An AI native-based industrial internet platform and a working method thereof

Abstract

The application discloses an AI original-based industrial internet platform and a working method thereof. The platform comprises an intelligent edge layer, which is used for collecting network data streams generated by devices of the industrial internet platform and analyzing the network data streams, identifying an industrial protocol to which each network data stream belongs, and assigning an MA identifier to the network data stream; a data intelligent management layer, which is used for constructing a knowledge graph of the devices; a low-code model production layer, which is used for training a plurality of AI models by using the network data collected in the intelligent edge layer, and obtaining a plurality of intelligent agents; a multi-agent collaboration layer, which is used for training a large language model in the industrial field as a central brain based on the knowledge graph and the network data of the devices, and decomposing a target production task into subtasks suitable for the intelligent agents; the intelligent agents perform information interaction and collaborative decision-making through intelligent interaction contracts, and output a global optimal solution for scheduling the intelligent agents; and a digital twin interaction layer, which is used for interacting with a user and returning an interaction result according to an input of the user.

Description

Technical Field

[0001] This invention belongs to the field of industrial internet technology, and relates to an AI-native industrial internet platform and its working method. Background Technology

[0002] Traditional industrial internet platforms primarily adopt an "Internet + Industry + AI plug-in" model, where users construct workflows by calling the SaaS layer through the PaaS layer. This model suffers from the following technical drawbacks: First, the architecture is fragmented. In existing platforms, AI capabilities are typically treated as independent modules. Data needs to be exported from the platform, preprocessed, trained to generate new models, and then repackaged and deployed via Docker before being used by the platform. Second, the interaction is complex. The platform focuses on visualization and reporting functions, relying on manual threshold setting and data querying, representing a passive mode of human information seeking and lacking proactive intelligence. Third, it lacks adaptability. Industrial platforms mainly rely on experienced technicians, and parameter settings are often deduced based on worker experience. The platform cannot dynamically optimize models and parameters according to changes in operating conditions.

[0003] Therefore, there is an urgent need for an AI-native design pattern to redefine the architecture and workflow of industrial internet platforms. Summary of the Invention

[0004] To address the problems existing in the prior art, the purpose of this invention is to provide an AI-native industrial internet platform and its working method, enabling a paradigm shift from "connectivity-driven" to "intelligent decision-driven" in industrial scenarios. This invention solves the problems of low integration between AI and the platform, insufficient interactive intelligence, and poor knowledge utilization efficiency in existing technologies. The resulting platform possesses the characteristics of a self-sensing, self-decision-making, self-optimizing, and self-interacting industrial intelligent agent, forming a complete and universally applicable underlying architecture in the industrial field.

[0005] The technical solution of this invention is as follows: An AI-native industrial internet platform is characterized by comprising an intelligent edge layer, a data intelligent governance layer, a low-code model production layer, a multi-agent collaboration layer, and a digital twin interaction layer. The intelligent edge layer is used to collect and parse the network data streams generated by various devices of the industrial Internet platform, identify the industrial protocol to which each network data stream belongs, and assign an MA identifier to the network data stream. The data intelligence governance layer is used to extract entity relationships from the unstructured data and structured time-series data of each device on the industrial internet platform and construct a knowledge graph for the corresponding device. The low-code model production layer is used to train several AI models using network data collected in the intelligent edge layer to obtain multiple intelligent agents. The multi-agent collaboration layer trains a large language model for the industrial field as a central brain based on the knowledge graph and network data of each device; and uses the central brain to decompose the target production task into sub-tasks applicable to each of the agents; each agent interacts with information and makes collaborative decisions through a set intelligent interaction contract, and outputs the global optimal solution for scheduling each agent according to the target production task. The digital twin interaction layer is used to interact with the user and return the interaction result based on the user's input.

[0006] Preferably, the smart contract pre-sets collaborative interaction rules, unified data specifications, and task collaboration logic among the smart agents, which are used to standardize the interaction process of each smart agent, verify interaction data, arbitrate task conflicts, coordinate and schedule the linkage and cooperation of each smart agent, realize standardized information interaction and collaborative operation, and output the global optimal solution.

[0007] Preferably, the central brain performs standardized verification of various input data according to the collaborative interaction rules and unified data specifications; coordinates and schedules information interaction among various intelligent agents, arbitrates multi-agent task conflicts, coordinates the synchronous advancement of sub-tasks, integrates the calculation results of various intelligent agents, carries out collaborative decision-making calculations, outputs standardized cross-agent interaction data and task scheduling instructions, sends them to the corresponding intelligent agents, and generates a globally optimal solution for multi-agent collaboration.

[0008] Preferably, the central brain adjusts the pipeline operation parameters in real time based on the stress and strain data of each device, and generates optimization instructions for each device and sends them to the corresponding device; the intelligent edge layer acquisition device updates the knowledge graph with the operation status and working condition data generated after executing the optimization instructions, and the low-code model production layer iteratively optimizes the intelligent agent based on the updated knowledge graph.

[0009] Preferably, the intelligent edge layer extracts multimodal physical data from the network data streams of each device; and uses the multimodal physical data from the network data streams of each device to train several AI models to obtain multiple intelligent agents.

[0010] Preferably, the digital twin interaction layer returns interaction results in natural language and visual charts.

[0011] A working method for an AI-native industrial internet platform, comprising the following steps: The intelligent edge layer collects and parses the network data streams generated by various devices on the industrial Internet platform, identifies the industrial protocol to which each network data stream belongs, and assigns an MA identifier to the network data stream. The data intelligence governance layer extracts entity relationships from the unstructured data and structured time-series data of each device on the industrial internet platform and constructs a knowledge graph for the corresponding device. The low-code model production layer uses network data collected in the intelligent edge layer to train several AI models, resulting in multiple intelligent agents. The multi-agent collaboration layer trains a large language model for the industrial field as a central brain based on the knowledge graph and network data of each device; and uses the central brain to decompose the target production task into sub-tasks applicable to each of the agents; each agent interacts with information and makes collaborative decisions through a set intelligent interaction contract, and outputs the global optimal solution for scheduling each agent according to the target production task. The digital twin interaction layer interacts with the user and returns the interaction result based on the user's input.

[0012] The advantages of this invention are as follows: This approach changes the traditional workflow model where the PaaS layer calls the SaaS layer. Microservices serve as the smallest, most basic unit of operation, no longer requiring container deployment. Microservices communicate with each other via HTTP / RESTful APIs.

[0013] This invention proposes a "data-knowledge-model-decision" cyclical design pattern, which can provide a foundation for the autonomous optimization and upgrading of intelligent agents, improve the performance of industrial internet platforms from the foundation, and enhance the integration between traditional industrial internet platforms and AI.

[0014] To develop a universally applicable architectural design pattern for the industrial field, which is perfected and can be promoted to form an industry standard. Attached Figure Description

[0015] Figure 1 Design a pattern diagram for the "data-knowledge-model-decision" cycle.

[0016] Figure 2 A diagram illustrating the design pattern for a six-layer architecture.

[0017] Figure 3 This is an example diagram of a knowledge graph.

[0018] Figure 4 A workflow diagram for optimizing energy consumption in industrial equipment. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] This invention proposes a "data-knowledge-model-decision" cyclical design pattern (such as...). Figure 1 (as shown) Data processing: This includes all the functions of the intelligent edge layer, mainly focusing on a series of data processing steps: from protocol parsing and acquisition to data acquisition and processing, from data collected from the device layer to the use of intelligent edge to achieve ubiquitous acquisition and feature extraction of multi-source heterogeneous data.

[0021] Knowledge: This includes data generated by the data intelligence governance layer after data preprocessing, and visualizes the data using knowledge graphs.

[0022] Model: Contains all the functionalities of the low-code model production layer, making the model generated by the low-code model production layer callable through a user-defined interface.

[0023] Decision-making: It includes all the functions of the multi-agent collaboration layer and the digital twin interaction layer. By utilizing multi-agent collaboration and digital twin, it transforms the model output into executable calling logic, executes it, and feeds the execution results back to the data layer, forming a closed-loop evolution.

[0024] The technical solution of this invention is a six-layer architecture design pattern with layered decoupling and embedded AI (e.g., Figure 2 As shown in the figure, it is characterized by comprising, from bottom to top: an intelligent edge layer, a data intelligent governance layer, a low-code model production layer, a multi-agent collaboration layer, a digital twin interaction layer, and an integrated secure and trustworthy identification chain that runs through the entire process.

[0025] The specific implementation plan is as follows: 1. Intelligent Edge Layer (AI-Driven Ubiquitous Connectivity) Traditional solution: only performs protocol conversion and data pass-through.

[0026] Invention solution: Multimodal AI Protocol Parsing: Introducing a multimodal AI model, this system not only collects protocol byte streams but also correlates them with multimodal physical data such as device vibration, temperature, and current. Through a pre-trained protocol syntax generation model and a CNN-based protocol classifier (which captures local byte patterns such as protocol headers and function code positions, identifies byte combination patterns at the lower level, field structures at the middle level, and overall protocol features at the higher level), it automatically identifies and parses various industrial protocols (such as Modbus and Profinet), thereby determining which industrial communication protocol the traffic belongs to. This eliminates the need for manual protocol parameter configuration, reducing human error.

[0027] Multimodal AI protocol parsing, for each raw network data stream of the industrial internet platform, no longer only collects the protocol byte stream, but also synchronously associates multimodal physical data such as equipment vibration, temperature, and current; relying on multimodal AI models, protocol syntax generation models, and hierarchical CNN protocol classifiers (bottom layer identifies byte combinations, middle layer identifies field structures, and high layer identifies overall protocol features), automatically identifies and fully parses the industrial communication protocol type of each data stream, accurately determines the industrial protocol such as Modbus and Profinet to which a single data stream belongs, without the need for manual configuration of protocol parameters and avoids human parsing errors, realizing intelligent automatic identification and parsing of protocols for the entire industrial internet data stream.

[0028] AI microservice mesh: Using AI inference engines and model services as basic middleware, any upper-layer application that needs AI capabilities can call them in real time via API, realizing "model as a service".

[0029] Lightweight edge inference: Compressed AI models are built into the edge gateway to enable local data processing. For example, anomaly feature extraction, cleaning, and tag assignment are performed before data is uploaded, uploading only "valuable data features" instead of "all raw data," reducing bandwidth pressure. Each uploaded data item is assigned an MA tag.

[0030] 2. Data Intelligence Governance Layer (Data Reshaping for AI) Traditional approach: Data governance is mainly for storage and retrieval, and the ETL (Extract, Transform, Load) process is fixed.

[0031] Invention solution: Feature repository: Establish an industrial feature repository for AI model training, enabling features to be computed once and reused repeatedly, ensuring feature consistency during training and inference.

[0032] Knowledge Graph Construction: Utilizing natural language processing technology, unstructured data such as equipment manuals, fault repair records, and process flow documents from the actual factory production environment are compared with structured time-series data to extract entity relationships, constructing a knowledge graph for each piece of equipment. The equipment is the root node, and other relevant information (such as the aforementioned fault repair records and process flow documents) are child nodes. Figure 3 As shown, the knowledge graph also provides a knowledge foundation for subsequent algorithm model iterations and upgrades. Data is identified by MA to ensure end-to-end security and traceability.

[0033] 3. Low-code model production layer (low-code model production) Traditional approach: requires professional algorithm engineers to write code, and model training cycle is long.

[0034] Invention solution: This layer encapsulates the AI ​​model or intelligent agent, using well-known technologies. It utilizes device data collected in the intelligent edge layer as input to train the well-known AI model, and outputs the trained model as an intelligent agent, which is stored in the low-code model production layer.

[0035] Automated Machine Learning: The built-in AutoML engine allows users to define data sources and labels via drag-and-drop, automatically handling feature engineering, algorithm selection, parameter tuning, and model evaluation. This lowers the barrier to entry for industrial AI applications.

[0036] Few-shot learning and data augmentation: To address the problem of limited fault samples in industrial scenarios, this paper integrates algorithms such as generative adversarial networks to generate simulated fault data through data augmentation, thereby improving the accuracy of the model in cases of insufficient samples.

[0037] Model lifecycle management: Manages the version of trained models, performs A / B testing and canary releases, and supports dynamic deployment of models on the edge and in the cloud.

[0038] 4. Multi-agent collaboration layer (core of the large-scale industrial model) Traditional approach: A single AI model solves a single problem, and the models cannot communicate with each other.

[0039] Invention solution: Industrial Big Model Foundation: Based on industrial knowledge graphs and network data from various devices, a large language model for the industrial field is trained as the "central brain".

[0040] Multi-agent collaboration: For complex production tasks (such as "responding to order changes"), the central brain breaks down the task into multiple sub-tasks applicable to existing agents: "production scheduling agent," "logistics agent," and "process parameter optimization agent," etc. Each agent interacts and makes collaborative decisions through intelligent interaction contracts, ultimately outputting the globally optimal solution. These intelligent contracts pre-define industrial production-specific collaborative interaction rules, unified data standards, and task collaboration logic for each agent. They can manage various independent agents such as production scheduling, logistics, and process parameter optimization, breaking down information barriers between modules. After breaking down complex production tasks such as order changes into multiple sub-tasks, the central brain of the industrial large-scale model standardizes the interaction processes of each agent, verifies interaction data, arbitrates task conflicts, and coordinates the coordinated cooperation of each agent. This achieves standardized information interaction and collaborative computation, ultimately merging and outputting the globally optimal solution for the coordinated agent, supporting dynamic system decision-making.

[0041] Its workflow is as follows: (1) Receive sub-tasks, task constraints and priority instructions from the central brain; synchronously access the real-time operating status of production scheduling, logistics and process optimization agents, as well as real-time industrial time-series data such as equipment stress and strain and pipeline conditions. (2) Based on the preset collaborative interaction rules and unified data specifications, standardize and verify various input data; coordinate and schedule various agents to carry out information interaction, arbitrate multi-agent task conflicts, coordinate the synchronous advancement of sub-tasks; integrate the calculation results of various agents and carry out collaborative decision-making calculations. (3) Output standardized cross-agent interaction data and task scheduling instructions, issue exclusive work instructions to each agent; generate the global optimal solution for scheduling multiple agents to collaboratively complete production tasks, and support the system to complete dynamic optimization decisions on business processes and equipment pipeline parameters.

[0042] "Standardized cross-agent interaction data and task scheduling instructions": Because each agent has preset collaborative interaction rules, unified data specifications and task collaboration logic, the default processing follows the preset rules and outputs scheduling instructions; according to the parameters in the scheduling instructions (which include agent IDs), each scheduling instruction is sent to the corresponding agent.

[0043] The parameters of a scheduling instruction consist of two parts: data and instructions. The data consists of the parameters from the previous scheduling instruction. If there are no error messages, it is used as input to the scheduling algorithm. If there are error messages, the error message and alarm are output.

[0044] Each agent calculates the optimal scheduling scheme (which device to call first and which device to call next) by calling the preset scheduling strategy algorithm in the agent according to the scheduling instructions.

[0045] 5. Digital Twin Interaction Layer (Natural Interaction and Virtual-Real Mapping) Traditional solution: 2D / 3D large screen display, requiring manual clicking to view details.

[0046] Invention solution: Generative Interaction: Introducing a conversational interactive interface driven by a Large Language Model (LLM). Managers can ask questions directly via voice or text, such as, "Analyze the reasons for the decline in yield on production line A yesterday and provide adjustment suggestions." The system automatically retrieves relevant data, drives model analysis, and returns results in natural language and visual charts, achieving "information finding people."

[0047] AI-driven digital twins: Digital twins are no longer static models, but are driven in real time by AI models. For example, a digital twin can predict the device status for the next 10 minutes based on real-time data and rehearse it in advance in a virtual space.

[0048] 6. Integrated Identification Secure and Trusted Chain Throughout all the above levels, distributed identity authentication and data encryption technology based on MA identifiers are used to ensure that the AI ​​model is not tampered with, the data source is trustworthy, and the decision-making process is traceable.

[0049] Example like Figure 4 As shown, the following example illustrates the "energy consumption optimization" scenario of a chemical company: 1. The intelligent edge layer collects data on the temperature, pressure, current, and ambient temperature of the reactor.

[0050] 2. The data intelligence governance layer calls the feature repository to extract historical data from the same period and associates it with the process parameter range in the equipment manual.

[0051] 3. Lightweight Model Factory Energy Consumption Prediction Model (LSTM-based): Based on the current operating conditions, predicts that energy consumption will reach its peak in the next hour.

[0052] 4. The "energy consumption optimization agent" and the "production scheduling agent" of the multi-agent collaboration layer negotiate and decide to shift some non-critical processes 30 minutes later without affecting output, and automatically adjust the PID controller parameters.

[0053] 5. The digital twin interaction layer highlights the equipment to be adjusted in the 3D digital factory and pushes a message to the duty manager via a chatbot: "The energy consumption peak shaving plan has been automatically executed, and it is expected to save 15% of electricity costs. Please confirm." The duty officer replies "confirm" in voice, and the system executes in a closed loop.

[0054] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An AI-native industrial internet platform, characterized in that, It includes an intelligent edge layer, a data intelligence governance layer, a low-code model production layer, a multi-agent collaboration layer, and a digital twin interaction layer; The intelligent edge layer is used to collect and parse the network data streams generated by various devices of the industrial Internet platform, identify the industrial protocol to which each network data stream belongs, and assign an MA identifier to the network data stream. The data intelligence governance layer is used to extract entity relationships from the unstructured data and structured time-series data of each device on the industrial internet platform and construct a knowledge graph for the corresponding device. The low-code model production layer is used to train several AI models using network data collected in the intelligent edge layer to obtain multiple intelligent agents. The multi-agent collaboration layer, based on the knowledge graph and network data of each device, trains a large language model for the industrial field as a central brain; and uses the central brain to decompose the target production task into sub-tasks applicable to each of the agents. Each of the aforementioned intelligent agents interacts with information and makes collaborative decisions through a set intelligent interaction contract, and outputs the globally optimal solution for scheduling each intelligent agent according to the target production task; The digital twin interaction layer is used to interact with the user and return the interaction result based on the user's input.

2. The industrial internet platform according to claim 1, characterized in that, The smart contract pre-sets collaborative interaction rules, unified data specifications, and task collaboration logic among the smart agents. This is used to standardize the interaction process of each smart agent, verify interaction data, arbitrate task conflicts, coordinate and schedule the linkage and cooperation of each smart agent, realize standardized information interaction and collaborative computation, and output the global optimal solution.

3. The industrial internet platform according to claim 2, characterized in that, The central brain standardizes and verifies various input data according to the collaborative interaction rules and unified data specifications; coordinates and schedules information interaction among various intelligent agents, arbitrates multi-agent task conflicts, coordinates the synchronous advancement of sub-tasks, integrates the calculation results of various intelligent agents, carries out collaborative decision-making calculations, outputs standardized cross-agent interaction data and task scheduling instructions, sends them to the corresponding intelligent agents, and generates the global optimal solution for multi-agent collaboration.

4. The industrial internet platform according to claim 3, characterized in that, The central brain adjusts the pipeline network operation parameters in real time based on the stress and strain data of each device, and generates optimization instructions for each device and sends them to the corresponding devices; the intelligent edge layer acquisition device updates the knowledge graph with the operation status and working condition data generated after executing the optimization instructions, and the low-code model production layer iteratively optimizes the intelligent agent based on the updated knowledge graph.

5. The industrial internet platform according to claim 1, 2, or 3, characterized in that, The intelligent edge layer extracts multimodal physical data from the network data streams of each device; it then uses this multimodal physical data to train several AI models, resulting in multiple intelligent agents.

6. The industrial internet platform according to claim 1, characterized in that, The digital twin interaction layer returns interaction results in natural language and visual charts.

7. A working method for an AI-native industrial internet platform, comprising the following steps: The intelligent edge layer collects and parses the network data streams generated by various devices on the industrial Internet platform, identifies the industrial protocol to which each network data stream belongs, and assigns an MA identifier to the network data stream. The data intelligence governance layer extracts entity relationships from the unstructured data and structured time-series data of each device on the industrial internet platform and constructs a knowledge graph for the corresponding device. The low-code model production layer uses network data collected in the intelligent edge layer to train several AI models, resulting in multiple intelligent agents. The multi-agent collaboration layer trains a large language model for the industrial field as a central brain based on the knowledge graph and network data of each device; and uses the central brain to decompose the target production task into sub-tasks applicable to each of the agents. Each of the aforementioned intelligent agents interacts with information and makes collaborative decisions through a set intelligent interaction contract, and outputs the globally optimal solution for scheduling each intelligent agent according to the target production task; The digital twin interaction layer interacts with the user and returns the interaction result based on the user's input.

8. The method according to claim 7, characterized in that, The smart contract pre-sets collaborative interaction rules, unified data specifications, and task collaboration logic among the smart agents. This is used to standardize the interaction process of each smart agent, verify interaction data, arbitrate task conflicts, coordinate and schedule the linkage and cooperation of each smart agent, realize standardized information interaction and collaborative computation, and output the global optimal solution.

9. The method according to claim 8, characterized in that, The central brain standardizes and verifies various input data according to the collaborative interaction rules and unified data specifications; coordinates and schedules information interaction among various intelligent agents, arbitrates multi-agent task conflicts, coordinates the synchronous advancement of sub-tasks, integrates the calculation results of various intelligent agents, carries out collaborative decision-making calculations, outputs standardized cross-agent interaction data and task scheduling instructions, sends them to the corresponding intelligent agents, and generates the global optimal solution for multi-agent collaboration.

10. The method according to claim 9, characterized in that, The central brain adjusts the pipeline network operation parameters in real time based on the stress and strain data of each device, and generates optimization instructions for each device and sends them to the corresponding devices; the intelligent edge layer acquisition device updates the knowledge graph with the operation status and working condition data generated after executing the optimization instructions, and the low-code model production layer iteratively optimizes the intelligent agent based on the updated knowledge graph.

Images