Discrete manufacturing production line process decision and optimization method and system based on digital twinning

Abstract

The invention provides a discrete manufacturing production line process decision and optimization method and system based on digital twinning, and relates to the field of production intellectualization and digitalization, and the method mainly comprises the steps: constructing a high-fidelity digital twinning model corresponding to a physical production line through the combination of mechanism modeling and data-driven modeling; establishing a virtual-real consistency evaluation index system, and realizing deviation identification and dynamic consistency maintenance of the digital twinborn model; production disturbance, equipment state change and process deviation information are sensed in real time; a part-process-equipment-quality multi-dimensional correlation model is constructed; an improved stochastic gradient descent algorithm is introduced to carry out dynamic updating and convergence control on intelligent agent strategy network parameters, and a knowledge base self-evolution updating mechanism is constructed; according to the method, dynamic updating and consistency maintaining of the twinborn model are achieved, the dynamic response capability of the digital twinborn model under the dynamic operation condition is remarkably improved, and the systematicness and reusability level of process knowledge are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent and digital production line technology, specifically to a method and system for process decision-making and optimization in discrete manufacturing production lines based on digital twins. Background Technology

[0002] Discrete manufacturing production lines are the core systems of modern equipment manufacturing, and their operational efficiency and production processes directly affect product performance and manufacturing costs. With the rapid development of technologies such as intelligent manufacturing, discrete manufacturing production lines are accelerating their evolution towards digitalization and intelligence. By introducing automated equipment, sensor networks, and information management platforms, production lines have significantly improved in terms of production efficiency, product quality, and resource utilization. However, in actual production processes, existing discrete manufacturing production lines still face the following problems: First, existing discrete manufacturing production line process plans are usually formulated by process engineers based on their own experience, lacking a systematic way of expressing knowledge and reasoning; second, when production conditions, equipment performance, or the external environment change, the process plan is difficult to adjust in a timely manner; finally, the coupling relationships between processes are complex, and there is a lack of data exchange and optimization coordination between different processes, making it difficult to achieve global process optimization, and the overall operation decision of the production line still exhibits the characteristics of "local optimum, global suboptimal".

[0003] To enhance the intelligence level of manufacturing systems, digital twin technology has been introduced into the field of discrete manufacturing. By constructing a high-fidelity model corresponding to the physical production line in virtual space, digital twins can achieve real-time mapping and dynamic simulation of the production process, providing support for process decision-making.

[0004] However, existing digital twin systems still have significant shortcomings: most systems focus on modeling and monitoring at the equipment and process levels, lacking deep integration with process knowledge and optimization algorithms, and have not yet formed a decision-making system based on mechanism-data-driven and knowledge reasoning integration; twin models are more used for state perception and result display, lacking the ability to model the correlation between complex processes and the ability to optimize multiple objectives; at the same time, in complex multi-process, multi-parameter discrete manufacturing scenarios, digital twins lack self-learning and online optimization mechanisms, making it difficult to form a closed-loop system from data perception to decision-making to intelligent optimization.

[0005] The aforementioned problems severely restrict the practical application of digital twin technology in discrete manufacturing production line process decision-making. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method and system for process decision-making and optimization in discrete manufacturing production lines based on digital twins. It deeply integrates the digital twin system with a process knowledge base and intelligent optimization mechanisms, constructing a closed-loop optimization system encompassing "perception-decision-optimization-evolution." By real-time acquisition of process parameters, equipment status information, and operational and collaborative mechanisms, a high-fidelity digital twin synchronized with the physical production line is constructed. Subsequently, the process knowledge base is used to semantically express and reason about process constraints, equipment capabilities, and empirical rules, generating process solutions tailored to production needs. Based on this, an adaptive optimization strategy for multiple devices, processes, and parameters is employed to generate process adjustment plans, and the optimization results are fed back to the physical production line for execution in real time. Quality data and equipment response information generated during production line operation are then fed back into the knowledge base for continuous updating of process rules and driving subsequent optimization strategy iterations, thereby achieving continuous learning, dynamic adjustment, and self-evolutionary upgrading of the process solutions.

[0007] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for process decision-making and optimization of discrete manufacturing production lines based on digital twins, comprising the following steps: S1. Based on the equipment operation characteristics and production process data of discrete manufacturing production lines, a high-fidelity digital twin model corresponding to the physical production line is constructed by combining mechanism modeling and data-driven modeling; a virtual-real consistency evaluation index system is established to realize the identification of deviations and dynamic consistency maintenance of the digital twin model; based on the virtual-real mapping mechanism of the digital twin model, production disturbances, equipment status changes and process deviation information are perceived in real time. S2. For multimodal process data of part features and process routes, key process elements are extracted using named entity recognition and feature template matching methods; semantic relationships between parts, processes, equipment and quality indicators are mined to construct a multidimensional association model of "part-process-equipment-quality"; based on a unified semantic modeling framework, a structured and searchable process knowledge base is constructed to generate an initial process plan. S3. Based on the initial process scheme, establish a process optimization and collaborative decision-making mechanism for multiple equipment, multiple processes, and multiple parameters; based on the disturbance information perceived in real time by digital twins, construct a multi-agent collaborative optimization mechanism, and combine deep reinforcement learning algorithms to realize the global coordination and dynamic optimization of the process scheme; introduce an improved stochastic gradient descent algorithm to dynamically update and converge the parameters of the agent policy network. S4. Construct a knowledge base self-evolution and update mechanism to transmit optimized decision-making schemes to the physical production line for execution; continuously perceive production execution information based on the digital twin model and feed it back to the process knowledge base; by analyzing the correlation between virtual and real mapping errors and process performance deviations, trigger knowledge node reconstruction and attribute updates; combine temporal correlation analysis and incremental learning algorithms to correct knowledge relationship weights, thereby achieving knowledge self-evolution and continuous optimization. In an optional implementation, step S1 involves: based on the equipment operation characteristics and production process data of the discrete manufacturing production line, using mechanistic modeling to characterize the physical laws of the process, and combining data-driven modeling to describe its dynamic evolution characteristics, constructing a high-fidelity digital twin model corresponding to the physical production line; based on the idea of ​​causal analysis, analyzing the influencing factors and changing patterns of the consistency between the virtual and real mapping of the digital twin model, establishing a virtual-real consistency evaluation index system under multi-dimensional attributes, and thereby realizing the deviation identification and consistency maintenance of the digital twin model; based on the virtual-real comparison and dynamic mapping mechanism of the digital twin model, through comparative analysis of the virtual model and the physical production line in terms of time series, state variables, and mechanistic response characteristics, realizing real-time perception and identification of information such as production disturbances, equipment state changes, and process deviations, specifically including: This research collects equipment operation data, process parameters, energy consumption characteristics, and processing quality indicators from discrete manufacturing production lines. It performs time synchronization, feature extraction, and noise filtering on multi-source heterogeneous data, and constructs a digital twin data model using algorithms such as neural networks. Based on mechanistic modeling methods, it describes key equipment and processes from multiple dimensions, including geometry, physics, behavior, and rules, constructing a digital twin mechanism model that reflects the mechanisms and laws of the processing process. Addressing the challenge of accurately depicting nonlinear and time-varying characteristics by mechanistic models, a multi-level model fusion strategy is developed. This strategy achieves complementary unification of the mechanistic model and the data-driven model at the input feature, state variable, and output layers, thereby obtaining a highly accurate digital twin model. A mechanism for maintaining consistency between the digital twin model and the physical production line is established. This involves constructing a consistency evaluation system based on multi-dimensional attributes such as accuracy, production cycle time, and dynamic response, to quantitatively analyze the differences between the digital twin model and the physical production line. Based on the results of the virtual-physical mapping comparison, and combined with causal analysis methods, key influencing factors causing deviations are identified. Then, techniques such as parameter self-calibration, incremental retraining, and structural correction are employed to dynamically update and continuously optimize the twin model, thereby maintaining consistent accuracy and stable response, achieving both consistency maintenance and continuous optimization. Based on the virtual-real mapping mechanism of the digital twin model, the differences between the virtual model and the physical system in terms of time series, state variables and mechanism response characteristics are dynamically compared to realize the real-time perception and identification of information such as operating condition disturbances, equipment performance degradation and process deviations. By combining dynamic threshold detection and time series feature analysis algorithms, the monitoring data is trend identified, and the state prediction model based on neural networks is used to realize the early perception and identification of potential anomalies, thereby providing data support for subsequent process optimization.

[0008] In an optional implementation, step S1 involves extracting and parsing key process elements using named entity recognition and feature template matching methods, targeting multimodal process data such as part features and process routes in the production process; combining statistical analysis and machine learning algorithms to mine the correlation patterns and mapping modes of process knowledge at different levels, analyzing the semantic relationships between parts, processes, equipment, and quality indicators, and constructing a multidimensional correlation model of "parts-process-equipment-quality"; and based on a unified semantic modeling framework, performing semantic modeling and unified management of nodes, attributes, and relationships to form a structured and searchable process knowledge base, generating process solutions oriented towards production needs, specifically including: For multimodal process data such as part features, process routes, equipment operation and quality inspection from sources such as expert knowledge, process technology manuals, and MES systems, named entity recognition and feature template matching algorithms are used to identify and extract key process elements, including part geometric features, process names, equipment models, tool parameters, machining accuracy and quality indicators, etc. Then, natural language processing and semantic segmentation technologies are used to realize semantic parsing and attribute mapping of the data to form a standardized process data expression structure. Based on the constructed triplet semantic structure, and using triplet statistical features and patterns, combined with clustering analysis and mutual information calculation methods, we identify the multidimensional mapping relationships between part features and process parameters, and between equipment capabilities and processing quality. We also explore the coupling patterns and dependency modes among different levels of process knowledge, such as part-level, feature-level, and process-level. Furthermore, through causal analysis and other methods, we construct a multidimensional association model of "part-process-equipment-quality" to realize the logical organization and hierarchical expression of knowledge from local processes to global processes. Based on a unified semantic modeling framework, the RDF / OWL language standard is used to semantically model and uniformly manage nodes, attributes, and relationships, constructing a reasonable process knowledge graph. Through ontology constraints and hierarchical inheritance mechanisms, semantic relationships and logical rules between parts, processes, equipment, and quality entities are defined, forming a knowledge structure with contextual understanding and semantic retrieval capabilities. Combining graph database indexing mechanisms and query optimization algorithms, efficient retrieval and dynamic updating of process knowledge are achieved. Finally, a structured and scalable process knowledge base is formed, providing knowledge support for subsequent process scheme generation, parameter optimization, and equipment scheduling decisions.

[0009] In an optional implementation, step S3 involves establishing a process optimization and collaborative decision-making mechanism for multiple devices, processes, and parameters based on the initial processing scheme generated by the knowledge base. Addressing the disturbances such as production disturbances, equipment state changes, and process deviations perceived in real-time by the digital twin model, a multi-agent collaborative optimization mechanism for the process scheme is constructed. Each agent performs local optimization based on its own state and learns and updates its strategy under the constraints of a system-level objective function. This is combined with deep reinforcement learning algorithms to achieve global coordination and dynamic optimization of the process scheme. During deep reinforcement learning training, an improved stochastic gradient descent (SGD) algorithm is introduced as the optimization core to dynamically update and converge to the agent's policy network parameters, improving the stability and global optimization capability of the strategy optimization. Based on this mechanism, each agent completes policy training and parameter updates in parallel, achieving adaptive optimization and continuous adjustment of the process scheme under different operating conditions, task switching, and equipment performance fluctuations. Specifically, this includes: To address the real-time perception of production disturbances and equipment status changes by digital twin systems, a multi-agent-based process optimization structure is established, modeling key equipment or processes in the production line as independent agents. Each agent uses its own process parameters, equipment load, and processing quality characteristics as input variables, combined with equipment constraints and process boundary conditions, with processing efficiency, energy consumption balance, and product accuracy as local optimization objectives. Based on this, a system-level comprehensive objective function is defined, covering multi-dimensional indicators such as production cycle time, resource utilization, and quality consistency, constructing a top-down global constraint optimization system to achieve global coordination and dynamic coupling optimization among multiple devices, processes, and parameters. Within a multi-agent optimization framework, a state-action-reward policy learning model is constructed using deep reinforcement learning algorithms. Simulation feedback from a digital twin environment is utilized to achieve policy iteration and knowledge sharing among agents. To improve the convergence efficiency and global optimization capability of policy optimization, an improved stochastic gradient descent (SGD) algorithm is introduced to perform distributed dynamic optimization of the agent policy network parameters. This algorithm introduces stochastic potential function constraints during the optimization process to adaptively control the gradient fluctuations of the loss function, thereby enhancing the stability and global optimization capability of the policy network under complex conditions. During training, a hierarchical policy update mechanism is employed, allowing local optimization results to be continuously corrected and iteratively converged under global constraints, effectively avoiding getting trapped in local optima. After the strategy training is completed, each agent dynamically updates the process plan and control strategy based on real-time production line status and digital twin feedback information, and achieves adaptive migration and re-optimization. By driving the online adjustment of the optimization strategy, the process plan is continuously optimized and updated, enabling the production process to maintain globally optimal execution and stable system operation under multi-objective constraints.

[0010] In an optional implementation, step S4 involves constructing a knowledge base self-evolution and update mechanism, optimizing decision-making schemes and transmitting them to the physical production line control system for execution, continuously sensing and analyzing information during the production execution phase based on a digital twin model, and feeding the sensed information back to the process knowledge base. By analyzing the correlation between virtual-real mapping errors and process performance deviations in the sensed information, knowledge node reconstruction and attribute updates are triggered. Combined with temporal correlation analysis and incremental learning algorithms, knowledge relationship weights are corrected to achieve data-driven knowledge self-evolution and continuous optimization. Specifically, this includes: After the process optimization results are sent to the physical production line, the equipment status, processing quality and process stability during the production execution stage are continuously perceived and analyzed based on the digital twin model. At the same time, in order to ensure the accuracy of the perception results, the twin model is dynamically updated and continuously optimized by combining the constructed virtual-real consistency maintenance mechanism. At the knowledge level, the correlation between virtual-real mapping deviation and process performance fluctuations is analyzed. Temporal correlation analysis is used to identify knowledge evolution trends, and incremental learning algorithms are combined to dynamically correct knowledge relationship weights and node attributes, thereby achieving self-evolution and adaptive optimization of the knowledge system. A two-way data-knowledge linkage mechanism is established to map real-time feedback data from the production process to knowledge nodes, enabling dynamic evolution and updating of the knowledge structure. The updated knowledge results are used to guide subsequent process optimization and decision-making reasoning, ensuring the integrity and timeliness of the process knowledge system during the continuous operation of the production line.

[0011] Secondly, the present invention provides a discrete manufacturing line process decision-making and optimization system based on digital twins, comprising: The production line digital twin model construction and sensing module is used to construct a high-fidelity digital twin model based on the equipment operation characteristics and production process data of discrete manufacturing production lines, establish a virtual-real consistency evaluation index system, realize deviation identification and consistency maintenance, and sense production disturbances, equipment status changes, and process deviation information in real time. Specifically, the production line digital twin model construction and sensing module is used to: collect equipment operation data, process parameters, energy consumption characteristics, and processing quality indicators; perform time synchronization, feature extraction, and noise filtering on multi-source heterogeneous data; construct a digital twin data model by combining neural network algorithms; and construct a digital twin mechanism based on mechanism modeling methods. The system employs a multi-level model fusion strategy to achieve complementarity and unification between the mechanistic model and the data-driven model; it establishes a virtual-real consistency evaluation system to quantitatively analyze the differences between the digital twin model and the physical production line; based on the virtual-real mapping comparison results, it identifies key influencing factors of deviations by combining causal analysis, and dynamically updates the twin model using parameter self-correction, incremental retraining, and structural correction methods; based on the virtual-real mapping mechanism of the digital twin model, it dynamically compares the differences between the virtual model and the physical system in terms of time series, state variables, and mechanistic response characteristics, enabling real-time perception and identification of operating condition disturbances, equipment performance degradation, and process deviation information. Based on the equipment operation characteristics and production process data of discrete manufacturing production lines, this study employs mechanistic modeling to depict the physical laws of the process and combines data-driven modeling to describe its dynamic evolution characteristics, constructing a high-fidelity digital twin model corresponding to the physical production line. Based on causal analysis, the study analyzes the influencing factors and changing patterns of the consistency between the virtual and physical mapping of the digital twin model, establishing a multi-dimensional attribute-based evaluation index system for virtual-physical consistency, and thereby achieving deviation identification and consistency maintenance of the digital twin model. Based on the virtual-physical comparison and dynamic mapping mechanism of the digital twin model, through comparative analysis of the virtual model and the physical production line in terms of time series, state variables, and mechanistic response characteristics, real-time perception and identification of information such as production disturbances, equipment state changes, and process deviations are achieved. The process knowledge base construction module is used to extract key process elements from multimodal process data, mine the semantic relationships between parts, processes, equipment, and quality indicators, construct a multidimensional association model of "parts-process-equipment-quality," build a structured and searchable process knowledge base based on a unified semantic modeling framework, and generate initial process plans. Specifically, it uses named entity recognition and feature template matching algorithms to identify and extract key process elements from multimodal process data, and forms a standardized process data expression structure through natural language processing and semantic segmentation techniques. Based on the triple semantic structure, combined with cluster analysis and mutual information calculation methods, it identifies the multidimensional mapping relationships between part features and process parameters, and between equipment capabilities and processing quality, and mines the coupling rules between process knowledge at different levels. It constructs a multidimensional association model of "parts-process-equipment-quality" through causal analysis. It uses the RDF / OWL language standard to perform semantic modeling and unified management of nodes, attributes, and relationships, and constructs a reasonable process knowledge graph. It defines the semantic relationships and logical rules between entities through ontology constraints and hierarchical inheritance mechanisms, forming a structured and scalable process knowledge base. For multimodal process data such as part features and process routes in the production process, key process elements are extracted and parsed using named entity recognition and feature template matching methods; combined with statistical analysis and machine learning algorithms, the correlation patterns and mapping modes of process knowledge at different levels are explored, and the semantic relationship between parts, processes, equipment and quality indicators is analyzed to construct a multidimensional correlation model of "parts-process-equipment-quality"; based on a unified semantic modeling framework, nodes, attributes and relationships are semantically modeled and uniformly managed to form a structured and searchable process knowledge base, and process solutions oriented towards production needs are generated; The dynamic optimization module for process schemes is used to establish a process optimization and collaborative decision-making mechanism for multiple equipment, processes, and parameters based on the initial process scheme. It constructs a multi-agent collaborative optimization mechanism based on real-time perception information from digital twins, and combines deep reinforcement learning algorithms and improved stochastic gradient descent algorithms to achieve global coordination and dynamic optimization of the process scheme. Specifically, key equipment or processes in the production line are modeled as independent agents. Each agent uses its own process parameters, equipment load, and processing quality characteristics as input variables, and performs local optimization in combination with equipment constraints and process boundary conditions. A system-level comprehensive objective function is defined to construct a global constraint optimization system. A state-action-reward policy learning model is constructed using deep reinforcement learning algorithms, and simulation feedback provided by the digital twin environment is used to achieve policy iteration and knowledge sharing among agents. An improved stochastic gradient descent algorithm is introduced to perform distributed dynamic optimization of the agent policy network parameters. Each agent dynamically updates the process scheme and control strategy based on real-time production line status and digital twin feedback information, achieving adaptive migration and re-optimization. Based on the initial processing scheme generated by the knowledge base, a process optimization and collaborative decision-making mechanism oriented towards multiple equipment, multiple processes, and multiple parameters is established. Addressing the disturbance information such as production disturbances, equipment state changes, and process deviations perceived in real time by the digital twin model, a multi-agent collaborative optimization mechanism for the process scheme is constructed. Each agent performs local optimization based on its own state and learns and updates its strategy under the constraints of the system-level objective function. Deep reinforcement learning algorithms are combined to achieve global coordination and dynamic optimization of the process scheme. During the deep reinforcement learning training process, an improved stochastic gradient descent (SGD) algorithm is introduced as the optimization core to dynamically update and converge to control the agent's policy network parameters, improving the stability and global optimization capability of the strategy optimization. Based on this mechanism, each agent completes policy training and parameter updates in parallel, achieving adaptive optimization and continuous adjustment of the process scheme under different operating conditions, task switching, and equipment performance fluctuations. The knowledge base self-evolution and update module is used to construct a knowledge base self-evolution and update mechanism. It transmits optimized decision-making schemes to the physical production line for execution, continuously senses production execution information based on a digital twin model, and feeds it back to the process knowledge base. It triggers knowledge node reconstruction and attribute updates by analyzing virtual-physical mapping errors and process performance deviations. Combined with temporal correlation analysis and incremental learning algorithms, it achieves knowledge self-evolution and continuous optimization. Specifically, after the process optimization results are sent to the physical production line, it continuously senses and analyzes the equipment status, processing quality, and process stability during the production execution phase based on the digital twin model. Combined with a virtual-physical consistency maintenance mechanism, it achieves dynamic updates of the twin model. It analyzes the correlation between virtual-physical mapping deviations and process performance fluctuations, uses temporal correlation analysis to identify knowledge evolution trends, and dynamically corrects knowledge relationship weights and node attributes using incremental learning algorithms. It establishes a data-knowledge bidirectional linkage mechanism, mapping real-time feedback data to knowledge nodes to achieve dynamic evolution and updates of the knowledge structure. A self-evolutionary update mechanism for the knowledge base is constructed, and the optimized decision-making scheme is transmitted to the physical production line control system for execution. Based on the digital twin model, information during the production execution stage is continuously perceived and analyzed. The perceived information is fed back to the process knowledge base. By analyzing the correlation between the virtual-real mapping error and the process performance deviation in the perceived information, knowledge node reconstruction and attribute updates are triggered. Combined with temporal correlation analysis and incremental learning algorithms, the knowledge relationship weights are corrected to achieve data-driven knowledge self-evolution and continuous optimization.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a high-fidelity digital twin model of a discrete manufacturing production line. By collecting equipment operation data and process parameters from the discrete manufacturing production line, and combining algorithms such as neural networks, a twin data model is constructed. Furthermore, a twin mechanism model is built from multiple dimensions, including geometric, physical, behavioral, and rule-based dimensions. A multi-level fusion strategy of the mechanism model and data model is implemented, achieving complementarity and unification between the two. Further, by establishing a virtual-real consistency evaluation system and a causal analysis mechanism, the differences between the twin model and the physical production line are quantitatively analyzed and dynamically corrected, enabling dynamic updates and consistency maintenance of the twin model. This maintains the model's consistency and stability, significantly improving the dynamic response capability of the digital twin model under dynamic operating conditions.

[0013] 2. This invention proposes a method for constructing a multi-dimensional semantic knowledge base oriented towards "parts-process-equipment-quality". Utilizing named entity recognition and feature template matching algorithms, core elements such as part features, process routes, equipment processing capabilities, and quality indicators are extracted from multi-source data. Combined with statistical analysis and machine learning algorithms, semantic association patterns and mapping relationships between different levels of process knowledge are mined, constructing a multi-dimensional association model. The RDF / OWL semantic modeling framework is adopted to uniformly manage and semantically represent knowledge nodes, attributes, and relationships, realizing the structuring, reasoning-based, and efficient retrieval of unstructured process experience knowledge, effectively improving the systematicity and reusability of process knowledge.

[0014] 3. This invention proposes a multi-agent dynamic process optimization method based on deep reinforcement learning. Key equipment or processes in the production line are modeled as independent agents, each with process parameters, equipment load, and processing quality characteristics as input variables. Under system-level objective function constraints, multi-objective optimization is achieved. By introducing an improved stochastic gradient descent (SGD) algorithm, distributed dynamic updates and adaptive convergence of policy network parameters are realized, improving the stability of policy optimization and global optimization capability. This method overcomes the limitations of traditional process optimization relying on static experience and single-objective decision-making, realizing multi-process collaboration and dynamic, collaborative optimization of process schemes in complex manufacturing systems.

[0015] 4. This invention establishes a virtual-real feedback and knowledge self-evolution mechanism to achieve continuous optimization and self-learning of process knowledge. Combining information perceived by the digital twin model, at the knowledge level, based on the correlation analysis between virtual-real mapping deviation and process performance fluctuations, time-series correlation and incremental learning algorithms are used to automatically correct knowledge relationship weights and node attributes, forming a self-evolving knowledge system. This mechanism enables the knowledge base to have continuous learning and optimization capabilities, ensuring the accuracy of process scheme formulation and stable execution under multiple operating conditions and complex tasks.

[0016] Furthermore, the design principle of this invention is reliable, the structure is simple, and it has a very wide range of application prospects. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention; Figure 2 This is a block diagram of a discrete manufacturing production line process decision-making and optimization method and system based on digital twin, which is an embodiment of the present invention. Figure 3 This is a flowchart of an improved stochastic gradient descent algorithm according to an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0021] The discrete manufacturing line process decision-making and optimization method based on digital twin provided in this embodiment of the invention is executed by computer equipment, and correspondingly, a discrete manufacturing line process decision-making and optimization system based on digital twin runs in computer equipment.

[0022] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The implementing entity can be a discrete manufacturing production line process decision-making and optimization method and system based on digital twins. Depending on different requirements, the order of steps in this flowchart can be changed, and some can be omitted.

[0023] like Figure 1 As shown, the method includes: Step 110: Based on the equipment operation characteristics and production process data of the discrete manufacturing production line, mechanistic modeling is used to characterize the physical laws of the process, and data-driven modeling is combined to describe its dynamic evolution characteristics, constructing a high-fidelity digital twin model corresponding to the physical production line; based on the idea of ​​causal analysis, the influencing factors and changing laws of the consistency between the virtual and real mapping of the digital twin model are analyzed, and a virtual-real consistency evaluation index system under multi-dimensional attributes is established, thereby realizing the deviation identification and consistency maintenance of the digital twin model; based on the virtual-real comparison and dynamic mapping mechanism of the digital twin model, through comparative analysis of the virtual model and the physical production line in terms of time series, state variables and mechanistic response characteristics, real-time perception and identification of information such as production disturbances, equipment state changes and process deviations are realized; Step 120: For multimodal process data such as part features and process routes in the production process, key process elements are extracted and parsed using the named entity recognition feature template matching method; combined with statistical analysis and machine learning algorithms, the correlation patterns and mapping modes of process knowledge at different levels are explored, and the semantic relationship between parts, processes, equipment and quality indicators is analyzed to construct a multidimensional correlation model of "parts-process-equipment-quality"; based on a unified semantic modeling framework, nodes, attributes and relationships are semantically modeled and uniformly managed to form a structured and searchable process knowledge base, and process solutions oriented towards production needs are generated; Step 130: Based on the initial processing scheme generated by the knowledge base, establish a process optimization and collaborative decision-making mechanism for multiple devices, processes, and parameters. Addressing the real-time disturbances, equipment state changes, and process deviations perceived by the digital twin model, a multi-agent collaborative optimization mechanism for the process scheme is constructed. Each agent performs local optimization based on its own state and learns and updates its strategy under the constraints of the system-level objective function. Deep reinforcement learning algorithms are combined to achieve global coordination and dynamic optimization of the process scheme. During deep reinforcement learning training, an improved stochastic gradient descent (SGD) algorithm is introduced as the optimization core to dynamically update and converge to the agent's policy network parameters, improving the stability and global optimization capability of the strategy optimization. Based on this mechanism, each agent completes policy training and parameter updates in parallel, achieving adaptive optimization and continuous adjustment of the process scheme under different operating conditions, task switching, and equipment performance fluctuations. Step 140: Construct a knowledge base self-evolution and update mechanism, optimize decision-making schemes and transmit them to the physical production line control system for execution, and continuously monitor and perceive the equipment operating status, processing quality and process stability during execution; feedback data is transmitted back to the digital twin model and process knowledge base, and the twin model is dynamically updated based on the virtual-real consistency maintenance mechanism; at the same time, the process knowledge base triggers knowledge node reconstruction and attribute update by analyzing the correlation between virtual-real mapping error and process performance deviation, and corrects the knowledge relationship weights by combining temporal correlation analysis and incremental learning algorithms, so as to realize data-driven knowledge self-evolution and continuous optimization.

[0024] To facilitate understanding of the present invention, the following description further illustrates the discrete manufacturing production line process decision-making and optimization method provided by the present invention, based on the principle of a discrete twin-based discrete manufacturing production line process decision-making and optimization method, and in conjunction with the process of discrete manufacturing production line process decision-making and optimization in the embodiments.

[0025] For details, please refer to Figure 2 The method for discrete manufacturing line process decision-making and optimization based on digital twins includes: S1. Based on the equipment operation characteristics and production process data of discrete manufacturing production lines, mechanistic modeling is used to characterize the physical laws of the process, and data-driven modeling is combined to describe its dynamic evolution characteristics, constructing a high-fidelity digital twin model corresponding to the physical production line; based on the idea of ​​causal analysis, the influencing factors and changing laws of the consistency between the virtual and real mapping of the digital twin model are analyzed, and a virtual-real consistency evaluation index system under multi-dimensional attributes is established, thereby realizing the deviation identification and consistency maintenance of the digital twin model; based on the virtual-real comparison and dynamic mapping mechanism of the digital twin model, through comparative analysis of the virtual model and the physical production line in terms of time series, state variables and mechanistic response characteristics, real-time perception and identification of information such as production disturbances, equipment state changes and process deviations are realized.

[0026] S101. Collect equipment operation data, process parameters, energy consumption characteristics, and processing quality indicators from discrete manufacturing production lines. Perform time synchronization, feature extraction, and noise filtering on multi-source heterogeneous data, and construct a digital twin data model by combining algorithms such as neural networks. Based on mechanism modeling methods, describe key equipment and processes from multiple dimensions such as geometry, physics, behavior, and rules, and construct a digital twin mechanism model that reflects the mechanism and laws of the processing process. To address the problem that mechanism models are difficult to accurately characterize nonlinear and time-varying characteristics, construct a multi-level model fusion strategy to achieve complementary unification of mechanism models and data-driven models at the input feature, state variable, and output layers, thereby obtaining a high-precision digital twin model. Real-time data acquisition is achieved by deploying sensors and interfaces with systems such as MES to obtain multi-source heterogeneous data from discrete manufacturing production lines. This data includes equipment operation data, process parameters, energy consumption, and quality information. For data from different sources and sampling frequencies, the data format is standardized and timestamps are synchronized. Interpolation algorithms and time alignment mechanisms are used to ensure the temporal consistency of multi-source data. Subsequently, the data is preprocessed to remove outliers and high-frequency noise interference, improving the accuracy and usability of the original data. Feature extraction is performed on the processed data to obtain its time-domain features, frequency-domain features, and time-frequency features. Principal component analysis and other algorithms are used to perform feature dimensionality reduction and selection to form a standardized feature vector set. Based on the processed multidimensional feature vectors, convolutional neural networks (CNN) or long short-term memory networks (LSTM) are used to capture the spatial and temporal correlation between process parameters and equipment status, construct a nonlinear mapping relationship between input features and processing results, and establish a digital twin data model. Simultaneously, based on mechanism modeling methods such as MBD and FEA, key equipment and processes are described from multiple dimensions including geometry, physics, behavior, and rules, constructing a digital twin mechanism model that reflects the mechanism and laws of the processing process. Specifically, geometric models of typical equipment such as machine tools and AGVs in the production line are established, and physical models reflecting their kinematic and dynamic characteristics are also established. Combined with finite element analysis or energy balance analysis, the thermal, force, and vibration responses during the processing are modeled to reveal the influence of key process parameters on processing quality and energy consumption. On this basis, the operating behavior of equipment and the state transition relationship between processes are described through process motion logic modeling methods, forming a unified expression from physical process to process behavior. To address the nonlinear and time-varying characteristics that are difficult to accurately describe using mechanistic models, a multi-level mechanism-data model fusion strategy is constructed to achieve complementary unification between the mechanism model and the data-driven model. First, at the input feature layer, feature mapping and normalization are used to unify the mechanism variables and data feature space. Second, at the state layer, Kalman filtering is introduced to achieve dynamic collaborative updating of the mechanism state and the data state. Finally, at the output layer, weighted integration and attention mechanisms are used to fuse the mechanism response and the data model prediction results to obtain a comprehensive output. ; in, Output for the data model, X For the input feature vector set; For the output of the mechanistic model, The input physical variables are α; α is the fusion weight coefficient, which can be dynamically adjusted by minimizing the difference between the virtual and real systems. This fusion model retains the interpretability of the mechanistic model while possessing the adaptive and real-time updating capabilities of the data model, thus forming a high-fidelity digital twin model with high accuracy, strong robustness, and dynamic response performance.

[0027] S102. Establish a mechanism for maintaining consistency between the digital twin model and the physical production line. Based on multi-dimensional attributes such as accuracy, production cycle time, and dynamic response, construct a consistency evaluation system to quantitatively analyze the differences between the digital twin model and the physical production line. Based on the results of the virtual-physical mapping comparison, identify the key influencing factors that cause deviations using causal analysis methods. Employ methods such as parameter self-calibration, incremental retraining, and structural correction to dynamically update and continuously optimize the twin model, thereby maintaining the model's accuracy consistency and response stability, and achieving model consistency maintenance and continuous optimization. A consistency evaluation system for digital twin models is constructed, establishing a unified set of evaluation indicators from multiple dimensions such as accuracy, production cycle time, and dynamic response. Calculation methods are defined for different indicators to quantify the degree of consistency between the digital twin model and the physical production line. Taking system response time as an example, used to measure the dynamic response performance of the digital twin model under control commands, it can be defined as: ; in, For the response time of the digital twin model, The actual response time measured on the physical production line. n The number of samples.

[0028] Similarly, other indicators also have corresponding calculation methods defined, forming a multi-dimensional set of virtual-real consistency indicators. The system automatically assesses the degree of virtual-real discrepancy based on the indicator calculation results and calculates the comprehensive evaluation value of virtual-real consistency using methods such as weighted average or analytic hierarchy process.

[0029] When the overall consistency evaluation falls below a preset threshold, the system automatically triggers the deviation diagnosis and model correction process. First, based on the time series trends of the evaluation indicators of each dimension, the system uses methods such as mutual information entropy and Granger causality test to identify the correlation and causal dependence between deviation changes and potential influencing variables, thereby determining the key influencing factors causing virtual-real deviations, such as equipment performance degradation, environmental disturbances, control delays, or sensor drift.

[0030] After deviation localization, the system performs dynamic model updates: a parameter self-calibration mechanism corrects key model parameters in real time within a sliding time window, gradually bringing model predictions closer to physical measurements; an incremental retraining strategy is employed to update gradients using small samples based on new data batches, preventing the forgetting of existing knowledge and improving model adaptability; when parameter correction fails to eliminate deviations, a structural correction mechanism is triggered to fine-tune the model structure or coupling terms to restore model stability. Through these mechanisms, the digital twin model can continuously correct deviations during changes in the physical production line operation, achieving dynamic preservation of the virtual-real mapping relationship.

[0031] S103. A virtual-real mapping mechanism based on a digital twin model dynamically compares the differences between the virtual model and the physical system in terms of time series, state variables, and mechanism response characteristics, enabling real-time perception and identification of information such as operating condition disturbances, equipment performance degradation, and process deviations. By combining dynamic threshold detection and time series feature analysis algorithms, trend identification is performed on the monitoring data, and a state prediction model based on neural networks is used to achieve early perception and identification of potential anomalies, thereby providing data support for subsequent process optimization. The simulation data and the data collected from the physical production line are aligned by the time synchronization module to establish a time-state correspondence; and differential calculation and trend analysis are performed on key state variables such as spindle speed and feed rate to extract dynamic features that characterize changes in the system's operating state. Based on this, a dynamic threshold detection algorithm is introduced to monitor key parameters in real time, and the threshold range is dynamically adjusted according to historical statistical distribution and state prediction results to achieve rapid identification of sudden disturbances and abnormal fluctuations; at the same time, the state sequence is predicted by combining time series feature analysis method to achieve real-time perception and identification of equipment performance degradation and process deviation. Furthermore, a state prediction model based on long short-term memory networks is constructed to model and predict equipment performance degradation and process deviation trends, enabling early detection and identification of potential anomalies. The prediction results and real-time sensing data are synchronously input into the digital twin optimization module, providing accurate data support for subsequent process parameter adjustments and optimization strategy updates.

[0032] S2. For multimodal process data such as part features and process routes in the production process, key process elements are extracted and parsed using named entity recognition and feature template matching methods. Combined with statistical analysis and machine learning algorithms, the correlation patterns and mapping modes of process knowledge at different levels are explored, and the semantic relationship between parts, processes, equipment and quality indicators is analyzed to construct a multidimensional correlation model of "parts-process-equipment-quality". Based on a unified semantic modeling framework, nodes, attributes and relationships are semantically modeled and uniformly managed to form a structured and searchable process knowledge base, and process solutions oriented towards production needs are generated.

[0033] S201. For multimodal process data such as part features, process routes, equipment operation and quality inspection from sources such as expert knowledge, process technology manuals, and MES systems, key process elements are identified and extracted using named entity recognition and feature template matching algorithms. These elements include part geometric features, process names, equipment models, tool parameters, machining accuracy and quality indicators. Natural language processing and semantic segmentation technologies are used to perform semantic parsing and attribute mapping on the data, forming a standardized process data expression structure. For multimodal process data from sources such as expert knowledge, process technology manuals, and MES systems, establish a process for extracting key process elements and expressing them semantically; First, data from different sources are preprocessed in a unified manner, including format parsing, noise filtering, and semantic tag normalization, in order to eliminate structural inconsistencies caused by differences in data sources. Subsequently, a named entity recognition model is used to extract process entity information, such as part geometric features, process names, equipment models, tool parameters, machining accuracy and quality indicators; and combined with a feature template matching algorithm, automatic recognition and attribute localization of specific structured expressions such as process cards and process specification tables are achieved.

[0034] Following entity recognition, dependency parsing, semantic segmentation, and contextual semantic disambiguation techniques from natural language processing are used to parse the semantic dependencies between different entities and label their semantic roles. Through semantic template matching and context window clustering, "action-object-constraint" relationships are identified, and the extracted results are uniformly mapped to a semantic structure of "entity-attribute-value" triples. A unified semantic mapping rule is established, defining the relationship types and logical constraints between entities based on the hierarchical structure of process knowledge, achieving semantic alignment and consistent expression of knowledge from different sources.

[0035] S202. Based on the constructed triplet semantic structure, and using the statistical characteristics and patterns of triples, combined with clustering analysis and mutual information calculation methods, we identify the multidimensional mapping relationships between part features and process parameters, and between equipment capabilities and processing quality. We also explore the coupling patterns and dependency modes among different levels of process knowledge, such as part-level, feature-level, and process-level. Furthermore, through causal analysis and other methods, we construct a multidimensional association model of "part-process-equipment-quality" to realize the logical organization and hierarchical expression of knowledge from local processes to global processes. Statistical analysis was performed on the triples extracted from the process dataset. By calculating the occurrence frequency and conditional probability of entities and attributes, a semantic frequency matrix was constructed. M Its elements are defined as:

[0036] in, Representing entities and The joint probability of occurrence in the same context , These represent their marginal probabilities. This matrix can be used to calculate the semantic association strength between different entities, providing a statistical basis for subsequent similarity calculations and cluster analysis; Furthermore, to quantify the nonlinear dependencies between different factors, a mutual information calculation method is introduced for any two random variables. X and Y Taking spindle speed and surface roughness as examples, their mutual information is defined as follows:

[0037] A higher mutual information value indicates a stronger dependency between variables. By analyzing the mutual information matrix of triplet combinations, key correlation features that significantly impact production line performance can be identified.

[0038] After obtaining the semantic frequency and mutual information matrix, a causal analysis model is further introduced to model and verify the influence direction and degree of key features, and a directed weighted graph is constructed to identify the hierarchical influence links of "part features - process parameters - equipment capabilities - processing quality", forming a knowledge logic organization structure from local processes to global processes. Ultimately, a multi-dimensional correlation model of "parts-process-equipment-quality" is constructed to achieve semantic coupling and hierarchical expression among complex process knowledge.

[0039] S203. Based on a unified semantic modeling framework, the RDF / OWL language standard is used to perform semantic modeling and unified management of nodes, attributes, and relationships, constructing a reasonable process knowledge graph. Through ontology constraints and hierarchical inheritance mechanisms, semantic relationships and logical rules between parts, processes, equipment, and quality entities are defined, forming a knowledge structure with contextual understanding and semantic retrieval capabilities. Combining graph database indexing mechanisms and query optimization algorithms, efficient retrieval and dynamic updating of process knowledge are achieved. Finally, a structured and scalable process knowledge base is formed, providing knowledge support for subsequent process scheme generation, parameter optimization, and equipment scheduling decisions. Based on the above multidimensional association model, a unified process knowledge semantic model and knowledge base structure are constructed. According to the unified semantic modeling framework, language standards such as RDF and OWL are adopted to transform the aforementioned triples and their logical relationships into semantic nodes and edges. Among them, nodes are used to represent entities such as parts, equipment, processes, and quality indicators, and edges are used to represent semantic relationships and constraint types between entities.

[0040] Based on this, by establishing ontology constraint rules and hierarchical inheritance mechanisms, a set of semantic relationships between "part-process-equipment-quality" is defined. Each relationship By triplet Formal representation is used to describe the logical connections and contextual dependencies between knowledge; and through rule-based reasoning and consistency verification algorithms, semantic conflicts between different knowledge sources are detected and resolved, thereby maintaining the semantic integrity and consistency of the knowledge system.

[0041] Furthermore, by combining the indexing mechanisms and query optimization algorithms of graph databases such as Neo4j, rapid retrieval of knowledge nodes and reasoning of relational paths can be achieved; Construct incremental update and dynamic insertion strategies to support online updates and association expansion of knowledge nodes, enabling the knowledge graph to have continuous evolution capabilities; Ultimately, a structured, scalable process knowledge base with reasoning capabilities was constructed, providing semantic-level knowledge support for subsequent process scheme generation, parameter optimization, and equipment scheduling decisions.

[0042] S3. Based on the initial processing scheme generated by the knowledge base, establish a process optimization and collaborative decision-making mechanism for multiple devices, multiple processes, and multiple parameters. Addressing the real-time disturbances, equipment state changes, and process deviations perceived by the digital twin model, a multi-agent collaborative optimization mechanism for the process scheme is constructed. Each agent performs local optimization based on its own state and learns and updates its strategy under the constraints of the system-level objective function. Deep reinforcement learning algorithms are combined to achieve global coordination and dynamic optimization of the process scheme. During the deep reinforcement learning training process, an improved stochastic gradient descent (SGD) algorithm is introduced as the optimization core to dynamically update and converge to the agent's policy network parameters, improving the stability and global optimization capability of the strategy optimization. Based on this mechanism, each agent completes policy training and parameter updates in parallel, achieving adaptive optimization and continuous adjustment of the process scheme under different operating conditions, task switching, and equipment performance fluctuations. S301. To address the real-time perception of production disturbance information and equipment status changes by the digital twin system, a process optimization structure centered on multiple agents is established. Key equipment or processes in the production line are modeled as independent agents. Each agent uses its own process parameters, equipment load, and processing quality characteristics as input variables, and combines equipment constraints and process boundary conditions to optimize local efficiency, energy consumption balance, and product accuracy. Based on this, a system-level comprehensive objective function is defined, covering multi-dimensional indicators such as production cycle time, resource utilization, and quality consistency. A top-down global constraint optimization system is constructed to achieve global coordination and dynamic coupling optimization among multiple devices, processes, and parameters. To enable digital twin systems to perceive production disturbance information and equipment status changes in real time, a process optimization structure centered on multiple agents is established. Key equipment, process units, and processing links in the production line are modeled as independent intelligent agents. Each intelligent agent makes decisions and optimizes based on its own state to achieve local optimization and global coordination. Each intelligent agent uses its own process parameters, processing cycle time, equipment load, and processing quality characteristics sensed by the digital twin system as input variables, and combines equipment constraints and process boundary conditions to construct a local optimization objective function:

[0043] in, For processing output, For the cycle time of the process, This indicates the equipment load rate or resource utilization rate. and These refer to the actual machining accuracy and the target accuracy, respectively. , , These are weighting coefficients used to balance efficiency, resource utilization, and quality objectives; To achieve global coordination and dynamic optimization among multiple devices and processes, considering factors such as production cycle time, resource load, capacity, and quality indicators, the system-level comprehensive objective function is defined as follows:

[0044] in, , , These represent the variances of production cycle time, resource load, and capacity distribution, respectively, and are used to measure the stability and coordination of the system. This refers to the overall product quality pass rate or comprehensive quality score index. ~ These are system-level weighting coefficients; The optimization direction of the objective function is to maximize. This means simultaneously achieving the comprehensive goals of rhythm coordination, resource balance, capacity matching, and optimal quality; The master control agent dynamically adjusts global constraints and task allocation based on the system-level objective function, and issues optimization instructions to each local agent, realizing hierarchical optimization and coordinated control from system-level constraints to process-level execution, thereby maintaining the stability, efficiency and quality consistency of the overall system under different production conditions.

[0045] S302. Within a multi-agent optimization framework, a state-action-reward policy learning model is constructed using deep reinforcement learning algorithms. Simulation feedback provided by a digital twin environment enables policy iteration and knowledge sharing among agents. To improve the convergence efficiency and global optimization capability of policy optimization, an improved stochastic gradient descent (SGD) algorithm is introduced to perform distributed dynamic optimization of the agent policy network parameters. This algorithm introduces stochastic potential function constraints during the optimization process to adaptively control the gradient fluctuations of the loss function, thereby enhancing the stability and global optimization capability of the policy network under complex conditions. During training, the system employs a hierarchical policy update mechanism, enabling local optimization results to be continuously corrected and iteratively converged under global constraints, effectively avoiding getting trapped in local optima. Within a multi-agent optimization framework, a state-action-reward policy learning model is constructed using deep reinforcement learning algorithms. In this embodiment, a deep Q-network (DQN) is selected to model the optimization policies of each agent. Each agent, based on its own state... Select Action And rewards based on feedback from the digital twin environment Update the policy network parameters. A high-fidelity twin simulation environment is used to implement policy iteration and experience playback mechanisms, promoting knowledge sharing and collaborative learning among agents. This constructs a distributed reinforcement learning optimization structure, enabling dynamic optimization and adaptive updating of the process scheme. The core update mechanism of DQN is:

[0046] in, For state-action value functions, η For learning rate, γ This is a discount factor used to balance immediate rewards and long-term benefits; To improve the convergence and global optimization capabilities of the algorithm under complex conditions, an improved stochastic gradient descent algorithm is introduced as the core mechanism for updating the parameters of the DQN network, such as... Figure 3 The diagram illustrates this method. This algorithm introduces a stochastic potential function constraint during parameter updates to adaptively control the gradient fluctuations of the loss function, effectively balancing convergence speed and stability. Its loss function is as follows:

[0047] in, This represents a stochastic potential term based on energy functionals, used to constrain the range of parameter perturbations. λ This is the adjustment coefficient; Each intelligent agent performs policy training and parameter updates in parallel in a digital twin simulation environment. The utilization rate of training samples is improved through an experience replay mechanism, and the target network's stable Q-value update process is utilized to ultimately achieve collaborative optimization among multiple processes and devices. To improve the convergence stability of policy network parameter updates in reinforcement learning, this study introduces an improved SGD optimization algorithm based on the stochastic potential function method. Its core idea is to place the SGD algorithm within a statistical physics framework and use stochastic differential equations to describe the parameter update process. In deep learning algorithms, the update of weight parameters in the SGD algorithm is controlled by the loss function; the smaller the loss function, the closer the estimated result is to the observed value. In this embodiment, within the framework of statistical physics, the parameter space of the deep learning model is abstracted as a non-equilibrium dynamic system. The update of the system's weight parameters can be described by a partial differential equation, as shown below: Where ω is the weight parameter of the entire neural network. It is the total loss function. α It's the learning rate. η (ω) is a random term; In this embodiment, the change in system state can be represented by a first-order stochastic differential equation:

[0048] Among them, state parameters x It is n Dimensional vector. Based on the stochastic potential function method, combining equations (9) and (10), the weight parameters are regarded as state variables of the system. Under near-steady-state conditions, they can be decomposed into:

[0049] Where F is the force matrix; according to the stochastic potential function method, the dynamic behavior of the system can be divided into three parts, namely: Potential matrix U: Represents the energy function stochastically related to the system; Symmetric positive semidefinite diffusion matrix D: used to describe the dissipative motion of the system; Antisymmetric transverse matrix Q: used to describe the conservative motion of the system; Construct the covariance matrix ∑ of the weight parameters, and calculate the relationship between ∑ and each matrix. Reconstruct the force matrix F and potential energy matrix U of the dynamic process based on the covariance matrix ∑ and the diffusion matrix D. Through the force matrix and loss function The relationship between the system and the potential energy matrix U is analyzed to obtain the relationship between the system's "inverse variance and flatness", and the regulatory effect of the energy function on the updating of the system's weight parameters is obtained. In this example, the neural network part of DQN is regarded as the above-mentioned non-equilibrium dynamic system. The dynamic matrices of each part of the system are constructed, and the dynamic characteristics of the SGD algorithm at the key point (the position of minimum loss function) are quantitatively predicted based on the system's diffusion matrix D, transverse matrix Q and potential energy matrix U, such as the existence of periodic oscillating vortex structures. The results show that at the minimum position of the loss function, the existence of the transverse matrix Q leads to the formation of a local "closed-loop" vortex structure in the parameter space, resulting in parameter update oscillation and convergence instability. Therefore, we analyze the influence of the transverse matrix Q on this vortex structure and quantitatively predict the existence of this structure in the parameter space of the system based on its distribution characteristics. We then optimize the SGD algorithm accordingly to solve the problems of weight parameter forgetting and local convergence in the multi-task learning process. The improved SGD algorithm, in terms of optimization approach, introduces stochastic potential function constraints to dynamically adjust the parameter update path, enabling the model to converge along the low curvature, flat minimum direction in the high-dimensional weight space. At the same time, it utilizes the potential energy-diffusion balance mechanism to suppress gradient oscillations, improving the stability of the convergence process and the global optimization capability, thereby significantly enhancing the robustness and decision accuracy of deep reinforcement learning algorithms under complex conditions.

[0050] S303. After the strategy training is completed, each agent dynamically updates the process plan and control strategy based on the real-time production line status and digital twin feedback information, and achieves adaptive migration and re-optimization. By driving the online adjustment of the optimization strategy, the process plan is continuously optimized and self-evolved, enabling the production process to maintain globally optimal execution and stable system operation under multi-objective constraints.

[0051] S4. Construct a knowledge base self-evolution and update mechanism, optimize the transmission of decision-making schemes to the physical production line control system for execution, and continuously perceive and analyze information during the production execution stage based on a digital twin model; the perceived information is fed back to the process knowledge base, and by analyzing the correlation between virtual-real mapping errors and process performance deviations in the perceived information, knowledge node reconstruction and attribute updates are triggered. Combined with temporal correlation analysis and incremental learning algorithms, the knowledge relationship weights are corrected to achieve data-driven knowledge self-evolution and continuous optimization. S401. After the process optimization results are sent to the physical production line, the equipment status, processing quality and process stability during the production execution stage are continuously perceived and analyzed based on the digital twin model. At the same time, in order to ensure the accuracy of the perception results, the twin model is dynamically updated and continuously optimized in combination with the constructed virtual-real consistency maintenance mechanism. S402. At the knowledge level, the correlation between the virtual-real mapping deviation and process performance fluctuations is analyzed by combining perception results. The temporal correlation analysis method is used to identify the knowledge evolution trend, and the knowledge relationship weights and node attributes are dynamically corrected by the incremental learning algorithm to realize the self-evolution and adaptive optimization of the knowledge system. A two-way linkage mechanism between data and knowledge is established to map real-time feedback data in the production process to knowledge nodes to realize the dynamic evolution and updating of the knowledge structure. The updated knowledge results are used to guide subsequent process optimization and decision reasoning to ensure the integrity and timeliness of the process knowledge system in the continuous operation of the production line.

[0052] Establish a matrix in the knowledge base to correspond the virtual-to-real mapping deviation with process performance indicators (such as machining accuracy, production cycle time, energy consumption, and equipment stability). ,in Representing knowledge nodes i The described process elements affect performance indicators j The influence coefficient; By comparing the predicted values ​​from the twin model with the actual measured values ​​from the physical production line, the deviation vector is calculated. And based on this, analyze the evolution trend of the deviation over time; Building upon this foundation, a temporal correlation analysis method is introduced to identify the temporal dependencies and evolutionary directions between knowledge nodes, forming a knowledge correlation chain in a time series; this is achieved by calculating the temporal correlation matrix between nodes. This study aims to determine the changes in knowledge activity and correlation strength under different time windows, providing a quantitative basis for judging the evolution of knowledge nodes. To achieve continuous knowledge updating and self-learning, an incremental learning method is used to dynamically adjust the weights and attributes of knowledge nodes. When new production data arrives, the principle of minimum perturbation is applied to update the parameters only for the affected nodes and their adjacency relationships. The update rule can be expressed as:

[0053] in, For knowledge nodes i The current weight, η For learning rate, This refers to the dynamic influence deviation between nodes. This mechanism ensures that the knowledge system can gradually absorb new technological rules while maintaining historical stability, thus achieving continuous evolution of knowledge.

[0054] Furthermore, a two-way linkage mechanism between data and knowledge is established to map real-time data collected by the digital twin system to knowledge nodes, thereby enabling data-driven knowledge triggering and reconstruction. When a specific process deviation or equipment status change is detected, the system automatically locates the affected knowledge node according to semantic matching rules, triggering corresponding attribute updates or structural reorganization; at the same time, the evolution results of the knowledge base have a reverse effect on the process optimization module, providing corrected process knowledge support for subsequent decision-making reasoning.

[0055] Through the above methods, the process knowledge base can achieve self-organization, self-evolution, and self-optimization in a dynamic production environment, thereby maintaining its high adaptability and timeliness to actual production line conditions, and providing continuous and reliable knowledge updates and intelligent decision support for the digital twin system.

[0056] In some embodiments, the discrete manufacturing line process decision-making and optimization system based on digital twins may include multiple functional modules composed of computer program segments. The computer programs of each program segment in the discrete manufacturing line process decision-making and optimization system may be stored in the memory of a computer device and executed by at least one processor to perform (see details). Figure 1 (Description) Functions for process decision-making and optimization in discrete manufacturing production lines.

[0057] In this embodiment, the discrete manufacturing line process decision-making and optimization system based on digital twins can be divided into multiple functional modules according to its functions, such as... Figure 2 As shown. The system's functional modules may include: a production line digital twin model construction and perception module 310, a process knowledge base construction module 320, a process scheme dynamic optimization module 330, and a knowledge base self-evolution and update module 340. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be detailed in subsequent embodiments.

[0058] The production line digital twin model construction and perception module: Based on the equipment operation characteristics and production process data of discrete manufacturing production lines, it uses mechanistic modeling to depict the physical laws of the process and combines data-driven modeling to describe its dynamic evolution characteristics, constructing a high-fidelity digital twin model corresponding to the physical production line; based on the idea of ​​causal analysis, it analyzes the influencing factors and changing laws of the consistency between the virtual and real mapping of the digital twin model, establishes a virtual-real consistency evaluation index system under multi-dimensional attributes, and realizes the deviation identification and consistency maintenance of the digital twin model accordingly; based on the virtual-real comparison and dynamic mapping mechanism of the digital twin model, through comparative analysis of the virtual model and the physical production line in terms of time series, state variables and mechanistic response characteristics, it realizes real-time perception and identification of information such as production disturbances, equipment state changes and process deviations; The process knowledge base construction module addresses multimodal process data such as part features and process routes in the production process. It utilizes named entity recognition and feature template matching methods to extract and parse key process elements. Combining statistical analysis and machine learning algorithms, it mines the correlation patterns and mapping models of process knowledge at different levels, analyzes the semantic relationships between parts, processes, equipment, and quality indicators, and constructs a multidimensional correlation model of "parts-process-equipment-quality." Based on a unified semantic modeling framework, it performs semantic modeling and unified management of nodes, attributes, and relationships, forming a structured and searchable process knowledge base, and generating process solutions tailored to production needs. The dynamic optimization module for process solutions establishes a process optimization and collaborative decision-making mechanism for multiple devices, processes, and parameters, based on the initial processing scheme generated by the knowledge base. Addressing real-time disturbances such as production disturbances, equipment state changes, and process deviations perceived by the digital twin model, a multi-agent collaborative optimization mechanism for process solutions is constructed. Each agent performs local optimization based on its own state and learns and updates its strategy under the constraints of the system-level objective function. Deep reinforcement learning algorithms are combined to achieve global coordination and dynamic optimization of the process solution. During deep reinforcement learning training, an improved stochastic gradient descent (SGD) algorithm is introduced as the optimization core to dynamically update and converge to the agent's policy network parameters, improving the stability and global optimization capability of the strategy optimization. Based on this mechanism, each agent completes policy training and parameter updates in parallel, achieving adaptive optimization and continuous adjustment of the process solution under different operating conditions, task switching, and equipment performance fluctuations. Knowledge base self-evolution and update module: Constructs a knowledge base self-evolution and update mechanism, optimizes decision-making schemes and transmits them to the physical production line control system for execution, continuously senses and analyzes information during the production execution stage based on a digital twin model; sensed information is fed back to the process knowledge base, and by analyzing the correlation between virtual-real mapping errors and process performance deviations in the sensed information, knowledge node reconstruction and attribute updates are triggered. Combined with temporal correlation analysis and incremental learning algorithms, knowledge relationship weights are corrected to achieve data-driven knowledge self-evolution and continuous optimization.

[0059] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. 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 also be covered within the protection scope of the present invention.

Claims

1. A method for process decision-making and optimization in discrete manufacturing lines based on digital twins, characterized in that, Includes the following steps: S1. Based on the equipment operation characteristics and production process data of discrete manufacturing production lines, a high-fidelity digital twin model corresponding to the physical production line is constructed by combining mechanism modeling and data-driven modeling; a virtual-real consistency evaluation index system is established to realize the identification of deviations and dynamic consistency maintenance of the digital twin model; based on the virtual-real mapping mechanism of the digital twin model, production disturbances, equipment status changes and process deviation information are perceived in real time. S2. For multimodal process data of part features and process routes, key process elements are extracted using named entity recognition and feature template matching methods; semantic relationships between parts, processes, equipment and quality indicators are mined to construct a multidimensional association model of "part-process-equipment-quality"; based on a unified semantic modeling framework, a structured and searchable process knowledge base is constructed to generate an initial process plan. S3. Based on the initial process plan, establish a process optimization and collaborative decision-making mechanism for multiple equipment, multiple processes, and multiple parameters; Based on the disturbance information perceived in real time by digital twins, a multi-agent collaborative optimization mechanism is constructed, and a deep reinforcement learning algorithm is combined to realize the global coordination and dynamic optimization of the process scheme; an improved stochastic gradient descent algorithm is introduced to dynamically update and converge to control the parameters of the agent policy network. S4. Construct a knowledge base self-evolution and update mechanism to transmit optimized decision-making solutions to the physical production line for execution; Based on the digital twin model, production execution information is continuously sensed and fed back to the process knowledge base. By analyzing the correlation between virtual and real mapping errors and process performance deviations, knowledge node reconstruction and attribute updates are triggered. Combined with temporal correlation analysis and incremental learning algorithms, knowledge relationship weights are corrected to achieve knowledge self-evolution and continuous optimization.

2. The method for discrete manufacturing line process decision-making and optimization based on digital twins according to claim 1, characterized in that, Step S1 specifically includes: S101, collecting equipment operation data, process parameters, energy consumption characteristics, and processing quality indicators; performing time synchronization, feature extraction, and noise filtering on multi-source heterogeneous data; constructing a digital twin data model using neural network algorithms; describing key equipment and processes from geometric, physical, behavioral, and rule dimensions based on mechanism modeling methods to construct a digital twin mechanism model; constructing a multi-level model fusion strategy to achieve complementary unification of the mechanism model and the data-driven model at the input feature, state variable, and output layers to obtain a high-precision digital twin model; S102, establishing a mechanism for maintaining the consistency between the virtual and real aspects of the digital twin model, from the perspectives of accuracy and production... A virtual-real consistency evaluation system is constructed based on the multi-dimensional attributes of cycle time and dynamic response to quantitatively analyze the differences between the digital twin model and the physical production line. Based on the virtual-real mapping comparison results, key influencing factors of deviation are identified by causal analysis. The twin model is dynamically updated and continuously optimized by means of parameter self-correction, incremental retraining and structural correction. S103, Based on the virtual-real mapping mechanism of the digital twin model, the differences between the virtual model and the physical system in terms of time series, state variables and mechanism response characteristics are dynamically compared. Combined with dynamic threshold detection and time series feature analysis algorithms, real-time perception and identification of operating condition disturbances, equipment performance degradation and process deviation information are realized.

3. The method for discrete manufacturing line process decision-making and optimization based on digital twins according to claim 1, characterized in that, Step S2 specifically includes: S201, for multimodal process data from expert knowledge, process technology manuals, and MES systems, using named entity recognition and feature template matching algorithms to identify and extract key process elements such as part geometric features, process names, equipment models, tool parameters, machining accuracy, and quality indicators, and forming a standardized process data expression structure through natural language processing and semantic segmentation techniques; S202, based on the triple semantic structure, combined with cluster analysis and mutual information calculation methods, identifying the multidimensional mapping relationship between part features and process parameters, equipment capabilities and machining quality, and mining the coupling rules and dependency patterns between different levels of process knowledge; constructing a multidimensional association model of "part-process-equipment-quality" through causal analysis; S203, using the RDF / OWL language standard to perform semantic modeling and unified management of nodes, attributes, and relationships, constructing a reasonable process knowledge graph; defining semantic relationships and logical rules between entities through ontology constraints and hierarchical inheritance mechanisms, and combining graph database indexing mechanisms and query optimization algorithms to form a structured and scalable process knowledge base.

4. The method for discrete manufacturing line process decision-making and optimization based on digital twins according to claim 1, characterized in that, Step S3 specifically includes: S301, modeling key equipment or processes in the production line as independent intelligent agents, each agent using its own process parameters, equipment load, and processing quality characteristics as input variables, combined with equipment constraints and process boundary conditions, with processing efficiency, energy consumption balance, and product accuracy as local optimization objectives; defining a system-level comprehensive objective function, covering multi-dimensional indicators such as production cycle time, resource utilization, and quality consistency, and constructing a global constraint optimization system; S302, constructing a state-action-reward policy learning model using deep reinforcement learning algorithms, and using simulation feedback provided by the digital twin environment to achieve policy iteration and knowledge sharing among agents; introducing an improved stochastic gradient descent algorithm to perform distributed dynamic optimization of the agent policy network parameters, and adaptively controlling the gradient fluctuation of the loss function by introducing stochastic potential function constraints during the optimization process; S303, each agent dynamically updates the process plan and control strategy based on real-time production line status and digital twin feedback information, realizing adaptive migration and re-optimization.

5. The method for discrete manufacturing line process decision-making and optimization based on digital twins according to claim 1, characterized in that, Step S4 specifically includes: S401, after the process optimization results are sent to the physical production line, continuously sensing and analyzing the equipment status, processing quality, and process stability during the production execution stage based on the digital twin model, and dynamically updating the twin model by combining the virtual-real consistency maintenance mechanism; S402, analyzing the correlation between virtual-real mapping deviation and process performance fluctuation, identifying knowledge evolution trends by using time-series correlation analysis, dynamically correcting knowledge relationship weights and node attributes by combining incremental learning algorithms; establishing a data-knowledge two-way linkage mechanism, mapping real-time feedback data to knowledge nodes, and realizing the dynamic evolution and updating of the knowledge structure.

6. A discrete manufacturing production line process decision-making and optimization system based on digital twins, characterized in that, include: The production line digital twin model construction and perception module (310) is used to construct a high-fidelity digital twin model based on the equipment operation characteristics and production process data of discrete manufacturing production lines, establish a virtual-real consistency evaluation index system, realize deviation identification and consistency maintenance, and perceive production disturbances, equipment status changes and process deviation information in real time; the process knowledge base construction module (320) is used to extract key process elements from multimodal process data, mine the semantic relationship between parts, processes, equipment and quality indicators, construct a multi-dimensional association model of "parts-process-equipment-quality", construct a structured and searchable process knowledge base based on a unified semantic modeling framework, and generate an initial process plan; the process plan dynamic optimization module (330) The first module is used to establish a process optimization and collaborative decision-making mechanism for multiple equipment, multiple processes, and multiple parameters based on the initial process scheme. It constructs a multi-agent collaborative optimization mechanism based on real-time perception information of digital twins and combines deep reinforcement learning algorithm and improved stochastic gradient descent algorithm to realize global coordination and dynamic optimization of process scheme. The second module is a knowledge base self-evolution update module (340), which is used to construct a knowledge base self-evolution update mechanism, transmit the optimization decision scheme to the physical production line for execution, continuously perceive production execution information based on the digital twin model and feed it back to the process knowledge base, and trigger knowledge node reconstruction and attribute update by analyzing virtual-real mapping error and process performance deviation. It combines temporal correlation analysis and incremental learning algorithm to realize knowledge self-evolution and continuous optimization.

7. The discrete manufacturing line process decision-making and optimization system based on digital twins according to claim 6, characterized in that, The production line digital twin model construction and perception module (310) is specifically used for: collecting equipment operation data, process parameters, energy consumption characteristics and processing quality indicators; performing time synchronization, feature extraction and noise filtering on multi-source heterogeneous data; constructing a digital twin data model in combination with neural network algorithms; constructing a digital twin mechanism model based on mechanism modeling methods; constructing a multi-level model fusion strategy to achieve the complementary unity of mechanism model and data-driven model; establishing a virtual-real consistency evaluation system to quantitatively analyze the differences between the digital twin model and the physical production line; and dynamically updating the twin model based on the virtual-real mapping comparison results, combined with causal analysis to identify key influencing factors of deviation, and using parameter self-correction, incremental retraining and structural correction methods. Based on the virtual-real mapping mechanism of the digital twin model, the differences between the virtual model and the physical system in terms of time series, state variables and mechanism response characteristics are dynamically compared, so as to realize the real-time perception and identification of operating condition disturbances, equipment performance degradation and process deviation information.

8. The discrete manufacturing line process decision-making and optimization system based on digital twins according to claim 6, characterized in that, The process knowledge base construction module (320) is specifically used for: identifying and extracting key process elements from multimodal process data using named entity recognition and feature template matching algorithms; forming a standardized process data expression structure through natural language processing and semantic segmentation techniques; identifying multidimensional mapping relationships between part features and process parameters, equipment capabilities and processing quality based on triple semantic structure, combined with cluster analysis and mutual information calculation methods; mining coupling rules between process knowledge at different levels; constructing a multidimensional association model of "part-process-equipment-quality" through causal analysis; using the RDF / OWL language standard to perform semantic modeling and unified management of nodes, attributes and relationships, and constructing a reasonable process knowledge graph; defining semantic relationships and logical rules between entities through ontology constraints and hierarchical inheritance mechanisms, forming a structured and scalable process knowledge base.

9. The discrete manufacturing line process decision-making and optimization system based on digital twins according to claim 6, characterized in that, The process dynamic optimization module (330) is specifically used for: modeling key equipment or processes in the production line as independent intelligent agents, with each intelligent agent using its own process parameters, equipment load and processing quality characteristics as input variables, and performing local optimization in combination with equipment constraints and process boundary conditions; defining a system-level comprehensive objective function and constructing a global constraint optimization system; constructing a state-action-reward policy learning model in combination with deep reinforcement learning algorithms, and using simulation feedback provided by the digital twin environment to realize policy iteration and knowledge sharing among intelligent agents; and introducing an improved stochastic gradient descent algorithm to perform distributed dynamic optimization of the policy network parameters of the intelligent agents. Each intelligent agent dynamically updates its process plan and control strategy based on real-time production line status and digital twin feedback information, achieving adaptive migration and re-optimization.

10. The discrete manufacturing line process decision-making and optimization system based on digital twins according to claim 6, characterized in that, The knowledge base self-evolution update module (340) is specifically used to: after the process optimization results are sent to the physical production line, continuously perceive and analyze the equipment status, processing quality and process stability of the production execution stage based on the digital twin model, and realize the dynamic update of the twin model by combining the virtual and real consistency maintenance mechanism; The correlation between virtual-real mapping deviation and process performance fluctuation is analyzed. The time-series correlation analysis method is used to identify knowledge evolution trends. The knowledge relationship weights and node attributes are dynamically corrected by combining incremental learning algorithms. Establish a two-way linkage mechanism between data and knowledge, mapping real-time feedback data to knowledge nodes to achieve dynamic evolution and updating of the knowledge structure.

Images