Comprehensive management and control system for digital factory

Abstract

The invention discloses a comprehensive management and control system for a digital factory, and relates to the field of management and control systems. A production index unit; the control instruction packet unit is used for defining decision variables according to the predicted production indexes, generating a candidate strategy set by using a particle swarm optimization algorithm, dynamically adjusting the candidate strategy set in combination with a causal reinforcement learning algorithm, and generating a structured control instruction packet based on a dynamic adjustment result; an instruction issuing unit; and a monitoring and optimizing unit. According to the method, the knowledge graph and the multi-source heterogeneous data are fused, so that the whole-process management and control from data acquisition to semantic understanding are realized, and the cognitive intelligence of a factory is improved; in combination with a multi-scale digital twinning and meta-learning prediction framework, a production process is dynamically simulated, a high-precision prediction index is generated, and reliable support is provided for generation and optimization of a control strategy.

Description

Technical Field

[0001] This invention relates to the field of control systems, and more specifically, to a comprehensive control system for digital factories. Background Technology

[0002] With the rapid development of technologies such as intelligent manufacturing, industrial internet, and artificial intelligence, digital factories are gradually becoming an important direction for the transformation and upgrading of modern manufacturing. By deeply integrating advanced technologies such as cyber-physical systems (CPS), industrial big data, cloud computing, and edge computing, digital factories can achieve comprehensive perception, collaborative optimization, and intelligent decision-making of key production factors such as "people, machines, materials, methods, and environment," thereby significantly improving production efficiency, resource utilization, and product quality stability.

[0003] However, most current digital factory systems still primarily focus on the digital upgrade of specific processes, such as Manufacturing Execution Systems (MES) at the shop floor level, Supervisory Control and Data Acquisition (SCADA) systems at the equipment level, and quality prediction models based on machine learning. While these systems possess a degree of intelligence within their respective domains, they still face numerous challenges overall, including inconsistent data standards, isolated models, closed knowledge bases, and a lack of cross-system collaboration in control strategies. This leads to delayed production scheduling responses and difficulties in achieving global convergence of optimization strategies, thus hindering further improvements in factory operational efficiency and intelligent decision-making. Furthermore, faced with multi-source, heterogeneous, and multi-level industrial data, traditional systems lack unified semantic modeling and causal reasoning capabilities, making it difficult to construct dynamic knowledge systems adapted to complex production scenarios. This limits the interpretability and generalization ability of data-driven optimization strategies.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] In view of the problems in related technologies, this invention proposes a comprehensive management and control system for digital factories to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows: A comprehensive control system for digital factories, comprising: The knowledge graph unit is used to collect multi-source heterogeneous data in the factory, establish a spatiotemporal benchmark data lake based on the multi-source heterogeneous data, and construct the factory knowledge graph by using entity relationship extraction and cognitive computing. The production index unit is used to generate working condition embedding vectors based on the factory knowledge graph. Using the working condition embedding vectors as input, a prediction model is constructed by combining search strategies and knowledge distillation techniques. The prediction model is then used to simulate the evolution of production status and output the predicted production index. The control instruction package unit is used to define decision variables based on predicted production indicators, generate a candidate strategy set using the particle swarm optimization algorithm, dynamically adjust the candidate strategy set using the causal reinforcement learning algorithm, and generate a structured control instruction package based on the dynamic adjustment results. The instruction issuing unit is used to load control instruction packages using a 3D visualization sandbox, perform multi-dimensional collision detection by combining physical rules and virtual scene library, and issue control instructions that pass the detection to the control system for execution. The monitoring and optimization unit is used to monitor the execution effect of control commands in real time and process the execution effect using online controlled experiment technology to obtain experimental data with causal significance in order to optimize the factory knowledge graph.

[0007] Furthermore, the production indicator units include: The digital twin module is used to construct multi-scale digital twin models based on the semantic structure and hierarchical relationships of the factory knowledge graph; The working condition embedding vector module is used to utilize a graph-based self-supervised learning framework to input the collected multi-source heterogeneous data into a preset multimodal graph encoder. Using the factory knowledge graph as a prior constraint, it learns the current production state in the semantic embedding space to generate semantic working condition embedding vectors. The prediction model module is used to construct a search framework for the prediction model based on the generated working condition embedding vectors, and to process the search framework by combining reinforcement learning algorithms and knowledge distillation mechanisms, and to construct the prediction model based on the processing results. The predictive production index module is used to integrate the constructed predictive model into the digital twin model to simulate the evolution of production status, and output the predicted production index based on the evolution simulation results.

[0008] Furthermore, the prediction model module includes: The search space submodule is used to identify multidimensional performance constraints in the current semantic scene based on the generated working condition embedding vector, and to use the multidimensional performance constraints as input to construct the search framework of the prediction model in order to define the search space. The candidate prediction model submodule is used to use reinforcement learning algorithms as a search strategy to progressively build and select model configurations for prediction models in a defined search space, and evaluate the performance of each model configuration in satisfying multidimensional performance constraints through a preset reward function, and select several candidate prediction models based on the evaluation results. The preliminary prediction model submodule is used to retrieve several seed models that best match the current semantic scene from the pre-built prediction model library based on the generated working condition embedding vector and the selected candidate prediction models. Then, using the seed models as the search starting point, the model configuration of the seed models is migrated to the search engine to generate the preliminary prediction model. The prediction model submodule is used to construct a prediction model that meets the current semantic scenario and multidimensional performance constraints by taking the output distribution of the preliminary prediction model and the intermediate layer features as targets and incorporating the distillation loss into the reward mechanism of reinforcement learning, based on the knowledge distillation technique.

[0009] Furthermore, the control instruction packet unit includes: The decision variable definition module is used to construct a multi-objective function vector based on the predicted production indicators, and define decision variables in combination with the physical boundary conditions of the production equipment operation; The candidate strategy set generation module is used to initialize the defined decision variables with the particle swarm optimization algorithm, generate an initial particle swarm, and establish a candidate strategy set through the fast non-dominated sorting algorithm and crowding calculation. The candidate strategy adjustment module is used to construct a multi-dimensional data matrix based on the collected historical multi-source heterogeneous data, use the causal discovery algorithm to analyze the causal weights of the multi-dimensional data matrix, and integrate the causal weights into the strategy gradient optimization process to dynamically adjust the candidate strategy. The control instruction package module is used to sort the dynamically adjusted candidate strategies based on the Pareto solution comprehensive score, map the sorting results to a preset control instruction template, verify the mapping results in conjunction with the syntax tree, and output a structured control instruction package.

[0010] Furthermore, the candidate strategy set generation module includes: The particle swarm initialization submodule is used to initialize the population using the particle swarm optimization algorithm based on the defined decision variables, and to encode the initialization results using a quantum mechanism to generate the initial particle swarm. The candidate strategy set construction submodule is used to determine the multi-objective dominance relationship of each particle in the generated initial particle swarm using a fast non-dominated sorting algorithm. Based on the determination results, the particles are divided into different non-dominated front levels, and the crowding degree of each particle is calculated to construct the candidate strategy set.

[0011] Furthermore, based on the generated initial particle swarm, a fast non-dominated sorting algorithm is used to determine the multi-objective dominance relationship of each particle in the initial particle swarm. Based on the determination results, the particles are divided into different non-dominated front levels, and the crowding degree of each particle is calculated to construct a candidate strategy set, including: Perform a quantum state observation operation on each particle in the initial particle swarm to obtain the quantum value of the particle, and map the quantum value to the corresponding solution; Each solution is evaluated using the constructed multi-objective function. Based on the evaluation results, the dominance relationship between solutions is analyzed using the fast non-dominated sorting algorithm. The particles are divided into different non-dominated frontier levels, and a target solution set containing multiple sub-objective solution sets is constructed. A pre-defined Gaussian mixture model is trained using the target solution set, and the mean and covariance matrix of each sub-target solution set are calculated using the expectation-maximization algorithm. Based on the calculation results, each solution is assigned to the sub-target solution set with the highest probability. Based on the allocation results of the sub-objective solution sets, several pairs of sub-objective solution sets are randomly selected for pairing, and exchange operations are performed between different sub-objective solution sets. Based on the results of the exchange operations, a hybrid solution set is generated. Calculate the crowding degree of each solution in the mixed solution set in its respective non-dominated layer, and sort the solutions by strategy of ascending non-dominated layer and descending crowding degree, and select the top few solutions as the candidate strategy set.

[0012] Furthermore, the formula for calculating congestion is: ; In the formula, CD ( i ) indicates the first i The crowding value of each solution; M Indicates the number of objective functions; m Indicates the index of the objective function; f m ( i+ 1) indicates that within the same non-dominated layer, the solution i The next solution i+1 is in the first position. m The values ​​of each objective function; f m ( i -1) indicates that within the same non-dominated layer, the solution... i The previous solution i-1 is in the first position m The values ​​of each objective function; This indicates that all solutions in the current non-dominated layer are at the th . m The maximum value of each objective function; This indicates that all solutions in the current non-dominated layer are at the th . m The minimum value on an objective function.

[0013] Furthermore, the candidate strategy adjustment module includes: The multidimensional data matrix submodule is used to extract historical decision variables from the collected historical multi-source heterogeneous data, and to preprocess and fuse the historical decision variables to construct a multidimensional data matrix. The causal graph structure submodule is used to analyze the constructed multidimensional data matrix using causal discovery algorithms, identify the causal structure between historical decision variables, extract the corresponding causal weights from the causal structure, and generate a causal graph structure. The candidate policy submodule is used to embed the extracted causal weights into a preset policy gradient optimization framework based on the generated causal graph structure and candidate policy set, to guide the optimization direction of the candidate policy on the causal path, and to dynamically adjust the candidate policy in combination with the entropy regularization algorithm.

[0014] Furthermore, the monitoring and optimization unit includes: The real-time data stream module is used to monitor the execution effect of control commands in the production process in real time. It evaluates the effect of control commands by collecting feedback data and converts the evaluation results into a structured real-time data stream. The experimental data acquisition module is used to process real-time data streams using online controlled experimental technology, design control experiments in conjunction with preset disturbance control commands, and compare and analyze the effects of control commands under different operating conditions to obtain experimental data with causal significance. The causal effect calculation module is used to construct a causal graph model between decision variables based on the acquired experimental data using causal analysis algorithms, identify the causal paths between control variables and control responses, and calculate the causal effects of each causal path. The knowledge graph optimization module is used to embed the calculated causal effects into the factory knowledge graph and update the node attributes and edge relationship weights in the factory knowledge graph to optimize the factory knowledge graph.

[0015] Furthermore, the formula for calculating causal effects is: ; In the formula, CE ( T→Y ) indicates from the control variable T To response variable Y The causal effect value; T Indicates control variables; Y Represents the response variable; N Indicates the number of experimental samples; i Indicates the sample index; E Expressing expectations; do ( T = t i ) indicates the control variable T Perform intervention procedures; X Indicates a mixed variable; x i Indicates the first i Confounding variables in each sample X The possible values ​​of ; do ( T = t 0 ) indicates the control variable T Perform baseline intervention procedures; t i Indicates the first iThe level of intervention applied in each sample; t 0 Indicates the baseline intervention level; w path This represents the causal path weight.

[0016] The beneficial effects of this invention are as follows: 1. This invention integrates knowledge graphs and multi-source heterogeneous data to achieve end-to-end control from data acquisition to semantic understanding, thereby enhancing factory cognitive intelligence. By combining multi-scale digital twins and meta-learning prediction frameworks, it dynamically simulates the production process and generates high-precision prediction indicators, providing reliable support for the generation and optimization of control strategies.

[0017] 2. This invention improves the global optimization capability of control strategies in complex production scenarios by integrating a multi-objective optimization engine and a causal reinforcement learning algorithm, combined with a quantum particle swarm optimization strategy; relying on three-dimensional sand table visualization pre-simulation and multi-dimensional collision detection technology, it realizes the simulation verification of control commands, thereby enhancing the safety, controllability and decision transparency of the actual control process.

[0018] 3. This invention achieves collaborative optimization of data, models, and control by constructing a closed-loop system that integrates real-time monitoring, online experiments, and causal analysis; it continuously improves strategy generation capabilities and model performance by dynamically updating the knowledge graph and decision engine through feedback information, possessing adaptive and self-optimizing characteristics, and ensuring the stable operation and continuous evolution of the integrated management and control system under complex working conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in 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.

[0020] Figure 1 This is a schematic diagram of a comprehensive control system for a digital factory according to an embodiment of the present invention; Figure 2 This is a flowchart of a predictive model module in an integrated management and control system for a digital factory according to an embodiment of the present invention.

[0021] In the picture: 1. Knowledge Graph Unit; 2. Production Indicator Unit; 3. Control Instruction Package Unit; 4. Instruction Issuance Unit; 5. Monitoring and Optimization Unit. Detailed Implementation

[0022] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0023] According to an embodiment of the present invention, a comprehensive management and control system for digital factories is provided.

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a comprehensive management and control system for a digital factory includes: Knowledge Graph Unit 1 is used to collect multi-source heterogeneous data from the factory, establish a spatiotemporal benchmark data lake based on the multi-source heterogeneous data, and construct a factory knowledge graph by using entity relationship extraction and cognitive computing.

[0025] It should be further explained that the process of collecting multi-source heterogeneous data from the factory, establishing a spatiotemporal benchmark data lake based on this data, and constructing a factory knowledge graph using entity relationship extraction and cognitive computing specifically includes: Collect multi-source heterogeneous data from the factory, including equipment sensor data, production logs, quality monitoring data, and environmental data. This data originates from various aspects of the factory, such as production lines, warehousing, and equipment operation, and has different structures and formats. The collected data is cleaned, denoised, and standardized to have a unified format, facilitating further analysis and modeling. Entity relation extraction technology is used to extract important entities (such as equipment, production processes, materials, and personnel) and their interrelationships (such as the relationship between equipment and processes, and the dependencies between various stages in the production process) from the processed data. Through cognitive computing, semantic labels are assigned to these entities and relationships to construct a knowledge graph of the factory. This allows various types of information in the factory operation to be intelligently represented in the graph, facilitating subsequent data mining and intelligent analysis.

[0026] Production Indicator Unit 2 is used to generate working condition embedding vectors based on the factory knowledge graph. Using the working condition embedding vectors as input, a prediction model is constructed by combining search strategies and knowledge distillation techniques. The prediction model is then used to simulate the evolution of production status and output the predicted production indicators.

[0027] In this optional embodiment, production indicator unit 2 includes: The digital twin module is used to construct multi-scale digital twin models based on the semantic structure and hierarchical relationships of the factory knowledge graph.

[0028] It should be further explained that, based on the semantic structure and hierarchical relationships of the factory knowledge graph, the construction of a multi-scale digital twin model specifically includes: Key entities and their hierarchical structures are extracted from the factory knowledge graph, including equipment, processes, materials, and personnel. Semantic relationships between these layers are constructed according to the factory's production process and management architecture. For example, a clear hierarchical structure is established from the equipment layer (e.g., sensors, robotic arms, production line equipment) to the process layer (e.g., cutting, assembly, packaging processes), then to the production line layer (e.g., various production sections, production line resource allocation), and finally to the overall factory layer (e.g., production scheduling, supply chain management). Based on the extracted semantic structure and hierarchical relationships, a corresponding digital twin model is constructed for each layer. For example, the digital twin for the equipment layer would include the real-time status of the equipment, The digital twins at the process level include operational parameters and operational health status. The process-level digital twin includes the process flow, material flow, and efficiency and quality parameters of each production link. The production line-level digital twin includes information such as production line scheduling, resource allocation, and production bottlenecks. The overall factory-level digital twin considers production targets, resource management, and production forecasting. By integrating the digital twins at each level, data and information can be shared and flow between different levels. Finally, by utilizing feedback from the digital twins and real-time updates to the knowledge graph, combined with sensor data and simulation results, the parameters of the digital twin model are dynamically adjusted to cope with changes in the production environment.

[0029] The working condition embedding vector module is used to utilize a graph-based self-supervised learning framework to input the collected multi-source heterogeneous data into a preset multimodal graph encoder. Using the factory knowledge graph as a prior constraint, it learns the current production state in the semantic embedding space to generate semantic working condition embedding vectors.

[0030] The prediction model module is used to construct a search framework for the prediction model based on the generated working condition embedding vectors, and to process the search framework by combining reinforcement learning algorithms and knowledge distillation mechanisms, and to construct the prediction model based on the processing results.

[0031] In this optional embodiment, the prediction model module includes: The search space submodule is used to identify multidimensional performance constraints in the current semantic scene based on the generated working condition embedding vector, and to use the multidimensional performance constraints as input to construct the search framework of the prediction model in order to define the search space. The candidate prediction model submodule is used to use reinforcement learning algorithms as a search strategy to progressively build and select model configurations for prediction models in a defined search space, and evaluate the performance of each model configuration in satisfying multidimensional performance constraints through a preset reward function, and select several candidate prediction models based on the evaluation results. The preliminary prediction model submodule is used to retrieve several seed models that best match the current semantic scene from the pre-built prediction model library based on the generated working condition embedding vector and the selected candidate prediction models. Then, using the seed models as the search starting point, the model configuration of the seed models is migrated to the search engine to generate the preliminary prediction model. The prediction model submodule is used to construct a prediction model that meets the current semantic scenario and multidimensional performance constraints by taking the output distribution of the preliminary prediction model and the intermediate layer features as targets and incorporating the distillation loss into the reward mechanism of reinforcement learning, based on the knowledge distillation technique.

[0032] It should be added that, such as Figure 2 As shown, the search space submodule accurately identifies multi-dimensional performance constraints in the current semantic scene using the generated work condition embedding vectors. These constraints are then used as input to construct a prediction model search framework with ≥12 dimensions (including 3-8 network layers, 4 activation functions, batch sizes of 32 / 64 / 128, etc.). This framework clearly defines the search boundary and feasible region, providing precise guidance for subsequent model searches. The candidate prediction model submodule uses reinforcement learning algorithms (such as Deep Q-Network, DQN) for searching, progressively optimizing the network structure and hyperparameters of the prediction model within the defined search space. The model is then evaluated using a preset reward function (comprehensively considering accuracy, efficiency, and resource consumption, with weights of 0.5, 0.3, and 0.2, respectively). The performance of each configuration is evaluated, and 10-15 candidate models with the best performance are selected to quickly identify potential high-quality models suitable for the current scenario. The preliminary prediction model submodule retrieves the Top-3 seed models that best match the current semantic scenario from a prediction model library containing 500+ industrial scenario models, based on the feature similarity between the working condition embedding vector and the candidate models. Starting with these seed models, their configurations (such as network structure and initial hyperparameters) are transferred to the search engine to generate a preliminary prediction model, thereby leveraging the experience of existing high-quality models to reduce search difficulty. The prediction model submodule uses knowledge distillation technology, taking the output distribution and intermediate layer features of the preliminary model as distillation targets, and integrating the distillation loss (with a weight of 0.3) into the reinforcement learning reward mechanism. Through 100-200 rounds of iterative training to optimize model parameters, a prediction model that meets the current semantic scenario and multi-dimensional performance constraints is finally constructed, balancing model accuracy, inference efficiency, and resource consumption.

[0033] The predictive production index module is used to integrate the constructed predictive model into the digital twin model to simulate the evolution of production status, and output the predicted production index based on the evolution simulation results.

[0034] It should be further explained that the constructed prediction model is integrated into the digital twin model for production state evolution simulation. The predicted production indicators output based on the evolution simulation results specifically include: The generated predictive models (such as production status prediction, equipment failure prediction, and energy consumption prediction) are embedded into a multi-scale digital twin. Using these integrated predictive models, the digital twin can simulate the evolution of production status based on real-time input data (including equipment operating status, production plans, and environmental conditions). The simulation process considers various disturbances, equipment failures, material flow, and scheduling adjustments during production. Through multi-dimensional data analysis, the digital twin generates predictions of key production indicators based on the simulation results, such as production efficiency, resource consumption, equipment health, and product quality, and uses these indicators as decision support. Ultimately, the simulation data is transformed into specific production indicators, providing a basis for optimizing, scheduling, and allocating resources in the production process.

[0035] Control instruction package unit 3 is used to define decision variables based on predicted production indicators, generate a candidate strategy set using particle swarm optimization algorithm, dynamically adjust the candidate strategy set using causal reinforcement learning algorithm, and generate a structured control instruction package based on the dynamic adjustment results.

[0036] In this optional embodiment, the control instruction packet unit 3 includes: The decision variable definition module is used to construct a multi-objective function vector based on the predicted production indicators, and define decision variables in combination with the physical boundary conditions of the production equipment operation.

[0037] It should be further explained that, based on the predicted production indicators, a multi-objective function vector is constructed, and, combined with the physical boundary conditions of the production equipment operation, the decision variables are defined as follows: Based on predicted production indicators (such as output, production efficiency, energy consumption, and quality control), a multi-objective function vector is constructed. For example, increasing output may increase energy consumption, while improving product quality may decrease production speed. By mathematically modeling these objective functions, the optimization objectives and constraints of each objective can be quantified, forming a multi-objective optimization problem. For each objective function, constraints on decision variables are defined to ensure that the production equipment meets physical boundary conditions during operation, such as equipment load limits, temperature limits, production speed ranges, and energy consumption limits. By modeling each objective function and considering these physical constraints, the actual operational range in the production process can be integrated into the optimization process, ensuring that all optimization operations are carried out within the practically feasible physical boundaries, avoiding unrealizable decision outcomes.

[0038] The candidate strategy set generation module is used to initialize the population of the defined decision variables using the particle swarm optimization algorithm, generate an initial particle swarm, and establish a candidate strategy set through the fast non-dominated sorting algorithm and crowding calculation.

[0039] In this optional embodiment, the candidate strategy set generation module includes: The particle swarm initialization submodule is used to initialize the population using the particle swarm optimization algorithm based on the defined decision variables, and to encode the initialization results using a quantum mechanism to generate the initial particle swarm.

[0040] The candidate strategy set construction submodule is used to determine the multi-objective dominance relationship of each particle in the generated initial particle swarm using a fast non-dominated sorting algorithm. Based on the determination results, the particles are divided into different non-dominated front levels, and the crowding degree of each particle is calculated to construct the candidate strategy set.

[0041] In this optional embodiment, based on the generated initial particle swarm, a fast non-dominated sorting algorithm is used to determine the multi-objective dominance relationship of each particle in the initial particle swarm. Based on the determination results, the particles are divided into different non-dominated front levels, and the crowding degree of each particle is calculated to construct a candidate strategy set, including: Perform a quantum state observation operation on each particle in the initial particle swarm to obtain the quantum value of the particle, and map the quantum value to the corresponding solution; Each solution is evaluated using the constructed multi-objective function. Based on the evaluation results, the dominance relationship between solutions is analyzed using the fast non-dominated sorting algorithm. The particles are divided into different non-dominated frontier levels, and a target solution set containing multiple sub-objective solution sets is constructed. A pre-defined Gaussian mixture model is trained using the target solution set, and the mean and covariance matrix of each sub-target solution set are calculated using the expectation-maximization algorithm. Based on the calculation results, each solution is assigned to the sub-target solution set with the highest probability. Based on the allocation results of the sub-objective solution sets, several pairs of sub-objective solution sets are randomly selected for pairing, and exchange operations are performed between different sub-objective solution sets. Based on the results of the exchange operations, a hybrid solution set is generated. Calculate the crowding degree of each solution in the mixed solution set in its respective non-dominated layer, and sort the solutions by strategy of ascending non-dominated layer and descending crowding degree, and select the top few solutions as the candidate strategy set.

[0042] In this optional embodiment, the formula for calculating congestion is: ; In the formula, CD ( i ) indicates the first i The crowding value of each solution; M Indicates the number of objective functions; m Indicates the index of the objective function; f m ( i+ 1) indicates that within the same non-dominated layer, the solution i The next solution i+1 is in the first position. m The values ​​of each objective function;f m ( i -1) indicates that within the same non-dominated layer, the solution... i The previous solution i-1 is in the first position m The values ​​of each objective function; This indicates that all solutions in the current non-dominated layer are at the th . m The maximum value of each objective function; This indicates that all solutions in the current non-dominated layer are at the th . m The minimum value on an objective function.

[0043] It should be further explained that, in the specific embodiment, the objective function includes production cost and production efficiency. The position and velocity of the particles are generated by the initial particle swarm, and a quantum state observation operation is performed on each particle to obtain its quantum value and map it to a feasible solution. Each solution is evaluated according to the production cost and efficiency objectives.

[0044] Fast non-dominated sorting divides the solutions into different non-dominated frontier levels. For example, the first level contains the solutions that perform best across all objectives. Crowding is calculated for each solution to assess the diversity of particles in the objective space. Using a Gaussian mixture model and the expectation-maximization algorithm, the mean and covariance of each sub-objective solution set are calculated, and the solutions are assigned to the most suitable sub-objective set. The mixed solution set is swapped by pairing parent solutions to generate a new candidate solution set. The solutions in the mixed solution set are then sorted, and the top solutions are selected based on the non-dominated level and crowding. Finally, the system generates an optimized candidate strategy set that optimally balances the objectives of production cost and production efficiency, which can be further used in multi-objective optimization engines or other decision-making systems.

[0045] The candidate strategy adjustment module is used to construct a multi-dimensional data matrix based on the collected historical multi-source heterogeneous data, use the causal discovery algorithm to analyze the causal weights of the multi-dimensional data matrix, and integrate the causal weights into the strategy gradient optimization process to dynamically adjust the candidate strategy.

[0046] In this optional embodiment, the candidate strategy adjustment module includes: The multidimensional data matrix submodule is used to extract historical decision variables from the collected historical multi-source heterogeneous data, and to preprocess and fuse the historical decision variables to construct a multidimensional data matrix. The causal graph structure submodule is used to analyze the constructed multidimensional data matrix using causal discovery algorithms, identify the causal structure between historical decision variables, extract the corresponding causal weights from the causal structure, and generate a causal graph structure. The candidate policy submodule is used to embed the extracted causal weights into a preset policy gradient optimization framework based on the generated causal graph structure and candidate policy set, to guide the optimization direction of the candidate policy on the causal path, and to dynamically adjust the candidate policy in combination with the entropy regularization algorithm.

[0047] It should be further explained that, in the specific embodiment, multi-source heterogeneous data such as factory equipment operation data, production line temperature and humidity sensor data, and employee working status were collected. This data was preprocessed and fused to construct a multi-dimensional data matrix, which included decision variables such as production efficiency, equipment health status, and process parameters. Based on this data, a causal discovery algorithm was used to identify the strong causal impact of equipment operation status on production efficiency and the indirect impact of temperature and humidity on equipment health status. By generating a causal graph structure, the causal weights between various decision variables were obtained. These causal weights were embedded in the strategy gradient optimization framework to guide the optimization of the scheduling strategy, ensuring that the optimization direction during production scheduling is focused on the strong causal relationship between equipment and production efficiency. At the same time, an entropy regularization strategy was used to ensure that the scheduling system could avoid premature convergence during the optimization process, continuously exploring different scheduling schemes until the optimal solution was found.

[0048] The control instruction package module is used to sort the dynamically adjusted candidate strategies based on the Pareto solution comprehensive score, map the sorting results to a preset control instruction template, verify the mapping results in conjunction with the syntax tree, and output a structured control instruction package.

[0049] Command issuing unit 4 is used to load control command packages using a 3D visualization sandbox, perform multi-dimensional collision detection by combining physical rules and virtual scene library, and issue control commands to the control system for execution after the detection is passed.

[0050] It should be further explained that the process involves using a 3D visualization sandbox to load control command packages, combining physical rules with a virtual scene library to perform multi-dimensional collision detection, and then issuing control commands to the control system for execution after successful detection. Specifically, this includes: Control command packages are generated and loaded as input into a 3D visualization sandbox. These packages typically consist of a series of control actions, scheduling instructions, and state adjustment information, affecting various entities in the virtual scene (such as equipment, production lines, and robots). The sandbox system presents a realistic production environment model using virtual reality technology, where each device, sensor, and its operating state is virtualized into a corresponding 3D object, forming a complete digital twin. The virtual scene library within the sandbox provides different environmental and physical scene configurations, such as production workshops, equipment states, and robot movement ranges. The virtual scene library allows for the simulation of various production conditions through scene presets. These scenes contain different physical rules, such as kinematic constraints, equipment interactions, environmental factors (such as temperature, humidity, and pressure), and equipment fault models. The system dynamically adjusts the state of the virtual environment based on the loaded control commands.

[0051] Control commands are loaded into the sandbox system, and multi-dimensional collision detection and effect pre-play begins. During this process, the sandbox system performs collision detection between various virtual objects based on physical rules and settings in the virtual scene. For example, it detects whether robots will collide with other equipment, whether materials on the production line will obstruct the operation of robots or conveyor belts, and whether equipment will malfunction due to overload or improper scheduling. In addition, it needs to consider multiple dimensions of physical interaction, such as time, space, mechanics, and thermodynamics. In this way, the system can pre-play the execution effect of control commands and discover potential risks or conflicts. The results of collision detection and effect pre-play generate an execution feedback report, which describes the effect of the control commands running in the virtual scene. These reports will indicate which commands caused unexpected behavior (such as equipment collisions, low operating efficiency, resource waste, etc.) and provide corresponding improvement suggestions. Only when the control command passes all pre-play tests and no major risks are detected will the command be marked as "verified successfully".

[0052] Once the control commands are validated and their feasibility confirmed in the virtual environment, these commands will be sent to the control system to begin execution in the real production environment. The system will then send optimized commands to the actual control devices, such as PLCs (Programmable Logic Controllers), SCADA (Supervisory and Data Acquisition Systems), or DCS (Distributed Control Systems), to ensure that the commands can be effectively executed in the actual production process, thereby improving production efficiency and safety.

[0053] For example, in a production process, suppose two devices on the production line (such as a conveyor belt and a robotic arm) need to work synchronously. The control instruction package requires the robotic arm to move to a specified position and cooperate with the conveyor belt. Through the verification process of the 3D sand table, it is found that there is a potential collision between the predetermined path of the robotic arm and the movement path of the conveyor belt, which leads to the system feedback that the path or the sequence of actions needs to be adjusted. This adjustment is detected and corrected in advance through collision detection, ensuring that the equipment is not damaged or the production efficiency is reduced due to uncoordinated actions during the production process.

[0054] The monitoring and optimization unit 5 is used to monitor the execution effect of control commands in real time and process the execution effect using online controlled experiment technology to obtain experimental data with causal significance in order to optimize the factory knowledge graph.

[0055] In this optional embodiment, the monitoring and optimization unit 5 includes: The real-time data stream module is used to monitor the execution effect of control commands in the production process in real time. It evaluates the effect of control commands by collecting feedback data and converts the evaluation results into a structured real-time data stream.

[0056] The experimental data acquisition module is used to process real-time data streams using online controlled experimental technology, design control experiments in conjunction with preset disturbance control commands, and compare and analyze the impact of control commands under different operating conditions to obtain experimental data with causal significance.

[0057] It should be further explained that the process of using online controlled experiment technology to process real-time data streams, designing control experiments in conjunction with preset disturbance control commands, and comparing and analyzing the impact of control commands under different operating conditions to obtain experimental data with causal significance specifically includes: By combining online controlled experiment technology to process real-time data streams from factories or production systems, and designing appropriate control experiments to evaluate the effectiveness of control commands, various production data, including equipment status, production progress, environmental monitoring, and process parameters, can be collected in real time through an interface with the factory control system. These data are then preprocessed. Based on the actual needs of the production system, appropriate disturbance control commands are designed. These commands typically involve appropriately disturbing the current system (such as adjusting equipment parameters, changing production plans, or introducing simulated faults) to explore the impact of different control strategies on the system. Control experiments involve applying different control commands (such as the original command and the disturbance command) under different operating conditions to compare system performance. By controlling the experimental conditions, the performance of the production system under the same initial state can be compared under different commands, obtaining more comprehensive data.

[0058] Furthermore, in control experiments, by comparing the relationship between control commands and system responses, causal inference algorithms (such as Granger causality tests and causal graph analysis) are used to analyze the causal impact of control commands on system behavior. This helps determine whether the control commands directly lead to performance improvements or declines, or whether other factors are at play. For example, under different disturbance commands, production line efficiency may improve due to certain control actions, but further causal analysis may be needed to confirm whether the command directly caused the result or whether it was influenced by other external factors. During the experiment, by combining real-time data streams with disturbance control commands, the effectiveness of control commands can be evaluated under multiple operating conditions, including different production loads, equipment states, and environmental factors (such as temperature and humidity). Control experiments further ensure the reliability and causality of the data by systematically comparing performance under different control conditions. Experimental data can be displayed through visualization methods (such as charts and dashboards) for real-time monitoring and subsequent analysis. Once causal data with significance is obtained through causal analysis, the system automatically generates an experimental report explaining the causal relationship of the control commands, the strength of their impact, and the system response under specific conditions. This provides a basis for further decision support and helps optimize control strategies.

[0059] For example, by using a data stream monitoring system to learn that there is a potential relationship between equipment temperature and production speed, the control system can adjust the equipment temperature through disturbance control commands and observe its impact on production efficiency during the experiment. Through a control experiment, the changes in production efficiency under high temperature and normal temperature conditions are tested respectively. Then, causal analysis is used to verify whether temperature changes directly affect production efficiency. This causal data can provide guidance for optimizing equipment scheduling strategies.

[0060] The causal effect calculation module is used to construct a causal graph model between decision variables based on the acquired experimental data using causal analysis algorithms, identify the causal paths between control variables and control responses, and calculate the causal effects of each causal path.

[0061] The knowledge graph optimization module is used to embed the calculated causal effects into the factory knowledge graph and update the node attributes and edge relationship weights in the factory knowledge graph to optimize the factory knowledge graph.

[0062] In this optional embodiment, the formula for calculating the causal effect is: ; In the formula, CE ( T→Y ) indicates from the control variable T To response variable Y The causal effect value; T Indicates control variables; Y Represents the response variable;N Indicates the number of experimental samples; i Indicates the sample index; E Expressing expectations; do ( T = t i ) indicates the control variable T Perform intervention procedures; X Indicates a mixed variable; x i Indicates the first i Confounding variables in each sample X The possible values ​​of ; do ( T = t 0 ) indicates the control variable T Perform baseline intervention procedures; t i Indicates the first i The level of intervention applied in each sample; t 0 Indicates the baseline intervention level; w path This represents the causal path weight.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A comprehensive control system for digital factories, characterized in that, The system includes: The knowledge graph unit is used to collect multi-source heterogeneous data in the factory, establish a spatiotemporal benchmark data lake based on the multi-source heterogeneous data, and construct the factory knowledge graph by using entity relationship extraction and cognitive computing. The production index unit is used to generate working condition embedding vectors based on the factory knowledge graph. Using the working condition embedding vectors as input, a prediction model is constructed by combining search strategies and knowledge distillation techniques. The prediction model is then used to simulate the evolution of production status and output the predicted production index. The control instruction package unit is used to define decision variables based on predicted production indicators, generate a candidate strategy set using the particle swarm optimization algorithm, dynamically adjust the candidate strategy set using the causal reinforcement learning algorithm, and generate a structured control instruction package based on the dynamic adjustment results. The instruction issuing unit is used to load control instruction packages using a 3D visualization sandbox, perform multi-dimensional collision detection by combining physical rules and virtual scene library, and issue control instructions that pass the detection to the control system for execution. The monitoring and optimization unit is used to monitor the execution effect of control commands in real time and process the execution effect using online controlled experiment technology to obtain experimental data with causal significance in order to optimize the factory knowledge graph. The control command packet unit includes: The candidate strategy set generation module is used to initialize the defined decision variables with the particle swarm optimization algorithm, generate an initial particle swarm, and establish a candidate strategy set through the fast non-dominated sorting algorithm and crowding calculation. The candidate strategy adjustment module is used to construct a multi-dimensional data matrix based on the collected historical multi-source heterogeneous data, use the causal discovery algorithm to analyze the causal weights of the multi-dimensional data matrix, and integrate the causal weights into the strategy gradient optimization process to dynamically adjust the candidate strategy.

2. The integrated control system for a digital factory according to claim 1, characterized in that, The production indicator unit includes: The digital twin module is used to construct multi-scale digital twin models based on the semantic structure and hierarchical relationships of the factory knowledge graph; The working condition embedding vector module is used to utilize a graph-based self-supervised learning framework to input the collected multi-source heterogeneous data into a preset multimodal graph encoder. Using the factory knowledge graph as a prior constraint, it learns the current production state in the semantic embedding space to generate semantic working condition embedding vectors. The prediction model module is used to construct a search framework for the prediction model based on the generated working condition embedding vectors, and to process the search framework by combining reinforcement learning algorithms and knowledge distillation mechanisms, and to construct the prediction model based on the processing results. The predictive production index module is used to integrate the constructed predictive model into the digital twin model to simulate the evolution of production status, and output the predicted production index based on the evolution simulation results.

3. The integrated control system for a digital factory according to claim 2, characterized in that, The prediction model module includes: The search space submodule is used to identify multidimensional performance constraints in the current semantic scene based on the generated working condition embedding vector, and to use the multidimensional performance constraints as input to construct the search framework of the prediction model in order to define the search space. The candidate prediction model submodule is used to use reinforcement learning algorithms as a search strategy to progressively build and select model configurations for prediction models in a defined search space, and evaluate the performance of each model configuration in satisfying multidimensional performance constraints through a preset reward function, and select several candidate prediction models based on the evaluation results. The preliminary prediction model submodule is used to retrieve several seed models that best match the current semantic scene from the pre-built prediction model library based on the generated working condition embedding vector and the selected candidate prediction models. Then, using the seed models as the search starting point, the model configuration of the seed models is migrated to the search engine to generate the preliminary prediction model. The prediction model submodule is used to construct a prediction model that meets the current semantic scenario and multidimensional performance constraints by taking the output distribution of the preliminary prediction model and the intermediate layer features as targets and incorporating the distillation loss into the reward mechanism of reinforcement learning, based on the knowledge distillation technique.

4. The integrated control system for a digital factory according to claim 1, characterized in that, The control command packet unit further includes: The decision variable definition module is used to construct a multi-objective function vector based on the predicted production indicators, and define decision variables in combination with the physical boundary conditions of the production equipment operation; The control instruction package module is used to sort the dynamically adjusted candidate strategies based on the Pareto solution comprehensive score, map the sorting results to a preset control instruction template, verify the mapping results in conjunction with the syntax tree, and output a structured control instruction package.

5. A comprehensive control system for a digital factory according to claim 4, characterized in that, The candidate strategy set generation module includes: The particle swarm initialization submodule is used to initialize the population using the particle swarm optimization algorithm based on the defined decision variables, and to encode the initialization results using a quantum mechanism to generate the initial particle swarm. The candidate strategy set construction submodule is used to determine the multi-objective dominance relationship of each particle in the generated initial particle swarm using a fast non-dominated sorting algorithm. Based on the determination results, the particles are divided into different non-dominated front levels, and the crowding degree of each particle is calculated to construct the candidate strategy set.

6. A comprehensive control system for a digital factory according to claim 5, characterized in that, Based on the generated initial particle swarm, a fast non-dominated sorting algorithm is used to determine the multi-objective dominance relationship of each particle in the initial particle swarm. Based on the determination results, the particles are divided into different non-dominated front levels, and the crowding degree of each particle is calculated to construct a candidate strategy set, including: Perform a quantum state observation operation on each particle in the initial particle swarm to obtain the quantum value of the particle, and map the quantum value to the corresponding solution; Each solution is evaluated using the constructed multi-objective function. Based on the evaluation results, the dominance relationship between solutions is analyzed using the fast non-dominated sorting algorithm. The particles are divided into different non-dominated frontier levels, and a target solution set containing multiple sub-objective solution sets is constructed. A pre-defined Gaussian mixture model is trained using the target solution set, and the mean and covariance matrix of each sub-target solution set are calculated using the expectation-maximization algorithm. Based on the calculation results, each solution is assigned to the sub-target solution set with the highest probability. Based on the allocation results of the sub-objective solution sets, several pairs of sub-objective solution sets are randomly selected for pairing, and exchange operations are performed between different sub-objective solution sets. Based on the results of the exchange operations, a hybrid solution set is generated. Calculate the crowding degree of each solution in the mixed solution set in its respective non-dominated layer, and sort the solutions by strategy of ascending non-dominated layer and descending crowding degree, and select the top few solutions as the candidate strategy set.

7. A comprehensive control system for a digital factory according to claim 6, characterized in that, The formula for calculating the congestion level is: ； In the formula, CD ( i ) indicates the first i The crowding value of each solution; M Indicates the number of objective functions; m Indicates the index of the objective function; f m ( i+ 1) indicates that within the same non-dominated layer, the solution i The next solution i+1 is in the first position. m The values ​​of each objective function; f m ( i -1) indicates that within the same non-dominated layer, the solution... i The previous solution i-1 is in the first position m The values ​​of each objective function; This indicates that all solutions in the current non-dominated layer are at the th . m The maximum value of each objective function; This indicates that all solutions in the current non-dominated layer are at the th . m The minimum value on an objective function.

8. A comprehensive control system for a digital factory according to claim 7, characterized in that, The candidate strategy adjustment module includes: The multidimensional data matrix submodule is used to extract historical decision variables from the collected historical multi-source heterogeneous data, and to preprocess and fuse the historical decision variables to construct a multidimensional data matrix. The causal graph structure submodule is used to analyze the constructed multidimensional data matrix using causal discovery algorithms, identify the causal structure between historical decision variables, extract the corresponding causal weights from the causal structure, and generate a causal graph structure. The candidate policy submodule is used to embed the extracted causal weights into a preset policy gradient optimization framework based on the generated causal graph structure and candidate policy set, to guide the optimization direction of the candidate policy on the causal path, and to dynamically adjust the candidate policy in combination with the entropy regularization algorithm.

9. A comprehensive control system for a digital factory according to claim 1, characterized in that, The monitoring and optimization unit includes: The real-time data stream module is used to monitor the execution effect of control commands in the production process in real time. It evaluates the effect of control commands by collecting feedback data and converts the evaluation results into a structured real-time data stream. The experimental data acquisition module is used to process real-time data streams using online controlled experimental technology, design control experiments in conjunction with preset disturbance control commands, and compare and analyze the effects of control commands under different operating conditions to obtain experimental data with causal significance. The causal effect calculation module is used to construct a causal graph model between decision variables based on the acquired experimental data using causal analysis algorithms, identify the causal paths between control variables and control responses, and calculate the causal effects of each causal path. The knowledge graph optimization module is used to embed the calculated causal effects into the factory knowledge graph and update the node attributes and edge relationship weights in the factory knowledge graph to optimize the factory knowledge graph.

10. A comprehensive control system for a digital factory according to claim 9, characterized in that, The formula for calculating the causal effect is: ； In the formula, CE ( T→Y ) indicates from the control variable T To response variable Y The causal effect value; T Indicates control variables; Y Represents the response variable; N Indicates the number of experimental samples; i Indicates the sample index; E Expressing expectations; do ( T=t i ) indicates the control variable T Perform intervention procedures; X Indicates a mixed variable; x i Indicates the first i Confounding variables in each sample X The possible values ​​of ; do ( T=t 0 ) indicates the control variable T Perform baseline intervention procedures; t i Indicates the first i The level of intervention applied in each sample; t 0 Indicates the baseline intervention level; w path This represents the causal path weight.

Images