Industrial operation system

Abstract

Relates to the technical field of operating systems, in particular to an industrial operating system which comprises an edge connection layer, an infrastructure layer, a service platform layer and an application layer, and the edge connection layer is used for achieving real-time mapping from a physical factory to a virtual factory; the infrastructure layer is used for processing mapping data and generating business logic codes, the business platform layer supports business hierarchical development and dynamic expansion based on the business logic codes, and the application layer is used for providing industrial App installation, unloading and updating functions based on a three-level expansion system and configuring a multi-level authorization management system; the decentralized networking module is used for supporting node autonomous operation and multi-level management unit configuration in a network disconnection environment, guaranteeing cross-node consistency of mapping data, supporting multiple networking modes and cross-industry universal data structures, and adapting to different industrial scenes and industry requirements.

Description

Technical Field

[0001] It relates to the technical field of operating systems. More specifically, it relates to an industrial operating system. Background Art

[0002] With the accelerating digital transformation in the industrial field, the industrial operating system is not only a key bridge connecting the physical factory and the virtual factory, but also the cornerstone for achieving comprehensive perception, real-time analysis, and intelligent decision-making of industrial data. However, in the face of the increasingly complex industrial production environment and growing intelligent requirements, the existing industrial operating systems are gradually showing significant technical bottlenecks and limitations, making it difficult to fully meet the urgent needs of industrial digital transformation and intelligent upgrading. Specifically, the existing technologies have deep flaws in the following key aspects:

[0003] In an industrial production environment, devices and systems, such as programmable logic controllers, computer numerical control machine tools, industrial robots, manufacturing execution systems (MES), etc., are usually provided by different manufacturers, and their data interfaces and communication protocols are complex and diverse, resulting in typical multi-source heterogeneity of industrial field data. Although the intelligent factory operating system based on the industrial Internet architecture disclosed in the patent CN115167292B promotes data integration to a certain extent by adopting a PaaS, SaaS, and IaaS layered architecture, when facing the huge differences in protocols between multi-vendor devices, its data interface configuration highly depends on manual configuration of data interfaces, leading to low data collection efficiency, and the real-time mapping of the virtual factory also faces significant delays due to manual intervention.

[0004] The industrial data sharing system disclosed in the patent CN114281891B builds a PaaS platform of the industrial Internet architecture through Kubernetes and Docker to support the convenient access of big data modules and cross-department data sharing and invocation modules, enabling elastic access of multi-source heterogeneous automation devices for industrial big data collection and realizing the connection between the underlying data and the upper-layer applications. However, this system relies on manually writing database fitting scripts and lacks an automated or low-code business layering design framework. This results in the need to modify the underlying code layer by layer when modifying and expanding business logic, with a long development cycle and prone to introducing compatibility errors, seriously dragging down the business response efficiency.

[0005] A manufacturing digital system based on a hybrid cloud architecture disclosed in the patent CN111932395A realizes elastic data storage and cross-platform migration through the cooperation of a private cloud sub-module (based on OpenStack, Kubernetes) and a public cloud sub-module, initially solving the problem of sharing heterogeneous data. However, its centralized architecture cannot guarantee autonomous operation in case of network interruption. Due to the lack of independent data caching and offline service capabilities of local nodes, key services are interrupted in the scenario of network disconnection.

[0006] Facing the above technical challenges, the industrial operating system urgently needs to make key breakthroughs in aspects such as data integration, application development, system autonomy, and network adaptability in order to truly meet the urgent needs of industrial digital transformation and intelligent upgrading. Summary of the Invention

[0007] For this reason, the purpose is to provide an industrial operating system that can support a new industrial operating system architecture for protocol dynamic adaptation, offline autonomous operation, and low-code agile development. To achieve the above purpose, the following technical solutions are provided:

[0008] An industrial operating system, comprising:

[0009] An edge connection layer for collecting the original data stream, performing protocol parsing and conversion according to the characteristics of the original data stream by matching preset parsing rules, supporting adaptive identification of unknown protocols and generation of parsing rules, as well as event orchestration and processing to generate mapped data, realizing real-time mapping from the physical factory to the virtual factory. The edge connection layer also supports dynamic task scheduling and resource allocation based on the real-time operating state, and the dynamic task scheduling and resource allocation are realized based on a learning model;

[0010] An infrastructure layer, including a technology base and a digital base. The technology base is used to process the mapped data and generate business logic code, and the digital base is used to manage cross-industry common data structures and store the mapped data;

[0011] A business platform layer for supporting business layered development and dynamic expansion based on the business logic code. The business platform layer includes a flexible design platform, low-code development tools, and a three-level extension system composed of a basic application layer, a plugin layer, and an extension layer;

[0012] An application layer for providing installation, uninstallation, and update functions of industrial applications based on the three-level extension system, and configuring a multi-level authorization management system. The multi-level authorization management system supports attribute-based access control or a zero-trust model. The application layer also supports real-time decision-making and control based on edge artificial intelligence;

[0013] A decentralized networking module for supporting node autonomous operation and multi-level management unit configuration in an offline environment, and ensuring cross-node consistency of the mapped data. The decentralized networking module also supports cross-node data analysis and model training based on federated learning; and

[0014] A digital twin collaborative optimization module for constructing a digital twin model of the physical factory, and performing collaborative optimization on each layer of the industrial operating system based on the digital twin model. The digital twin model is used to predict system performance and optimize business processes.

[0015] Further configured as: The edge connection layer includes:

[0016] A protocol parsing and conversion unit, configured to perform protocol parsing and conversion according to a preset parsing rule matched with the characteristics of the original data stream; and

[0017] An adaptive recognition and parsing rule generation unit, configured to support the adaptive recognition of unknown protocols and the generation of parsing rules. The adaptive recognition and parsing rule generation unit is implemented by a protocol recognition module based on machine learning. The protocol recognition module based on machine learning automatically recognizes the type of unknown protocol according to the characteristics of the original data stream, and automatically generates a protocol parsing rule corresponding to the type of unknown protocol.

[0018] Further configured as: The edge connection layer includes a dynamic task scheduling and resource allocation unit, configured to perform dynamic task scheduling and resource allocation based on the real-time operating state. The dynamic task scheduling and resource allocation unit is implemented based on a learning model, and the learning model is a reinforcement learning model, configured to adaptively adjust the priority of data acquisition tasks and / or resource allocation according to the real-time operating state including device status, data type, network congestion degree, and system load, so as to optimize the real-time performance and efficiency of data acquisition.

[0019] Further configured as: The digital twin collaborative optimization module uses the digital twin model to predict the performance bottlenecks and potential faults of the industrial operating system, and feeds back the prediction results to each layer of the industrial operating system to guide system optimization.

[0020] Further configured as: The digital twin collaborative optimization module is further configured to simulate and optimize the business processes in the business platform layer based on the digital twin model.

[0021] Further configured as: The decentralized networking module includes:

[0022] A node autonomous operation unit, configured to support the autonomous operation of nodes in a network-disconnected environment and the configuration of multi-level management units;

[0023] A data consistency guarantee unit, configured to guarantee the cross-node consistency of the mapped data; and

[0024] A cross-node data analysis and model training unit, configured to support cross-node data analysis and model training based on federated learning. The cross-node data analysis and model training unit is configured to coordinate multiple nodes to perform joint model training without sharing the original data.

[0025] Further configured as: The application layer includes an edge artificial intelligence decision and control unit for supporting real-time decision-making and control based on edge artificial intelligence. The edge artificial intelligence decision and control unit is implemented by deploying an artificial intelligence-based inference model in the edge connection layer and the node autonomous controller. The inference model is used to perform real-time analysis and decision-making on locally collected data and is used to directly control industrial equipment or adjust data collection strategies.

[0026] Further configured as: The technology base includes an automated code generation tool. The automated code generation tool is used to generate business logic code according to the mapping data and business requirements. The automated code generation tool supports an artificial intelligence-based code generation model to generate code in multiple target languages and optimize it.

[0027] Further configured as: The digital base supports data modeling and dimension expansion according to the dimensions of space, material, time, and finance.

[0028] Further configured as: The industrial operating system has a module hot-swap function, supports the dynamic installation, uninstallation, and update of industrial applications, plugins, and extensions, and supports personalized selection and configuration according to the actual application scenario.

[0029] Compared with the deficiencies of the prior art, the beneficial effects are:

[0030] Through the edge connection layer, efficient collection and protocol conversion of multi-source heterogeneous data are achieved, the manual configuration burden is reduced, and the mapping delay from the physical factory to the virtual factory is significantly reduced.

[0031] The business platform layer supports hierarchical business development and dynamic expansion through low-code development tools and a three-level extension system, shortens the development cycle, reduces compatibility errors, and improves business response speed.

[0032] The decentralized networking module supports autonomous operation in case of network disconnection, uses local caching and blockchain technology to ensure data consistency, and enhances the reliability of the system in the case of network disconnection.

[0033] Supports multiple networking modes and cross-industry common data structures, and adapts to different industrial scenarios and industry requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 : Schematic diagram of the overall architecture;

[0035] Figure 2 : Schematic diagram of the edge connection layer;

[0036] Figure 3 : Schematic diagram of the infrastructure layer;

[0037] Figure 4 : Schematic diagram of the business platform layer;

[0038] Figure 5 : Schematic diagram of the application layer;

[0039] Figure 6 : Schematic diagram of the decentralized networking module;

[0040] Figure 7 : Schematic diagram of the digital twin collaborative optimization module;

[0041] Figure 8 : Flowchart of the AI-based protocol recognition and generation method;

[0042] Figure 9 : Flowchart of the reinforcement learning-based dynamic scheduling method;

[0043] Figure 10 : Flowchart of the federated learning-based cross-node model training method;

[0044] Figure 11 : Flowchart of the edge AI-based real-time decision-making and control method;

[0045] Figure 12 : Flowchart of the digital twin-based collaborative optimization method;

[0046] Figure 13 : Flowchart of the AI-based code generation method. Detailed implementation manner

[0047] As Figure 1 shown, an industrial operating system provided maps the real factory digitally to a virtual factory and conducts efficient management and applications based on this virtual factory. The entire system adopts a hierarchical decoupled design, mainly including an edge connection layer, an infrastructure layer, a business platform layer, and an application layer. To enhance the distributed collaborative ability and intelligent level of the system, a decentralized networking module and a digital twin collaborative optimization module are introduced. These two modules are tightly coupled with each layer to achieve cross-layer collaborative work. The data stream collects the original data stream from the physical factory, processes it through the edge connection layer to generate mapped data, enters the infrastructure layer for storage and management, and supports various business logics and applications on the business platform layer and the application layer. The digital twin module uses the data of each layer to build a virtual model and optimize it, and feeds back the optimization results to each layer. The decentralized networking module ensures communication and data consistency between nodes and supports distributed intelligence.

[0048] As Figure 2 shown, the edge connection layer is a bridge connecting the physical world and the digital world. It is deployed on edge computing units, industrial gateways, or industrial control computers close to devices and control systems. Its functions include:

[0049] Multi-source heterogeneous data collection: The edge connection layer has extensive connection capabilities and supports a variety of industrial communication protocols and transmission methods. For example, for the ModBusTCP protocol, the acquisition module can establish a TCP connection, send ModBus request messages, and receive and parse response messages. For the OPCUA protocol, it can act as an OPCUA client to connect to the OPCUA server, browse the address space, and subscribe to data point changes. In addition, it also supports connecting devices that support the RS232 / 485 protocol through a serial port and connecting wireless sensors through wireless modules (such as LTE-M, NB-IoT, Wi-Fi).

[0050] Protocol parsing, conversion, and adaptive recognition and generation: The original data stream collected is a byte sequence or a data packet in a specific format. The protocol parsing and conversion unit parses this data according to the preset parsing rules and extracts meaningful data points (such as sensor readings, device status, control instructions, etc.). As Figure 8 shown, when encountering an unknown protocol data stream that cannot match the preset rules, the protocol recognition module based on machine learning deeply analyzes the original data stream. In order for the machine learning model to identify the protocol type, it is necessary to extract discriminative features from the original data stream. These features include:

[0051] Statistical features: The length of the data packet, average length, maximum and minimum lengths, data packet arrival interval time, frequency distribution of bytes, entropy of byte values, etc.

[0052] Payload features: The byte sequence of the payload part of the data packet (after Padding or truncated to a fixed length), the byte value or pattern at a specific position (head or tail) in the payload. These original byte sequences themselves can be used as feature inputs. To standardize the input, the variable-length data payload is padded (filled with specific bytes, such as 0x00) or truncated to a preset maximum length (256 bytes). This processing ensures that the data input to the model has a fixed dimension.

[0053] Sequential features: The order of bytes in the data packet, the pattern of consecutive bytes, the frequency of occurrence and position of specific bytes or sequences in the data stream. These can be extracted from the original byte sequence through methods such as sliding window, N-gram analysis, etc.

[0054] Temporal features: The time interval sequence between multiple data packets, the sequence of data packet sizes changing over time. These temporal features are obtained by recording the data packet arrival timestamps and calculating the time differences, forming time series data.

[0055] Handshake or connection features: The three-way handshake pattern of the TCP connection, the initial communication packet sequence of a specific protocol. These features are identified by tracking the network connection status and analyzing a small number of data packets at the initial stage of connection establishment.

[0056] Considering the characteristics of industrial protocol data streams (such as sequentiality, specific patterns, and limitations on real-time performance or computing resources), the following algorithms are adopted to achieve protocol recognition:

[0057] Convolutional Neural Network (CNN): Considering that the payload part of industrial protocol data streams usually contains specific byte patterns or structures. After encoding the original data stream (or its key payload part) (for example, one-dimensional vectorization or conversion into a pseudo-image), it is used as input to construct a one-dimensional or two-dimensional convolutional neural network. Through convolutional layers and pooling layers, the model can automatically learn the feature representations in the protocol payload, and finally classify the protocol types through fully connected layers.

[0058] Long Short-Term Memory Network (LSTM) or Gated Recurrent Unit (GRU): For protocols whose protocol behaviors or features have obvious temporal or sequential dependencies (such as certain handshake processes, state machine-driven protocols), recurrent neural networks such as LSTM or GRU can effectively capture long-distance dependencies in the data stream. Taking the original data stream or its sequential features as time steps and inputting them into the LSTM / GRU network, the model can learn the dynamic features and state transitions of the data stream for protocol classification.

[0059] Hybrid Model: To utilize both the local patterns and temporal dependencies of the data stream simultaneously, first use CNN to extract spatial features from the payload of the data packet, and then input these feature sequences into the LSTM / GRU network to model the data packet sequence, so as to more comprehensively identify the protocol.

[0060] The structure of the hybrid model is to use the CNN structure to extract features. The output of the last pooling layer is used as the feature for each time step of the input sequence and input into a layer of LSTM network. The output of the last time step of the LSTM network is connected to a fully connected layer containing neurons, and finally connected to the Softmax output layer. The training method and hyperparameter settings of the entire model are similar to those of a single model, using the Adam optimizer, the cross-entropy loss function, and an early stopping strategy.

[0061] Select or adjust the above models according to the specific characteristics of industrial protocol data and the computing resource limitations of edge nodes: For edge nodes with limited computing resources, relatively simple CNN models or LSTM / GRU models with fewer parameters can be given priority, or the models can be optimized such as quantization and pruning.

[0062] During model training and deployment, select an appropriate model structure and the number of parameters according to the computing power limitations of edge nodes (CPU model, memory size, number of computing cores). In addition, model quantization or model pruning can be adopted to further reduce the model size and computing overhead, so as to ensure low-latency inference (inference latency less than 50 ms) on resource-constrained edge nodes.

[0063] Based on the identified protocol type, the adaptive recognition and parsing rule generation unit can automatically generate protocol parsing rules. For the identified ModBusTCP protocol, the system can infer the positions and formats of each field (such as transaction identifier, protocol identifier, length, unit identifier, function code, data) in the protocol frame and generate corresponding parsing rules by combining the known ModBusTCP protocol specification template with further analysis of unknown data stream samples. These rules can be stored in the form of structured data (such as JSON or XML) and dynamically loaded into the protocol parsing and conversion unit.

[0064] The parsed data will undergo event orchestration and processing, performing operations such as cleaning, format conversion, unit conversion, aggregation, and calculation according to preset or dynamically generated rules. Average the readings of multiple sensors within a short period of time, or convert analog signals to actual physical quantities.

[0065] Dynamic task scheduling and resource allocation: As Figure 9 shown, to optimize the performance and efficiency of edge nodes, a dynamic scheduling unit based on reinforcement learning is utilized. This unit continuously monitors the real-time operating status of edge nodes, including CPU utilization, memory usage, network bandwidth, the priorities of each data acquisition task, device response time, data type (alarm data usually has a higher priority than regular acquisition data), and the overall system load, etc. These status information constitute the state space of the reinforcement learning environment, and the action space of the reinforcement learning environment is designed as a series of operations that can directly act on the key performance parameters of the edge connection layer. The actions are not only to adjust the priorities, but can also dynamically adjust the parameters of the data acquisition protocol (for example, change the quality of service level QoS of MQTT messages), control the frequency limit of data reporting, or trigger local data aggregation and filtering operations. These actions directly affect the data flow and resource consumption, and are unique control points for industrial data acquisition and processing processes.

[0066] The dynamic task scheduling and resource allocation unit is implemented based on a reinforcement learning model. The reinforcement learning model uses the operating state information of edge nodes as the input state space. The operating state information includes at least the resource utilization data of edge nodes, the relevant states of data collection tasks, the device state information, and the key metrics of the industrial edge computing environment such as network state metrics (e.g., network latency, packet loss rate, bandwidth availability). These raw data are preprocessed and transformed into numerical or discrete state representations acceptable to the reinforcement learning model. The action space of the reinforcement learning model includes a series of feasible resource regulation and task scheduling operations. These operations can be discrete actions, such as increasing or decreasing the priority level of a specific data collection task by one level; or selecting different data transmission protocols / channels, or continuous actions, such as slightly adjusting the data collection frequency of a task within a certain range or adjusting the proportion of CPU time slices allocated to a task. These actions are actually executed by calling the task scheduling and service quality (QoS) control interfaces inside the edge connection layer, thereby dynamically changing the resource allocation and task execution order of data collection and processing.

[0067] The reward function aims to quantify the impact of the current action on the system performance. It comprehensively considers the conflicting and dynamically changing optimization goals in the industrial scenario. The real-time reward (giving a positive reward if high-priority data is collected and processed within the specified time, and a penalty for delay), the efficiency reward (giving a positive reward based on resource utilization and task throughput), the stability penalty (giving a penalty based on the system error rate and task failure rate), and the energy consumption cost (such as considering resource consumption). The form of the reward function can be expressed as:

[0068] ;

[0069] where 、 、 、 ]>are weights set according to actual business requirements. By maximizing the cumulative reward, the reinforcement learning model learns the optimal scheduling strategy. These weights are set according to factors such as the actual business scenario, the importance of key tasks, real-time requirements, and system resource constraints, and can be adjusted and optimized through the system configuration interface.

[0070] The structure of the reinforcement learning model is an Actor-Critic network. The Actor network is a feedforward neural network, and the input is the state vector , representing the running status information of edge nodes, including but not limited to the following dimensions: CPU utilization rate (0 - 100%); memory usage rate (0 - 100%); current network bandwidth utilization rate (up / down); priority of each data collection task (1 - 5); length of the task queue to be processed; latency of the last device response (unit: ms); key device status flag (normal / abnormal / maintenance); network quality metrics (such as average ping value, packet loss rate), etc.

[0071] The Actor network uses the ReLU activation function. Its output layer defines the policy function , that is, given the current state action probability distribution.

[0072] For a discrete action space (such as "increase the priority of task X by one level", "pause task Y", "select network channel Z", etc.): the output layer assigns a probability value to each possible action through the softmax function: ;

[0073] where is the raw output value corresponding to the th action in the output layer of the Actor network; represents the th optional action;

[0074] For a continuous action space (such as "set the data collection frequency of task A to ", "allocate CPU for task B", etc.), the Actor network outputs the parameters of a certain probability distribution, such as the mean and standard deviation of a Gaussian distribution. The specific action value is sampled from this distribution:

[0075] ;

[0076] represents the expected value of the action output in state ; represents the standard deviation of the action, controlling the exploration intensity; represents the normal distribution; action is sampled from this distribution and used for actual execution of scheduling operations.

[0077] When adjusting the data collection frequency, the Actor outputs and , and the actual frequency value is sampled from ; the same applies to CPU time slice percentage or bandwidth allocation.

[0078] The Critic network is also a feedforward neural network with the ReLU activation function. The Critic takes the state vector as input and outputs a scalar value , representing the expected cumulative reward value following the policy in the state :

[0079]

[0080] where is the expected value in the state ; is the immediate reward obtained at the time step ; is the discount factor; it is used to balance the importance of the current reward and future rewards.

[0081] Training process:

[0082] The training of the Actor-Critic model is carried out iteratively, usually in an offline environment, using a simulation environment (based on historical data or digital twin models) to complete.

[0083] The Actor selects an action in the state according to the current policy :

[0084] ; After executing the action, the environment returns the next state +1 and the immediate reward .

[0085] The Critic is used to evaluate the value of the current state and updates its parameters through the temporal difference error.

[0086] The TD error is defined as: ;

[0087] The goal of the Critic is to minimize the expected value of the square of the TD error, and the optimization objective is: ;

[0088] Use the gradient descent method to update .

[0089] The update of the Actor depends on the advantage function estimation . In the A2C algorithm, the advantage estimation is: ;

[0090] While in the PPO algorithm, the generalized advantage estimation is used: ;

[0091] Among them are the GAE parameters.

[0092] The update objective of the Actor is to maximize the advantage-weighted policy likelihood. In PPO, the clipped surrogate objective function is used:

[0093] ;

[0094] Among them is the clip parameter of PPO, usually set to 0.2.

[0095] The hyperparameters are set as follows: optimizer: Adam; learning rate: 0.0005; discount factor : 0.99; GAE parameter : 0.95; PPO clip parameter : 0.2.

[0096] The data used during training is the desensitized or aggregated real-time operation status data collected and processed by the edge connection layer. After the model training is completed, the trained policy model file will be deployed to the edge node.

[0097] On the edge node, the deployed reinforcement learning model runs in inference mode. An independent monitoring agent or module is responsible for periodically (e.g., every 100 ms) or event-drivenly collecting the real-time status data of the edge node, constructing a state vector, and inputting it into the deployed reinforcement learning model for inference. The output of the model inference is the optimal action instruction in the current state (e.g., the specific value or percentage for adjusting the task priority). After receiving the action instruction, the scheduling execution module at the edge end actually performs the corresponding scheduling and resource allocation operations by calling the underlying operating system API or the control interface of the edge connection layer.

[0098] The edge connection layer transmits the processed mapping data to the infrastructure layer through a secure and reliable channel. The MQTT protocol is used to achieve real-time data transmission using its publish / subscribe model, or the HTTP / HTTPS protocol is used for bulk data upload.

[0099] Such as Figure 3 shown, the infrastructure layer is the core foundation of the operating system, including a unified technology base and a unified digital base.

[0100] The unified technology base provides general technical services to support the operation of various systems; microservice governance adopts a microservice architecture, and each functional module is provided in the form of a service. The availability and elasticity of the system are ensured through mechanisms such as service registration and discovery, load balancing, circuit breaking, and degradation; the general engine service provides various engines to process business logic and data streams. For example, a rule engine can trigger specific operations according to preset rules, and a process engine can orchestrate complex business processes.

[0101] The structured modeling tool is used to model the entities (equipment, production lines, workshops), relationships (connection relationships, parent-child relationships), and processes (process flow, logistics flow) of the physical factory, providing a basis for digital twin and business logic development.

[0102] Business logic code automatic generation tool: As Figure 13 shown, this tool receives the data model description and business requirement description in the unified digital base as inputs, and introduces an artificial intelligence-based code generation model to assist and accelerate the business logic development process. The business requirement description can be structured data (such as an abstract representation of a business process diagram) or unstructured data (such as a natural language description). Based on the artificial intelligence code generation model (a model based on the Transformer architecture), this model is trained to understand specific terms, data structures, and business logic patterns in the industrial field, and generate a first draft or code snippet of the source code in the target programming language (such as Python, Java, Golang, or a domain-specific language DSL) according to the input. The generation process involves parsing the input, understanding the business intent, retrieving relevant code templates or patterns, and synthesizing new code. The model can generate functions, classes, modules, or even the skeleton of a complete application.

[0103] The data set sources for training the code generation model include public industrial software code repositories; existing industrial control and business logic code libraries within enterprises; and code examples corresponding to specific industrial business flowcharts or natural language descriptions generated manually or by semi-automatic tools. The training process uses the method of fine-tuning based on large-scale pre-trained models. The generated code will be integrated into the subsequent development process for verification and optimization. The key link among them is static code analysis. Static code analysis is independent of the artificial intelligence model. By analyzing the code structure and syntax, based on preset industrial software security specifications, coding standards, and common vulnerability pattern libraries, it automatically detects potential errors and security vulnerabilities in the code, including syntax errors, type mismatches, null pointer references, memory leaks, resource contention, insecure external input processing, access control logic defects, etc. The static analysis tool will generate a detailed analysis report, indicating the specific location and type of the problem, further ensuring the code quality and security. The system supports manual review or manual intervention. Domain experts or senior developers review the static analysis report and manually read and verify the code generated by the AI. For complex logic errors or business-related security risks that the AI fails to identify, manual modification and improvement are carried out. Only the code that passes static analysis and manual review will be approved for deployment to the business platform layer. In addition, the generated business logic code will also be incorporated into the automated test process, including unit tests, integration tests, and end-to-end tests. At the same time, the code generation tool or subsequent optimization module can also optimize the generated code according to the performance characteristics and resource limitations of the target deployment platform.

[0104] Data model description: This part of the input corresponds to specific data types such as device data models, process flow data models, factory topology data models, and business entity data models in industrial scenarios. For example, the data model description can include the names, data types, units, and data collection frequencies of the sensor data points of the device (such as temperature, pressure, vibration); the definitions of key parameters in the industrial process (such as boiler temperature set value, flow rate, chemical substance concentration); the hierarchical structure relationships of factory workshops, production lines, and equipment; and the attributes and relationships of business-related entities (such as work orders, batches, materials). These descriptions usually adopt structured formats, such as based on the OPCUA information model, or using JSONSchema, ProtocolBuffers, or custom metadata formats. The data model description is directly derived from the standardized data model in the unified digital foundation.

[0105] Business requirement description: It can be an abstract representation of the business flowchart visually constructed through the flexible design platform (such as the types of nodes, connection relationships, data input and output); or it can be the application component attributes and interaction logics configured by the user through the low-code tool; these requirement descriptions capture the core functions, conditional judgments, and action executions of the business logic.

[0106] The artificial intelligence-based code generation method comprises the following steps:

[0107] Input processing and feature engineering: Receive structured data model descriptions and business requirement descriptions. Convert these descriptions into a format that can be understood by an artificial intelligence model. For structured inputs, directly construct their graph structure or tree structure representations. For natural language descriptions, use embedding techniques to convert them into vector representations, which capture the structural and semantic features of industrial data and business logic.

[0108] The artificial intelligence model (based on Transformer) analyzes the input feature representations to understand the user's business intent, data flow, processing logic, and the association with the data model. The model can identify specific functional modules to be implemented, conditional logic, loop structures, data transformation operations, and interaction points with other modules or services.

[0109] Code sequence generation: The model generates a source code token sequence in the target programming language according to the understood requirements and data model. The model uses an attention mechanism to focus on the parts of the input that are most relevant to the currently generated code snippet, or uses decoding strategies such as beam search to generate multiple possible code sequences and evaluate them. During the generation process, the model refers to common programming patterns and code templates in the industrial field.

[0110] Code structure and logic verification: The generated code sequence undergoes preliminary structure and logic verification, such as checking for syntax errors, correct variable references, and reasonable control flow. This can be achieved by integrating the parser of the programming language or static analysis tools. If problems are found, feedback can be given to the code generation model for iterative correction.

[0111] Handling of integration points with external services: When generating code, the model generates code to call data storage services to read or write data and code to call general engine services to perform specific calculations according to business requirements and data models. These integration points ensure that the generated code can interact correctly through predefined API interface specifications.

[0112] Code optimization, verification, and post-processing: Format the generated code and add comments. Perform error detection and security vulnerability scanning through independent static code analysis tools. Conduct manual reviews to confirm code logic and security. Execute automated test cases to verify functional correctness. Further optimize according to the performance characteristics of the target platform.

[0113] Output: Business logic code that can be directly used for deployment and execution at the business platform layer and complies with specific programming language specifications.

[0114] The unified digital base is the data core of the system, responsible for the unified management and modeling of data.

[0115] Data modeling by dimension: The digital base supports data modeling in multiple dimensions such as space (factory, workshop, production line, equipment location), material (raw material, semi-finished product, finished product), time (production time, equipment operation time, maintenance time), finance (cost, revenue), etc. This multi-dimensional modeling method enables data to be analyzed and organized from different perspectives, supporting business models across industries and fields. For example, the operation data of equipment can be associated with production orders, personnel information, material batches, etc. for more comprehensive analysis.

[0116] Data storage: Stores the mapped data transmitted from the edge connection layer. Various data storage technologies can be adopted, such as relational databases, time series databases (for storing time series data), distributed file systems, etc., and the most suitable storage solution is selected according to the data type and access mode.

[0117] As Figure 4 shown: The business platform layer is built on top of the infrastructure layer, providing a platform and tools to support the development, deployment, and extension of business logic.

[0118] Flexible design platform: This platform provides a set of visual tool sets, aiming to accelerate the transformation of the business requirements of industrial enterprises into executable software functions. For example, users can define business processes, data conversion rules, interface elements, etc. by dragging and dropping. The platform converts these visual definitions into business logic code or configurations, which can be generated by the automated code generation tools in the infrastructure layer or written manually. The platform supports the life cycle management of functional modules, including installation, uninstallation, upgrade, backup, and migration.

[0119] Low-code development tools (Digital Matrix): As part of the flexible design platform or an independent tool, through a graphical interface and pre-set components, users can quickly build simple business scenario applications without deeply writing code. For example, forms, reports, workflow approval processes, data dashboards, etc. can be easily designed, and these tools convert the visual operations of users into executable business logic.

[0120] Three-level extension system (App Matrix): Supports a three-level extension system composed of a basic application layer, a plugin layer, and an extension layer.

[0121] The basic application layer provides core business applications common across industries, such as equipment management, production monitoring, data analysis, etc. These applications are usually provided or certified by the core development team of the operating system.

[0122] The plugin layer allows third-party developers or users to enhance or customize the functions of the basic application. The plugin interacts with the basic application through the reserved interfaces, providing additional functions or modifying the existing behaviors. The plugin can be installed and uninstalled dynamically without modifying the source code of the basic application.

[0123] The extension layer allows for deeper customization and the development of new applications. Users can develop brand-new industrial applications or make substantial modifications to existing applications based on the capabilities provided by the infrastructure layer and the tools in the business platform layer.

[0124] Application and authorization management: The business platform layer provides unified management functions for industrial applications, plugins, and extensions.

[0125] Multi-level authorization management system: Supports hierarchical management of users, such as super administrators, administrators, and ordinary users. Permissions can be controlled in a fine-grained manner down to the functional modules, interface elements, and even data field levels of the application.

[0126] Attribute-based access control (ABAC): Authorization decisions are based on the attributes of the requester (such as user role, department, location), the attributes of the resource being accessed (such as data type, device), the attributes of the operation (such as read, write, execute), and environmental attributes (such as time, network location). For example, a user can perform a write operation only when within their affiliated workshop and operating on the device data of this workshop.

[0127] Zero-trust model: Does not trust any user or device, and all access requests must undergo strict authentication and authorization. Even internal users accessing internal resources need to be verified and authorized. Zero-trust can be achieved by implementing mechanisms such as multi-factor authentication, device authentication, continuous monitoring, and dynamic authorization.

[0128] As Figure 5 shown, the application layer is the interface for direct user interaction, providing access and management of various industrial applications.

[0129] Industrial application (App) management: Users can browse, search, and install various industrial applications through the application store. The system provides unified management of the installed applications, including functions such as start, stop, update, and uninstall. Application updates can adopt strategies such as hot updates or gray releases to ensure the stable operation of the system.

[0130] Real-time decision-making and control based on edge artificial intelligence: As Figure 11 shown, it supports the deployment of artificial intelligence-based inference models on the node autonomous controllers in the edge connection layer and / or the decentralized networking module. These models utilize the real-time data collected locally for rapid analysis and decision-making.

[0131] Pre-trained AI models are deployed to edge nodes. These models can be classification models for equipment failure prediction, image recognition models for product quality inspection, or regression models for process parameter optimization.

[0132] The model on the edge node receives real-time sensor data as input. It analyzes this data to output decision results, such as predicting equipment anomalies, identifying anomaly types, evaluating product quality, or determining the optimal adjustment of process parameters.

[0133] Based on these inference results, the edge AI decision-making and control unit can directly send control instructions to industrial equipment or adjust data collection strategies. For example, if an impending failure is predicted, it can command an emergency stop; for detected substandard products, it can trigger a rejection mechanism or directly adjust the controller set point to optimize process parameters.

[0134] The primary optimization objective is to achieve desired operational goals by converting sensor inputs into effective control outputs. This includes maintaining key process variables (such as temperature, pressure, and flow rate) within specified optimal ranges to ensure product quality and process efficiency. Unplanned downtime is minimized by predicting failures and enabling proactive control measures (such as reducing loads and initiating safety shutdown procedures). The optimization objectives can be broken down into:

[0135] Minimize the mean square error or mean absolute error between CPV and the set point, maximize the percentage of time that CPV is within the acceptable operating window, and minimize the fluctuation in unit energy consumption or unit production time for producing a specific product.

[0136] By analyzing equipment sensor data, potential failures can be predicted in advance, and control instructions can be output to avoid risks or guide maintenance, thereby extending the equipment's trouble-free operation time and ensuring the safety of personnel and equipment.

[0137] Detailed training process:

[0138] Input data: Real-time mapping data collected and processed by the edge connection layer, primarily time series data from various sensors on industrial equipment. Examples include vibration, temperature, pressure, current, voltage, flow rate, speed, gas concentration, and image / video streams.

[0139] The format is usually a multivariate time series, for example, containing N sensor readings within a time window T, forming an input tensor of shape (T, N).

[0140] Tag data: Sources synchronized with input data, known "optimal" control actions or desired system state parameters. These tags are obtained through:

[0141] Expert experience data: Records of control operations successfully taken by senior operators under similar working conditions (such as equipment parameters adjusted manually, intervention measures).

[0142] The actual control parameter sequence selected from the historical database, which has led to ideal production results (such as high quality, high efficiency) under specific input conditions.

[0143] According to the specific task, it can be continuous values (such as percentage of valve opening, motor speed setting value, heater power setting value), or discrete values or categories (such as control mode switching instructions, fault type labels).

[0144] Data cleaning: Handling missing values (such as interpolation method, mean / median filling), outliers (such as statistical-based rejection, smoothing).

[0145] Data synchronization and alignment: Ensuring strict temporal correspondence between the input sensor sequence and the output control labels.

[0146] Time-domain features: Calculating statistical features (mean, variance, peak value, kurtosis, margin, etc.) for the original time series.

[0147] Frequency-domain features: Extracting the spectral features of the signal through Fourier transform (FFT), wavelet transform, etc.

[0148] Combined features: Constructing new cross features based on domain knowledge.

[0149] Data standardization / normalization: Scaling sensor data and label data with different dimensions to a similar numerical range (such as 0 - 1 or mean = 0, variance = 1) to facilitate the stability and convergence speed of model training.

[0150] Data augmentation: For the case of insufficient sample size, data augmentation techniques (such as adding noise, time series deformation) can be used to expand the training set.

[0151] The inference process receives the real-time mapping data provided by the edge connection layer, uses it as the input of the inference model, the model performs forward calculation, and outputs decision results (for example, abnormal alarm signals, or specific control parameter values). These decision results are directly sent to the actuator or device through the edge connection layer or the underlying control interface to achieve closed-loop real-time control. When the model detects an anomaly, it will trigger an alarm event, which can be subscribed and processed by the business platform layer, such as sending a notification to the operator or triggering a higher-level fault handling process.

[0152] Output layer design:

[0153] Regression task (predicting continuous control parameters): The number of neurons in the output layer is equal to the number of control parameters to be predicted. The activation function is usually a linear function (no activation) or is selected according to the parameter range (e.g., Sigmoid for the 0-1 range).

[0154] Classification task (predicting discrete control instructions or states): The number of neurons in the output layer is equal to the number of classes. The activation function is Softmax (for mutually exclusive multi-classes) or Sigmoid (for binary classification or multi-label).

[0155] Definition of loss function:

[0156] Mean Squared Error (MSE): ;

[0157] Mean Absolute Error (MAE): ;

[0158] where, represents the value of the loss function, which is an indicator to measure the prediction performance of the model. Our goal is usually to minimize this value; represents the summation over all samples, from the first sample to the last sample ; represents the true value or actual observed value of the -th sample in the model, which usually refers to the known optimal control parameter or the desired system state; represents the predicted value of the model for the -th sample, which is the control parameter or state estimate output by the model based on the input sensor data; represents the error between the true value and the predicted value of the -th sample.

[0159] Cross-entropy loss: Binary cross-entropy for binary classification and categorical cross-entropy for multi-classification.

[0160] Model training and optimization:

[0161] Initialization: Reasonably initialize the model weights (such as Xavier, He initialization).

[0162] Optimizer: Select a suitable gradient descent optimization algorithm, such as Adam.

[0163] Training strategy: Divide the training data into small batches for iterative training; adopt a preset learning rate decay strategy (such as step decay, exponential decay) or adaptive learning rate adjustment.

[0164] Prevent overfitting, such as L1 / L2 regularization, Dropout, Early Stopping.

[0165] Hyperparameter Tuning: Use the validation set to tune the hyperparameters of the model (such as learning rate, batch size, number of network layers, number of neurons, regularization coefficient, etc.). Common methods include grid search, random search, and Bayesian optimization.

[0166] Evaluation Metrics: Monitor the performance metrics on the validation set during the training process. For regression tasks, R² score, RMSE, and MAE are commonly used; for classification tasks, accuracy, precision, recall, F1 score, ROC-AUC, etc. are commonly used.

[0167] Adjustment Strategy: Edge AI can also adjust the data collection strategy of the edge connection layer according to the analysis results. For example, if a device is detected to be in a sub-healthy state, the data collection frequency of that device can be increased for more refined monitoring.

[0168] Such as Figure 6 、 Figure 10 As shown, the decentralized networking module is the key to achieving the flexibility and resilience of the system.

[0169] Node Autonomous Operation: The system supports the configuration of multi-level management units. For example, area management nodes, factory management nodes, workshop management nodes, etc. can be configured. The edge nodes under each management unit can still be used offline in a network-disconnected environment, forming self-contained working units. By preloading the core business logic and rules (implemented through event orchestration and rule engines in the edge connection layer) and local data caching on the edge nodes, even when disconnected from the upper-layer network, the edge nodes can continue to execute key tasks, such as data collection, local processing, and control instruction execution. The local cache can use algorithms such as Least Recently Used (LRU) or First In First Out (FIFO) to manage the storage space.

[0170] Data Consistency Guarantee: After the network connection is restored, the decentralized networking module ensures the eventual consistency of cross-node data through a data synchronization mechanism. To achieve this goal, the decentralized networking module adopts a data synchronization mechanism based on distributed logs or version control. For example, each node maintains a local data update log or uses Vector Clocks to track data versions. When the network connection is restored, nodes exchange their data update logs or version information with each other through peer-to-peer or Gossip protocols to identify the data that needs to be synchronized. For possible data conflicts (e.g., the same data item is modified on different offline nodes), the system adopts strategies such as Conflict-free Replicated Data Types (CRDTs) or based on timestamps, version numbers, and preset conflict resolution rules (e.g., taking the latest modification as the standard, or merging according to business rules) for automatic or semi-automatic conflict resolution and data merging. Through these mechanisms, it is ensured that the data generated on each node during the network disconnection can be reliably merged and synchronized, and finally reach the cross-node data consistency state.

[0171] Cross-node Data Analysis and Model Training Based on Federated Learning: As Figure 10 shown, cross-node data analysis and model training are carried out using federated learning without sharing the original data. This is especially applicable to industrial scenarios involving sensitive or private data. The process of federated learning usually includes multiple communication rounds, and each round involves the following main steps:

[0172] Global Model Distribution: The central coordination node (which can be a cloud server or a designated edge node) distributes the current global model (or its parameters) to multiple edge nodes or management units (clients) participating in federated learning.

[0173] Local Model Training: After receiving the global model, each client trains the model using its private original dataset locally. The training process is similar to traditional machine learning training, calculating the loss using local data and updating the model parameters. The key point is that the client does not upload the local original data to the central node or any other node.

[0174] Upload of Local Updates: After the client completes local training, it uploads the model parameter updates obtained from local training (e.g., the difference in model parameters or directly upload the new model parameters) to the central coordination node. To protect privacy, the uploaded data can be encrypted or processed with differential privacy.

[0175] Global Model Aggregation: The central coordination node collects the local updates from multiple clients and uses an aggregation algorithm to merge these updates to generate a new global model. A commonly used aggregation algorithm is federated averaging, whose basic idea is to calculate the weighted average of the model parameters uploaded by all clients (the weights are usually related to the data volume or model performance of the clients).

[0176] Model distribution and iteration: The central coordination node redistributes the aggregated new global model to the clients again to start the next training round. Repeat the steps until the model performance meets the requirements or reaches the preset number of training rounds.

[0177] Through this collaborative training method, the model can learn richer and more generalizable knowledge from data distributed across multiple nodes, while protecting the privacy and security of the data. This can be applied to joint modeling across factories, workshops, and even enterprises.

[0178] Appendix Figure 7 As shown: The digital twin collaborative optimization module runs through the entire industrial operating system and realizes system prediction, analysis, and optimization by constructing and utilizing the digital twin model of the physical factory.

[0179] Constructing the digital twin model of the physical factory: As Figure 12 shown, the digital twin model is a real-time mapping of the physical factory in the digital world. It is constructed based on the real-time mapping data collected by the edge connection layer, the historical data stored in the infrastructure layer, and the structured modeling tools provided by the technology base. The model can include:

[0180] Geometric models (3D models of equipment, layout models of factories, etc.); physical models (physical laws of equipment operation, physical and chemical laws of process processes, etc.); behavior models (control logics of equipment, scheduling rules of production processes, etc.); data models (real-time operation data, historical data, maintenance records, production plans, etc.).

[0181] These models can be constructed manually or automatically by integrating CAD / CAM data, sensor data, PLC programs, MES / ERP data, and using modeling tools.

[0182] Collaborative optimization based on the digital twin model: The digital twin collaborative optimization module establishes information channels and interaction interfaces with each level of the industrial operating system to achieve collaborative optimization.

[0183] Predicting System Performance and Potential Failures: Use the digital twin model to perform high-fidelity simulation on the operation of the physical plant. By inputting the real-time operation parameters collected by the edge connection layer and the historical data provided by the infrastructure layer, the model can predict the future state of equipment, the evolution of the process, the performance indicators of the system (such as output, energy consumption, quality), as well as the probability and time of potential failures. For example, by simulating the operation state of the equipment under the current load, predict the remaining life of its key components. The prediction results (for example, it is predicted that equipment A will have a bearing failure in 3 days, or it is predicted that network congestion is about to occur) are formatted into standardized event or data structures and fed back to each layer of the industrial operating system through the internal message bus or API interface of the system. For example, when an equipment anomaly or performance bottleneck is predicted, an instruction can be triggered to send to the dynamic scheduling unit of the edge connection layer, and accordingly adjust the data collection frequency or task priority of the relevant equipment; the prediction results can notify the edge AI model at the application layer to pre-load the corresponding diagnostic model or adjust the control strategy in advance; the predicted potential system performance bottleneck or resource shortage can be fed back to the infrastructure layer for resource elastic scaling or load balancing adjustment; the suggestions for predictive maintenance can be generated as workflow tasks at the business platform layer.

[0184] Simulating and Optimizing Business Processes: Use the digital twin model to simulate business processes built based on business logic code in the business platform layer, such as different production scheduling plans, different equipment maintenance strategies, different material distribution plans, etc. Run these scenarios in the digital twin environment, simulate their execution processes and results in the real physical plant, and evaluate their performance in terms of output, cost, efficiency, resource consumption, etc. By comparing the simulation results of different scenarios, find the optimal business processes and strategies. The optimized business process or strategy parameters can be deployed and applied to the actual physical plant through the configuration management interface or low-code tools at the business platform layer.

[0185] In addition to the above core modules and functions, there are also multiple features that enhance the overall system capabilities. These features are the result of the collaborative work of each layer and module. For example, the module hot-swap function supports the dynamic installation, uninstallation, and update of industrial applications, plugins, and extensions. This means that users can flexibly customize and expand the system functions according to actual needs without stopping the entire system operation, greatly improving the flexibility and usability of the system. This function is achieved through the service registration and discovery mechanism and the loose coupling interface between modules.

[0186] The above are only the preferred embodiments and are not intended to limit. Those skilled in the art should include the usual changes and substitutions within the scope of the technical solution in the protection scope.

Claims

1. An industrial operating system, characterized in that, It includes: An edge connection layer, which is used to collect the original data stream, perform protocol parsing and conversion according to the characteristics of the original data stream by matching preset parsing rules, support adaptive recognition of unknown protocols and generation of parsing rules, as well as event orchestration and processing, generate mapping data, and realize real-time mapping from the physical factory to the virtual factory. The edge connection layer also supports dynamic task scheduling and resource allocation based on the real-time operation status, and the dynamic task scheduling and resource allocation are realized based on a learning model; An infrastructure layer, including a technology base and a digital base. The technology base is used to process the mapping data and generate business logic code, and the digital base is used to manage cross-industry common data structures and store the mapping data; A business platform layer, which is used to support business hierarchical development and dynamic expansion based on the business logic code. The business platform layer includes a flexible design platform, low-code development tools, and a three-level expansion system composed of a basic application layer, a plug-in layer, and an extension layer; An application layer, which is used to provide installation, uninstallation, and update functions for industrial applications based on the three-level expansion system, and configure a multi-level authorization management system. The multi-level authorization management system supports attribute-based access control or a zero-trust model. The application layer also supports real-time decision-making and control based on edge artificial intelligence; A decentralized networking module, which is used to support autonomous operation of nodes and configuration of multi-level management units in a disconnected environment, and ensure cross-node consistency of the mapping data. The decentralized networking module also supports cross-node data analysis and model training based on federated learning; and A digital twin collaborative optimization module, which is used to build a digital twin model of the physical factory and perform collaborative optimization on each layer of the industrial operating system based on the digital twin model. The digital twin model is used to predict system performance and optimize business processes.

2. An industrial operating system according to claim 1, characterized in that, The edge connection layer includes: A protocol parsing and conversion unit, which is used to perform protocol parsing and conversion according to the characteristics of the original data stream by matching preset parsing rules; and An adaptive recognition and parsing rule generation unit, which is used to support adaptive recognition of unknown protocols and generation of parsing rules. The adaptive recognition and parsing rule generation unit is implemented through a machine learning-based protocol recognition module. The machine learning-based protocol recognition module automatically recognizes the type of unknown protocol according to the characteristics of the original data stream and automatically generates a protocol parsing rule corresponding to the unknown protocol type.

3. An industrial operating system according to claim 2, characterized in that, The edge connection layer includes a dynamic task scheduling and resource allocation unit, which is used to perform dynamic task scheduling and resource allocation based on the real-time operation status. The dynamic task scheduling and resource allocation unit is realized based on a learning model, and the learning model is a reinforcement learning model, which is used to adaptively adjust the priority of data collection tasks and / or resource allocation according to the real-time operation status including device status, data type, network congestion degree, and system load, so as to optimize the real-time performance and efficiency of data collection.

4. An industrial operating system according to claim 3, characterized in that, The digital twin collaborative optimization module uses the digital twin model to predict the performance bottlenecks and potential faults of the industrial operating system, and feeds back the prediction results to each layer of the industrial operating system to guide system optimization.

5. An industrial operating system according to claim 4, characterized in that, The digital twin collaborative optimization module is further configured to simulate and optimize the business processes in the business platform layer based on the digital twin model.

6. An industrial operating system according to claim 5, characterized in that, The decentralized networking module includes: A node autonomous operation unit, configured to support the autonomous operation of nodes and the configuration of multi-level management units in a disconnected network environment; A data consistency guarantee unit, configured to guarantee the cross-node consistency of the mapping data; and A cross-node data analysis and model training unit, configured to support cross-node data analysis and model training based on federated learning. The cross-node data analysis and model training unit is configured to coordinate multiple nodes to perform joint model training without sharing the original data.

7. An industrial operating system according to claim 1, characterized in that, The application layer includes an edge artificial intelligence decision and control unit, configured to support real-time decision and control based on edge artificial intelligence. The edge artificial intelligence decision and control unit is implemented by deploying an artificial intelligence-based inference model in the edge connection layer and the node autonomous controller. The inference model is configured to perform real-time analysis and decision on locally collected data, and is configured to directly control industrial devices or adjust data collection strategies.

8. An industrial operating system according to claim 1, characterized in that, The technical foundation includes an automated code generation tool, which is configured to generate business logic code according to the mapping data and business requirements. The automated code generation tool supports an artificial intelligence-based code generation model to generate code in multiple target languages and optimize it.

9. An industrial operating system according to claim 8, characterized in that, The digital foundation supports data modeling and dimension expansion according to the dimensions of space, material, time, and finance.

10. An industrial operating system according to claim 9, characterized in that, The industrial operating system has a module hot-swap function, supports the dynamic installation, uninstallation, and update of industrial applications, plugins, and extensions, and supports personalized selection according to actual application scenarios.

Images