domain and intelligent solutions, many technical challenges remain. Traditional language-based large models lack independent memory capabilities, requiring continuous accumulation of computational examples. Furthermore, relying solely on language and textual modeling results in outputs that cannot accurately represent the industrial physical world, leading to high inference latency and making it difficult to adapt to the needs of real-time early warning and dynamic control in industrial scenarios. Simultaneously, industrial data lacks a unified and efficient storage and retrieval system, making the reuse of non-linguistic industrial knowledge difficult and resulting in high time and computing costs for decision optimization.

[0003] Existing industrial intelligent systems have independent functional modules, resulting in serious data silos, a lack of a unified scheduling and coordination center, unreasonable allocation of computing power and tasks, and a lack of global hard constraint control. This can easily lead to situations where local optimization results in a loss of overall operational efficiency. Modules such as carbon energy intelligent control, digital twins, and process optimization cannot achieve deep linkage, making it difficult to form globally optimal industrial decision-making and execution solutions.

[0004] Industrial digital twin technology still suffers from modeling biases, low consistency between virtual and real-world mapping, and an inability to achieve full-element real-world matching from components to the entire factory area. It also lacks dynamic iteration capabilities throughout the entire lifecycle. Equipment operation and maintenance, as well as quality inspection, often remain at localized levels, lacking full lifecycle data support, resulting in insufficient accuracy in early warning and control. Furthermore, industrial interaction modes are simplistic, failing to provide personalized services for different roles, and a robust feedback mechanism is lacking. The system cannot autonomously evolve based on on-site execution results, making it difficult to adapt to the dynamic changes in industrial scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide an industrial intelligent agent interaction system and implementation method driven by a fully integrated large model, which solves the problems of shortcomings in traditional large models, isolated modules in industrial systems, deviations in digital twins, and lack of autonomous evolution in existing technologies.

[0006] To achieve the above objectives, this invention provides an industrial intelligent agent interaction system driven by a global integrated large model, including a global world large model training and inference module, a super memory storage module, an intelligent retrieval engine module, four core functional pillar modules, a multimodal interaction portal module, and a data acquisition and execution control module.

[0007] The global world large model training and inference module is used for non-verbal modeling, multi-scale training, lightweight output and low-latency inference. The super memory storage module generates a super memory library to store the entire lifecycle data of industrial scenarios; it is also used to store the lightweight training results of a global world-scale model. The intelligent search engine module supports multi-dimensional keyword retrieval in industrial scenarios, combines search results with the reasoning capabilities of a global world model, and outputs optimization and adjustment solutions. The four core functional pillars include the carbon energy intelligent control and optimization submodule, the holographic digital twin collaboration submodule, the hyperdimensional knowledge insight and decision-making submodule, and the process AI autonomous evolution submodule. Each submodule is coordinated and scheduled through a collaborative hub to achieve intelligent control and optimization of the entire carbon energy chain, holographic digital twin real-scene mapping and simulation collaboration in industrial scenarios, non-verbal knowledge insight and intelligent decision-making inference of all elements, and AI autonomous evolution management and control of business processes and equipment status. The multimodal interactive portal module provides personalized KPI, alerts, tasks and knowledge displays for different industrial roles, supports multimodal interaction, and embeds a unified intelligent assistant in the interfaces of various business modules, supporting real-time conversational information acquisition and operation execution. The data acquisition and execution control module collects data on all elements of the physical world, issues inference and decision-making instructions and intelligent control commands, and provides feedback on the execution results.

[0008] This invention also provides a method for implementing an industrial intelligent agent interaction system driven by a large-scale, fusion-based model, comprising the following steps: S1. Full-element data collection and preprocessing to generate standardized data sources; S11. Collect data on all elements of the physical world. The data on all elements of the physical world includes four dimensions: geospatial, industrial IoT, business logic, and physical rules. S12. The collected data is cleaned, deduplicated, standardized, and missing value imputed. It is then categorized and coded by dimension to form a standardized data source. , , Represents a set of geospatial features. Represents the time-series feature set of the Industrial Internet of Things. Represents a set of business logic features. Represents a set of physical rule features; S2. Construct a global world-wide model for the industrial physical world; S3. Based on standardized data sources, train and optimize the global world model; S4, lightweight processing of the global world model; S5, Construction and Data Storage of the Supermemory; S6, intelligent search engine matching and data retrieval; S7, Global World Model Reasoning and Industrial Intelligent Decision Generation; S8, the coordinated execution of the four core functional pillars; S9. Establish a multimodal interactive two-way closed loop, collect feedback data of the entire life cycle of industrial objects and complete analysis and optimization, drive the super memory and model iteration, and form an infinite closed loop of the entire life cycle of the autonomous evolution of industrial intelligent agents.

[0009] Therefore, the present invention employs the above-mentioned industrial intelligent agent interaction system and implementation method driven by a large-scale, fusion-based model, which has the following beneficial effects: (1) The industrial intelligent agent interaction system driven by the full-domain fusion large model breaks through the limitations of traditional language-based large models. It realizes the real restoration of the industrial physical world through non-linguistic modeling and multi-scale training. After lightweight processing and cloud-edge-device distributed deployment, it achieves low-latency inference and meets the core needs of industrial real-time early warning and dynamic control. At the same time, the independent super memory bank is equipped with an intelligent retrieval engine to realize efficient storage and multi-dimensional accurate retrieval of industrial full life cycle data. There is no need to repeatedly accumulate calculation cases, which greatly improves the efficiency of industrial decision-making and inference. (2) By establishing a collaborative hub as the central control hub for the four core functional pillars, the data barriers between modules are effectively broken down, and the multi-source data fusion scheduling, dynamic allocation of computing power and tasks and global hard constraint control are realized. This promotes the deep linkage, parallel execution and data interoperability of modules such as carbon energy intelligent control and digital twins, optimizes industrial decision-making and execution schemes from the global level, solves the problem of poor coordination and local optimization of modules in traditional industrial systems, and ensures overall operational efficiency. (3) The implementation of core modules such as holographic digital twin and process AI autonomous evolution has enabled precise mapping of all elements in the industrial scene and intelligent management and control of equipment throughout its entire life cycle. It can not only rely on virtual simulation to complete cross-temporal and spatial collaboration and solution verification, but also realize predictive maintenance of equipment, adaptive optimization of processes and full-process quality inspection. The multimodal interactive portal can also provide personalized services for different industrial roles, adapt to diverse operational needs, and improve the intelligence and convenience of industrial production and operation. (4) A complete multimodal interactive two-way closed loop and industrial full life cycle feedback closed loop have been established. By collecting feedback data in all stages and conducting deviation analysis, the super memory is incrementally updated and the model is lightweighted and iterated, enabling the industrial intelligent agent to have continuous autonomous evolution capabilities. At the same time, the full life cycle of industrial knowledge and data is accumulated and reused, helping enterprises to achieve the goals of refined carbon energy management, cost reduction and efficiency improvement, and adapting to the long-term development needs of dynamic changes in industrial scenarios.

[0010] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0011] Figure 1This is a schematic diagram of the structure of an industrial intelligent agent interaction system driven by a large-scale, fusion-based model according to the present invention. Figure 2 This is an overall flowchart of the implementation method of an industrial intelligent agent interaction system driven by a large-scale fusion model according to the present invention. Detailed Implementation

[0012] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0013] Please see Figure 1 A global integrated large model-driven industrial intelligent agent interaction system includes a global world large model training and inference module, a super memory storage module, an intelligent retrieval engine module, four core functional pillar modules, a multimodal interactive portal module, and a data acquisition and execution control module.

[0014] The global world large model training and inference module is used for non-verbal modeling, multi-scale training, lightweight output and low-latency inference. The super memory storage module generates a super memory library to store the entire lifecycle data of industrial scenarios; it is also used to store the lightweight training results of a global world-class large model. The intelligent search engine module supports multi-dimensional keyword retrieval in industrial scenarios, combining search results with the reasoning capabilities of a global world model to output optimization and adjustment solutions; The four core functional pillars include a carbon energy intelligent control and optimization submodule, a holographic digital twin collaboration submodule, a hyperdimensional knowledge insight and decision-making submodule, and a process AI autonomous evolution submodule. Each submodule is coordinated and scheduled through a collaborative hub, achieving intelligent control and optimization of the entire carbon energy chain, holographic digital twin real-scene mapping and simulation collaboration in industrial scenarios, non-verbal knowledge insight and intelligent decision-making simulation of all elements, and AI autonomous evolution control of business processes and equipment status. The collaborative hub is the overall control and data interaction hub among the four core functional pillars. It receives multi-source data, dynamically allocates computing power and tasks, verifies the consistency of virtual and real mapping, manages cross-module hard constraints, integrates the execution results of each submodule, and promotes a closed-loop data system throughout the entire lifecycle, realizing intelligent industrial decision-making and execution. The system achieves global optimization and efficient collaboration; the carbon energy intelligent control optimization submodule integrates energy management system, carbon trading market data, and process AI optimization to realize a panoramic carbon energy profile, AI energy-saving optimization, and intelligent carbon asset management; the holographic digital twin collaboration submodule constructs a multi-scale, high-fidelity holographic digital twin, providing simulation as a service and supporting cross-temporal and spatial collaborative operations; the hyperdimensional knowledge insight and decision-making submodule constructs an enterprise-level super knowledge graph; through natural language interaction, it enables one-click inquiry of the entire plant; based on historical data and real-time status, it performs multi-scenario simulations, provides multi-dimensional simulations and risk assessments, and realizes intelligent decision-making simulations; the process AI autonomous evolution submodule realizes adaptive optimization of business processes, autonomous predictive maintenance of equipment, and intelligent quality control.

[0015] The multimodal interactive portal module provides personalized KPIs, alerts, tasks and knowledge displays for different industrial roles, supports multimodal interaction, and embeds a unified intelligent assistant in the interfaces of various business modules, supporting real-time conversational information acquisition and operation execution. The data acquisition and execution control module collects data on all elements of the physical world, issues inference and decision-making and intelligent control commands, and provides feedback on the execution results.

[0016] Please see Figure 2 A method for implementing an industrial intelligent agent interaction system driven by a large-scale, fusion-based model includes the following steps: S1. Full-element data collection and preprocessing to generate standardized data sources; S11. Collect data on all elements of the physical world. The data on all elements of the physical world includes four dimensions: geospatial, industrial IoT, business logic, and physical rules. Geospatial data includes GIS elevation data, satellite imagery, and climate data; industrial IoT data includes time-series data from equipment sensors, with a sampling frequency of 1Hz-100Hz; business logic data includes structured data from ERP, MES, and SCM systems; and physical rule data includes finite element simulation libraries and material property libraries. S12. The collected data is cleaned, deduplicated, standardized, and missing value imputed. It is then categorized and coded by dimension to form a standardized data source. , , Represents a set of geospatial features. Represents the time-series feature set of the Industrial Internet of Things. Represents a set of business logic features. Represents a set of physical rule features.

[0017] S2. Construct a global world-wide model for the industrial physical world; S21. Construct feature encoding layers for four dimensions: geospatial, industrial IoT, business logic, and physical rules. The geospatial feature encoding layer transforms the three-dimensional geographic information of industrial scenes, such as factory terrain and equipment spatial location, into spatial feature vectors that the model can recognize. It is the core encoding layer for building the "Digital Earth" base, and its expression is: ; In the formula, This represents a multidimensional feature encoding function, responsible for standardizing and vectorizing the data of geospatial feature sets, transforming them into high-dimensional feature vectors that can be recognized by machines; Represents the spatial convolutional coding function; Indicates a spatial location encoding function; The feature encoding layer of the Industrial Internet of Things (IIoT) transforms device IoT time-series data into time-series feature vectors, capturing device operating trends and abnormal patterns, supporting real-time early warning and dynamic mapping. The expression is: ; In the formula, This represents a temporal feature modeling and fusion function that uses a temporal network structure to uncover patterns in data changes over time and identify the operating status of equipment. This represents the time series normalization function, which normalizes time series data. This represents a timing Transformer encoding function; The feature encoding layer for business logic transforms industrial business rules into graph-structured features, enabling the model to understand production and operation logic and output decisions that balance technical feasibility and business economics. The expression is as follows: ; In the formula, This represents the business logic encoding and fusion function, which transforms the enterprise's business processes and management rules into decision logic, giving the model the ability to "understand management"; This represents the business graph construction function; This represents the encoding function of a graph neural network; The feature encoding layer for physical rules transforms industrial physical laws into features that the model can understand, ensuring that the model output conforms to common physical sense. The expression is: ; In the formula, This represents the physical rule mapping function, which transforms physical rules into model weights to ensure that the model's inference results conform to objective physical common sense and do not produce erroneous outputs that violate physical laws. Represents the physical symbol embedding function; Represents the physical constraint encoding function; Indicates the feature fusion operator; S22. Based on the feature encoding layers of four dimensions, the basic framework of the model is built according to the deep learning network, and the core expression of the global world model is constructed: ; In the formula, This represents the core feature matrix of the global world model.

[0018] S3, Training and Optimization of the Global World Model; Training and optimization of the global world model based on standardized data sources; S31, By predicting losses Physical rules penalty loss and business logic penalty losses Constructing a joint loss function The forced model conforms to industry rules, and the expressions are as follows: ; ; ; In the formula, Indicates the weighting coefficients of the prediction model. ; Indicates the mean square error loss; Represents cross-entropy loss; Indicates training parameters The output features of the global world model; Represents the physical threshold function; This represents a business threshold function; Indicates the weight of physical rule constraints; Indicates the weight of business rule constraints; S32. Dynamic feature fusion is achieved through a cross-modal attention mechanism, and the AdamW optimizer is used to iteratively update the model parameters. The expression is: ; In the formula, Indicates the iteration round index; Indicates the updated parameters; Indicates the first Wheel parameters; Indicates the learning rate; This represents the weight decay coefficient; This represents the gradient of the total loss; S33. Set optimization goals, if , This indicates the convergence threshold for the industrial scenario, at which point training stops and the optimized global world model is output. The expression for the optimization objective is: ; In the formula, Represents the optimal model parameter set; This represents the operator for minimizing parameters.

[0019] S4, lightweight processing of the global world model; S41. Quantize the model parameters using the feature quantization function. Bit integer, This indicates quantization precision and initially reduces storage size; the expression is: ; In the formula, This represents the model after feature quantization; Represents the characteristic quantization function; This represents a fully trained global world model. S42. Eliminate redundant parameters and retain core inference features by using model pruning functions; ; In the formula, This represents the model after pruning and quantization. Represents the model pruning function; Indicates the pruning rate; S43. Using the trained global world model as the teacher model, knowledge distillation is performed through the knowledge distillation function to compensate for the accuracy loss and obtain a lightweight target model, which is then stored in the super memory bank for low-latency inference. ; In the formula, This represents the lightweight target model; This represents the knowledge distillation function.

[0020] S5, Construction and Data Storage of the Supermemory; S51. Construct a super memory bank; S511, super memory bank architecture planning and deployment; physical / logical decoupling, independent allocation of computing power, storage and network resources, without occupying large model training computing power, designing a three-layer distributed architecture, namely cloud central master library, edge node slave library and terminal cache library; cloud central master library is deployed in industrial cloud brain, edge node slave library is deployed in production line and factory, and terminal cache library is deployed in interactive devices; S512. Build a hybrid storage engine: Deploy a time-series database to store time-series data; deploy a vector database to store lightweight large model parameters and high-dimensional feature vectors; deploy a graph database to store geospatial data and business logic data; establish a unified data scheduling entry point to enable communication between the three databases; S513. Deploy the super memory module, including installing the super memory scheduling unit, configuring the data standardization processing component and the index service component; the super memory scheduling unit is responsible for data association, feature extraction and read / write scheduling; the data standardization processing component is a tool that integrates cleaning, normalization and format conversion; the index service component supports vector index, spatiotemporal index, topological index and inverted index; The S514 is equipped with standardized write and high-speed read interfaces to meet the low latency, high concurrency, and high reliability read and write requirements of industrial scenarios. S515, configure multi-replica backup and off-site disaster recovery mechanism, enable data encryption and access control, and complete the configuration of data consistency verification rules; S52, Data storage; S521, Multi-source data access: Unified reception of core data, completion of entry point aggregation, including lightweight global world model results and standardized data sources; S522. The data entering the database is processed uniformly through the super memory module; S523. Accurately allocate the processed data. Industrial IoT time-series data are stored in the time-series database, lightweight global world model results are stored in the vector database, and geospatial data, business logic data, and physical rule data are all stored in the graph database. S524. According to access frequency and timeliness, the data is divided into hot data, warm data and cold data, and stored in layers; hot data is stored in the edge node slave database, warm data is stored in the cloud center master database, and cold data is stored in the terminal cache database. S525. Add scenario, spatiotemporal, dimension, and type tags to each data entry, build vector indexes, spatiotemporal indexes, topological indexes, and inverted indexes, establish globally unique IDs, and achieve data association and binding across multiple databases; S526. Assign a unique version number to the lightweight global world model, business logic data and physical rule data, record the update time, update content and related data version, complete the formal entry into the database, write to the main database in the cloud center and synchronize the copy. S527. Real-time incremental synchronization of hot data from edge nodes to the cloud center master database.

[0021] S6, intelligent search engine matching and data retrieval; S61. Parse industrial multimodal retrieval requests and bind them to four dimensions: geospatial, industrial IoT, business logic, and physical rules. S62. The optimal result is selected after parallel matching of keywords, vector similarity, topological relationships and spatiotemporal constraints, and hard constraint verification. S63. Based on the hybrid storage structure of supermemory, realize the linkage and hierarchical retrieval of time-series, vector, and graph databases; S64. After standardizing and encapsulating the retrieved data, return it to the global world large model training and inference module, and simultaneously complete cache optimization and log recording to form a closed-loop mechanism of retrieval, retrieval and feedback.

[0022] S7, Global World Model Reasoning and Industrial Intelligent Decision Generation; S71. Load a lightweight global world model and standardized data sources in four dimensions, complete feature encoding, cross-modal fusion and forward computation, and output structured judgment results after verification by physical rules and hard constraints of business logic. Use the structured judgment results as the reasoning results. S72. Analyze the reasoning results, extract decision elements and bind hard constraints, generate hierarchical decision schemes through multi-objective optimization, generate standardized industrial intelligent decision schemes after secondary verification and issue them for execution. S73. By transmitting execution results back and analyzing deviations, the model is driven to iteratively optimize and form a closed loop for the entire industrial intelligence process.

[0023] S8, the coordinated execution of the four core functional pillars; S81, the collaborative hub of the four core functional pillar modules receives standardized industrial intelligent decision-making schemes and verifies their completeness and compliance; according to the functional positioning of each sub-module in the four core functional pillar modules, it decomposes and classifies the standardized industrial intelligent decision-making schemes, and clarifies the core execution objectives, execution parameters, spatiotemporal range and constraints of each sub-module; the collaborative hub is initialized, the data interaction interfaces of each sub-module are opened, and the real-time data synchronization frequency and execution result feedback rules are configured; according to the priority of industrial scenarios, computing resources and execution sequence are allocated to the execution tasks of each sub-module. S82. The coordination center issues standardized execution instructions to each sub-module, adds association markers to cross-sub-module linkage tasks, and clarifies the support relationship between sub-modules. After receiving the task, each sub-module completes the parameter initialization of its own execution environment, sends the task reception confirmation back to the coordination center, and forms a task execution ledger. S83. Under the support of the collaborative hub, each sub-module executes its own core tasks in parallel, and at the same time completes one-way and two-way data support according to the association mark, so as to achieve division of labor without isolation. The carbon energy intelligent control and optimization submodule performs full-chain monitoring and intelligent optimization of carbon energy; based on industrial IoT time-series feature sets. It collects real-time energy consumption and carbon emission data across the entire supply chain, updating the panoramic carbon energy profile and dynamic carbon footprint map; receives carbon energy optimization targets from decision-making schemes, combines them with digital twin scenario simulation data from the holographic digital twin collaborative submodule, performs AI dynamic optimization of energy-consuming equipment and production schedules, and generates carbon energy optimization schemes; it connects to real-time data from the carbon trading platform, completes automatic carbon quota calculation, and generates AI strategy suggestions for carbon asset buying, selling, and exchanging; it synchronizes real-time carbon energy consumption data and optimization control parameters to the collaborative hub, providing energy consumption constraints for the process AI autonomous evolution submodule. The holographic digital twin collaborative submodule performs virtual-real mapping simulation and cross-temporal collaborative verification; it loads multi-scale holographic digital twins of industrial scenes, based on geospatial feature sets. Industrial IoT time-series feature sets It achieves real-time and accurate mapping between digital twins and the physical world; undertakes simulation verification tasks in decision-making schemes, conducts virtual simulation verification of carbon energy optimization schemes, equipment operation and maintenance and process adjustment schemes in a cloud finite element environment, and outputs simulation feasibility results; supports global cross-regional engineers to conduct remote inspections, scheme reviews and fault diagnosis in virtual twin scenarios, and generates human collaborative verification opinions; and synchronizes simulation results, virtual-real mapping deviation data and human collaborative verification opinions to the collaborative hub to provide a basis for the correction of execution schemes in other sub-modules. The Hyperdimensional Knowledge Insight and Decision-Making submodule provides comprehensive business knowledge support and decision-making simulations. Based on an enterprise-level super knowledge graph, it retrieves industrial knowledge matching the decision-making tasks, including R&D documents, process drawings, fault records, supply chain data, and industry cases. It performs intelligent decision-making simulations for multiple contingency plans for carbon energy optimization, equipment operation and maintenance, and process adjustment, analyzing the risks, benefits, and feasibility of different execution plans. It responds to natural language interactive queries from various submodules and users, enabling one-click access to knowledge support across the entire plant. It synchronizes knowledge graph retrieval results and decision-making simulation reports to the collaboration hub, providing theoretical and case support for the execution plans of other submodules. The process AI autonomous evolution submodule performs intelligent control and optimization of business processes and equipment status; based on industrial IoT time-series feature sets. By integrating the entire lifecycle data of equipment with a super memory, self-service predictive maintenance is performed, generating equipment operation and maintenance plans, including equipment fault early warning work orders, automatic parts allocation, and notification to operation and maintenance personnel; core business processes are monitored, including orders, production scheduling, delivery, and logistics coordination, and combined with carbon energy consumption constraints and knowledge support, process bottlenecks are automatically identified and resource allocation and workflow are dynamically adjusted to generate process adjustment plans; combined with visual AI and process parameters, real-time quality inspection and root cause analysis are performed throughout the entire production process to continuously optimize process parameters; and equipment operating status, process optimization data, and quality inspection results are synchronized to the collaborative hub to provide real-time data for carbon energy optimization and holographic digital twin updates; S84. Based on the collaborative hub, realize the real-time sharing of execution data and execution results of each sub-module, and complete the first round of cross-sub-module result fusion; S85. Led by the collaborative hub, the unified integration and final review of the compliance and feasibility of the execution results of the four sub-modules are completed to ensure that the execution plan is the optimal one globally. S86. The finalized integrated execution plan will be distributed to the execution end in a hierarchical manner according to whether the industrial scenario is automatically executed by machines or manually coordinated, so as to realize the standardization and personalized implementation of instructions; S87, four sub-modules work together to monitor the execution process, ensuring that instructions are delivered without deviation and enabling dynamic adjustment of the execution process; S88 collects the execution results of the entire process and synchronizes them to the global world model and super memory bank, forming a closed loop of reasoning, decision-making, execution and feedback, driving the four sub-modules and the global world model to evolve together.

[0024] S9, multimodal interaction and full lifecycle closed-loop feedback; S91. Initialize the multi-terminal interactive environment and configure role permissions. Encapsulate the fusion execution scheme of the four sub-modules into role-based, multi-terminal adapted multimodal content and push it accurately. Receive user interaction commands in multiple forms such as voice, gesture, and touch. After parsing and verification, distribute them to the corresponding modules for execution. At the same time, collect interaction logs and synchronize them to the super memory bank to achieve a two-way interactive closed loop of "accurate push - convenient operation - command landing - log retention". S92 defines the feedback data collection nodes and dimensions for the entire process of industrial object design, production / construction, operation and maintenance, and disposal. Data from multiple sources is collected across all dimensions and aggregated into the data acquisition and execution control module. After standardization and quality verification, valid data is stored in a super memory to form a full lifecycle data archive. Deviation analysis is performed by comparing actual data with decision-making objectives to pinpoint root causes. Feedback data and deviation data then drive incremental updates to the super memory, performing lightweight incremental iterations on the global world model and the four core functional pillars. Finally, the iterative optimization results are fed back to the user in a multimodal manner and applied to subsequent interactions and decisions, forming an infinite closed loop of "data acquisition - deviation analysis - memory update - model iteration - re-interaction optimization" throughout the entire lifecycle, enabling the autonomous evolution of the industrial intelligent agent.

[0025] Therefore, this invention adopts the aforementioned industrial intelligent agent interaction system and implementation method driven by a fully integrated large model. By constructing a fully integrated large model to drive the industrial intelligent agent interaction system, it breaks through the limitations of traditional language-based large models. It restores the industrial scene through non-linguistic modeling, achieves low-latency inference through lightweight processing and cloud-edge-device deployment, and is equipped with an independent super memory and intelligent retrieval engine. It eliminates the need for repeated accumulation of calculation cases, thereby improving decision-making efficiency. The system takes a collaborative hub as the overall control, breaks down the data barriers between modules, realizes multi-source data fusion and dynamic allocation of computing power, promotes the linkage and optimization of each module, and ensures overall operational efficiency. At the same time, relying on holographic digital twins and process AI autonomous evolution modules, it realizes real-world mapping of all industrial elements, full life-cycle management of equipment, and personalized multimodal interaction. It establishes a dual closed-loop mechanism to drive the system's autonomous evolution, accumulates and reuses industrial knowledge, helps enterprises to achieve refined carbon energy management and cost reduction and efficiency improvement, and adapts to the dynamic development needs of industrial scenarios.

[0026] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An industrial intelligent agent interaction system driven by a large-scale, fusion-based model, characterized in that, include: The global world large model training and inference module is used for non-verbal modeling, multi-scale training, lightweight output and low-latency inference. The super memory storage module generates a super memory library to store data throughout the entire lifecycle of industrial scenarios; Used to store lightweight training results of a global world-scale model; The intelligent search engine module supports multi-dimensional keyword retrieval in industrial scenarios, combines search results with the reasoning capabilities of a global world model, and outputs optimization and adjustment solutions. The four core functional pillars include the carbon energy intelligent control and optimization submodule, the holographic digital twin collaboration submodule, the hyperdimensional knowledge insight and decision-making submodule, and the process AI autonomous evolution submodule. Each submodule is coordinated and scheduled through a collaborative hub to achieve intelligent control and optimization of the entire carbon energy chain, holographic digital twin real-scene mapping and simulation collaboration in industrial scenarios, non-verbal knowledge insight and intelligent decision-making inference of all elements, and AI autonomous evolution management and control of business processes and equipment status. The multimodal interactive portal module provides personalized KPI, alerts, tasks and knowledge displays for different industrial roles, supports multimodal interaction, and embeds a unified intelligent assistant in the interfaces of various business modules, supporting real-time conversational information acquisition and operation execution. The data acquisition and execution control module collects data on all elements of the physical world, issues inference and decision-making instructions and intelligent control commands, and provides feedback on the execution results.

2. The industrial intelligent agent interaction system driven by a large-scale, fusion-based model according to claim 1, characterized in that, Among the four core functional pillar modules: The Collaboration Hub is the central control and data interaction hub among the four core functional pillar modules. It receives multi-source data, dynamically allocates computing power and tasks, verifies the consistency of virtual and real mapping, manages cross-module hard constraints, integrates the execution results of each sub-module, and promotes the closed loop of data throughout the entire life cycle, so as to achieve global optimization and efficient collaboration in industrial intelligent decision-making and execution. The carbon energy intelligent control and optimization submodule integrates energy management system, carbon trading market data and process AI optimization to achieve a panoramic carbon energy profile, AI energy-saving optimization and intelligent management of carbon assets. The holographic digital twin collaboration submodule constructs multi-scale, high-fidelity holographic digital twins, provides simulation as a service, and supports cross-temporal and spatial collaborative operations; The Hyperdimensional Knowledge Insight and Decision-Making submodule constructs an enterprise-level super knowledge graph; Through natural language interaction, it enables one-click inquiry of the entire factory; based on historical data and real-time status, it conducts simulations of multiple scenarios, provides multi-dimensional simulations and risk assessments, and realizes intelligent decision-making simulations. The AI-driven autonomous evolution submodule for processes enables adaptive optimization of business processes, autonomous predictive maintenance of equipment, and intelligent quality control.

3. A method for implementing a fully integrated large-scale model-driven industrial intelligent agent interaction system as described in any one of claims 1-2, characterized in that, Includes the following steps: S1. Full-element data collection and preprocessing to generate standardized data sources; S11. Collect data on all elements of the physical world. The data on all elements of the physical world includes four dimensions: geospatial, industrial IoT, business logic, and physical rules. S12. The collected data is cleaned, deduplicated, standardized, and missing value imputed. It is then categorized and coded by dimension to form a standardized data source. , , Represents a set of geospatial features. Represents the time-series feature set of the Industrial Internet of Things. Represents a set of business logic features. Represents a set of physical rule features; S2. Construct a global world-wide model for the industrial physical world; S3. Based on standardized data sources, train and optimize the global world model; S4, lightweight processing of the global world model; S5, Construction and Data Storage of the Supermemory; S6, intelligent search engine matching and data retrieval; S7, Global World Model Reasoning and Industrial Intelligent Decision Generation; S8, the coordinated execution of the four core functional pillars; S9. Establish a multimodal interactive two-way closed loop, collect feedback data of the entire life cycle of industrial objects and complete analysis and optimization, drive the super memory and model iteration, and form an infinite closed loop of the entire life cycle of the autonomous evolution of industrial intelligent agents.

4. The implementation method of an industrial intelligent agent interaction system driven by a large-scale, fusion-based model according to claim 3, characterized in that, In S2: S21. Construct feature encoding layers for four dimensions: geospatial, industrial IoT, business logic, and physical rules. The feature encoding layer of geospatial space is expressed as follows: ； In the formula, Represents a multidimensional feature encoding function; Represents the spatial convolutional coding function; Indicates a spatial location encoding function; The feature encoding layer of the Industrial Internet of Things (IIoT) is expressed as follows: ； In the formula, Represents the temporal feature modeling and fusion function; Represents the time-normalized function; This represents a timing Transformer encoding function; The feature encoding layer for business logic transforms industrial business rules into graph-structured features, enabling the model to understand production and operation logic and output decisions that balance technical feasibility and business economics. The expression is as follows: ； In the formula, This represents the business logic encoding and fusion function; This represents the business graph construction function; This represents the encoding function of a graph neural network; The feature encoding layer of physical rules is expressed as follows: ； In the formula, This represents the physical rule mapping function, which transforms physical rules into model weights; Represents the physical symbol embedding function; Represents the physical constraint encoding function; Indicates the feature fusion operator; S22. Based on the feature encoding layers of four dimensions, the basic framework of the model is built according to the deep learning network, and the core expression of the global world model is constructed: ； In the formula, This represents the core feature matrix of the global world model.

5. The implementation method of an industrial intelligent agent interaction system driven by a large-scale, fusion-based model according to claim 4, characterized in that, In S3: S31, By predicting losses Physical rules penalty loss and business logic penalty losses Constructing a joint loss function The forced model conforms to industry rules, and the expressions are as follows: ； ； ； In the formula, Indicates the weighting coefficients of the prediction model. ; Indicates the mean square error loss; Represents cross-entropy loss; Indicates training parameters The output features of the global world model; Represents the physical threshold function; This represents a business threshold function; Indicates the weight of physical rule constraints; Indicates the weight of business rule constraints; S32. Dynamic feature fusion is achieved through a cross-modal attention mechanism, and the AdamW optimizer is used to iteratively update the model parameters. The expression is: ； In the formula, Indicates the iteration round index; Indicates the updated parameters; Indicates the first Wheel parameters; Indicates the learning rate; This represents the weight decay coefficient; This represents the gradient of the total loss; S33. Set optimization goals, if... , This indicates the convergence threshold for the industrial scenario, at which point training stops and the optimized global world model is output. The expression for the optimization objective is: ； In the formula, Represents the optimal model parameter set; This represents the operator for minimizing parameters.

6. The implementation method of an industrial intelligent agent interaction system driven by a large-scale, fusion-based model according to claim 5, characterized in that, In S4: S41. Quantize the model parameters using the feature quantization function. Bit integer, This indicates quantization precision and initially reduces storage size; the expression is: ； In the formula, This represents the model after feature quantization; Represents the feature quantization function; This represents a fully trained global world model. S42. Eliminate redundant parameters and retain core inference features by using model pruning functions; ； In the formula, This represents the model after pruning and quantization. Represents the model pruning function; Indicates the pruning rate; S43. Using the trained global world model as the teacher model, knowledge distillation is performed through the knowledge distillation function to compensate for the accuracy loss and obtain the lightweight target model, which is then stored in the super memory bank. ； In the formula, This represents the lightweight target model; This represents the knowledge distillation function.

7. The implementation method of an industrial intelligent agent interaction system driven by a large-scale, fusion-based model according to claim 6, characterized in that, In S5: S51. Construct a super memory bank; S511, super memory database architecture planning and deployment; physical / logical decoupling, independent allocation of computing power, storage and network resources, design of a three-layer distributed architecture, namely cloud central master database, edge node slave database and terminal cache database; cloud central master database is deployed in industrial cloud brain, edge node slave database is deployed in production line and factory, and terminal cache database is deployed in interactive devices; S512, Build a hybrid storage engine: Deploy a time-series database to store time-series data; Deploy a vector database to store lightweight large model parameters and high-dimensional feature vectors; Deploy a graph database to store geospatial data and business logic data; establish a unified data scheduling entry point to enable communication between the three databases; S513. Deploy the super memory module, including installing the super memory scheduling unit, configuring the data standardization processing component and the index service component; The S514 is equipped with standardized write and high-speed read interfaces to meet the low latency, high concurrency, and high reliability read and write requirements of industrial scenarios. S515, configure multi-replica backup and off-site disaster recovery mechanism, enable data encryption and access control, and complete the configuration of data consistency verification rules; S52, Data storage; S521, Multi-source data access: Unified reception of core data, completion of entry point aggregation, including lightweight global world model results and standardized data sources; S522. The data entering the database is processed uniformly through the super memory module; S523. Accurately allocate the processed data. Industrial IoT time-series data are stored in the time-series database, lightweight global world model results are stored in the vector database, and geospatial data, business logic data, and physical rule data are all stored in the graph database. S524. According to access frequency and timeliness, the data is divided into hot data, warm data and cold data, and stored in layers; hot data is stored in the edge node slave database, warm data is stored in the cloud center master database, and cold data is stored in the terminal cache database. S525. Add scenario, spatiotemporal, dimension, and type tags to each data entry, build vector indexes, spatiotemporal indexes, topological indexes, and inverted indexes, establish globally unique IDs, and achieve data association and binding across multiple databases; S526. Assign a unique version number to the lightweight global world model, business logic data and physical rule data, record the update time, update content and related data version, complete the formal entry into the database, write to the main database in the cloud center and synchronize the copy. S527. Real-time incremental synchronization of hot data from edge nodes to the cloud center master database.

8. The implementation method of an industrial intelligent agent interaction system driven by a large-scale, fusion-based model according to claim 7, characterized in that, In S6: S61. Parse industrial multimodal retrieval requests and bind them to four dimensions: geospatial, industrial IoT, business logic, and physical rules. S62. The optimal result is selected after parallel matching of keywords, vector similarity, topological relationships and spatiotemporal constraints, and hard constraint verification. S63. Based on the hybrid storage structure of supermemory, realize the linkage and hierarchical retrieval of time-series, vector, and graph databases; S64. After standardizing and encapsulating the retrieved data, return it to the global world large model training and inference module, and simultaneously complete cache optimization and log recording to form a closed-loop mechanism of retrieval, retrieval and feedback.

9. The implementation method of an industrial intelligent agent interaction system driven by a large-scale, fusion-based model according to claim 8, characterized in that, In S7: S71. Load a lightweight global world model and standardized data sources in four dimensions, complete feature encoding, cross-modal fusion and forward computation, and output structured judgment results after verification by physical rules and hard constraints of business logic. Use the structured judgment results as the reasoning results. S72. Analyze the reasoning results, extract decision elements and bind hard constraints, generate hierarchical decision schemes through multi-objective optimization, generate standardized industrial intelligent decision schemes after secondary verification and issue them for execution. S73. By transmitting execution results back and analyzing deviations, the model is driven to iteratively optimize and form a closed loop for the entire industrial intelligence process.

10. The implementation method of a fully integrated large-scale model-driven industrial intelligent agent interaction system according to claim 9, characterized in that, In S8: S81, the collaborative hub of the four core functional pillar modules receives standardized industrial intelligent decision-making schemes and verifies their completeness and compliance; according to the functional positioning of each sub-module in the four core functional pillar modules, it decomposes and classifies the standardized industrial intelligent decision-making schemes, and clarifies the core execution objectives, execution parameters, spatiotemporal range and constraints of each sub-module; the collaborative hub is initialized, the data interaction interfaces of each sub-module are opened, and the real-time data synchronization frequency and execution result feedback rules are configured; according to the priority of industrial scenarios, computing resources and execution sequence are allocated to the execution tasks of each sub-module. S82. The coordination center issues standardized execution instructions to each sub-module, adds association markers to cross-sub-module linkage tasks, and clarifies the support relationship between sub-modules. After receiving the task, each sub-module completes the parameter initialization of its own execution environment, sends the task reception confirmation back to the coordination center, and forms a task execution ledger. S83. Under the support of the collaborative hub, each sub-module executes its own core tasks in parallel, and at the same time completes one-way and two-way data support according to the associated tags; The carbon energy intelligent control and optimization submodule performs full-chain monitoring and intelligent optimization of carbon energy; based on industrial IoT time-series feature sets. Real-time collection of energy consumption and carbon emission data across the entire energy chain, updating a panoramic carbon energy profile and dynamic carbon footprint map; The system receives carbon energy optimization targets from decision-making schemes, combines digital twin scenario simulation data from the holographic digital twin collaborative submodule, performs AI dynamic optimization of energy-consuming equipment and production schedules, and generates carbon energy optimization schemes; it connects to real-time data from the carbon trading platform to complete automatic carbon quota calculation and generate AI strategy suggestions for carbon asset buying, selling and exchanging; and it synchronizes real-time carbon energy consumption data and optimization control parameters to the collaborative hub to provide energy consumption constraints for the process AI autonomous evolution submodule. The holographic digital twin collaborative submodule performs virtual-real mapping simulation and cross-temporal collaborative verification; Loading multi-scale holographic digital twins of industrial scenarios, based on geospatial feature sets Industrial IoT time-series feature sets It completes the real-time and accurate mapping between digital twins and the physical world; undertakes the simulation verification tasks in decision-making schemes, and conducts virtual simulation verification of carbon energy optimization schemes, equipment operation and maintenance and process adjustment schemes in the cloud finite element environment, and outputs simulation feasibility results; It supports engineers from around the world to conduct remote inspections, solution reviews, and fault diagnosis in virtual twin scenarios, generating collaborative verification opinions; and it synchronizes simulation results, virtual-real mapping deviation data, and collaborative verification opinions to the collaborative hub. The Hyperdimensional Knowledge Insight and Decision-Making submodule provides full-business knowledge support and decision-making simulation. Based on an enterprise-level super knowledge graph, retrieve industrial knowledge that matches the decision-making task; The system performs intelligent decision-making simulations of multiple contingency plans for carbon energy optimization, equipment operation and maintenance, and process adjustment; analyzes the risks, benefits, and feasibility of different implementation plans; responds to natural language interactive queries from various sub-modules and users, enabling one-click access to knowledge support across the entire plant; and synchronizes knowledge graph retrieval results and decision-making simulation reports to the collaborative hub. The process AI autonomous evolution submodule performs intelligent control and optimization of business processes and equipment status; based on industrial IoT time-series feature sets. By integrating the entire lifecycle data of equipment with a super memory database, self-service predictive maintenance is performed, generating equipment operation and maintenance plans, including equipment fault early warning work orders, automatic parts allocation, and notification to operation and maintenance personnel; core business processes are monitored, including orders, production scheduling, delivery, and logistics coordination, and combined with carbon energy consumption constraints and knowledge support, process bottlenecks are automatically identified and resource allocation and workflow are dynamically adjusted to generate process adjustment plans; combined with visual AI and process parameters, real-time quality inspection and root cause analysis are performed throughout the entire production process to continuously optimize process parameters; and equipment operating status, process optimization data, and quality inspection results are synchronized to the collaboration center. S84. Based on the collaborative hub, realize the real-time sharing of execution data and execution results of each sub-module, and complete the first round of cross-sub-module result fusion; S85. Led by the collaborative hub, the unified integration and final review of the compliance and feasibility of the execution results of the four sub-modules are completed to ensure that the execution plan is the optimal one globally. S86. The finalized integrated execution plan will be distributed to the execution end in a hierarchical manner according to whether the industrial scenario is automatically executed by machines or manually coordinated, so as to realize the standardization and personalized implementation of instructions; S87, four sub-modules work together to monitor the execution process, ensuring that instructions are delivered without deviation and enabling dynamic adjustment of the execution process; S88 collects the execution results of the entire process and synchronizes them to the global world model and super memory, driving the four sub-modules to evolve together with the global world model.

Images