Digital twinning-based multi-mode adaptive industrial control system and method

Abstract

The invention relates to the technical field of intelligent manufacturing, and discloses a multi-mode adaptive industrial control system and method based on digital twinning. The problem that a traditional industrial control system is insufficient in dynamic adaptability under complex working conditions can be solved to a certain extent. The system comprises a data acquisition and preprocessing module used for acquiring multi-modal original data and preprocessing all the acquired data to obtain various feature information; the digital twinborn model module is used for constructing a composite digital twinborn body based on a physical law, historical operation data and industry expert experience, and the control decision module is used for generating an initial control strategy through a multi-agent cooperation unit according to the composite digital twinborn body. Performing compliance verification and correction on the initial control strategy through a fuzzy logic unit, and outputting a final control strategy; and the execution and feedback module is used for executing control operation according to the final control strategy and transmitting state information to the digital twinborn model module.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and relates to, but is not limited to, a multimodal adaptive industrial control system and method based on digital twins. Background Technology

[0002] With the deepening development of industry and intelligent manufacturing, industrial control systems are accelerating their transformation from automation to intelligence. In complex and ever-changing industrial scenarios, traditional adaptive control algorithms are unable to meet the control requirements in dynamic environments. Parameter fluctuations, process adjustments, sudden failures, and the complexity of information processing brought about by multimodal data during equipment operation have all caused control strategies based on fixed models to gradually expose problems such as response lag and decreased accuracy.

[0003] While digital twin technology offers new ideas for industrial system modeling and optimization, existing digital twin models have significant shortcomings in dynamic adaptability. They cannot reflect real-time changes in physical entities in a timely manner, leading to a continuous accumulation of mapping deviations between the virtual and the real world, which affects the accuracy and effectiveness of control decisions.

[0004] Furthermore, industrial production places increasingly stringent demands on the real-time performance, reliability, and security of systems. Insufficient edge computing capabilities, low efficiency of cloud-edge collaboration, and data security risks also constrain the development of industrial intelligence. In key industries such as chemical, energy, and automobile manufacturing, insufficient adaptability of control systems frequently leads to decreased production efficiency, fluctuations in product quality, energy waste, and even safety accidents.

[0005] Therefore, there is a need for an industrial control system that can effectively integrate multi-source data, dynamic optimization models, and intelligent decision-making control to improve the intelligence level and core competitiveness of industrial systems. Summary of the Invention

[0006] In view of this, in order to solve the problem of insufficient dynamic adaptability of traditional industrial control systems under complex working conditions, this application provides a multimodal adaptive industrial control system based on digital twin.

[0007] The specific technical solutions of this invention are as follows: In a first aspect, embodiments of the present invention provide a multimodal adaptive industrial control system based on digital twins, comprising: The data acquisition and preprocessing module is used to acquire multimodal raw data from the industrial site in real time and preprocess all acquired data to obtain various feature information. The multimodal raw data includes equipment operating status, production process parameters, environmental changes, physical laws, historical operating data and industry expert experience. The digital twin model module is used to construct a composite digital twin based on physical laws, historical operating data, and industry expert experience. The composite digital twin includes a physical model, a behavioral model, and a rule model. The control decision module is used to generate a final control strategy based on the composite digital twin through a multi-agent collaborative unit and a fuzzy logic unit, wherein the multi-agent collaborative unit includes at least two agents trained through reinforcement learning. The execution and feedback module is used to execute control operations according to the final control strategy and transmit status information to the digital twin model module. The status information includes actual operating parameters, environmental status and abnormal signals.

[0008] In one possible implementation, the data acquisition and preprocessing module further includes: The multi-source data acquisition unit is used to comprehensively cover industrial scenarios and acquire multimodal raw data through multi-dimensional deployment of acquisition points; The preprocessing unit is used to clean, denoise, convert formats, and extract features from the multimodal raw data to obtain various feature information.

[0009] In one possible implementation, the digital twin model module is further configured to fuse the various feature information and state information, and to dynamically calibrate and optimize the composite digital twin based on the fusion result.

[0010] In one possible implementation, the generation of the final control strategy through a multi-agent collaborative unit and a fuzzy logic unit further includes: The initial control strategy for each industrial control task subsystem is generated by a multi-agent collaborative unit. The subsystem's operating status is evaluated using fuzzy logic units, and the initial control strategy is verified and corrected based on the evaluation results to output the final control strategy.

[0011] In one possible implementation, the multimodal adaptive industrial control system further includes: The interaction module is used to receive adjustment instructions from the operator and convert the adjustment instructions into executable digital signals and transmit them to the execution and feedback module.

[0012] In one possible implementation, the execution and feedback module further includes: The visualization unit is used to monitor the equipment operation trend, production process progress and system health status in real time, and will determine whether to issue an abnormal alarm based on the abnormal signals in the status information.

[0013] In one possible implementation, each agent learns through a deep Q-network and collaboratively optimizes control decisions by sharing information. Each agent is responsible for a specific control task, including equipment operation, logistics scheduling, and quality monitoring.

[0014] In one possible implementation, the fuzzy logic unit includes: The fuzzification interface is used to map accurate input to the corresponding fuzzy set and calculate the membership value through the membership function; A fuzzy rule base is used to predefine fuzzy rules based on expert experience to indicate the correspondence between inputs and outputs; A fuzzy inference engine is used to activate corresponding rules based on the fuzzy set corresponding to the input, and obtain the membership degree of the fuzzy set corresponding to the output through fuzzy logic operations. The deblurring interface is used to convert fuzzy output into precise values.

[0015] In one possible implementation, the construction of the composite digital twin based on feature information derived from physical laws, historical operational data, and industry expert experience further includes: Based on the equipment's structural design drawings, material property parameters, and dynamic principles, the mechanical motion, energy transfer, and thermodynamic changes of the physical entity are modeled to obtain a physical model. Based on data-driven machine learning algorithms, a behavioral model is constructed according to historical operating data and the physical model. The behavioral model includes the operating mode during normal operation and abnormal signs before the occurrence of failure. Based on process standards, safety regulations, and industry laws, a rule model is constructed, which includes decision constraints. Secondly, this application provides a multimodal adaptive industrial control method based on digital twins, including: Acquire multimodal raw data, which includes equipment operating status, production process parameters, environmental changes, physical laws, historical operating data, and industry expert experience; All acquired data are preprocessed to obtain various feature information; A composite digital twin is constructed based on physical laws, historical operational data, and industry expert experience. The composite digital twin includes a physical model, a behavioral model, and a rule model. Based on the integrated analysis of characteristic and status information of equipment operating status, production process parameters and environmental changes, the composite digital twin is dynamically calibrated and optimized. Based on the composite digital twin, an initial control strategy is generated through a multi-agent collaborative unit, and then the initial control strategy is verified and corrected through a fuzzy logic unit to output the final control strategy. Control operations are executed according to the final control strategy, and status information is transmitted to the digital twin model module. The status information includes actual operating parameters, environmental status, and abnormal signals.

[0016] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: The data acquisition and preprocessing module provided in this application integrates multiple types of data, such as equipment operating parameters, visual images, and process logs, through a multi-acquisition module. Combined with the cleaning, feature extraction, and standardization processing of the preprocessing module, it solves the problems of data fragmentation and strong noise interference in traditional systems, improving data availability by more than 40%.

[0017] The digital twin model module provided in this application constructs a full-dimensional virtual image covering equipment structure, operating rules, and process constraints through composite modeling of physical models, behavioral models, and rule models. The model update module dynamically calibrates model parameters based on real-time feedback data (such as parameter drift caused by equipment wear), keeping the state deviation between the virtual model and the physical entity within 3%, which is more than 50% more accurate than traditional static models. It breaks through the limitations of traditional digital twins in "visual reproduction" and realizes the pre-simulation and dynamic optimization of the production process through bidirectional interaction between the model layer and the control decision layer.

[0018] The control decision module provided in this application achieves dynamic control of multivariable, strongly coupled industrial systems by using a deep Q-network-driven agent to autonomously learn control strategies and combining them with rule constraints from a fuzzy logic module. Furthermore, the fuzzy logic module can effectively handle uncertainties in industrial environments (such as sensor noise and process fluctuations). By dynamically adjusting the reward function and exploration strategy of reinforcement learning, the system maintains stable control even under scenarios such as equipment wear and raw material composition fluctuations. The convergence speed of the control strategy is improved by 30%, and the anti-interference capability is significantly enhanced.

[0019] The execution and feedback module provided in this application forms a complete closed loop of "decision-execution-feedback-correction" through real-time status acquisition and deviation calculation, which shortens the control cycle and improves the response speed to abnormal conditions. In addition, the visualization unit presents the system status in real time in the form of three-dimensional twin models and dynamic curves. Combined with the alarm triggering mechanism, it improves the operator's efficiency in identifying abnormal conditions by 70%. At the same time, it supports flexible switching between manual parameter adjustment and automated control. In flexible production scenarios, the work order switching time is shortened, which significantly improves the adaptability of the production line. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A schematic diagram of the system structure of a multimodal adaptive industrial control system based on digital twins provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a multimodal adaptive industrial control method based on digital twins, provided as an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0023] It should be noted that the terms "first, second, and third" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0024] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments of the invention pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0025] Figure 1This is a schematic diagram of the system structure of a multimodal adaptive industrial control system based on digital twins, provided as an embodiment of the present invention. Figure 1 As shown, the system may include: The data acquisition and preprocessing module is used to acquire multimodal raw data and preprocess all acquired data to obtain various feature information. The multimodal raw data includes equipment operating status, production process parameters, environmental changes, physical laws, historical operating data, and industry expert experience.

[0026] The data acquisition and preprocessing module further includes: The multi-source data acquisition unit, through multi-dimensional deployment of acquisition points, provides comprehensive coverage of industrial scenarios and acquires multimodal raw data.

[0027] In some embodiments, the multi-source data acquisition unit acts as a data acquisition interface for the system, providing comprehensive coverage of industrial scenarios through multi-dimensionally deployed acquisition points, and efficiently acquiring multi-modal raw data including equipment operating status, production process parameters, and environmental changes.

[0028] In real industrial environments, equipment operating status data can reflect the wear and performance changes of mechanical parts, while production process parameters record key information such as material ratios and process temperatures. Environmental parameters such as temperature, humidity, and dust concentration also have a significant impact on the production process. Multi-source data acquisition units focus on overcoming the challenges of dispersed data sources and complex and heterogeneous data types, achieving accurate capture of all elements of industrial production and providing rich raw data resources for system operation.

[0029] The preprocessing unit cleans, denoises, converts formats, and extracts features from the multimodal raw data to obtain feature information.

[0030] In some embodiments, the preprocessing unit acts as the core of data processing, performing in-depth processing on the raw data acquired by the multi-source data acquisition module. In response to common problems in the raw data, such as noise interference, outliers, and inconsistent formats, the preprocessing module sequentially performs operations such as cleaning, denoising, format conversion, and feature extraction.

[0031] In the data cleaning process, logical errors and invalid data are removed by setting reasonable data verification rules.

[0032] The noise reduction process uses specialized algorithms to remove noise caused by sensor fluctuations or transmission interference.

[0033] Format conversion unifies data from different protocols and standards into a standardized format that the system can recognize.

[0034] Feature extraction utilizes mathematical transformations and machine learning algorithms to extract representative key features from massive amounts of data.

[0035] This process allows the raw data to be refined and its quality and usability to be significantly improved. It provides standardized, high-quality data support for the accuracy of subsequent digital twin model construction and the scientific nature of control decisions, ensuring that the system can operate efficiently and make intelligent decisions based on reliable data.

[0036] The digital twin model module is used to construct a composite digital twin based on physical laws, historical operating data, and industry expert experience. The composite digital twin includes a physical model, a behavioral model, and a rule model.

[0037] The physical laws include the equipment's dynamic equations, thermodynamic laws, and material mechanical properties, while the industry expert experience includes the standard range of process parameters, best practices for equipment maintenance, and expert methods for troubleshooting.

[0038] In some embodiments, the digital twin model module includes a model building unit, which serves as the basis for mapping physical entities in the virtual world. Its construction process is like a precise digital jigsaw puzzle, requiring deep integration of multi-dimensional information. This module integrates physical laws, historical operating data, and industry expert experience. The historical operating data includes equipment monitoring data (such as temperature, pressure, vibration, etc. collected by sensors in real time), operation and maintenance records (fault repair logs, component replacement time, etc.), production reports (output statistics, energy consumption data, etc.), and historical simulation data (past model calculation results) to build a composite digital twin that includes a physical model, a behavioral model, and a rule model.

[0039] The construction steps of the composite digital twin further include the following steps S31 to S33: S31. Based on the equipment's structural design drawings, material property parameters, and dynamic principles, a refined model is created to model the mechanical motion, energy transfer, and thermodynamic changes of the physical entity, resulting in a physical model.

[0040] Specifically, a physical model is a precise digital replica of a physical entity. Based on the structural design drawings, material property parameters, and dynamic principles of the equipment, it uses finite element analysis, multibody dynamics simulation, and other technical means to finely model the mechanical motion, energy transfer, thermodynamic changes, and other characteristics of the physical entity. Taking an industrial steam turbine as an example, the physical model can not only simulate the stress distribution of the blades under high-speed rotation, but also predict the impact of steam flow on efficiency.

[0041] S32, based on a data-driven machine learning algorithm, constructs a behavioral model according to historical operating data and the physical model. The behavioral model includes the operating mode during normal operation and abnormal signs before a failure occurs.

[0042] Specifically, the behavior model serves as the behavioral standard for the digital twin. Relying on data-driven machine learning algorithms, it extracts patterns from massive amounts of time-series data of device operation. Through deep learning models such as Long Short-Term Memory Network (LSTM) and Convolutional Neural Network (CNN), the behavior model can learn the operating patterns of the device during normal operation and abnormal signs before a failure occurs.

[0043] S33. Based on process standards, safety specifications and industry regulations, a rule model is constructed. The rule model includes decision constraints, and an early warning mechanism is automatically triggered when a threshold is reached.

[0044] Specifically, rule models act as operational constraints for digital twins. They integrate knowledge from fields such as process standards, safety regulations, and industry regulations to form decision-making constraints. For example, in chemical production, rule models compare process parameters such as upper limits of reaction temperature and pressure safety thresholds with real-time monitoring data. Once a threshold is triggered, an early warning mechanism is automatically activated, and suggestions for standard operating procedures are provided.

[0045] The physical model, behavioral model, and rule model interact and collaborate through a data interface. The physical model provides basic structural constraints for the behavioral model, the behavioral model provides dynamic decision-making basis for the rule model, and the rule model in turn optimizes the parameter settings of the physical and behavioral models, jointly constructing a virtual image that can comprehensively reflect the state and behavior of physical entities.

[0046] In some embodiments, the digital twin model module further includes a model update unit. After construction, the model update unit integrates feature information and status information based on equipment operating status, production process parameters, and environmental changes. Based on the integration results, the composite digital twin is dynamically calibrated and optimized, and the data of the composite digital twin is updated in real time. This model update unit is the guarantee for ensuring the real-time performance and accuracy of the digital twin. Through the millisecond-level real-time feature data transmitted by the data acquisition and preprocessing module, and the full-dimensional status information returned by the execution and feedback module, the constructed digital twin model is dynamically calibrated and deeply optimized.

[0047] The model update unit deeply integrates the real-time feature data from the data acquisition and preprocessing module with the status information returned by the execution and feedback module using existing multi-source heterogeneous data fusion algorithms. The specific steps are as follows: The two types of data are time-stamp aligned and format standardized to eliminate data dimensional differences; spatiotemporal correlation analysis algorithms are used to explore the dynamic mapping relationship between the data; the fused data is used as input, and by analyzing these integrated data, the differences between the digital twin model and the actual situation are found, and the model parameters and structure are adjusted accordingly; so as to achieve dynamic calibration and optimization of the model, so that the digital twin always fits the real state.

[0048] Specifically, the model update unit dynamically corrects model parameters by combining a wear prediction model and a thermal deformation compensation model constructed using machine learning algorithms. The wear prediction model can predict the wear of equipment components in advance based on the received parameters, providing data support for the formulation of maintenance plans. The thermal deformation compensation model can predict and compensate for physical deformation caused by heat based on changes in parameters such as temperature, ensuring production accuracy. The input parameters of the wear prediction model include multi-dimensional sensor data such as equipment running time, load pressure, ambient temperature, and lubrication status, and the output is a predicted value of the wear degree of key equipment components in a future period.

[0049] During training, a Long Short-Term Memory (LSTM) network is used to form a time-series dataset by combining historical wear data with corresponding operating parameters. The model weights are optimized through backpropagation to minimize the prediction error. The thermal deformation compensation model takes real-time parameters such as feed rate, cutting force, and material thermal expansion coefficient during the machining process as input and outputs thermal deformation compensation parameters of the target workpiece (such as tool offset and temperature compensation coefficient). The model is trained using a Convolutional Neural Network (CNN). A training set is constructed by collecting a large amount of machining experimental data. The model parameters are iteratively adjusted using the mean square error loss function to achieve accurate prediction and compensation of thermal deformation trends.

[0050] In some embodiments, during continuous high-intensity cutting operations, the microscopic wear of the cutting edge of the CNC machine tool in the smart factory, the change in the clearance of the spindle bearing, and even the thermal expansion of the material caused by fluctuations in ambient temperature and humidity can all lead to slight deviations between the actual state of the physical entity and the virtual model.

[0051] When the system detects that the tool wear exceeds the preset threshold, the model update unit will automatically call the adaptive parameter adjustment algorithm to synchronously modify key indicators in the virtual model, such as the cutting force coefficient and surface roughness parameters (obtained through measurement / simulation calculation / historical data accumulation), so that the virtual model can follow the dynamic evolution of the physical entity with sub-millimeter accuracy.

[0052] This enables real-time mapping of digital twins to physical entities. Furthermore, by constructing a closed-loop control system of "monitoring-analysis-optimization-feedback," the virtual model possesses self-learning and self-adaptive capabilities, providing the control decision-making layer with multi-dimensional decision-making basis, including fault warning, performance prediction, and energy consumption optimization, truly realizing precise simulation and intelligent control of industrial production processes.

[0053] The control decision module is used to generate an initial control strategy based on the composite digital twin through a multi-agent collaborative unit, and then perform compliance verification and correction on the initial control strategy through a fuzzy logic unit to output the final control strategy.

[0054] In some embodiments, the multi-agent collaborative unit includes multiple agents, each of which learns through a deep Q-network and collaboratively optimizes control decisions by sharing information. Each agent is responsible for a specific control task, which includes equipment operation, logistics scheduling, and quality monitoring.

[0055] The agent learns using a deep Q-network (DQN). Taking a device-operating agent as an example, its state space includes real-time operating parameters (temperature, pressure, speed, etc.), historical fault records, and maintenance information; its action space contains various control commands (adjusting speed, opening / closing valves, etc.). The agent selects actions based on its current state and observes the reward signals from the environment after executing the actions (positive rewards for improved device stability, reduced energy consumption, and improved product quality, and negative rewards for faults and production interruptions). By continuously iterating and updating the Q-value, the agent gradually learns the optimal control strategy (equivalent to the aforementioned initial control strategy). Multiple agents share information through communication mechanisms and collaboratively optimize control decisions. For example, when planning material transportation routes, the logistics scheduling agent refers to the equipment busyness information provided by the equipment operation agent to avoid conflicts between transportation routes and equipment maintenance periods, thereby improving overall production efficiency.

[0056] The data input for the multi-agent collaborative unit originates from the feature data (such as equipment operating parameters, environmental indicators, and prediction deviations of digital twin models) obtained by the data acquisition and preprocessing module. This data is then structured and encapsulated according to the agents' roles. For example, in a chemical reaction control scenario, each agent (such as a temperature control agent or a pressure regulation agent) collects state data for its corresponding dimension (such as reactor temperature values ​​and pressure fluctuation curves). After normalization, this data forms a state space. ,in The standardized one-dimensional feature value is n, where n is the feature dimension; the action space A is defined as discrete control commands (such as valve opening adjustment level, motor speed range), which are converted into an input format that the model can recognize through one-hot encoding.

[0057] The reinforcement learning model of this deep Q-network architecture includes: The input layer receives the agent's state vector S, with the same dimension as the number of features.

[0058] The hidden layer uses a 2-3 layer fully connected neural network combined with the ReLU activation function to extract higher-order correlations of state features.

[0059] The output layer has the number of nodes equal to the dimension of the action space and outputs the Q-value (expected cumulative reward) of each action, i.e., where is the network parameter.

[0060] Meanwhile, during the multi-agent collaborative training process, an experience replay strategy is adopted. The agent stores the samples generated by interacting with the environment into the experience pool. Random sampling reduces data correlation and improves training stability.

[0061] Specifically, the following method is used: The current network parameters are periodically copied to the target network using a target network update method. Used to calculate the target value: ; in Indicates the agent's state Execute action Then, transition to state Instant rewards obtained at that time Discount factor (0≤ ≤1), used to weigh the importance of immediate rewards versus future rewards. The closer to 0, the more the agent focuses on immediate rewards; the closer to 1, the more it values ​​long-term gains. Indicates the state Next, from all possible actions Selecting from the middle can enable The action with the highest value Target network State The next action to be performed Value estimation: This method effectively alleviates the value estimation bias problem.

[0062] In the optimization phase, the mean squared error is used as the loss function to measure the difference between the predicted Q-value and the target Q-value. This is combined with the Adam optimizer, and the learning rate is set... The network parameters are updated at a certain scale to optimize the training effect.

[0063] This intelligent agent adopts In the reasoning stage, probability is used. Randomly explore the action space, with 1- The action with the highest current Q value is selected to balance exploration and exploitation. Meanwhile, each agent achieves policy coordination by sharing some state information or reward signals (such as global production efficiency indicators) based on a distributed training framework (such as A3C) to avoid local optima.

[0064] The fuzzy logic unit includes: a fuzzification interface, which maps accurate inputs to corresponding fuzzy sets and calculates membership values ​​using a membership function; a fuzzy rule base, which predefines fuzzy rules based on expert experience and describes the input-output relationship in "IF-THEN" form; a fuzzy inference engine, which activates corresponding rules based on the fuzzy set corresponding to the input and calculates the membership degree of the fuzzy set corresponding to the output through fuzzy logic operations; and a defuzzification interface, which converts the fuzzy output into precise values.

[0065] In some embodiments, to compensate for the shortcomings of reinforcement learning in handling fuzzy and uncertain information, fuzzy logic is introduced into the control algorithm. The fuzzy logic unit consists of a fuzzification interface, a fuzzy rule base, a fuzzy inference engine, and a defuzzification interface.

[0066] Taking equipment fault diagnosis as an example, the vibration and noise signals collected by the data acquisition and preprocessing modules are often fuzzy and difficult to describe with precise numerical values. Through the fuzzification interface, these signals are converted into fuzzy linguistic variables (such as "violent vibration" and "high noise"). A series of fuzzy rules based on expert experience and historical data are predefined in the fuzzy rule base, such as "if the vibration is severe and the noise is high, the equipment may have a serious fault". The fuzzy inference engine infers the probability of equipment fault based on the input fuzzy linguistic variables and the fuzzy rules, and finally, the fuzzy conclusion is converted into a precise fault probability value through the defuzzification interface, providing a decision-making basis for the equipment operation agent. In terms of adjusting control strategies, fuzzy logic can dynamically adjust the weights of the reward function of reinforcement learning based on the fuzzy evaluation of the system's operating state. For example, when the system is on the edge of instability, the weights of stability-related rewards can be increased to guide the agent to prioritize control actions that stabilize the system.

[0067] In the fusion mechanism of the multi-agent collaborative module and the fuzzy logic module provided in this application, the agent first generates an initial control policy based on the current system state through a reinforcement learning algorithm. Then, the fuzzy logic system performs a fuzzy evaluation of the system state to determine the complexity and uncertainty of the current operating condition. If the operating condition is relatively stable, the initial control policy generated by reinforcement learning is executed. If the operating condition is complex or has significant uncertainty, the fuzzy logic system dynamically adjusts the weight of the reinforcement learning reward function according to pre-set rules to guide the agent to regenerate a more reasonable final control policy.

[0068] Specifically, during the machine tool processing in the smart workshop, when fluctuations in the hardness of the workpiece material are detected, the fuzzy logic system determines that the complexity of the working condition has increased, increases the weight of the rewards related to processing accuracy, and prompts the intelligent agent of the equipment to adjust the cutting parameters to ensure processing quality.

[0069] Multi-agent reinforcement learning and fuzzy logic work together to form a closed-loop optimization mechanism. Fuzzy logic provides reinforcement learning with more reasonable state descriptions and decision guidance, helping the agent to converge to the optimal control strategy more quickly. Reinforcement learning, on the other hand, optimizes the parameters and rules of the fuzzy logic unit through continuous trial and error learning. To improve algorithm efficiency, a distributed computing architecture is adopted, distributing the learning and computing tasks of each agent to edge computing nodes and cloud servers. Edge computing nodes are responsible for handling tasks with high real-time requirements and small data volumes (such as generating local control commands for devices), while cloud servers undertake large-scale data storage, complex model training, and global optimization tasks. At the same time, optimization techniques such as genetic algorithms are introduced to optimize the network structure of multi-agent reinforcement learning, the rules and parameters of fuzzy logic systems, thereby improving the algorithm's global search capability and convergence speed.

[0070] By adjusting parameters through a feedforward mechanism, the fuzzy logic unit dynamically adjusts the hyperparameters of the reinforcement learning model based on real-time system conditions (such as equipment operational stability indicators and data noise levels). When significant data fluctuations are detected, a "increase exploration rate" signal is output through fuzzy rules to enhance the agent's adaptability to new operating conditions. When the system enters a steady state, the exploration rate is reduced and the weight of long-term returns in the reward function is increased to accelerate policy convergence. The fuzzy weighted reward function is the original reward signal of reinforcement learning (such as product quality pass rate and energy consumption indicators). After processing by the fuzzy logic module, a dynamic reward value is generated. 0.

[0071] in, The dynamic reward value, adjusted by fuzzy weighting, serves as a new reward signal for the reinforcement learning algorithm, guiding the agent to adjust its decisions and thus optimizing the industrial collaborative control process. This is a fuzzy weighting coefficient, the value of which is dynamically determined by the fuzzy logic module based on the current system state (such as process parameters, equipment maintenance cycles, etc.) through fuzzy inference. This coefficient serves to amplify or reduce the original reward signal. 0 represents the initial reward signal for reinforcement learning, which can be a quantitative indicator that directly reflects the achievement of industrial control objectives, such as product quality pass rate or energy consumption index.

[0072] Specifically, if the current process parameters are close to the safety threshold, the fuzzy logic module will output a larger value. Value, strengthen the reward for "risk avoidance" behavior; if the equipment is in a maintenance cycle, The value was reduced to prioritize ensuring production continuity.

[0073] In some embodiments, the specific implementation process of the multi-agent collaboration module and the fuzzy logic module working together is as follows: The multi-agent system generates an initial control strategy (such as valve opening recommendations based on DQN), and the fuzzy logic module performs compliance verification and correction on the strategy (such as limiting the valve opening to no more than the safety limit).

[0074] Complementary outputs are used in high-dimensional complex scenarios (such as multi-device collaborative control). The agent is responsible for handling the dynamically changing continuous state space, while the fuzzy logic module handles discrete rule constraints (such as safety procedures). Finally, the final control strategy is generated through linear combination or a voting mechanism, as follows: ; in, This is the final control strategy generated, used to guide the actual operation of the industrial system. The initial control strategy is generated for agents such as Deep Q-Networks (DQNs) based on a dynamically changing continuous state space. This is a modified control strategy generated by the fuzzy logic module based on discrete rule constraints. This is a weighting coefficient, with a value range of [0,1]. It is dynamically adjusted by the complexity of the operating conditions and is used to balance... and The extent of their contribution.

[0075] In the above embodiments, reinforcement learning's data-driven capability solves the problem of autonomous decision-making under complex working conditions; fuzzy logic's rule-constraining capability enhances the security and interpretability of decisions; the combination of the two achieves collaborative control of "learning strategies from data and regulating behavior in rules," significantly improving the adaptability and reliability of industrial systems in uncertain environments.

[0076] The execution and feedback module is used to execute control operations according to the final control strategy and transmit status information to the digital twin model module. The status information includes actual operating parameters, environmental status and abnormal signals.

[0077] In some embodiments, the execution and feedback module undertakes the dual tasks of real-time status acquisition and error calculation. Through various sensors (such as current sensors and position encoders) deployed in industrial equipment and production lines, it continuously captures the actual operating parameters (such as motor speed and valve opening), environmental conditions (such as temperature and humidity fluctuations), and abnormal signals (such as vibration over-limit alarms) after the equipment executes the final control strategy. It then compares these data with the command parameters issued by the control decision layer, calculates the deviation value (such as the difference between the temperature setpoint and the measured value), and after standardizing the deviation data, feeds it back to the data and preprocessing layer in real time in the form of digital signals, forming a closed-loop control link. This enables the system to dynamically adjust the model parameters and the final control strategy according to the actual execution effect, ensuring the accurate achievement of the control target.

[0078] The execution and feedback module further includes: The visualization unit is used to monitor the equipment operation trend, production process progress and system health status in real time, and will determine whether to issue an abnormal alarm based on the status information.

[0079] This visualization unit is the core window for presenting the system status. It mainly displays the real-time data collected by the feedback module, the dynamic operating status of the digital twin model module, and the execution process of the final control strategy in an intuitive graphical form in multiple dimensions.

[0080] This unit uses visualization components such as dynamic graphs (e.g., trends in equipment operating parameters), interactive dashboards (e.g., real-time values ​​of key indicators), and 3D virtual scenes (e.g., dynamic mapping of digital twins of production line equipment) to transform the operating status of industrial systems into easily readable visual information.

[0081] Specifically, it can display the waveform changes of equipment vibration amplitude in real time, the dynamics of material conveying path in the production process, and the prediction results of equipment failure by the digital twin model module (such as remaining life countdown). At the same time, the visualization unit integrates an alarm triggering mechanism. When the system detects abnormal data (such as temperature exceeding the threshold or equipment operating status deviating from the normal range), it issues alarm signals in the form of color flashing, pop-up prompts, etc., to help maintenance personnel quickly locate the problem.

[0082] Its core value lies in lowering the threshold for data understanding through visualization technology, making the operating status, potential risks, and optimization space of complex industrial systems perceptible and insightful, providing intuitive basis for the strategic adjustments of the control decision-making level, and improving the transparency and management efficiency of industrial production processes.

[0083] Furthermore, some embodiments of this application provide a multimodal adaptive industrial control system based on digital twins, which further includes: The interaction module is used to receive adjustment instructions and convert them into executable digital signals for transmission to the execution and feedback module.

[0084] In some embodiments, operators can monitor equipment operating trends (such as energy consumption curve fluctuations), production process progress (such as work order completion rate), and system health status (such as equipment remaining life prediction) in real time through a visualization unit, while also receiving abnormal alarms issued by the system (such as pressure over-limit warnings).

[0085] At the same time, it also supports operators to manually input adjustment instructions (such as temporarily modifying process parameters), and convert these instructions into executable digital signals and transmit them to the control decision layer, so as to realize flexible switching between manual intervention and automated control.

[0086] Through the coordinated operation of the interaction module and the execution and feedback module, not only is the closed-loop integrity of the system control process ensured, but also intuitive monitoring and management tools are provided to operators, enhancing the transparency and controllability of the industrial production process.

[0087] The modules in the aforementioned digital twin-based multimodal adaptive industrial control system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0088] Corresponding to the aforementioned embodiments of the multimodal adaptive industrial control system based on digital twins, and employing the same technical concept, this application also provides embodiments of the multimodal adaptive industrial control method based on digital twins.

[0089] Figure 2 This is a flowchart illustrating a multimodal adaptive industrial control method based on digital twins, provided as an embodiment of the present invention.

[0090] In one exemplary embodiment, such as Figure 2 As shown, this multimodal adaptive industrial control method based on digital twins may include: Step S100: Obtain multimodal raw data, which includes equipment operating status, production process parameters, environmental changes, physical laws, historical operating data, and industry expert experience.

[0091] Step S200: Preprocess all acquired data to obtain various feature information.

[0092] Step S300: Construct a composite digital twin based on physical laws, historical operating data, and feature information from industry experts' experience. The composite digital twin includes a physical model, a behavioral model, and a rule model.

[0093] Step S400: Based on the characteristic information and status information of equipment operating status, production process parameters and environmental changes, the composite digital twin is dynamically calibrated and optimized.

[0094] Step S500: Based on the composite digital twin, an initial control strategy is generated through a multi-agent collaborative unit, and then the initial control strategy is verified and corrected through a fuzzy logic unit to output the final control strategy.

[0095] Step S600: Execute control operations according to the final control strategy and transmit the status information to the digital twin model module. The status information includes actual operating parameters, environmental status, and abnormal signals.

[0096] For specific limitations on the digital twin-based multimodal adaptive industrial control method, please refer to the limitations on the digital twin-based multimodal adaptive industrial control system mentioned above, which will not be repeated here.

[0097] As can be seen, the multimodal adaptive industrial control system and method based on digital twins provided in some embodiments of this application, on the basis of closed-loop control, introduces an innovative adaptive control algorithm that integrates multi-agent reinforcement learning and fuzzy logic. This algorithm senses parameter fluctuations and disturbances in real time, dynamically adjusts the final control strategy, and improves system stability and control accuracy. It no longer relies on the precise mathematical model requirement of PID control. Through multi-agent reinforcement learning driven by a deep Q-network, it autonomously learns the optimal control strategy, exhibiting stronger adaptability to changes in system parameters. The fuzzy logic module introduces expert experience and rule constraints, dynamically adjusting the reinforcement learning strategy, further enhancing the algorithm's ability to handle uncertain information and improving the system's robustness and anti-interference capability under uncertain environments. This solves the problem of poor adaptability of traditional control algorithms under complex working conditions. It also reduces the dependence of MPC on the model: utilizing the dynamic update mechanism of the digital twin model layer, combined with multi-agent reinforcement learning and fuzzy logic, the digital twin model is optimized and corrected in real time.

[0098] The system also includes a data acquisition and preprocessing module, which performs deep cleaning and feature extraction on multi-source heterogeneous data. At the control decision level, it integrates multi-agent reinforcement learning and fuzzy logic to achieve efficient fusion and processing of multi-modal data and fully explore the potential value of the data.

[0099] Furthermore, the digital twin model module provided in this application innovatively integrates physical models, behavioral models, and rule models to construct a full-dimensional virtual mirror. The digital twin model can not only visualize physical entities, but also deeply intervene and optimize the physical system through innovative adaptive control algorithms. It can also dynamically calibrate model parameters based on real-time feedback data to maintain a high degree of consistency between the virtual model and the physical entity, providing the system with accurate prediction and optimization basis. This breaks through the limitations of traditional digital twin models that are static and have low accuracy, and breaks the limitation of one-way mapping in traditional digital twin systems, realizing two-way interaction between the physical system and the virtual model.

[0100] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0101] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0102] In the several embodiments provided by this invention, it should be understood that the disclosed methods can be implemented in other ways. The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments. The features disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0103] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-modal adaptive industrial control system based on digital twinning, characterized in that, The application relates to an industrial control system, comprising: a data acquisition and preprocessing module for acquiring multi-modal original data of an industrial site in real time and preprocessing all acquired data to obtain various feature information, wherein the multi-modal original data comprises equipment running state, production process parameters, environmental change, physical law, historical running data and industry expert experience; a digital twin model module for constructing a composite digital twin based on the physical law, the historical running data and the industry expert experience, wherein the composite digital twin comprises a physical model, a behavior model and a rule model; a control decision module for generating a final control strategy according to the composite digital twin through a multi-agent collaborative unit and a fuzzy logic unit, wherein the multi-agent collaborative unit comprises at least two agents trained through reinforcement learning; an execution and feedback module for executing a control operation according to the final control strategy and transmitting state information to the digital twin model module, wherein the state information comprises actual running parameters, environmental state and abnormal signals.

2. The digital-twin-based multi-modal adaptive industrial control system of claim 1, wherein, The data acquisition and preprocessing module further comprises: a multi-source data acquisition unit for acquiring multi-modal original data through multi-dimensional deployment of acquisition points to comprehensively cover the industrial scene; a preprocessing unit for cleaning, denoising, format conversion and feature extraction of the multi-modal original data to obtain various feature information.

3. The digital-twin-based multi-modal adaptive industrial control system of claim 1, wherein, The digital twin model module is further used for fusing the various feature information and state information and dynamically calibrating and optimizing the composite digital twin based on the fusion result.

4. The digital-twin-based multi-modal adaptive industrial control system of claim 1, wherein, The final control strategy is generated through the multi-agent collaborative unit and the fuzzy logic unit, and the generation further comprises: generating an initial control strategy of each industrial control task subsystem through the multi-agent collaborative unit; evaluating the subsystem running state through the fuzzy logic unit and correcting the initial control strategy based on the evaluation result to output the final control strategy.

5. The digital-twin-based multi-modal adaptive industrial control system of claim 1, wherein, The application further comprises: an interaction module for receiving an adjustment instruction of an operator and converting the adjustment instruction into an executable digital signal to be transmitted to the execution and feedback module.

6. The digital-twin-based multi-modal adaptive industrial control system of claim 1, wherein, The execution and feedback module further comprises: a visualization unit for monitoring equipment running trend, production process progress and system health state in real time and judging whether to issue an abnormal alarm according to the abnormal signal in the state information.

7. The digital-twin-based multi-modal adaptive industrial control system of claim 4, wherein, Each agent learns through a deep Q network and cooperatively optimizes control decision through shared information, and each agent is responsible for a specific control task, wherein the control task comprises equipment running, logistics scheduling and quality monitoring.

8. The digital-twin-based multi-modal adaptive industrial control system of claim 4, wherein, The fuzzy logic unit comprises: a fuzzification interface for mapping accurate input to a corresponding fuzzy set and calculating membership value through a membership function; a fuzzy rule base for predefining fuzzy rules based on expert experience to indicate the corresponding relationship between input and output; a fuzzy inference engine for activating corresponding rules according to the fuzzy set corresponding to the input and obtaining the membership of the fuzzy set corresponding to the output through fuzzy logic operation; a defuzzification interface for converting fuzzy output into an accurate value.

9. The digital-twin-based multi-modal adaptive industrial control system according to any of claims 1 or 3, characterized in that, The composite digital twin body based on the characteristic information of physical law, historical operation data and industry expert experience further comprises: Modeling mechanical movement, energy transmission and thermodynamic change characteristics of the physical entity based on structural design drawings of the equipment, material attribute parameters and kinetic principles, to obtain a physical model; Constructing a behavior model based on a data-driven machine learning algorithm according to historical operation data and the physical model, wherein the behavior model comprises an operating mode in normal operation and abnormal signs before a fault occurs; Constructing a rule model according to process standards, safety specifications and industry regulations, wherein the rule model comprises decision constraint conditions.

10. A multi-modal adaptive industrial control method based on digital twinning, characterized in that, The method comprises: Obtaining multi-modal raw data, wherein the multi-modal raw data comprises equipment operating states, production process parameters, environmental change conditions, physical laws, historical operation data and industry expert experience; Preprocessing all obtained data to obtain various characteristic information; Constructing a composite digital twin body based on the characteristic information of physical law, historical operation data and industry expert experience, wherein the composite digital twin body comprises a physical model, a behavior model and a rule model; Integrating and analyzing characteristic information and state information of equipment operating states, production process parameters and environmental change conditions to dynamically calibrate and optimize the composite digital twin body; Generating an initial control strategy through a multi-agent collaborative unit based on the composite digital twin body, and performing compliance verification and correction on the initial control strategy through a fuzzy logic unit to output a final control strategy; Performing control operations according to the final control strategy, and transmitting state information to a digital twin model module, wherein the state information comprises actual operating parameters, environmental states and abnormal signals.

Images