Internet of Things processing method and device based on intention recognition and digital twinning

Abstract

The invention discloses an Internet of Things processing method and system based on intention recognition and digital twinning, and belongs to the technical field of Internet of Things. The method comprises the following steps: receiving a natural language intention representing a user business target; automatically generating an initial workflow based on the intention and a preset domain knowledge graph; performing predictive verification on the initial workflow based on a digital twin model associated with equipment in the workflow; and after deployment, on the basis of the real-time operation data and a preset reward function, online adaptive optimization is carried out on the operation parameters of the workflow through a reinforcement learning algorithm. According to the scheme, intelligence, automation and optimization of the whole life cycle of the workflow of the Internet of Things are achieved, and the application configuration efficiency, the system operation reliability and the resource utilization rate are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Things, and in particular to an Internet of Things processing method and device based on intent recognition and digital twinning. BACKGROUND

[0002] With the deep integration of 5G communication, artificial intelligence (AI) and edge computing technology, the Internet of Things (IoT) is evolving from simple device connection to a complex ecosystem of intelligent connection of all things. In this context, how to quickly, efficiently and reliably build and manage large-scale and high-complexity Internet of Things applications for different industries (such as intelligent manufacturing, smart city, smart home, etc.) has become a key to promoting industrial digital transformation.

[0003] In order to reduce the development threshold of Internet of Things applications, some visual configuration schemes are proposed in the prior art. For example, the prior art discloses an Internet of Things processing device, which provides a graphical interface. Users can intuitively configure a target workflow representing the working process of Internet of Things devices by dragging "device icons" representing physical devices and "logic icons" representing processing logic on the editing interface and connecting them with connectors. The scheme also supports simulation verification of the configured process and association of real feature data of Internet of Things devices in the process to monitor the actual working state. This visual, "building block" type of configuration simplifies the development process to some extent, enabling personnel without professional programming background to build Internet of Things applications, laying the foundation for rapid configuration of Internet of Things scenarios.

[0004] However, the inventors have found in their in-depth research and practice of the above-mentioned prior art scheme that it still has the following technical bottlenecks when dealing with increasingly complex intelligent Internet of Things scenarios: 1. The configuration process relies on human experience and lacks intelligence. Although this scheme realizes visual configuration, for a complex system (such as an intelligent factory containing hundreds of devices and complex logic), the building of the workflow completely relies on the deep understanding of the business logic by the implementers and manual dragging operation. The configuration process is tedious and inefficient, and the quality of the final process is highly dependent on human experience, lacking intelligent guidance and automatic generation capabilities, making it difficult to quickly respond to changing business needs.

[0005] 2. Simulation verification dimension is single, and depth prediction ability is lacked. The simulation verification in the scheme mainly focuses on the logical connectivity and functional correctness of the process, and cannot predict the comprehensive influence of the workflow running in the real physical world. For example, it cannot estimate the total energy consumption that a workflow will generate before deployment, the physical wear and tear caused to critical equipment (such as compressors, motors), and cannot quantify the expected degree of achievement of business goals (such as cost savings, production efficiency). The limitations of such verification can lead to hidden problems such as resource waste and premature aging of equipment when the configured process is actually run.

[0006] 3. The workflow is static and lacks adaptive optimization capability. Once the workflow configured by the existing scheme is deployed, its internal logic and parameters (such as delay, threshold) are usually fixed. When the external environment (such as seasonal changes, production scheduling adjustments, equipment aging) changes, the original static workflow may no longer be the optimal solution, and may even become inefficient or unsuitable. The system cannot learn and dynamically adjust itself according to real-time running data to continuously approach the optimal working state, and requires manual intervention for reconfiguration and optimization, resulting in high operation and maintenance costs and difficulty in ensuring system robustness and efficiency.

[0007] Therefore, how to provide an Internet of Things application processing scheme that can not only automatically generate a workflow from a user's business intent, but also perform deep predictive verification of its physical and business impact before deployment, and adaptively optimize it according to actual environmental changes after deployment, is a key technical problem that needs to be solved by those skilled in the art. SUMMARY

[0008] The present application aims to provide an Internet of Things processing system and method based on intent recognition and digital twinning to solve the deficiencies in the prior art. The technical problems to be solved by the present application are solved by the following technical solutions.

[0009] According to one aspect of the present application, an adaptive Internet of Things processing method based on intent recognition and digital twinning is provided, comprising: receiving a natural language intent representing a user's business goal; performing semantic analysis on the natural language intent to extract key entities, behaviors and constraint conditions, querying a pre-set domain knowledge graph for devices, logical relationships and best practices associated with the key entities, behaviors and constraint conditions, and automatically generating an initial workflow containing at least one device icon and a logic icon; performing predictive verification on the initial workflow based on a digital twinning model associated with the device icon in the initial workflow to output a verification result; the predictive verification comprises: predict at least one of resource consumption, physical wear, or business goal achievement of the initial workflow within a preset running period; and after deployment of the initial workflow, perform online adaptive optimization of running parameters of the initial workflow based on real-time running data and a preset reward function associated with the user business goal through a reinforcement learning algorithm.

[0010] Further, the semantic analysis of the natural language intent includes: using a natural language processing model based on a Transformer or a recurrent neural network to analyze the natural language intent to identify the key entities, behaviors, and constraint conditions.

[0011] Further, the preset reward function is specifically as follows: Reward = -w1 x E - w2 x Risk wherein E represents an energy consumption index, Risk represents a risk index, and w1 and w2 are adjustable weight coefficients.

[0012] Further, after the predictive verification of the initial workflow, the method further includes: displaying the verification result on a display interface; and in response to a user deployment instruction for the verification result, deploying the initial workflow to an actual Internet of Things device corresponding to the device icon.

[0013] Another aspect of the present application provides an adaptive Internet of Things processing device based on intent recognition and digital twinning, comprising: a receiving module configured to receive a natural language intent representing a user business goal; a generating module configured to perform semantic analysis of the natural language intent to extract key entities, behaviors, and constraint conditions, query a preset domain knowledge graph for devices, logical relationships, and best practices associated with the key entities, the behaviors, and the constraint conditions, and automatically generate an initial work containing at least one device icon and a logical icon; a verification module configured to perform predictive verification of the initial workflow based on a digital twinning model associated with the device icon in the initial workflow to output a verification result; the predictive verification includes predicting at least one of resource consumption, physical wear, or business goal achievement of the initial workflow within a preset running period; and an optimization module configured to, after deployment of the initial workflow, perform online adaptive optimization of running parameters of the initial workflow based on real-time running data and a preset reward function associated with the user business goal through a reinforcement learning algorithm.

[0014] Further, the generating module is specifically used for: performing semantic analysis on the natural language intention to extract key entities, behaviors and constraint conditions; and querying devices, logical relationships and best practices associated with the key entities, behaviors and constraint conditions in the domain knowledge graph to construct the initial workflow.

[0015] Further, the performing semantic analysis on the natural language intention comprises: analyzing the natural language intention by using a natural language processing model based on a Transformer or a recurrent neural network to identify the key entities, behaviors and constraint conditions.

[0016] Further, the preset reward function is specifically as follows: Reward=-w1×E-w2×Risk wherein E represents an energy consumption index, Risk represents a risk index, and w1 and w2 are adjustable weight coefficients.

[0017] According to another aspect of the present application, an electronic device is provided, which comprises a processor and a memory connected to the processor in communication; the memory stores computer program instructions; and the processor executes the computer program instructions stored in the memory to implement the method as described above.

[0018] According to still another aspect of the present application, a computer readable storage medium is provided, which stores a computer program; and the computer program is executed by a processor to implement the method as described above.

[0019] Compared with the prior art, the technical solution provided by the present application has at least the following beneficial effects: In view of the problem that the configuration process of the prior art relies on manual experience and has insufficient intelligence, the present application introduces an intention recognition module, which can receive the natural language intention of a user and automatically generate a workflow based on a domain knowledge graph. This scheme changes the complex configuration process from manual construction to intelligent generation, which not only greatly improves the configuration efficiency and reduces the dependence on the professional skills of the operator, but also ensures the initial quality and logical completeness of the generated process.

[0020] In view of the problem of single simulation verification dimension and lack of deep prediction capability in the prior art, the digital twin verification module is introduced, high-fidelity simulation is carried out before workflow deployment, and energy consumption, physical loss and business target achievement degree in future operation are accurately predicted and quantified. The scheme upgrades the traditional "function verification" to "comprehensive performance prediction", helps users to avoid potential design defects and operation risks, and realizes data-driven optimization decision.

[0021] In view of the problem of static solidification of the workflow and lack of self-adaptive optimization capability in the prior art, the self-adaptive optimization module is introduced, and the deployed workflow can continuously learn and fine-tune parameters according to the change of the real environment by using the reinforcement learning algorithm. The scheme realizes the fundamental change from "static execution" to "dynamic evolution", can independently cope with external disturbance, continuously approaches the optimal operation state, greatly reduces the artificial operation and maintenance cost, and guarantees the long-term efficient operation of the system in the dynamic environment. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a network architecture diagram of an application scenario of an embodiment of the present application; Figure 2 is a flowchart of an adaptive Internet of Things processing method based on intent recognition and digital twinning according to an embodiment of the present application; Figure 3 is a semantic analysis and workflow automatic construction process schematic diagram based on a domain knowledge graph according to an embodiment of the present application; Figure 4 is a structural block diagram of an adaptive Internet of Things processing device based on intent recognition and digital twinning according to an embodiment of the present application; Figure 5 is a hardware structure schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the present application clearer and more complete, specific embodiments of the present application will be described in detail below with reference to the drawings. It should be understood that the specific embodiments described herein are only to explain the present application, and do not constitute any form of limitation on the present application. Any modification and equivalent replacement made by those skilled in the art based on the disclosure herein, without creative labor, should fall within the scope of the present application.

[0024] The technical solution proposed in this invention can be applied to various computer systems that require efficient configuration and management of Internet of Things (IoT) applications, including but not limited to 5G, industrial internet (IoT), smart cities, intelligent manufacturing, and smart homes. For clarity, the following embodiments will primarily be described using an industrial internet application scenario integrating edge computing as a background.

[0025] Figure 1 A network architecture diagram illustrating an application scenario of an embodiment of the present invention is shown. Figure 1 As shown, the architecture includes: User Equipment (UE) 10, IoT terminal device 20, IoT gateway 30, edge computing node 40, and core network / cloud platform 50. The technical solution of the present invention can be executed in a distributed or centralized manner by one or more servers deployed on edge computing node 40 and / or core network / cloud platform 50.

[0026] In a specific application scenario, such as a smart factory, user device 10 can be the Human-Machine Interface (HMI) in the factory's central control room. IoT terminal devices 20 consist of various sensors (such as temperature sensors and vibration sensors) and actuators (such as motor controllers and alarm lights) on the production line. The IoT gateway 30 is responsible for aggregating data from field devices and communicating with terminal devices 20 via industrial protocols (such as Profinet and Modbus), while also communicating with edge computing nodes 40 via an IP network. Edge computing nodes 40 are servers deployed within the factory campus, responsible for low-latency real-time data processing. The core network / cloud platform 50 is a remote data center responsible for compute-intensive tasks and large-scale data storage.

[0027] Figure 2 A flowchart illustrating an adaptive Internet of Things (IoT) processing method based on intent recognition and digital twins according to an embodiment of the present invention is shown. In this embodiment, the method can be collaboratively executed by the edge computing node 40 and the core network / cloud platform 50 in the aforementioned application scenario architecture. The method specifically includes the following steps: Step S201: Receive natural language intent representing the user's business objective.

[0028] This step is the starting point of the whole process, aiming to take the user's high-level, unstructured business requirements as the input of the system. In one embodiment, the user (e.g., an operations engineer of a factory) can input a piece of natural language text describing his business goal through a text input box on the graphical user interface (GUI) of the user device 10. For example, the user can input: "Monitor the spindle motor of all CNC machines, if the motor temperature exceeds 85 degrees Celsius for 5 consecutive minutes, reduce the load of the machine by 50%, and notify the equipment maintenance team." This piece of text is the natural language intent. The electronic device (e.g., the edge computing node 40) executing this method receives the data packet containing this natural language intent sent by the user device 10 through the network interface.

[0029] Step S202, based on the natural language intent and a pre-set domain knowledge graph, automatically generating an initial workflow containing at least one device icon and a logic icon.

[0030] The purpose of this step is to automatically and structurally convert the natural language intent that is difficult for machines to directly understand into a visual and executable workflow. In one specific implementation, this step specifically includes: Performing semantic analysis on the natural language intent to extract key entities, actions, and constraint conditions. The processor executing this step can call a pre-trained natural language processing (NLP) model. This model analyzes the received text and identifies predefined semantic elements therein. For example, for the aforementioned example intent, the following can be extracted: Key entities: "CNC machine" (device type), "spindle motor" (device component), "equipment maintenance team" (target role).

[0031] Actions: "monitor" (data collection), "reduce load" (device control), "notify" (message push).

[0032] Constraint conditions: "temperature exceeding 85 degrees Celsius" (threshold), "5 consecutive minutes" (time window), "load reduction by 50%" (control parameter).

[0033] In a preferred implementation, a deep learning-based NLP model can be used to perform the above semantic analysis, such as a Transformer-based bidirectional encoding model (e.g., BERT) or a recurrent neural network model (e.g., bidirectional LSTM + CRF sequence labeling), to more accurately identify entities, actions, and constraint information in the user's intent.

[0034] Knowledge graph query and workflow construction: query the devices, logical relationships and best practices associated with the above key entities, behaviors and constraint conditions in the domain knowledge graph, and then construct the initial workflow. The domain knowledge graph is pre-stored in the memory, which can adopt a directed graph data structure, including Internet of Things device nodes (representing specific physical devices), logical function nodes (representing data processing or control logic), and their associated relationships (directed edges in the graph can represent data flow or control relationship between devices and logic). In an implementation manner, the data of the knowledge graph can be stored in the form of triples (subject-relation-object) in the graph database, or cached in the form of adjacency table in the memory. Each device node can be attached with attribute information describing device type, performance parameters, etc.; each logic node represents a reusable function module (such as "data acquisition", "threshold judgment", "alarm notification", etc.), and the edge connects the feasible connection between devices and logic, logic and logic, or the process order of industry best practice. Based on the elements obtained by semantic analysis, the processor performs query matching in the knowledge graph. For example: the query "monitor temperature" can match the "temperature sensor" device node and the "data acquisition" logic node; the query "more than 85 degrees Celsius for 5 minutes" can match the "time window aggregation" logic node and the "threshold judgment" logic node; the query "reduce load" will match an executable action supported by the "CNC machine tool" device node (such as reducing the spindle power). The processor instantiates these device nodes and logic nodes obtained by query into corresponding visual device icons and logic icons, and automatically connects the corresponding icons with connecting lines according to their pre-defined association relationship in the knowledge graph, thereby generating a structured and visual initial Internet of Things workflow.

[0035] Figure 3 The flowchart of the semantic analysis and automatic workflow construction based on the domain knowledge graph in one embodiment of the present application. As shown in the figure, the knowledge graph query and workflow construction module of the present embodiment mainly includes three functional areas, which are: domain knowledge graph area, semantic analysis and query matching processor area, and initial workflow generation area.

[0036] In the left domain knowledge graph area, the knowledge graph is stored in the form of a directed graph or triples, and is composed of multiple device nodes and logic nodes. The device node is used to represent the physical device in the Internet of Things system, for example, the device node includes "temperature sensor" and "CNC machine tool"; the logic node is used to represent data processing or control logic, for example, the logic node includes "data acquisition", "time window aggregation", "threshold judgment" and "alarm notification".

[0037] The nodes are connected by directed edges to represent data flow or control relationship. For example, the temperature sensor node supports the data collection logic node, the output of the data collection logic node serves as the input of the time window aggregation node; the output of the time window aggregation node is further transmitted to the threshold judgment node, and when the judgment condition is met, the threshold judgment node can trigger the alarm notification node and the execution behavior of the CNC machine tool device node.

[0038] The knowledge graph is pre-stored in the memory of the system in a structured form, and constitutes an associated network of knowledge in the field of Internet of Things.

[0039] In the semantic analysis and query matching processor region in the middle, the business intention of the user is first received through the natural language input interface, for example, "monitor the temperature, if the temperature exceeds 85℃ for 5 minutes in succession, then reduce the load and notify the maintenance personnel". The natural language intention is represented by an intention input unit.

[0040] The processor first performs extraction of entities, behaviors and constraint conditions, converts the natural language into structured semantic elements; then searches for the device nodes and logic nodes corresponding to the extraction results in the knowledge graph through a query module (for example, a matching query in the form of subject-relation-object); finally, the matched nodes are generated into corresponding device icons and logic icons through a mapping and instantiation module, which are used for subsequent workflow construction. The processor realizes semantic bridging and automatic mapping from natural language intention to structured graph nodes.

[0041] In the initial workflow generation region on the right side, the system automatically generates an initial workflow which is visualized and executable according to the output results of the processor.

[0042] The workflow includes a plurality of device icons and logic icons: the icons are sequentially associated through directed connection lines, forming a complete logic link from data collection to control execution.

[0043] For example, the data output by the temperature sensor is collected by the logic node, and then enters the time window aggregation node for 5-minute average processing, and then the threshold judgment node judges whether the temperature exceeds 85℃. If the judgment result is "yes", the action node executes the power reduction operation of the CNC machine tool, and at the same time, the action node sends a notification to the maintenance personnel. The automatically generated workflow can be directly used as a deployable configuration scheme of the Internet of Things system, realizing automatic construction from semantic intention to execution process.

[0044] In summary, Figure 3 The complete technical process of "semantic analysis-graph query-workflow construction" in the embodiment of the application is reflected, the automatic mapping from natural language to Internet of Things execution logic is realized, and the system configuration efficiency and intelligent degree are significantly improved.

[0045] Step S203, based on a digital twin model associated with the device icon in the initial workflow, the initial workflow is predicted to output a check result.

[0046] The purpose of this step is to deeply simulate and evaluate the comprehensive impact of the workflow before it is actually deployed to the physical system. In one embodiment, one or more digital twin models of each type of Internet of Things device are pre-stored in the memory. The digital twin model is a virtual mapping of the physical device in the digital space, which contains information such as physical characteristic parameters, performance curves, and failure mechanisms of the device. For example, for a motor device, the digital twin model can include its power consumption model (such as a function or curve of power with load), thermodynamic model (such as a differential equation of temperature rise and heat dissipation), and fatigue damage model of mechanical components, etc. These models can be constructed and calibrated based on the specification parameters and historical operation data provided by the device manufacturer to realistically simulate the actual behavior of the device.

[0047] When the processor executes this step, it loads the corresponding digital twin model for each device icon in the initial workflow and simulates the execution of the workflow in a discrete event simulation environment. For example, the simulator can simulate inputting a continuously rising motor temperature data stream to the digital twin model of the temperature sensor, and observe whether the entire workflow can trigger the execution action of "reducing load" under the correct conditions. In this way, the correctness of the workflow logic in the simulation environment can be verified.

[0048] Optionally, in a preferred embodiment, the predictive check also includes predicting the multi-dimensional performance indicators of the initial workflow within a preset operation period, such as resource consumption, physical wear and tear, and business target achievement. Specifically, during the simulation process, the processor not only verifies the correctness of the logical function, but also uses the algorithms built into the digital twin model to calculate the following indicators: Resource consumption: for example, by calling the power consumption calculation module in the motor twin model, the total power consumed in a day by executing the workflow (such as the frequent start-stop motor load reduction strategy) is predicted.

[0049] Physical wear and tear: for example, by calling the fatigue damage accumulation module in the motor twin model, the impact of the workflow on the service life of the motor bearing in a year of operation is predicted.

[0050] Business target achievement: for example, the expected success rate or contribution of the early warning load reduction strategy to the business goal of "avoiding non-planned downtime of the machine tool" is evaluated.

[0051] After the simulation is completed, the processor integrates all the prediction results to form a check result. For example, a comprehensive report containing the above quantitative indicator data and potential risk prompts can be generated.

[0052] Optionally, after performing the step, the method further comprises: presenting the verification result on the display interface, and deploying the initial workflow to the actual Internet of Things device corresponding to the device icon for running according to a deployment instruction issued by the user for the verification result. Specifically, the processor sends the generated verification report to the user device 10 for the user to view. After the user reviews the report, if the user considers that the result is acceptable, the user can send a confirmation deployment instruction through the interface. After receiving the instruction, the processor will issue the configuration of the workflow to the actual Internet of Things system (for example, to the Internet of Things gateway 30 or the edge computing node 40 for execution), so that the workflow logic is associated with the real physical device and starts online running.

[0053] Step S204, after the initial workflow is deployed, the running parameters of the initial workflow are online self-adaptive optimized based on real-time running data and a preset reward function through a reinforcement learning algorithm.

[0054] This step aims to enable the deployed workflow to have the ability of self-learning and continuous optimization, so as to adapt to the dynamically changing actual environment. In one embodiment, after the workflow is deployed and starts running, the processor starts one or more reinforcement learning agents to monitor and optimize the workflow. The agent continuously obtains real-time running data (such as real motor temperature, load, power consumption, environmental temperature, and other sensor data) from the actual Internet of Things device 20, and these data constitute the state space of the reinforcement learning algorithm. At the same time, the system pre-defines a reward function associated with the user's business goal in the memory. Optionally, the preset reward function is associated with the user's business goal. For example, if the user's intention is "to save energy as much as possible on the premise of ensuring safety", the reward function can be designed as: Reward=-w1×E-w2×Risk, wherein E represents an energy consumption index, Risk represents a risk index (such as the probability of device overheating leading to failure or a penalty value), and w1 and w2 are adjustable weight coefficients. The reward function gives a score to the instant performance of the system: the higher the energy consumption, the lower the reward (actually negative), and the higher the risk, also makes the reward decrease, thereby guiding the system to minimize energy consumption and running risk at the same time.

[0055] The processor executes a predetermined reinforcement learning algorithm (e.g., Q-learning algorithm or deep Q network, DQN algorithm) to continuously train the agent according to the current state and reward function. The agent takes the real-time state as input, selects a small adjustment to the adjustable parameters in the workflow as an action in the action space (these adjustable operating parameters constitute the action space of the algorithm, such as the temperature threshold for alarm triggering, the percentage of load reduction, and other policy parameters), and observes the reward change of the adjusted system. The reinforcement learning agent can use a certain exploration strategy (such as the ε-greedy strategy) to try different combinations of parameter values, iteratively learn through a large number of “trial and error”, and gradually update the state-action value (Q value) estimate. For the Q-learning algorithm, the agent stores the learned values in the Q table; for the DQN algorithm, a deep neural network is used to approximate the value function under a complex state space. Finally, the agent can automatically converge to a set of optimal parameters that maximize the long-term cumulative reward, thereby realizing online adaptive optimization of the workflow operation. In other implementations, the reinforcement learning algorithm can also use policy gradient-based methods (such as PPO, etc.), and these changes do not affect the core idea of the present application.

[0056] In combination with the aforementioned example scenario of intelligent factory machine tool cooling management, at the initial deployment of the workflow, the operating parameters depend on the default configuration in the knowledge graph (e.g., the temperature threshold is initially set to 85°C, and the load reduction ratio is 50%). After a period of online learning, the reinforcement learning agent automatically adjusts and optimizes these parameters. For example, the agent may lower the temperature threshold to 82°C and increase the load reduction ratio to 60% to more aggressively prevent overheating and reduce energy consumption. The actual running results show that the optimized workflow reduces the average energy consumption by about 10% compared to the initial configuration, and no device overheating shutdown event occurs, achieving better energy saving effect while ensuring device safety. As can be seen, through the above steps, the business intent (machine tool cooling and safe operation) proposed by the user in this embodiment is automatically executed and optimized end-to-end: the system first converts the user's natural language requirements into executable IoT workflows, then uses digital twin technology to perform in-depth prediction and verification before deployment, and finally uses the reinforcement learning agent to continuously optimize parameter settings during operation, ultimately achieving the user's desired business goal.

[0057] Figure 4 A structural block diagram of an adaptive IoT processing device 400 based on intent recognition and digital twin according to an embodiment of the present application is shown. The device can implement the method embodiments described above. As shown in Figure 4 The device includes a receiving module 401, a generating module 402, a verification module 403, and an optimization module 404.

[0058] The receiving module 401 is configured to perform step S201 in the method embodiments described above, i.e., receiving a natural language intent representing a user service target.

[0059] The generating module 402 is configured to perform step S202 in the method embodiments described above, i.e., automatically generating an initial workflow containing at least one device icon and a logic icon based on the natural language intent and a preset domain knowledge graph. Optionally, in an embodiment, the generating module 402 is specifically configured to: perform semantic analysis on the natural language intent to extract key entities, behaviors, and constraint conditions; and query the domain knowledge graph for devices, logic relationships, and best practices associated with the key entities, behaviors, and constraint conditions to construct the initial workflow.

[0060] The checking module 403 is configured to perform step S203 in the method embodiments described above, i.e., performing predictive checking on the initial workflow based on a digital twin model associated with the device icon in the initial workflow to output a checking result. Optionally, in an embodiment, the checking module 403 is specifically configured to: predict at least one of resource consumption, physical loss, or service target achievement degree of the initial workflow within a preset running period.

[0061] The optimization module 404 is configured to perform step S204 in the method embodiments described above, i.e., after deployment of the initial workflow, performing online adaptive optimization on the running parameters of the initial workflow based on real-time running data and a preset reward function through a reinforcement learning algorithm.

[0062] Figure 5 A hardware structure schematic diagram of an electronic device 500 according to an embodiment of the present application is shown. The electronic device 500 can be a server, a personal computer, an industrial control computer, or a node in a distributed computing system. As shown, the electronic device includes a processor 501, a memory 502, and a bus 503 connecting them. Optionally, a communication interface can also be included. Figure 5

[0063] The memory 502 stores computer program instructions, which can be a volatile memory (such as RAM) and / or a non-volatile memory (such as ROM, solid state disk). The computer program instructions include instructions for implementing any of the foregoing method embodiments. In addition, the memory 502 can also store the digital twin model and the domain knowledge graph.

[0064] ​The processor 501 can be a central processing unit (CPU), a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), or other processing core capable of executing a program. The processor 501 implements the above method when executing the computer program instructions stored in the memory 502.

[0065] The application further provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement any of the above methods. The storage medium can be non-volatile, such as a hard disk, a U disk, an optical disk, etc.

[0066] In the above detailed description, reference was made to the accompanying drawings, which form a part of the detailed description. In the drawings, like numerals typically identify like components, unless the context clearly dictates otherwise. The illustrative embodiments described in the detailed description, drawings, and claims are not meant to be limiting. Other embodiments can be used, and other changes can be made, without departing from the spirit or scope of the subject matter presented herein.

[0067] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An Internet of Things (IoT) processing method based on intent recognition and digital twins, characterized in that, include: Receive natural language intents that represent the user's business objectives; Semantic parsing is performed on the natural language intent to extract key entities, behaviors, and constraints. The devices, logical relationships, and best practices associated with the key entities, behaviors, and constraints are queried in a preset domain knowledge graph. An initial workflow containing at least one device icon and a logical icon is automatically generated. Based on a digital twin model associated with the device icon in the initial workflow, the initial workflow is predictively validated to output a validation result; The predictive verification includes: Predict at least one of the following: resource consumption, physical wear and tear, or degree of achievement of business objectives within a preset operating cycle of the initial workflow; as well as After the initial workflow is deployed, the operating parameters of the initial workflow are adaptively optimized online using a reinforcement learning algorithm based on real-time operating data and a preset reward function associated with the user's business objectives.

2. The method according to claim 1, characterized in that, The semantic parsing of the natural language intent includes: The natural language intent is analyzed using a natural language processing model based on Transformer or recurrent neural networks to identify the key entities, behaviors, and constraints.

3. The method according to claim 1, characterized in that, The preset reward function is as follows: Reward = -w1×E - w2×Risk Where E represents the energy consumption index, Risk represents the risk index, and w1 and w2 are adjustable weighting coefficients.

4. The method according to claim 1, characterized in that, After performing predictive validation on the initial workflow, the method further includes: The verification result is displayed on the display interface; and In response to a user deployment instruction for the verification result, the initial workflow is deployed to the actual IoT device corresponding to the device icon.

5. An Internet of Things (IoT) processing device based on intent recognition and digital twin, characterized in that, include: A receiving module is used to receive natural language intents that represent the user's business objectives; A generation module is used to perform semantic parsing on the natural language intent to extract key entities, behaviors and constraints, query devices, logical relationships and best practices associated with the key entities, behaviors and constraints in a preset domain knowledge graph, and automatically generate an initial workflow containing at least one device icon and a logical icon. A verification module is used to perform predictive verification on the initial workflow based on a digital twin model associated with the device icon in the initial workflow, so as to output a verification result; The predictive verification includes: predicting at least one of the following: resource consumption, physical loss, or degree of achievement of business objectives within a preset operating cycle of the initial workflow; as well as An optimization module is used to perform online adaptive optimization of the running parameters of the initial workflow after the initial workflow is deployed, based on real-time running data and a preset reward function associated with the user's business objectives, using a reinforcement learning algorithm.

6. The apparatus according to claim 5, characterized in that, The semantic parsing of the natural language intent includes: The natural language intent is analyzed using a natural language processing model based on Transformer or recurrent neural networks to identify the key entities, behaviors, and constraints.

7. The apparatus according to claim 5, characterized in that, The preset reward function is as follows: Reward = -w1×E - w2×Risk Where E represents the energy consumption index, Risk represents the risk index, and w1 and w2 are adjustable weighting coefficients.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer program instructions; When the processor executes the computer program instructions stored in the memory, it implements the method as described in any one of claims 1-4.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-4.

Images