An industrial intelligent platform system supporting industrial mechanism and large model fusion

Abstract

This invention discloses an industrial intelligent platform system supporting the fusion of industrial mechanisms and large-scale models, relating to the field of industrial internet technology. It includes an industrial data management module, an industrial mechanism knowledge and reasoning unit, an industrial large-scale model service unit, a dual-wheel collaborative reasoning core module, a low-code application development unit, and a full-process security control module. By setting up a dual-wheel collaborative reasoning core module, this invention performs semantic parsing and feature quantification of industrial problems, automatically completing problem decomposition and intelligent task allocation, and supporting flexible switching between multiple modes such as division of labor collaboration, enhanced collaboration, verification collaboration, and complementary collaboration. Through an intermediate result sharing pool, it achieves information interaction and format conversion between the two types of units, and then generates the final decision through multi-algorithm fusion and mandatory mechanism verification. This fully leverages the advantages of the rigorous and interpretable nature of the mechanism model while effectively constraining the output anomalies of the large-scale model, improving the accuracy and security of industrial decision-making.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, specifically to an industrial intelligent platform system that supports the integration of industrial mechanisms and large-scale models. Background Technology

[0002] With the digital and intelligent upgrading of the manufacturing industry, industrial intelligent systems have become core tools for enterprises to optimize production and support decision-making. Traditional solutions rely on physical laws and expert experience, which yield rigorous results but have long construction cycles and poor adaptability to complex scenarios. In recent years, large-scale models have been gradually applied to the industrial field, with outstanding understanding and inductive capabilities, but they lack industrial rule constraints and are prone to conclusions that do not conform to reality. The shortcomings of these two types of technologies limit the further improvement of industrial intelligence.

[0003] However, the application of combining traditional regularity models with large models in the industrial field is still in its initial stage. Most existing solutions are simply sequential calls or result splicing, lacking mature deep collaboration mechanisms. Task allocation and cooperation methods are basically set manually, and it is impossible to automatically match the optimal collaboration strategy according to specific problems. Numerical results and textual conclusions are difficult to convert and reuse, and there is a general lack of final verification of industrial rules, making it difficult to adapt to the differentiated needs of different industrial scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide an industrial intelligent platform system that supports the integration of industrial mechanisms and large-scale models, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Figure 1 An industrial intelligent platform system that supports the fusion of industrial mechanisms and large-scale models includes: The industrial data management module is used to process the collection, cleaning, labeling, storage and management of multi-source heterogeneous industrial data, providing unified data support for the entire platform; The Industrial Mechanism Knowledge and Reasoning Unit is used to store and manage various mechanism knowledge assets in the industrial field, and to provide standardized and callable mechanism reasoning services. The Industrial Large Model Service Unit is used to access, manage, and run various general and industrial vertical large models, providing natural language understanding, knowledge reasoning, and pattern recognition. The dual-wheel collaborative reasoning core module is connected to the industrial data management module, the industrial mechanism knowledge and reasoning unit, and the industrial large model service unit, respectively, to realize the automatic decomposition of industrial problems, task scheduling, collaborative reasoning, and result fusion. The low-code application development unit, connected to the dual-wheel collaborative reasoning core module, is used to quickly build, deploy, and maintain various industrial intelligent applications. The end-to-end security management module connects to all the above modules to provide security protection covering the entire process of data, models, applications, and operation.

[0006] like Figure 2 and Figure 3 The dual-wheel collaborative reasoning core module constructs a standardized, configurable, and self-evolving mechanism and large-scale model deep integration framework, which can intelligently select the optimal collaborative strategy according to the dynamic characteristics of industrial problems and achieve seamless collaboration between the two. It includes a collaborative reasoning controller, a problem decomposition and intelligent scheduling unit, an intermediate result sharing pool unit, a multi-mode result fusion and decision-making unit, a closed-loop feedback learning unit, and a collaborative mode configuration library unit.

[0007] Preferably, the collaborative reasoning controller is the central control unit of the dual-wheel collaborative reasoning core module, responsible for coordinating the lifecycle management of the entire reasoning process. Its composition and functions are as follows: Ⅰ. Inference Session Manager: Creates an independent session context for each inference request, records all states, intermediate results and operation logs during the inference process, and supports pausing, resuming and terminating the inference process; II. Resource Scheduler: Based on the complexity, real-time requirements, and security level of the inference task, it dynamically allocates computing resources such as CPU, GPU, and memory, prioritizing the execution of high-priority tasks. III. Anomaly Handling and Fault Tolerance Mechanism: Real-time monitoring of the operation status of the industrial mechanism knowledge and reasoning unit and the industrial large model service unit. When a model timeout, crash or output anomaly is detected, a degradation strategy is automatically executed, such as switching to a backup model, returning the previous valid result or triggering manual intervention. IV. Reasoning Process Visualization Unit: The complete reasoning process is displayed graphically, including the problem decomposition results, task allocation, execution status of each model, intermediate results, and final decision basis, which facilitates debugging by developers and understanding by industrial personnel.

[0008] Preferred, such as Figure 4 The problem decomposition and intelligent scheduling unit is based on multi-dimensional feature quantization of problem representation and hybrid decision-making for task allocation. By transforming natural language industrial problems into computable feature vectors, and combining rule-based hard constraints with machine learning-based soft prediction, optimal task allocation is achieved. Its components and functions are as follows: I. Industrial Problem Semantic Analysis Subunit: Utilizes a large industrial model to perform deep semantic understanding of user-input natural language problems, extracting the core intent, key entities (equipment, processes, products, parameters, etc.), constraints (time, cost, quality, safety, etc.), and output requirements of the problem; It supports multimodal question input, including text, tables, images, and time-series data curves, and can uniformly convert inputs of different modalities into the platform's internal standard question representation format; It has built-in industry-specific dictionaries and grammar rules, which can accurately identify industrial terms, abbreviations and industry-specific expressions.

[0009] II. Problem Feature Quantization Subunit: Each industrial problem Q is represented as a five-dimensional feature vector: ; In the formula: The degree of certainty of the problem; Calculate the complexity of the problem; To meet the real-time requirements of the problem; The level of security risk of the problem; This refers to the data dependency type in the problem.

[0010] Degree of certainty of the problem : ; In the formula: The number of explicitly measurable parameters contained in the problem; The number of fuzzy or uncertain parameters contained in the problem; The coverage of the mechanism model corresponding to the problem (0-1, the higher the value, the more mature the mechanism model). The uncertainty of the mechanism model corresponding to the problem (0-1, the higher the value, the greater the error of the mechanism model). , Let be the weighting coefficient, satisfying Typical values ​​in industrial scenarios , .

[0011] Problem computational complexity : ; In the formula: The number of parameters required to solve the problem; The number of data samples required to solve the problem; The number of computational steps required to solve the problem; , , These are weighting coefficients; typical values ​​for industrial scenarios. , , = .

[0012] Real-time requirements for the problem Discrete quantization is used, and the response time is divided into four levels: millisecond level for response time less than 1 second, corresponding to value 1; second level for response time greater than or equal to 1 second and less than 60 seconds, corresponding to value 2; minute level for response time greater than or equal to 60 seconds and less than 3600 seconds, corresponding to value 3; and hour level for response time greater than or equal to 3600 seconds, corresponding to value 4.

[0013] Problem security risk level Also using discrete quantization, it is divided into four levels: low risk (1) for erroneous decisions with no safety impact; medium risk (2) for erroneous decisions leading to minor losses; high risk (3) for erroneous decisions leading to equipment damage or production stoppage; and extremely high risk (4) for erroneous decisions leading to casualties or major accidents.

[0014] Problem Data Dependency Type Represented using one-hot encoding as [ ],in An equal value of 1 indicates that the problem primarily relies on structured data. A value of 1 indicates that the problem primarily relies on unstructured data.

[0015] III. Intelligent Task Allocation Algorithm Subunit: A hybrid decision model is used for task allocation, with the decision function being: ; In the formula: For the question The optimal allocation result, To represent the assignment to the mechanistic model, To represent the assignment to the large model; For the model Problem Solving Expected accuracy; For the model Problem Solving The expected probability of failure.

[0016] Rule-based hard constraints: when satisfied or or Under any condition, it must be assigned to the mechanistic model.

[0017] Machine learning-based soft prediction: For problems that do not meet hard constraints, a gradient boosting tree (XGBoost) model trained on historical data is used for prediction. and : ; ; in: , For XGBoost regression model; For the model Historical performance data, including accuracy, response time, and failure rate.

[0018] IV. Task Execution and Monitoring Subunit: This subunit packages the decomposed subtasks into standardized task requests and sends them to the Industrial Mechanism Knowledge and Reasoning Unit and the Industrial Large Model Service Unit, respectively. It monitors the execution status of each subtask in real time, including queued, running, completed, and failed. It supports parallel and sequential execution of tasks and can automatically adjust the execution order based on the dependencies between subtasks. Preferably, the intermediate result sharing pool unit is based on a multimodal data representation and distributed publish-subscribe mechanism using a unified data model. By defining a data exchange format commonly used in the industrial field, it achieves seamless conversion and sharing between numerical results from mechanistic models and semantic results from large models. Its composition and functions are as follows: I. Unified Data Presentation Layer: Defines Industrial Common Data Objects (IGDOs): ; in: A unique identifier for a data object; Data types include numeric, array, matrix, time series, text, image, and knowledge graph types; For data values, different storage formats are used depending on the data type; Generate timestamps for the data; For data source; Metadata includes data units, precision, and confidence level.

[0019] II. Distributed Storage Layer: A hybrid storage architecture combining in-memory database and persistent storage is adopted. The in-memory database is used to store the intermediate results of the currently executed inference task, providing millisecond-level access speed. The persistent storage is used to store the intermediate results of historical inference tasks for subsequent analysis and model training. It supports data version management and can record the generation time, generated model, and version number of each intermediate result.

[0020] III. Data Access and Synchronization Layer: Provides standardized read and write interfaces, supporting the reading and writing of intermediate results at any time during the inference process by the industrial mechanism knowledge and reasoning unit and the industrial large model service unit. It implements a data publish-subscribe mechanism, which can automatically notify all subtasks that depend on the updated intermediate result, ensuring data consistency and isolation. Data between different inference sessions is isolated from each other and will not interfere with each other.

[0021] IV. Intermediate Result Preprocessing Subunit: This subunit performs preliminary filtering, aggregation, and transformation of intermediate results to make them more suitable for use by another unit. It transforms mechanistic numerical results into semantic descriptions of the larger model: This describes the numerical results output by the mechanistic model. A template-based natural language generation method is used: ; in: For the generated natural language text; A predefined semantic template library; This is a template fill function that fills the numerical results into the corresponding positions in the template.

[0022] Numerical parameters for the semantic transformation mechanism of large model outputs: for the text results of large models Numerical parameters are extracted using methods based on Named Entity Recognition (NER) and relation extraction. ; in: These are the extracted numerical parameters; For industrial knowledge graphs; This function extracts entity and relation names from text and their corresponding values, mapping them to concepts in a knowledge graph.

[0023] Preferred, such as Figure 5 This system integrates multi-mode results fusion and decision-making units based on credibility assessment, combining multi-source information fusion and mechanism consistency verification. By quantifying the credibility of mechanistic models and large-scale model results, a suitable fusion algorithm is used to generate comprehensive results. These results are then rigorously verified using mechanistic knowledge to ensure the safety and reliability of the decision-making process. Its components and functions are as follows: I. Result Standardization and Preprocessing Subunit: This subunit converts the results from the two subunits, which have different formats and structures, into a standard decision representation format within the platform. It also performs a preliminary validity check on the results and filters out obviously unreasonable or invalid results.

[0024] II. Result Credibility Assessment Sub-unit: This sub-unit assesses the credibility of the output results for each unit from multiple dimensions, generating a credibility score (between 0 and 1). Reliability of mechanistic model results : ; in: The historical verification accuracy of the mechanistic model is (0-1). Assign a quality score (0-1, calculated based on data completeness, accuracy, and timeliness) to the input data. The working condition matching degree is 0-1, representing the similarity between the current working condition and the working condition used in the training of the mechanism model. , , Let be the weighting coefficient, satisfying =1, a typical value; , , .

[0025] Reliability of large model results : ; in: The confidence score (0-1, provided by the large model itself) outputs the large model. To ensure consistency of output results for large models (0-1, similarity calculation based on results of repeated inferences); The degree of matching between the results and the mechanistic knowledge (0-1, the higher the value, the more the results conform to the physical laws); , , Let be the weighting coefficient, satisfying Typical values , , .

[0026] It can identify inconsistencies between the results of two units and mark them as high-risk areas.

[0027] III. Multi-algorithm Fusion Subunit: Contains multiple built-in fusion algorithms, automatically selecting the most suitable fusion algorithm based on problem type and result characteristics. Weighted average fusion method: Suitable for scenarios where the results are continuous numerical values ​​(such as parameter optimization and remaining useful life prediction). Weights are assigned based on the credibility scores of the two unit results. ; in: The result after fusion; This is the output of the mechanistic model; This is the output of the large model.

[0028] DS evidence theory fusion method: suitable for classification problems with high uncertainty (such as fault diagnosis and anomaly detection), and can effectively handle conflicts between pieces of evidence. Define the recognition framework ,in Indicates the first One possible outcome.

[0029] For each model, construct the Basic Probability Assignment (BPA) function: ; ; in: For the model Regarding the results The basic probability distribution; For the model Regarding the results The predicted probability; For the model Credibility; Let be the uncertainty of model a.

[0030] The BPA of the two models is fused using Dempster's combination rule: ; in: For the results after fusion The basic probability distribution; , These are the BPA functions for the mechanistic model and the large-scale model, respectively. This represents an empty set, i.e., a situation where two pieces of evidence completely conflict.

[0031] The final decision choice has the result of the maximum basic probability allocation: ; Voting fusion method: Applicable to multi-classification problems, it involves voting between the mechanistic model and multiple large models, with the result receiving the most votes being the final result.

[0032] Analytic Hierarchy Process (AHP) Fusion Method: Applicable to multi-objective decision-making problems (such as production scheduling and process optimization), it can comprehensively consider multiple evaluation indicators. It supports user-defined fusion algorithms to meet the needs of specific scenarios.

[0033] IV. Mechanism Verification and Correction Subunit: A crucial step in ensuring the safe and reliable output of the large model. Regardless of the fusion algorithm used, the final fusion result must be verified by the mechanism model. The verification content includes: physical conservation law verification, process parameter boundary verification, safety specification verification, and logical consistency verification; Define the mechanism constraint set Each constraint This represents a physical law or process specification. (Regarding the fusion result...) Perform the following verification: ; in: To verify the results, True This indicates that the verification has passed. False This indicates that the verification failed. Indicate the result Does the constraint satisfy? .

[0034] For results that fail the validation, a correction method based on gradient descent is used: ; in: This is the corrected result; The learning rate; The loss function is defined as the degree to which the result violates the constraints. This is the gradient of the loss function.

[0035] The decision generation and interpretability enhancement subunit generates final decision recommendations based on the fused results, including specific operational instructions, parameter values, execution time, and precautions. It automatically generates an explanation report for the decisions, clearly stating: which parts of the decision were made by the mechanistic model and their basis (e.g., "calculated according to the first law of thermodynamics"), and which parts were made by the larger model and their basis (e.g., "derived from the analysis of 100 similar historical cases"); it also includes a credibility score for the decisions and potential risk warnings. The subunit supports converting the decision explanation report into natural language for easier understanding and execution by industry personnel.

[0036] Preferred, such as Figure 6 The closed-loop feedback learning unit is based on a continuous optimization mechanism of "reasoning, decision-making, execution, and effect." It collects actual production effect data to quantitatively evaluate decision quality and continuously optimizes problem decomposition strategies, task allocation algorithms, and fusion parameters using reinforcement learning and online learning methods. Its components and functions are as follows: I. Actual Effect Data Collection Subunit: Automatically collects actual production effect data after decision execution, including product quality, energy consumption, output, equipment operating status, etc.; supports manual feedback interface, allowing operators to evaluate the accuracy and effectiveness of decisions; and correlates actual effect data with the platform's prediction results to form a complete closed-loop data chain of "reasoning-decision-execution-effect".

[0037] II. Effectiveness Evaluation Sub-unit: The effectiveness of reasoning and decision-making is quantitatively evaluated from four dimensions: accuracy, timeliness, economy, and security. A comprehensive score for decision-making effectiveness is then calculated. The calculation formula is: ; in: Rate the accuracy (0-1); Timeliness is rated (0-1); Score the economic efficiency (0-1); Assign a safety score (0-1); , , , Let be the weighting coefficient, satisfying Typical values , , =0.2, .

[0038] Accuracy rating Calculation formula: For regression problems: ; For classification problems: ; in: This is the platform's predicted value; This is the actual value; , These are the maximum and minimum values ​​of the parameter, respectively; At the same time, it is necessary to analyze the reasons for the poor decision-making effect and pinpoint whether it is due to incorrect problem decomposition, improper task allocation, insufficient model accuracy, or unreasonable fusion algorithm.

[0039] III. Policy Optimization Subunit: This subunit employs reinforcement learning to optimize the collaborative reasoning strategy. The collaborative reasoning process is modeled as a Markov Decision Process (MDP). state space The set of all possible problem feature vectors; Action space A is the set of all possible cooperative strategies (including problem decomposition methods, task allocation schemes, fusion algorithm selection, etc.). reward function In the state Execute action The immediate reward obtained afterward is defined as the decision effectiveness score. ; Transition probability In the state Execute action After transitioning to state The probability of.

[0040] The optimal strategy is learned using the Q-learning algorithm: ; in: The state-action value function represents the state... Execute action Long-term expected rewards; This is the learning rate (typically 0.1). This is the discount factor (typically 0.9). For instant rewards; To perform the action The next state after; The maximum expected reward for the next state.

[0041] The specific optimizations include: Problem decomposition strategy optimization: Based on the effect evaluation results, adjust the granularity and rules of problem decomposition to improve the accuracy of problem decomposition; Task allocation algorithm optimization: The intelligent task allocation model is retrained using new closed-loop data to improve the rationality of task allocation; Fusion algorithm optimization: Adjust the parameters and weights of the fusion algorithm to improve the accuracy of the fusion results; Collaboration mode optimization: Optimize the configuration parameters of the preset collaboration mode based on the actual effects of different scenarios.

[0042] IV. Model Iteration Trigger Subunit: When the model's performance is detected to have dropped to a preset threshold, the model retraining or fine-tuning process is automatically triggered; new production data and cases can be automatically added to the training dataset to continuously improve the accuracy of the mechanistic model and the large model.

[0043] Preferably, the collaborative mode configuration library unit performs collaborative mode matching and recommendation based on scene features. By establishing a mapping relationship between collaborative modes and scene features, it achieves automatic selection and optimization of collaborative modes. Its components and functions are as follows: I. Preset Collaboration Mode Set: The platform has four built-in core collaboration modes, each with a clearly defined applicable scenario, workflow, and default configuration parameters: ① Division of labor and collaboration model.

[0044] Applicable scenarios: The problem can be clearly decomposed into multiple independent sub-problems, and different sub-problems are suitable for processing by mechanistic models or large models respectively; Typical applications: predictive maintenance of equipment, integrated production scheduling; Workflow: Problem decomposition, parallel allocation of subtasks, independent reasoning, result fusion, mechanism verification, and output decision.

[0045] ② Enhance collaborative models.

[0046] Applicable scenarios: Mechanistic models can generate basically feasible solutions, but large models need to be optimized and refined under mechanistic constraints; Typical applications: process parameter optimization, formulation design; Workflow: Generate initial solution / parameter range from the mechanistic model, optimize the large model under constraints, verify the mechanism through simulation, perform feedback iteration, and output the optimal solution.

[0047] ③ Verify the collaborative mode.

[0048] Applicable scenarios: Large models can quickly generate preliminary results, but mechanistic models are needed to verify the correctness and safety of the results; Typical applications: abnormal event handling, fault diagnosis; Workflow: Generate a preliminary analysis report from the large model, verify the causes of anomalies one by one using the mechanistic model, correct unreasonable parts, and generate the final treatment plan.

[0049] ④ Complementary and synergistic model.

[0050] Applicable scenarios: When the problem is relatively complex, neither the mechanistic model nor the large model alone can yield satisfactory results, and the two need to provide complementary information from different perspectives; Typical applications: root cause analysis of complex equipment failures, new product process development; Workflow: Two units process the same problem simultaneously, generate results separately, integrate multi-dimensional results, verify the mechanism, and output a comprehensive decision.

[0051] II. Custom Collaborative Mode Editor: Provides a visual drag-and-drop editor that allows users to customize collaborative reasoning modes according to their business needs; users can freely define problem decomposition rules, task allocation strategies, model execution order, and result fusion methods; supports testing, verification, and version management of custom modes.

[0052] III. Collaborative Mode Recommendation Subunit: Based on the problem's feature vector and historical performance data, automatically recommends the most suitable collaborative mode to the user. Calculate the problem's feature vector. With collaborative mode feature vector Similarity: ; in: The first feature vector of the problem One dimension; For collaborative mode The eigenvector of the first One dimension; , The first The maximum and minimum values ​​of each feature dimension.

[0053] Overall workflow of the dual-wheel collaborative reasoning core module: S1. Receive inference requests automatically triggered by users or the system; S2, the collaborative inference controller creates inference sessions and allocates computing resources; S3, the problem decomposition and intelligent scheduling unit performs semantic parsing and feature quantification on the problem, decomposes it into multiple sub-tasks, and assigns them to the corresponding reasoning units; S4, the industrial mechanism knowledge and reasoning unit and the industrial large model service unit execute their respective sub-tasks in parallel or serially, and the intermediate results are exchanged through the intermediate result sharing pool unit. S5, the multi-mode result fusion and decision-making unit fuses the output results of the two units and verifies the mechanism to generate the final decision and interpretation report; S6. The collaborative reasoning controller returns the decision results to the user or sends them to the production system for execution. S7, the closed-loop feedback learning unit collects actual production effect data to optimize collaborative reasoning strategies and models.

[0054] Preferably, the industrial data management module is the data foundation of the platform, responsible for the unified access and processing of multi-source heterogeneous data from the entire industrial scenario, providing high-quality data input for mechanistic models and large models.

[0055] The industrial data management module includes: The data acquisition unit supports the acquisition of real-time data, production process data, quality inspection data, environmental monitoring data, and business management data from equipment via mainstream industrial protocols such as OPCUA, Modbus, MQTT, HTTP, and S7. The data processing unit is used for data cleaning, noise reduction, normalization, feature extraction, time-series data processing, and multi-source data association and fusion. The data annotation unit provides semi-automated and automated data annotation tools, supporting the annotation of industrial text, images, videos, and time-series data, and supporting the review and version management of annotation results; The data storage unit adopts a hybrid storage architecture, storing time-series data through a time-series database, structured data through a relational database, unstructured data through an object database, and knowledge graph data through a graph database. The data service unit provides standardized data access API interfaces, supporting data querying, subscription, push and batch export, and enabling secure data sharing among various modules of the platform.

[0056] Preferably, the industrial mechanism knowledge and reasoning unit is a unified carrier of industrial knowledge, which standardizes and encapsulates various mechanism assets accumulated by enterprises, so that they can be uniformly scheduled and called by the dual-wheel collaborative reasoning core module.

[0057] The Industrial Mechanism Knowledge and Reasoning Unit includes: The mechanism model library stores mathematical and simulation models that have been validated in industrial fields, such as physical models, chemical models, process models, equipment models, and energy models. It supports model version management, parameter configuration, and online updates. Industry knowledge graphs construct semantic networks covering elements such as equipment, processes, materials, products, personnel, faults, and safety, clearly expressing the causal, relational, and constraint relationships between these elements; The process rule library stores operating procedures, process parameter ranges, quality standards, safety regulations, alarm thresholds, and interlocking logic during the production process. The expert experience database stores the fault diagnosis experience, process optimization experience, anomaly handling experience and operation skills of industrial experts in the form of structured cases. The mechanism reasoning service unit encapsulates all mechanism knowledge and models into standardized RESTful APIs and gRPC interfaces, supporting remote calls, batch execution, parallel computing, and result return.

[0058] Preferably, the industrial large model service unit is the platform's capability engine, specifically designed for collaboration with mechanistic models, providing large model management and service capabilities throughout the entire lifecycle.

[0059] The industrial large-scale model service unit includes: The model management unit supports the access, registration, version management, and lifecycle management of various open-source large models (such as Llama, Qwen, and Baichuan) and commercial large models. The model training unit provides large-scale incremental pre-training, instruction fine-tuning, efficient parameter fine-tuning (LoRA, QLoRA) and RLHF functions, and supports customized training using industrial private data; The model optimization unit provides model quantization, pruning, distillation, and inference acceleration functions, and supports low-precision quantization such as INT4 and INT8 to meet the real-time requirements of industrial scenarios. The model deployment unit supports cloud deployment, edge deployment, and cloud-edge collaborative deployment modes, and supports elastic scaling and automatic fault recovery of the model; The unified inference interface provides standardized large-model inference services for upper-layer applications, supporting various capabilities such as text generation, semantic understanding, question answering, code generation, and image recognition.

[0060] Preferably, the low-code application development unit is an application building tool for industrial users on the platform, which can significantly reduce the development threshold of industrial intelligent applications and shorten the application launch cycle.

[0061] The low-code application development unit includes: The visual designer provides a drag-and-drop application development interface that supports process design, interface design, and logic design, allowing application development to be completed without writing a lot of code. The component library contains pre-built industrial components, mechanism model components, large model components, and collaborative reasoning components, supporting component reuse and customization; The application template library provides templates for typical industrial applications such as predictive equipment maintenance, process optimization, quality inspection, production scheduling, and anomaly handling, based on four collaborative modes. Integrated interfaces support seamless integration with existing industrial systems such as MES, ERP, SCADA, PLM, and LIMS; The application deployment and operation unit supports one-click deployment, containerized operation, real-time monitoring, automatic upgrades, and version rollback of applications.

[0062] Preferably, the full-process security management module is the guarantee for the secure operation of the platform. It is specially designed to address the new security risks introduced by the industrial large model and provides security protection covering the entire process of data, model, application and operation.

[0063] The end-to-end security management module includes: The identity authentication and access control unit supports role-based access control (RBAC) and attribute-based access control (ABAC) to achieve fine-grained access control. The model security unit provides functions such as model watermarking, model encryption, adversarial sample detection, and large model output auditing to prevent model leakage and malicious use. The data security unit provides functions such as data classification and grading, data desensitization, data encryption, data access control, and data leakage detection to ensure the security and confidentiality of industrial data. The security unit provides system monitoring, anomaly detection, intrusion prevention, load balancing, and disaster recovery functions to ensure the stable operation of the platform. The audit log unit records all user operations and system operation logs, and supports log querying, statistics, and audit traceability.

[0064] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves deep collaborative reasoning between industrial mechanism knowledge and large-scale models by setting up a dual-wheel collaborative reasoning core module. This overcomes the shortcomings of existing technologies, such as simple splicing of two types of units, a single fusion mechanism, and an inability to adapt to diverse industrial problems. This module can perform semantic parsing and feature quantification of industrial problems, automatically complete problem decomposition and intelligent task allocation, and support flexible switching between multiple modes such as division of labor collaboration, enhanced collaboration, verification collaboration, and complementary collaboration. Through an intermediate result sharing pool, it realizes information interaction and format conversion between the two types of units, and then generates the final decision through multi-algorithm fusion and forced mechanism verification. This fully leverages the advantages of the rigorous and interpretable mechanism model, effectively constrains the output anomalies of the large-scale model, and significantly improves the accuracy and security of industrial decision-making.

[0065] 2. In this invention, through closed-loop feedback learning units and standardized inference service design, continuous optimization of platform capabilities and efficient reuse of existing enterprise mechanistic knowledge assets are achieved, effectively reducing the implementation cost and development threshold of industrial intelligent applications. The platform can directly access existing enterprise mechanistic models, process rules, expert experience, etc., without the need for repeated development and construction; by collecting actual production effect data, it continuously optimizes problem decomposition strategies, task allocation algorithms, fusion parameters, and collaborative mode configurations, driving iterative improvement of inference capabilities; with supporting low-code application development unit components and templates, various industrial applications can be quickly built, possessing good scalability and scenario adaptability. Attached Figure Description

[0066] Figure 1 This is an overall architecture diagram of an industrial intelligent platform system that supports the fusion of industrial mechanisms and large models according to the present invention; Figure 2 This is an architecture diagram of the core module of dual-wheel collaborative reasoning in an industrial intelligent platform system that supports the fusion of industrial mechanisms and large models according to the present invention. Figure 3 This is a flowchart illustrating the overall workflow of the dual-wheel collaborative reasoning core module in an industrial intelligent platform system that supports the fusion of industrial mechanisms and large models, as described in this invention. Figure 4 This is a flowchart of the problem decomposition and intelligent scheduling unit in an industrial intelligent platform system that supports the fusion of industrial mechanisms and large models according to the present invention. Figure 5 This is a flowchart of a multi-mode result fusion and decision-making unit in an industrial intelligent platform system that supports the fusion of industrial mechanisms and large models, according to the present invention. Figure 6 This is a flowchart of a closed-loop feedback learning unit in an industrial intelligent platform system that supports the fusion of industrial mechanisms and large models, as described in this invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] This invention discloses an industrial intelligent platform system that supports the fusion of industrial mechanisms and large-scale models. Its overall architecture is as follows: Figure 1 As shown, it mainly consists of six core parts: an industrial data management module, an industrial mechanism knowledge and reasoning unit, an industrial large-scale model service unit, a dual-wheel collaborative reasoning core module, a low-code application development unit, and a full-process security control module. The detailed architecture of the dual-wheel collaborative reasoning core module is as follows: Figure 2 As shown.

[0069] Example 1: Taking the predictive maintenance of centrifugal pumps in a large chemical enterprise as an example, this illustrates the workflow of the platform of the present invention using a division of labor and collaboration mode: Step 1: Data Acquisition and Preprocessing: The industrial data management module collects real-time operating parameters of the centrifugal pump, such as vibration, temperature, pressure, current, and flow rate, via the OPC UA protocol, at a sampling frequency of 1Hz. Simultaneously, it retrieves historical maintenance records, operating logs, design parameters, and spare parts information from the ERP system. The data processing unit cleans, denoises, normalizes, and extracts features from the collected data to generate a standardized dataset.

[0070] Step 2: Problem Reception and Decomposition: Users submit the question "Predict the probability of failure and possible types of failure for centrifugal pump A within the next 7 days" through the platform interface. The problem decomposition and intelligent scheduling unit of the dual-wheel collaborative reasoning core module performs semantic analysis and feature quantification on the problem. Problem feature vector =[0.7,0.5,3,3,[1,0]] Degree of certainty =0.7, computational complexity =0.5, real-time requirements =3 (minute level), security risk level =3 (High Risk), mainly relies on structured data.

[0071] Based on the intelligent task allocation algorithm, the problem is broken down into three sub-problems: Sub-problem 1: Calculate the health status index of a centrifugal pump based on equipment operating parameters. =0.9, allocated to the mechanistic model) Sub-problem 2: Analyze the historical failure modes and development trends of centrifugal pumps. =0.3, allocated to the large model) Sub-question 3: Generate fault repair suggestions and spare parts demand plans ( =0.2, allocated to the large model) Step 3: Task Allocation and Parallel Reasoning The industrial mechanism knowledge and reasoning unit invokes the failure mechanism models of centrifugal pumps, including bearing failure models, impeller failure models, and seal failure models. Based on real-time operating data, the health status indices of each component are calculated: bearing health index = 0.62, impeller health index = 0.85, and seal health index = 0.78.

[0072] The industrial large model service unit analyzed the historical operating data of the centrifugal pump over the past 3 years, 12 maintenance records, and failure cases of 50 similar devices. It identified bearing wear as the most common failure mode for this type of equipment, and the current vibration trend had a similarity of 85% with historical bearing failure cases.

[0073] The large model combines the equipment's health status, failure modes, maintenance procedures, and spare parts inventory information to generate preliminary maintenance recommendations: "It is recommended to arrange a shutdown for maintenance within 3 days to replace the bearing assembly, which is expected to take 4 hours. The required spare parts include 2 sets of SKF6205 bearings and 1 set of seals."

[0074] Step 4: Result Fusion and Output Result reliability assessment: Reliability of mechanistic model results =0.85, reliability of large model results =0.78.

[0075] The failure probability is calculated using the weighted average fusion method: bearing failure probability = 0.62 × 0.85 / (0.85 + 0.78) + 0.85 × 0.78 / (0.85 + 0.78) = 0.73.

[0076] Mechanism verification: Verify that the maintenance recommendations comply with the maintenance procedures and safety regulations for centrifugal pumps.

[0077] The final predictive maintenance report reads: "Centrifugal pump A has a 73% probability of bearing failure within the next 7 days. It is recommended to schedule a shutdown for maintenance and replacement of the bearing assembly within 3 days. Basis: 1. The bearing health index calculated by the mechanistic model is 0.62; 2. The vibration trend analysis of the large model shows a similarity of 85% with historical bearing failure cases."

[0078] Feedback and Optimization: Three days later, maintenance personnel inspected the centrifugal pump and confirmed that the bearings were indeed severely worn. After the maintenance results were entered into the platform, the closed-loop feedback learning unit calculated the effectiveness score of this decision. =0.92, and this data was used to optimize the task allocation algorithm and the fault prediction capability of the large model.

[0079] Example 2: Taking the optimization of hot rolling process parameters in a steel company as an example, this illustrates the workflow of the platform of the present invention using the enhanced collaborative mode: Problem Definition: The user raised the question of "optimizing rolling process parameters for Q235 steel grade, 3.0mm thick hot-rolled plate, to reduce energy consumption while ensuring product quality." The platform obtains information such as the current steel composition (C=0.18%, Si=0.22%, Mn=0.45%), raw material temperature (1180℃), and equipment status from the MES system.

[0080] Initial solution calculation for the mechanism: The dual-wheel collaborative reasoning core module first calls upon industrial mechanism knowledge and reasoning units. Based on the principles of thermodynamics, metallurgy, and rolling mechanics, the mechanism reasoning engine calculates the theoretical range and initial solutions for key process parameters. Final rolling temperature: 850-900℃, initial solution = 875℃; Rolling speed: 8-12 m / s, initial solution = 10 m / s; Total reduction: 70%-75%, initial solution = 72.5%; Winding temperature: 600-650℃, initial solution = 625℃; Large-scale model optimization: The parameter range and initial solution generated by the mechanistic model are input as constraints into the industrial large-scale model service unit. Within this parameter range, the large-scale model, combined with production, quality, and energy consumption data from 5000 batches of products of the same specification over the past two years, searches for the optimal combination of process parameters using a reinforcement learning algorithm. After 100 iterations, the large-scale model generates the optimal parameter combination: Final rolling temperature = 862℃; Rolling speed = 11.2 m / s; Total reduction = 73.8%; Winding temperature = 618℃; The predicted energy consumption reduction is 4.2%, and the product's mechanical properties meet the GB / T700-2006 standard. Mechanism simulation verification: The dual-wheel collaborative inference core module again invoked the mechanism inference engine to perform simulation verification of this parameter combination. The mechanism model simulated the rolling process, predicting the product's yield strength = 245MPa, tensile strength = 380MPa, elongation = 28%, and energy consumption = 128kWh / t. Simulation results show that the product quality meets the requirements, and energy consumption is reduced by 4.5%, consistent with the prediction results of the large model.

[0081] Process issuance and execution: The verified optimal process parameters are automatically issued to the PLC control system on the production site for execution through the integration interface between the platform and the MES system.

[0082] Effect Evaluation and Iteration: After production is completed, the platform automatically collects product quality inspection data (yield strength = 242MPa, tensile strength = 375MPa, elongation = 27.5%) and actual energy consumption data (126kWh / t). The effectiveness score of this decision is calculated. =0.95. The closed-loop feedback learning unit uses this data to optimize the reward function of the large model, making subsequent optimization results more accurate.

[0083] Example 3: Taking the abnormal event handling of a 300MW generator unit in a thermal power plant as an example, the workflow of the platform of this invention using the verification collaboration mode is illustrated: Step 1: Anomaly Detection and Triggering: The industrial data management module collects various operating parameters of the generator set in real time. When it detects that the main steam temperature abnormally rises from 540℃ to 565℃ within 10 minutes and exceeds the alarm threshold of 560℃, the abnormal event handling process is automatically triggered.

[0084] Step 2: Preliminary Analysis of the Large Model: The dual-wheel collaborative reasoning core module first sends abnormal data, relevant operation logs, and alarm information to the industrial large model service unit. The large model generates a preliminary analysis report within 30 seconds. "Possible cause 1: The desuperheating water regulating valve is stuck, resulting in insufficient desuperheating water flow (confidence level 75%)"; "Possible cause 2: Unstable combustion in the furnace, with the flame center shifting upwards (confidence level 60%)"; "Possible cause 3: Scale buildup in the superheater, leading to decreased heat transfer efficiency (45% confidence level)"; "Initial handling recommendations: Immediately increase the flow rate of the desuperheating water and check the status of the desuperheating water regulating valve."

[0085] Step 3: Mechanism Model Verification: The dual-wheel collaborative reasoning core module sends the preliminary analysis report generated by the large model to the industrial mechanism knowledge and reasoning unit for verification. The desuperheating water system mechanism model was used to verify the relationship between the desuperheating water flow rate and the main steam temperature. The calculation results show that when the desuperheating water flow rate is at its current value, the main steam temperature should be around 545℃, which deviates significantly from the actual value of 565℃, supporting the judgment that the desuperheating water regulating valve is stuck.

[0086] By using the furnace combustion mechanism model and analyzing the furnace negative pressure, oxygen content, and coal quantity data, the results showed that the combustion was stable, ruling out the possibility of unstable furnace combustion.

[0087] The heat transfer model of the superheater was called to calculate the heat transfer coefficient of the superheater. The results showed that the heat transfer coefficient was normal, ruling out the possibility of superheater scaling.

[0088] Step 4: Result Revision and Decision Making Results Summary: The cause of the desuperheating water regulating valve sticking was retained, while the other two causes were deleted. Mechanism verification: The verification process is recommended to comply with the "Operating Regulations for Thermal Power Plants". The final abnormal event handling plan is as follows: "Cause of abnormality: The desuperheating water regulating valve is stuck, resulting in insufficient desuperheating water flow. Handling steps: 1. Immediately manually increase the opening of the desuperheating water regulating valve to 80%; 2. Notify maintenance personnel to check the desuperheating water regulating valve; 3. Closely monitor the main steam temperature. If it continues to rise to 570℃, immediately reduce the unit load to 200MW; 4. Emergency shutdown if necessary. Safety precautions: During operation, take precautions to prevent water hammer caused by a sudden drop in steam temperature."

[0089] Step 5: Decision Support and Execution: The platform presents the final treatment plan to the operators, along with relevant operating procedures and historical case studies. Operators follow the plan, and the main steam temperature returns to normal within 5 minutes.

[0090] Step 6: Knowledge Accumulation: After an anomaly is handled, the platform automatically records the event's detailed information, processing procedure, and results in the expert experience database as a reference for handling similar events in the future. Simultaneously, the closed-loop feedback learning unit uses this case study to optimize the anomaly analysis capabilities of the large model.

[0091] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An industrial intelligent platform system supporting the fusion of industrial mechanisms and large-scale models, characterized in that, include: The industrial data management module is used to process the collection, cleaning, labeling, storage and management of multi-source heterogeneous industrial data, providing unified data support for the entire platform; The Industrial Mechanism Knowledge and Reasoning Unit is used to store and manage various mechanism knowledge assets in the industrial field, and to provide standardized and callable mechanism reasoning services. The Industrial Large Model Service Unit is used to access, manage, and run various general and industrial vertical large models, providing natural language understanding, knowledge reasoning, and pattern recognition. The dual-wheel collaborative reasoning core module is connected to the industrial data management module, the industrial mechanism knowledge and reasoning unit, and the industrial large model service unit, respectively, and is used to realize the automatic decomposition of industrial problems, task scheduling, collaborative reasoning, and result fusion. It includes a collaborative reasoning controller, a problem decomposition and intelligent scheduling unit, an intermediate result sharing pool unit, a multi-mode result fusion and decision-making unit, a closed-loop feedback learning unit, and a collaborative mode configuration library unit; The low-code application development unit, connected to the dual-wheel collaborative reasoning core module, is used to quickly build, deploy, and maintain various industrial intelligent applications. The end-to-end security management module connects to all the above modules to provide security protection covering the entire process of data, models, applications, and operation.

2. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large-scale models according to claim 1, characterized in that, The collaborative reasoning controller in the dual-wheel collaborative reasoning core module serves as the central control unit, responsible for coordinating the lifecycle management of the entire reasoning process. Its composition and functions are as follows: Inference Session Manager: Creates an independent session context for each inference request, records all states, intermediate results and operation logs during the inference process, and supports pausing, resuming and terminating the inference process; Resource scheduler: Dynamically allocates CPU, GPU, and memory computing resources based on the complexity, real-time requirements, and security level of the inference task, prioritizing the execution of high-priority tasks; Anomaly handling and fault tolerance mechanism: Real-time monitoring of the operation status of the industrial mechanism knowledge and reasoning unit and the industrial large model service unit. When a model timeout, crash or output anomaly is detected, a degradation strategy is automatically executed, including switching to a backup model, returning the previous valid result or triggering manual intervention. Reasoning Process Visualization Unit: The complete reasoning process is displayed graphically, including the problem decomposition results, task allocation, execution status of each model, intermediate results, and the basis for the final decision.

3. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large-scale models according to claim 1, characterized in that, The problem decomposition and intelligent scheduling unit is based on multi-dimensional feature quantification of problem representation and hybrid decision-making task allocation. Its composition and functions are as follows: Industrial Problem Semantic Parsing Subunit: Utilizes a large industrial model to perform deep semantic understanding of user-input natural language problems, extracting the core intent, key entities, constraints, and output requirements of the problem; supports multimodal problem input and uniformly converts it into the platform's standard problem representation format; includes built-in industry-specific dictionaries and grammar rules; Problem Feature Quantification Subunit: Each industrial problem is represented as a five-dimensional feature vector, including the degree of problem determinism, computational complexity, real-time requirements, safety risk level, and data dependency type. The intelligent task allocation algorithm subunit employs a hybrid decision model for task allocation, combining rule-based hard constraints with machine learning-based soft predictions. When any of the following conditions are met—extremely high safety risk level, problem certainty greater than or equal to 0.8, or real-time requirement at the millisecond level—tasks must be allocated to the mechanism model. For problems that do not meet the hard constraints, a gradient boosting tree model trained on historical data is used to predict the expected accuracy and expected failure probability, and this is used to determine the optimal allocation result. Task Execution and Monitoring Subunit: Packages the decomposed subtasks into standardized task requests and sends them to the Industrial Mechanism Knowledge and Reasoning Unit and the Industrial Large Model Service Unit, respectively; monitors the execution status of each subtask in real time, supports parallel and serial execution of tasks, and automatically adjusts the execution order according to the dependencies between subtasks.

4. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large-scale models according to claim 1, characterized in that, The intermediate result sharing pool unit is based on a unified data model for multimodal data representation and a distributed publish-subscribe mechanism. Its composition and functions are as follows: Unified Data Representation Layer: Defines common industrial data objects, including unique identifiers, data types, data values, data generation timestamps, data sources, and metadata; Distributed storage layer: It adopts a hybrid storage architecture that combines in-memory database and persistent storage. The in-memory database is used to store the intermediate results of the currently executing inference task, and the persistent storage is used to store the intermediate results of historical inference tasks, supporting data version management; Data Access and Synchronization Layer: Provides standardized read and write interfaces, supporting the industrial mechanism knowledge and reasoning unit and the industrial large model service unit to read and write intermediate results at any time during the reasoning process, realizing a data publish-subscribe mechanism; when an intermediate result is updated, it automatically notifies all subtasks that depend on that result; data between different reasoning sessions is isolated from each other; Intermediate result preprocessing subunit: Filters, aggregates, and transforms intermediate results; for numerical results output by the mechanistic model, a template-based natural language generation method is used to convert them into semantic descriptions; for text results output by the large model, numerical parameters are extracted using a method based on named entity recognition and relation extraction.

5. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large models according to claim 1, characterized in that, The multi-mode result fusion and decision-making unit, based on credibility assessment, multi-source information fusion, and mechanism consistency verification, consists of the following components and functions: The results standardization and preprocessing subunit converts the results of the two units with different formats and structures into the standard decision representation format within the platform, and performs a preliminary validity check on the results. The results credibility assessment subunit evaluates the credibility of the mechanistic model results and the large model results from multiple dimensions and generates credibility scores. Identify inconsistencies between the results of two units and mark them as high-risk areas; Multi-algorithm fusion subunit: It has multiple built-in fusion algorithms, including weighted average fusion method, DS evidence theory fusion method, voting fusion method and analytic hierarchy process fusion method. It automatically selects the most suitable fusion algorithm for result fusion based on the problem type and result characteristics. Mechanism Verification and Correction Subunit: Verifies the mechanism model of the final fusion result. The verification content includes physical conservation laws, process parameter boundaries, safety specifications, and logical consistency. For results that fail the verification, an automatic correction method based on gradient descent is used. Decision generation and interpretability enhancement subunit: Generates final decision recommendations based on the fused results, including operation instructions, parameter values, execution time and precautions; Automatically generate decision explanation reports, clearly distinguishing the decision basis, credibility score, and potential risk warnings of mechanistic models and large models.

6. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large models according to claim 1, characterized in that, The closed-loop feedback learning unit is based on a continuous optimization mechanism of reasoning, decision-making, execution, and effect closed loop, and its composition and functions are as follows: Actual effect data collection subunit: Automatically collects actual production effect data after decision execution, supports manual feedback interface, and correlates actual effect data with the platform's prediction results to form a complete closed-loop data chain; Effectiveness evaluation sub-unit: The effectiveness of reasoning and decision-making is quantitatively evaluated from four dimensions: accuracy, timeliness, economy and security, and a comprehensive score for decision effectiveness is calculated; Analyze the reasons for poor decision-making results and pinpoint the problematic aspects; The strategy optimization subunit: uses reinforcement learning to optimize the collaborative reasoning strategy, models the collaborative reasoning process as a Markov decision process, and uses the Q-learning algorithm to learn the optimal strategy; The optimizations include problem decomposition strategies, task allocation algorithms, fusion algorithms, and collaborative mode configuration parameters. Model iteration triggering subunit: When the model's performance is detected to have dropped to a preset threshold, the model retraining or fine-tuning process is automatically triggered. New production data and cases are automatically added to the training dataset.

7. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large-scale models according to claim 1, characterized in that, The collaborative mode configuration library unit is based on scene feature-based collaborative mode matching and recommendation, and its composition and functions are as follows: Preset Collaboration Modes: The platform has four built-in core collaboration modes, including: a division-of-labor collaboration mode suitable for problems that can be clearly decomposed into independent sub-problems; an enhanced collaboration mode suitable for generating initial solutions from mechanistic models and optimizing them by large models; a verification collaboration mode suitable for large models to quickly generate results and verify them by mechanistic models; and a complementary collaboration mode suitable for complex problems that require both parties to complement each other from different perspectives. Each mode includes a clear applicable scenario, workflow, and default configuration parameters. Custom Collaborative Reasoning Mode Editor: Provides a visual drag-and-drop editor that allows users to customize collaborative reasoning modes, freely defining problem decomposition rules, task allocation strategies, model execution order, and result fusion methods; Collaborative Mode Recommendation Subunit: Based on the feature vector of the problem and historical performance data, automatically recommend the most suitable collaborative mode to the user.

8. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large models according to claim 1, characterized in that, The industrial data management module includes: The data acquisition unit supports the acquisition of real-time equipment data, production process data, quality inspection data, environmental monitoring data, and business management data via mainstream industrial protocols such as OPCUA, Modbus, MQTT, HTTP, and S7. The data processing unit is used for data cleaning, noise reduction, normalization, feature extraction, time-series data processing, and multi-source data association and fusion. The data annotation unit provides semi-automated and automated data annotation tools, supporting the annotation of industrial text, images, videos, and time-series data; The data storage unit adopts a hybrid storage architecture, storing time-series data through a time-series database, structured data through a relational database, unstructured data through an object database, and knowledge graph data through a graph database. The data service unit provides standardized data access API interfaces, supporting data querying, subscription, push, and batch export.

9. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large-scale models according to claim 1, characterized in that, The industrial mechanism knowledge and reasoning unit includes: The mechanism model library stores physical, chemical, process, equipment, and energy models that have been validated in industrial fields, and supports model version management, parameter configuration, and online updates. Industry knowledge graph: Construct a semantic network covering equipment, process, materials, products, personnel, faults, and safety elements to express the causal, correlation, and constraint relationships between elements; The process rule library stores operating procedures, process parameter ranges, quality standards, safety regulations, alarm thresholds, and interlocking logic during the production process. The expert experience database stores the fault diagnosis experience, process optimization experience, anomaly handling experience and operation skills of industrial experts in the form of structured cases. The mechanism reasoning service unit encapsulates all mechanism knowledge and models into standardized RESTful APIs and gRPC interfaces, supporting remote calls, batch execution, parallel computing, and result return.

10. The industrial intelligent platform system supporting the fusion of industrial mechanisms and large models according to claim 1, characterized in that, The end-to-end security management module includes: The identity authentication and access control unit supports role-based access control and attribute-based access control, enabling fine-grained access control. The model security unit provides functions such as model watermarking, model encryption, adversarial sample detection, and large model output review. The data security unit provides functions such as data classification and grading, data de-identification, data encryption, data access control, and data leakage detection. The security unit provides system monitoring, anomaly detection, intrusion prevention, load balancing, and disaster recovery functions. The audit log unit records all user operations and system operation logs, and supports log querying, statistics, and audit traceability.

Images