A copper-clad plate production line collaborative regulation method based on digital twinning

Abstract

The application relates to a copper-clad plate production line collaborative regulation method based on digital twinning, which comprises the following steps: through full-process digital twinning modeling, process data normalization and multi-level feature labeling, a multi-level data set of dynamic virtual-real mapping is constructed, high consistency of physical production line working conditions and simulation systems is realized, combined with autonomous scheduling intelligent agents, real-time sensing and joint feature mining of multi-dimensional data such as equipment load, health degree and energy consumption are realized, real-time adaptive scheduling capability of the production line to sudden load, equipment aging and other states is improved, scheduling target weights are periodically and adaptively generated, optimal immediate balance of multi-objective benefits is realized, evolutionary game and swarm intelligence optimization are adopted, optimal resource scheduling schemes across devices under multi-objective constraints are obtained, through continuous feedback evolution of real-time production data, a closed-loop mechanism of scheduling-execution-calibration-redistribution is formed, and the self-learning and long-period stability of the scheduling system are significantly improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a collaborative control method for copper clad laminate production lines based on digital twins. Background Technology

[0002] The main problems with existing technologies include:

[0003] First, the multi-objective scheduling trade-off mechanism lacks flexibility. Most scheduling systems use fixed target weights or simple weighting methods, which cannot quickly and adaptively adjust scheduling targets according to real-time operating conditions on the production site (order dynamics, equipment aging, energy efficiency and electricity price fluctuations). As a result, when faced with changing process scenarios and resource bottlenecks, the optimal solution for production line scheduling is delayed or even fails, and the overall capacity and energy efficiency potential cannot be fully released.

[0004] Second, the integration of digital twins and scheduling optimization is limited. Currently, most digital twins perform monitoring and visualization functions, lacking in-depth support for multi-objective scheduling collaborative optimization, evolutionary learning, and real-time feedback self-correction capabilities.

[0005] Third, the global optimization of cross-equipment and cross-process scheduling lacks an efficient collaborative mechanism. Distributed autonomous agents mainly rely on static goal collaboration or local game theory, making it difficult to achieve global adaptive resource allocation and iterative evolution of scheduling strategies under complex constraints, dynamic processes, and frequent migration of production line bottlenecks.

[0006] Therefore, there is an urgent need for a new multi-objective collaborative control method based on evolutionary digital twins and adaptive intelligent agents, which can dynamically quantify and adjust the weight of each scheduling objective, realize distributed negotiation and evolutionary optimization among various equipment and processes, adapt to the changing needs of processes and orders, and effectively tap into and release the overall flexible manufacturing efficiency of the production line. Summary of the Invention

[0007] To address the problems existing in the prior art, this application aims to provide a collaborative control method for copper clad laminate production lines based on digital twins.

[0008] The method for collaborative control of a copper clad laminate production line based on digital twins described in this application includes the following steps:

[0009] S1. Obtain real-time operation data of each equipment and process in the copper clad laminate production line, including multi-dimensional process parameters such as load, operating status, equipment health, energy consumption, and capacity, in order to establish a data input basis for a multi-level digital twin.

[0010] S2. Normalize the acquired real-time operation data and label the normalized data features based on different equipment types and process attributes to obtain a labeled dataset, so as to distinguish the operating condition variables that cover different equipment characteristics and process differences.

[0011] S3. Based on the normalized and labeled dataset, a multi-level digital twin modeling structure is constructed for each equipment type and process step to realize dynamic virtual-real mapping of physical production line operating conditions and multi-objective indicators.

[0012] S4. Input the multi-level digital twin modeling structure into the autonomous scheduling intelligent agent. The autonomous scheduling intelligent agent continuously senses and analyzes the load level, equipment health status and energy consumption parameters of the local and adjacent units to generate autonomous scheduling sensing characteristics under specific operating conditions.

[0013] S5. Based on the perception characteristics of the autonomous scheduling intelligent agent, combined with order priority, real-time electricity price changes and equipment aging trends, calculate the global target weight vector of the current production line to achieve periodic adaptive generation of multi-objective optimization weights;

[0014] S6. Input the global objective weight vector into the evolutionary game and swarm intelligence optimization algorithm, execute the resource scheduling collaborative game and scheduling scheme evolution among multiple intelligent agents, and generate the optimal solution set for cross-device scheduling under the current multi-objective constraints;

[0015] S7. Based on the generated cross-device scheduling optimal solution set, conduct parallel simulation verification of multiple schemes on the digital twin production line, and prioritize the selection of the final scheduling control parameter set that achieves the optimal overall goal and meets the requirements for flexible process switching.

[0016] S8. The final scheduling control parameter set is sent to the physical production line for scheduling execution, and the actual production response status is recorded online. Through the digital twin modeling structure and feedback data, the evolution model is continuously calibrated to achieve adaptive parameter correction.

[0017] S9. Monitor key process indicators after scheduling execution, including equipment health, energy consumption curve and capacity achievement rate, to determine whether there are abnormal fluctuations. If performance is detected to deviate from the set threshold, the scheduling parameter adaptive adjustment mechanism is automatically triggered to achieve closed-loop optimization under abnormal conditions.

[0018] Furthermore, in step S1, real-time operational data of each piece of equipment and process in the copper-clad laminate production line is acquired, including multi-dimensional process parameters such as load, operating status, equipment health, energy consumption, and capacity, to establish a data input foundation for a multi-level digital twin, including:

[0019] A multi-level intelligent sensor network method is used to collect real-time process parameters of all key production equipment in the copper clad laminate production line in order to achieve high-precision data acquisition of various process nodes.

[0020] The system is equipped with multimodal intelligent sensors that collect physical quantities and operating status parameters, including but not limited to current, voltage, temperature, vibration, pressure, and power. Through industrial-grade acquisition terminals, it achieves multi-channel synchronous data acquisition, comprehensively covering process data dimensions such as equipment load, operating status, health, energy consumption, and production capacity.

[0021] Further, in step S2, the acquired real-time operation data is normalized, and the normalized data features are labeled based on different equipment types and process attributes to obtain a labeled dataset, in order to distinguish operating condition variables covering different equipment characteristics and process differences, including:

[0022] The acquired real-time operation data is normalized based on industrial data standardization principles to generate a standardized set of process parameters for equipment and process conditions.

[0023] Based on a standardized set of process parameters, a dictionary of equipment and process attributes is constructed according to the type of production line equipment and the attributes of the process, forming a unified equipment type label and process link label.

[0024] For the normalized set of process parameters, feature vectors are assembled according to labels, and multiple characteristic labels are assigned by splicing and mapping.

[0025] A feature consistency verification algorithm is used on the structured feature-labeled samples to eliminate the distribution bias caused by potential heterogeneous data and generate a consistent feature-labeled sample set.

[0026] Process differences are identified in the standardized feature-labeled sample set. Based on the process fluctuation identification model, the distribution of variables across equipment and processes is screened, and a multi-objective scheduling optimization variable set is extracted.

[0027] Furthermore, in step S3, based on the normalized and labeled dataset, a multi-level digital twin modeling structure is constructed for each equipment type and process step to achieve dynamic virtual-real mapping of the physical production line operating conditions and multi-objective indicators, including:

[0028] Based on the label-based dataset, the equipment and process condition sub-modules are split;

[0029] A multi-objective factor mapping method is applied to correlate indicators such as load, energy consumption, and production capacity with state variables;

[0030] Use time-series modeling algorithms to process dynamic behavior and output real-time operating parameters;

[0031] Through a distributed virtual-physical mapping mechanism, physical production line data is synchronized to a digital twin to build a dynamic data mirror;

[0032] The algorithm performs a consistency calibration of the performance indicators, aligns the physical data with the twin prediction output, and generates mapping consistency parameters.

[0033] Further, in step S4, the multi-level digital twin modeling structure is input into the autonomous scheduling intelligent agent. This agent continuously senses and analyzes the load levels, equipment health status, and energy consumption parameters of local and adjacent units to generate autonomous scheduling sensing characteristics under specific operating conditions, including:

[0034] Interface encapsulation is performed on the instantiated process state parameters in the constructed multi-level digital twin modeling structure;

[0035] Instantiated process state parameters based on a multi-level digital twin modeling structure activate intelligent agents to monitor local load fluctuations in real time.

[0036] The autonomous scheduling intelligent agent instantiates the process state parameters of the input multi-level digital twin modeling structure to determine the health status of the equipment, and evaluates the health status and failure probability of the equipment through embedded algorithms.

[0037] An autonomous scheduling intelligent agent is used to perform real-time energy efficiency analysis and calculation on energy consumption parameters from a multi-level digital twin modeling structure, and outputs energy efficiency risk indicators based on the correlation between process load and equipment health status.

[0038] The autonomous scheduling intelligent agent obtains autonomous scheduling perception features from local process units and combines them with perception features from adjacent units to synthesize and calculate;

[0039] Based on load, health, and energy consumption correlation indicators, the final autonomous scheduling perception feature set is generated through feature importance evaluation.

[0040] Furthermore, in step S5, based on the perception characteristics of the autonomous scheduling intelligent agent, and combined with order priority, real-time electricity price changes, and equipment aging trends, the global target weight vector of the current production line is calculated to achieve periodic adaptive generation of multi-objective optimization weights, including:

[0041] In the collaborative control process of copper clad laminate production line driven by digital twin, it is responsible for taking autonomous scheduling perception characteristics as input, integrating multi-source business factors such as order priority, real-time electricity price changes and equipment aging trends, and calculating the global target weight vector of multi-objective optimization of the production line at the current stage through high-frequency calculation and adaptive weight generation mechanism.

[0042] As a core component of flexible production line scheduling and multi-objective optimal strategy generation, it directly affects the global evaluation baseline of subsequent scheduling algorithms, providing parameter basis for cross-device intelligent optimization and dynamic resource allocation.

[0043] Further, in step S6, the global objective weight vector is input into the evolutionary game and swarm intelligence optimization algorithm to perform resource scheduling collaborative game and scheduling scheme evolution among multiple intelligent agents, generating the optimal cross-device scheduling solution set under the current multi-objective constraints, including:

[0044] Input global weights, weight normalize indicators such as process and energy consumption, and construct an evaluation matrix;

[0045] By integrating the global objective with the local state of each intelligent agent through distributed knowledge sharing, cross-agent state synchronization is achieved, forming a complete game input;

[0046] Based on the resource strategy encoding of the cooperative state initialization, a cooperative game model is established; through multiple rounds of evolution, including state transition and fitness functions, a group of globally advantageous strategies is selected.

[0047] Based on the strategy population, the particle swarm optimization algorithm in the intelligent algorithm is used to perform parameter mutation and evolution to find the optimal solution set for cross-equipment and process scheduling.

[0048] By combining multi-objective performance and adaptability criteria, the solution set is evaluated both subjectively and objectively, and the optimal cross-device scheduling scheme is output.

[0049] Furthermore, in step S7, based on the generated cross-device scheduling optimal solution set, multi-scheme parallel simulation verification is performed on the digital twin production line. The final scheduling control parameter set that prioritizes achieving the optimal overall objective and meeting the requirements for flexible process switching is selected first, including:

[0050] The optimal solution set for cross-device scheduling is loaded into the digital twin platform to fully map the resource allocation logic and provide input for parallel simulation.

[0051] Driven by scheduling parameter groups, the twin engine simulates key scenarios such as equipment load, process switching and energy consumption fluctuations in parallel, and reproduces the real production line operation mechanism.

[0052] For the process data output from the simulation, the response values ​​of multi-dimensional indicators such as capacity, energy efficiency, and health are calculated, and a feature evaluation matrix is ​​constructed after normalization.

[0053] For the feature matrix, we examine the response latency, flexibility and other indicators of the scheduling scheme during process switching, and screen the parameter set that supports dynamic switching.

[0054] For the parameter set that passes the test, a global weighting algorithm is applied to select the optimal solution for multi-objective balance, and the structured results are output for production line execution.

[0055] Furthermore, in step S8, the final scheduling control parameter set is sent to the physical production line for scheduling execution, and the actual production response status is recorded online. Through the digital twin modeling structure and feedback data, the evolution model is continuously calibrated to achieve adaptive parameter correction, including:

[0056] The scheduling parameter group is serialized and encapsulated, and then synchronously distributed to the distributed device controller via an industrial protocol.

[0057] The controller drives the device to execute tasks according to scheduling constraints and generates device response data streams in real time.

[0058] Cascaded acquisition of equipment status, energy consumption, production capacity and other characteristic data to support twin model calibration;

[0059] Align real-time data with the digital twin model and generate dynamic accuracy correction factors through error analysis;

[0060] Based on the mapping relationship between decision parameters and targets in the adaptive optimization scheduling model with correction factors, the virtual-real deviation is corrected;

[0061] The effectiveness of the strategy is re-evaluated using a revised model, and feedback reports are generated to drive the continuous evolution of the model.

[0062] Furthermore, in step S9, key process indicators after scheduling execution are monitored, including equipment health, energy consumption curves, and capacity achievement rate, to determine whether there are abnormal fluctuations. If a performance deviation from a set threshold is detected, an adaptive adjustment mechanism for scheduling parameters is automatically triggered to achieve closed-loop optimization under abnormal conditions, including:

[0063] Based on the digital twin model, real-time data collection of process indicators such as equipment health, energy consumption, and production capacity is used to construct a full-process monitoring dataset;

[0064] Using threshold determination and time series analysis algorithms, abnormalities such as declining health, energy consumption jumps, and insufficient production capacity are automatically identified and classified.

[0065] Extract abnormal feature parameters, correlate them with historical model behavior, and generate an abnormal causal report based on the relationship between events;

[0066] Reinforcement learning in self-learning rules generates scheduling parameter correction suggestions that are adapted to abnormal features;

[0067] The suggested improvements are fed back to the digital twin model, driving the scheduling strategy to be updated in real time;

[0068] Monitoring the recovery trend of indicators verifies the positive impact of optimization measures on health, energy consumption, and production capacity, thereby enhancing system robustness.

[0069] The collaborative control method for copper clad laminate production lines based on digital twins described in this application has the following advantages: It constructs a hierarchical digital twin model through the normalization and feature labeling of multi-source process data; ensures consistency between the virtual and real worlds through industrial sensor networks; achieves high-precision mapping of the entire operating condition chain; configures autonomous intelligent agents for equipment and processes; integrates load identification and health assessment capabilities; achieves holographic perception through distributed feature aggregation; innovates the operating condition perception weight module; integrates order priority, real-time electricity price, and equipment aging trends; dynamically optimizes multi-objective weight allocation; breaks through static limitations; introduces evolutionary game theory and swarm intelligence algorithms to drive agent collaborative evolution; improves the global convergence efficiency of the scheduling scheme; conducts parallel simulation pre-verification based on digital twins to reduce the failure rate; and achieves continuous model evolution through a "scheduling-execution-feedback" closed loop. Attached Figure Description

[0070] Figure 1 This application describes a method for collaborative control of a copper clad laminate production line based on digital twins. Figure 1 ;

[0071] Figure 2 This application describes a method for collaborative control of a copper clad laminate production line based on digital twins. Figure 2 . Detailed Implementation

[0072] like Figures 1-2 As shown, the method for collaborative control of a copper clad laminate production line based on digital twins described in this application includes:

[0073] like Figures 1-2 As shown, S1, acquire real-time operation data of each equipment and process in the copper clad laminate production line, including multi-dimensional process parameters such as load, operating status, equipment health, energy consumption, and capacity, in order to establish the data input basis for a multi-level digital twin.

[0074] Further, in step S1,

[0075] S1.1 equips all key production equipment in the copper clad laminate production line with intelligent sensors and industrial data acquisition terminals. Through the acquisition module, production process parameters such as load, operating status, equipment health, energy consumption and capacity are collected in real time to obtain the original process data stream.

[0076] By utilizing industrial buses (such as Modbus, PROFINET, OPC UA, etc.) and edge data acquisition modules, the output of heterogeneous sensors is standardized to realize the synchronization interface of multi-source signals, and the acquisition time base is unified to form the original process data stream for key equipment.

[0077] Employing an embedded sampling calibration algorithm, based on a set sampling frequency and sliding window, the real-time data acquisition cycle is dynamically adjusted to ensure accurate real-time tracking of process parameters with high change rates; sampling frequency f s The data window N satisfies the following constraints:

[0078]

[0079] Among them, T sampling f is the window duration. s Where N is the sampling frequency and N is the number of sampling points in the window, which can be flexibly set according to the characteristics of the device.

[0080] Redundant channels and breakpoint resume mechanisms are adopted to perform real-time data packet integrity verification on critical data paths, and automatically switch to backup channels when the acquisition link fails, ensuring the reliable and continuous transmission of raw process data.

[0081] By using high-bandwidth data buffering and local caching mechanisms, batch raw process parameters are stored in an industrial database or edge storage module in timestamp format, laying the foundation for subsequent data association modeling and process parameter synchronization.

[0082] By using multimodal intelligent sensors and industrial bus data acquisition and calibration processing, comprehensive, real-time, and standardized raw process data stream acquisition is achieved, providing high-quality and traceable underlying data support for multi-level digital twin modeling.

[0083] For example, in a high-end copper-clad laminate intelligent production line, a three-phase current transformer (range 0-400A, accuracy 0.5%), a pressure sensor (0-10MPa, 0.1%FS), and a temperature sensor (-20℃-300℃, accuracy 0.05℃) are configured in the main control box of the press equipment, and are respectively connected in parallel to OPC. The UA data acquisition terminal has a data sampling frequency of 10Hz for each device. For baking ovens, it uses a combination of environmental temperature and humidity sensors and high-precision thermocouples to collect operating parameters such as process temperature and equipment cavity humidity. All acquisition channels are connected to the edge acquisition server through Ethernet switches. Each process batch is assigned a unique timestamp tag, and a dual-channel data verification algorithm is used to improve data reliability. During 24 / 7 automated operation, a single device collects approximately 864,000 data points per day, with an overall packet loss rate of less than 0.001%. It has achieved comprehensive raw process data stream acquisition for different equipment and processes. Its output results are multi-level, time-segmented raw equipment process datasets with high precision and low packet loss data quality, supporting the entire data foundation for subsequent equipment process tag association, data preprocessing, and digital twin modeling.

[0084] S1.2 Based on the equipment ledger information and process node configuration table, the collected original process data stream is labeled with equipment process association tags to realize multi-dimensional mapping and archiving of multiple equipment, multiple process variables and production nodes.

[0085] The original process data stream is used as the input basis, and subsequent processing is carried out in combination with equipment ledger information and process node configuration table.

[0086] Using a basic equipment attribute parsing algorithm, each data entry in the collected raw process data stream is associated and matched with each piece of equipment based on the unique equipment code, equipment type, installation location information, and factory asset label recorded in the equipment ledger, generating a mapping matrix between the unique equipment identifier and the corresponding data.

[0087] Furthermore, by using a process node mapping parsing algorithm (parameters include process number, section function, and node topology table), pattern feature discrimination is performed on the process trigger signals, operation events, and production batch numbers in the original data waveform, thereby realizing automatic identification of process links and dynamic binding of process node labels.

[0088] By using a multi-dimensional label overlay mechanism for equipment and processes, unique equipment identifiers and process node labels are integrated into each original process data record in a multi-level structure, forming a structured label template that covers multiple equipment, multiple processes, and multiple levels of process variables.

[0089] The algorithm for verifying the consistency of associated labels is adopted. The algorithm checks each labeled multidimensional label data sample based on the rules of label set integrity, temporal synchronization and process topology consistency. It identifies anomalies such as missing labels, mislabeling and process breakpoints, automatically backtracks and corrects them and outputs process associated label labeling results with high accuracy.

[0090] By using equipment process association labeling process, the original process data stream is transformed into a structured data archive with multi-equipment, multi-process variables and multi-dimensional mapping capabilities of production nodes, realizing multi-angle indexing of original data and accurate data traceability.

[0091] For example, in a high-end copper clad laminate production line, the raw process data stream is collected and recorded from the OPC UA terminal at a sampling frequency of 10Hz. Each data entry includes raw current, temperature, pressure, and other signals. In this step, the equipment asset numbers such as press 001 and baking oven A in the equipment ledger are used, combined with "pressing process - node 05" and "heat treatment process - node 09" in the process node configuration table, and each data entry is labeled as follows using a hash mapping method: the equipment label is #PR001, the process label is #Step05, and the production batch label is Batch20240601_1. After the equipment-process-batch triple label is embedded, the data entry is upgraded to the following structure: {timestamp, current, temperature, pressure, equipment = PR001, process = Step05, batch = Batch20240601_1};

[0092] For 864,000 sets of raw data per day, process label consistency verification is used to check for label breakpoints in 5% of the data, and unlabeled nodes are automatically corrected through process event tracing algorithm;

[0093] By labeling multiple equipment processes with related tags, the resulting structured multidimensional dataset significantly improves the efficiency of subsequent data retrieval, process mapping, and scheduling modeling, with a label accuracy rate of 99.8%.

[0094] S1.3 applies real-time data preprocessing algorithms to the raw process data stream with equipment process association labels, including industrial noise filtering, missing value imputation, and outlier detection and correction, to generate a noise-free and complete standardized process cube dataset.

[0095] S1.4 inputs the standardized process multidimensional dataset into the data synchronization and timing calibration module. Based on timestamps and trigger events, it performs timing alignment and synchronization aggregation on data involving different equipment and process links, generating a dynamically consistent multi-equipment time series process parameter matrix.

[0096] S1.5 uses a dynamic consistency-based multi-device time-series process parameter matrix. Through a multi-level data structure generation module, it outputs a structured, hierarchical (equipment-level, process-level, production line-level) named dataset, which is then labeled as the data input basis for a multi-level digital twin to support subsequent modeling and optimization.

[0097] like Figures 1-2 As shown in step S2, the acquired real-time operation data is normalized, and the normalized data features are labeled based on different equipment types and process attributes to obtain a labeled dataset, so as to distinguish the operating condition variables that cover different equipment characteristics and process differences.

[0098] Further, in step S2,

[0099] S2.1 performs normalization processing on the acquired real-time operation data based on industrial data standardization principles. Multi-dimensional process parameters such as load, energy consumption, operating status, equipment health, and capacity are uniformly processed using algorithms such as minimum-maximum normalization and Z-score normalization to generate a standardized set of process parameters for equipment and process conditions.

[0100] S2.2 Based on the standardized process parameter set, according to the production line equipment type (such as press, copper stripping equipment, baking oven, etc.) and process attributes (such as feeding, pressing, stripping, final inspection, etc.), a dictionary of equipment and process attributes is constructed to form a unified equipment type label and process link label.

[0101] S2.3 For the normalized set of process parameters, feature vectors are assembled according to the equipment type label and process step label. Through feature splicing and attribute mapping, each data sample is assigned a corresponding multi-feature label to generate structured feature-labeled samples.

[0102] S2.4 employs a feature consistency verification algorithm on the structured feature-labeled samples to verify the standardization consistency of the feature vector dimensions and value ranges of different equipment and processes in the series data, eliminate the distribution bias caused by potential heterogeneous data, and generate a consistent feature-labeled sample set.

[0103] S2.5 performs process difference identification processing on the standardized feature-labeled sample set, and further screens the distribution of key variables across equipment and processes based on the process fluctuation identification model, extracting the set of operating condition variables for multi-objective scheduling optimization, providing a highly identifiable data foundation for subsequent multi-objective optimization weight generation and virtual-real mapping steps.

[0104] like Figures 1-2 As shown, S3, based on the normalized and labeled obtained label dataset, constructs a multi-level digital twin modeling structure for each equipment type and process link to realize dynamic virtual-real mapping of physical production line operating status and multi-objective indicators.

[0105] Further, in step S3,

[0106] S3.1 uses normalized and labeled data as input conditions and employs an industrial knowledge-driven hierarchical decomposition algorithm to decompose each equipment type and process step into functional units to generate sub-module description parameters of the working status of each level of equipment and process, ensuring that the digital twin modeling structure has the ability to distinguish process attributes.

[0107] S3.2 describes the parameters of the generated operating condition sub-modules by applying a multi-objective factor mapping method to construct a one-to-one correspondence between multi-objective indicators such as load, energy consumption, production capacity, and equipment health and operating condition variables, thereby realizing the dynamic association expression of operating condition sub-modules to multi-objective indicators.

[0108] Using the operating status submodule description parameters of each device and process obtained through step S3.1 as input, perform multi-objective factor mapping processing.

[0109] By employing linear or nonlinear mapping methods (parameters: load, energy consumption, production capacity, equipment health, and other indicator sets), a corresponding target factor specification set is established for each operating condition state variable, thereby realizing the parameter expansion from a single operating condition variable to a multi-objective indicator space.

[0110] Furthermore, through a factor weight allocation algorithm (parameters: equipment type, process attribute, historical process performance), multi-objective factor weights are assigned to the parameters of each working condition sub-module, generating parameter expressions with multi-objective attributes, and weight normalization processing is performed on each objective factor.

[0111] A dynamic correlation feature extraction algorithm (parameters: time series data window, factor cross-correlation coefficient) is used to analyze the real-time correlation strength between parameters of each working condition sub-module and target indicators. Highly correlated factors are selected and included in the main factor mapping structure, while weakly correlated factors are set as auxiliary levels.

[0112] By applying a set of multi-objective factor action functions, the relationship between each operating condition parameter and the target index is expressed one-to-one using formulas. For example, the following linear model is used for the load-energy consumption relationship:

[0113] C i，j (t)=α i，j L i (t)+β i，j +∈ i，j (t)

[0114] Among them, C i，j (t) represents the value of the j-th target factor of the i-th device at time t, L i (t) represents the core operating condition variable (e.g., load) of the i-th device, α i，j With β i，j The factor relationship coefficient, ∈ i,j (t) represents the noise correction term.

[0115] For higher-order or nonlinear relationships, further modeling can be achieved using polynomial or neural network mapping functions as follows:

[0116]

[0117] in, X is a nonlinear mapping function. i (t) is the vector of all operating conditions of the i-th device.

[0118] A multi-objective dynamic calibration and optimization mechanism is adopted, which combines historical label data with real-time feedback to periodically update the factor mapping relationship parameters, ensuring that the mapping relationship can dynamically reflect the actual working conditions of the production line.

[0119] Through this chain-like mapping process, a one-to-one correspondence and dynamically adjustable mapping network is established between the description parameters of the working condition sub-module and the multi-objective scheduling factors. This enables the digital twin modeling structure to express the multi-objective state of the physical production line in a refined manner, thereby improving the accuracy of subsequent virtual-real synchronization and multi-objective scheduling optimization.

[0120] For example, in a copper clad laminate production line, the operating status submodule of the press equipment includes: instantaneous load current L press (t), Current production rate P press (t) and the health maintenance index H press (t). For the energy consumption target, α is set using the above linear mapping model. press,energy =1.2, β press，energy =0.5. The mapping formula between the operating condition submodule and the energy consumption target factor is:

[0121] C press,energy (t)=1.2×L press (t)+0.5

[0122] Assuming instantaneous detection L press If (t) = 10, then the current value of the energy consumption target factor is:

[0123] C press,energy (t) = 1.2 × 10 + 0.5 = 12.5

[0124] For the comprehensive objective of production capacity and health, an interpolation-type nonlinear function model is used. Historical data regression shows that the best fit is Q. press，pref (t)=[P press (t) / H press (t)] 0.8 If P press (t)=22,H press If (t) = 0.91, then the comprehensive objective factor is:

[0125] Q press，perf (t)=(22 / 0.91) 0.8 ≈16.57

[0126] When multiple objective factors are updated dynamically in real time, compared with historical feedback, if the fitting residual e fit If e < 1, the mapping relationship remains unchanged. fitIf the value is greater than or equal to I, then the self-correction of the mapping function parameters is initiated. This process realizes dynamic and high-precision mapping from the equipment operating condition submodule to multiple target indicators, enabling downstream collaborative scheduling optimization to have sufficient data support and state differentiation capabilities.

[0127] S3.3 describes the working condition submodule based on multi-objective factor mapping, and uses time-series modeling algorithms to model the dynamic behavior of each equipment type and process link, outputting time-series working condition modeling parameters that can reflect the characteristics of real-time production process changes, thereby improving the dynamic response capability of the digital twin structure.

[0128] Using the parameters describing the working condition submodule after multi-objective factor mapping as input, and for the dynamic working condition sampling sequences of each equipment type and process link, a combined time series modeling algorithm (including multivariate autoregressive model, recurrent neural network model RNN / LSTM, time series clustering algorithm, etc., with parameter settings depending on the number of working condition variables and sampling frequency) is adopted to realize the time dynamic behavior modeling of key working condition variables in the production process.

[0129] By using the multivariate autoregressive (VAR) modeling method (parameter: p-order lag length, automatically selected according to the AIC criterion), a joint dynamic prediction relationship is established for the time series variables such as load, energy consumption, capacity, and health of each equipment and process. The autocorrelation and cross-correlation dynamic characteristic coefficient matrix between the operating condition variables is obtained, and the complete set of equipment-level time series state parameters is output.

[0130] Furthermore, deep time series models such as recurrent neural networks (RNN) and long short-term memory networks (LSTM) are applied (parameters: input window length, number of hidden layer units, learning rate, etc.) to input the standardized multi-objective working condition variable sequence into the network, train the working condition dynamic behavior prediction model, extract complex cross-time series nonlinear relationship features, and output process-level working condition change trend and prediction confidence index.

[0131] Dynamic Time Warping (DTW) and temporal clustering algorithms (parameters: distance metric type, number of clusters) are used to perform cluster analysis on the historical operating conditions of different equipment or processes, identify typical operating condition evolution patterns, and construct representative parameters for patterns, which are used to improve model generalization and capture abnormal behavior.

[0132] Based on the time-series state characteristics and parameters obtained above, principal component analysis (PCA) is used (parameter: the number of principal components accounts for more than 95% of the cumulative variance) to reduce dimensionality and extract key time-series influencing factors, generating a representative time-series operating condition modeling parameter set.

[0133] Through a dynamic feedback calibration mechanism (parameters: online residual threshold, adaptive regression window), the difference between real-time operating data and model prediction output is analyzed. If the model prediction residual is found to exceed the set threshold, the model parameters are automatically triggered for adaptive correction, thereby improving the response speed and accuracy of the modeling structure to the dynamic behavior of the production line.

[0134] Through the above chain-derived combined temporal modeling method, the condition state sub-module description parameters after factor mapping are transformed into dynamic temporal condition modeling parameters that can fully reflect the evolution law of each piece of equipment, each process, the entire production line condition, and the change characteristics of multiple target indicators, thereby realizing the high-resolution dynamic response of the digital twin modeling structure to real-time production line condition changes.

[0135] For example, for a copper-clad laminate press, the time-series sampling period of the operating condition submodule is set to 1 second. Load current, energy consumption, production rate, and health index data are collected over 30 minutes, forming 1800 sets of multi-objective time-series sequences. A VAR model with p=3 is used to jointly model the four-dimensional variables. The optimal lag order is selected using the AIC criterion, and the autoregressive and cross-regressive coefficients of each variable are output, revealing the coupling relationship between short-cycle load fluctuations and energy consumption and health. Further training of the LSTM model is then performed, with the input window length set to 60 (1-minute sliding window), and the output... For one-time production rate and health prediction, the number of hidden layer units is set to 128, and the average prediction confidence reaches over 97%. The DTW method is used to cluster all 30-minute operating condition sequences, with the number of categories set to 3. The parameters of three operating condition evolution modes—typical impact load, smooth operation, and intermittent failure—are output. PCA is used to select the top two principal components with a cumulative contribution rate of 98% from all time series features. After feature dimensionality reduction, the dynamic parameter base coordinates of the digital twin are fed back. The average residual between real-time operating condition data and model prediction output is less than 0.8 (below the threshold of 1), and no correction is required.

[0136] When the health index of the machine drops suddenly after maintenance, causing the residual to be greater than 1, the adaptive correction mechanism of the model parameters is automatically activated. The new model iteration is completed within 2 minutes, and the residual is restored to the low value range. Finally, the high-dimensional time-series operating condition modeling parameters that fully cover the dynamic operating condition changes of the equipment are output, providing a high-timeliness and high-accuracy data foundation for the subsequent virtual-real synchronization and collaborative control algorithm calls of the digital twin.

[0137] S3.4 integrates the above-mentioned time-series operating condition modeling parameters and operating condition status sub-module structure, and adopts a distributed virtual-real mapping mechanism to synchronize the operating condition data of each level of the physical production line to the digital twin structure in real time, thereby obtaining a dynamic data mirror with local process details and global collaborative relationships.

[0138] Using multi-objective time-series working condition modeling parameters and working condition state sub-module structure as input conditions, it covers all equipment-level and process-level data nodes that have been normalized and feature-labeled.

[0139] A distributed virtual-physical mapping mechanism (parameters: node hierarchical topology -- equipment / process / production line, data synchronization cycle, network message collaboration protocol) is adopted to realize real-time data synchronization between each process and equipment node of the physical production line and its corresponding working condition module in its digital twin.

[0140] Furthermore, through a data timing alignment algorithm (parameters: multi-level timestamp, event triggering window, packet loss retransmission mechanism), fine-grained time calibration and synchronous error correction are performed on the dynamic working condition data sequence collected across devices and processes to ensure that the status parameters of the physical end and the digital twin end are consistent and aligned throughout the entire link, and a dynamic timing mapping relationship table with high real-time performance is generated.

[0141] Furthermore, by applying message queues and buffering mechanisms (parameters: node message queue length, aggregation batch window), raw dynamic operating condition data packets from multiple devices and across stages are grouped, cached, and aggregated for admission screening. This enables stable transmission of high-concurrency, low-latency data streams and automatically identifies and prioritizes the recovery of abnormal or delayed data packets, thereby improving the integrity rate of virtual-real mapping.

[0142] Furthermore, based on the data hierarchical aggregation strategy (parameters: equipment level, process level, production line level mapping hierarchy), the synchronized data is assembled according to the structured hierarchy to form a multi-level process parameter mirror chain. For each level of data, the aggregation operator is used to extract the core dynamic indicators, so as to realize the high-fidelity synchronization of local process details and the time-series linkage mapping of global collaborative relationships.

[0143] Finally, a distributed mapping error monitoring mechanism (parameters: inter-node residual threshold, synchronization delay timeout threshold) is adopted to perform online residual analysis on the entire virtual-physical mapping process. The mechanism performs real-time monitoring and correction of errors in the subdivided indicators of physical working condition data and digital twin working condition status, and outputs a highly consistent dynamic data mirror to support subsequent indicator calibration and collaborative scheduling optimization.

[0144] Through a distributed hierarchical virtual-real mapping mechanism, dynamic sampling data of equipment and process conditions at all levels are synchronously mapped with simulation state parameters in the digital twin in a full-link, highly consistent manner. This results in a dynamic data mirror with local process detail accuracy and global state coordination capabilities, providing fundamental data support for flexible production line scheduling and multi-objective adaptive optimization.

[0145] For example, in the actual deployment of the copper clad laminate production line, the press and baking oven are configured as distributed virtual-physical mapping nodes. The press node samples the load current L_press(t) and production rate P_press(t) in real time with a sampling period of 1 second. Time alignment uses millisecond-level event timestamps. The baking oven node collects the inlet temperature T_oven_in(t) and energy consumption C_oven(t) with a sampling period of 2 seconds. An independent message buffer queue is used for data aggregation. The system sets the data synchronization period to 5 seconds and uses the MQTT protocol to ensure reliable distributed communication. The mapping hierarchy includes equipment level and process level. In each synchronization cycle, the data packets of all equipment nodes are aggregated through the message queue. After special delay packets are marked, they are automatically completed in sequence. The hierarchical aggregation function outputs the 1-minute average of P_press and L_press for the press submodule to obtain the equipment-level dynamic indicators. Finally, after distributed mapping error monitoring, the residual of the same data window between the physical and twin ends is less than 0.05, achieving a data mirror consistency of more than 97%.

[0146] The implementation of the above mechanism ensures the real-time, complete, and highly consistent synchronous output of multi-source dynamic operating condition data from the production line to the digital twin structure, providing a solid state-aware foundation for subsequent multi-objective scheduling weight generation and intelligent optimization scheduling strategies.

[0147] Based on the existing multi-level digital twin structure, S3.5 executes an index consistency calibration algorithm to align the physical production line operating data with the multi-objective indicators predicted by the digital twin, outputting index mapping consistency parameters to ensure the accuracy of virtual-real mapping and provide highly reliable modeling support for subsequent intelligent scheduling perception and collaborative optimization.

[0148] like Figures 1-2 As shown in step S4, the multi-level digital twin modeling structure is input into the autonomous scheduling intelligent agent. The autonomous scheduling intelligent agent continuously senses and analyzes the load level, equipment health status and energy consumption parameters of the local and adjacent units to generate autonomous scheduling sensing characteristics under specific operating conditions.

[0149] Further, in step S4,

[0150] S4.1 encapsulates the instantiated process state parameters in the constructed multi-level digital twin modeling structure to achieve efficient data docking and parsing of the autonomous scheduling intelligent agent, providing standardized input for the subsequent autonomous perception mechanism.

[0151] Based on the instantiated process state parameters output by the preceding multi-level digital twin modeling structure, it covers dynamic operating condition variables and their multi-objective index mapping results at all levels from equipment level, process level to production line level.

[0152] By adopting a standardized interface encapsulation method (parameters: data protocol type, field mapping table, timing synchronization mechanism), the format of instantiated process status parameters is standardized and the structure is aligned, ensuring that data at each level of equipment, process and production line can be read at high speed by the autonomous scheduling intelligent agent.

[0153] Furthermore, through data field serialization and hierarchical tag injection mechanism (parameters: field sequence, type tag, process node identifier), a general attribute tag is packaged for each instantiated process parameter to realize the automatic identification of operating parameters and context analysis function, and the data priority field is dynamically marked according to the proxy requirements.

[0154] By employing a data access interface adaptation algorithm (parameters: API abstraction level, access caching mechanism, maximum fault tolerance latency), and deploying RESTful API or industrial OPC UA protocol interfaces, the autonomous scheduling intelligent agent can ensure batch calls, asynchronous parsing, and high-concurrency reception of instantiated process status parameters, thereby achieving generalized access to data streams and protection against packet loss.

[0155] Furthermore, through protocol consistency verification and real-time reception verification algorithms (parameters: protocol version number, data packet timestamp, tolerance threshold), the data consistency and timing synchronization of the autonomous scheduling intelligent agent receiver and the digital twin output are automatically compared, and abnormal or overlapping data packets are screened and corrected in real time to ensure the standardization and integrity of the input data.

[0156] Finally, an interface status report generation algorithm (parameters: data loss rate, format anomaly ratio, latency statistics) is adopted to periodically output health status reports of data connection, providing underlying support for the autonomous scheduling intelligent agent self-diagnosis and operation and maintenance optimization module.

[0157] Through interface encapsulation and data adaptation algorithms, the instantiated process state parameters in the multi-level digital twin modeling structure are transformed into standardized data inputs that are formatted, automatically tagged, and seamlessly integrated, thereby achieving efficient, stable, and low-latency access to operational data by the autonomous scheduling intelligent agent.

[0158] For example, for the press equipment node in the copper clad laminate production line, the instantiated process status parameters include: instantaneous load current (L_press), remaining health index (H_press), unit cycle energy consumption (E_press), and current capacity rate (P_press), which are output as a structured JSON data stream and split into equipment-level (press) and process-level (pressing) tag fields;

[0159] The interface encapsulation method adopts the OPC UA protocol, and the field mapping table is defined as {“L_press”:float,“H_press”:float,“E_press”:float,“P_press”:float}, with the tags “NodeType=Device” and “Stage=Compression” appended. The data packet period is 500ms. The autonomous scheduling intelligent agent calls the API in batches, capturing two data packets per second. A caching mechanism is used to ensure that the maximum latency is less than 300ms. Through protocol consistency verification, occasional field spelling anomalies (such as “H_press” being mistakenly filled as “Hpre”) are detected and automatically corrected before being input into the formal parameter stream. The process health report shows that the data packet loss rate is less than 0.05% and the format anomaly rate is 0.03% within 1 hour. There is no latency exceeding the standard throughout the entire connection, and the autonomous agent access success rate reaches 99.9%. Finally, standardized, labeled, and error-free device operating parameters are output, providing a high-quality input guarantee for the autonomous scheduling perception link to achieve high-frequency data perception and intelligent scheduling empowerment.

[0160] Autonomous scheduling intelligent agents refer to an autonomous decision-making system based on artificial intelligence technology. It achieves intelligent scheduling and optimization of resources in dynamic environments by integrating a closed-loop mechanism of perception, planning, execution and feedback. Its core capabilities include real-time perception of multi-source data (such as equipment status and task requirements), dynamic decision-making using AI models such as reinforcement learning or genetic algorithms, and distributed task allocation through a multi-agent collaborative architecture (such as master-slave agent). In industrial scenarios, such agents usually rely on digital twin technology to build virtual mappings and achieve millisecond-level response through cloud-edge-device collaborative computing. Typical applications include flexible manufacturing scheduling and multi-robot path planning. This concept is essentially an extension of "intelligent agent" technology in the field of scheduling. Although it is not listed as a proper noun in encyclopedias, it has been widely described in academic literature as a dedicated intelligent agent system with autonomy, adaptability and continuous learning capabilities.

[0161] S4.2 instantiates process state parameters based on a multi-level digital twin modeling structure, executes an autonomous scheduling perception activation process, and uses an autonomous scheduling intelligent agent to collect and monitor load level data covering local process units in real time to dynamically capture production load change scenarios.

[0162] Based on the standardized and encapsulated digital twin modeling structure with multi-level labels, process state parameters are instantiated, and high-frequency dynamic operating condition perception is implemented for local process units of autonomous scheduling intelligent agents.

[0163] An adaptive load data real-time acquisition method (parameters: sampling period, equipment type, data buffer length) is adopted. Through an autonomous scheduling intelligent agent, periodic or event-triggered sampling operations are performed on the key load sensors configured in the local process unit to achieve high-density capture of the load raw data stream.

[0164] Furthermore, a real-time data denoising and signal smoothing algorithm (parameters: filter window, smoothing type, outlier discrimination threshold) is used to filter out industrial noise and smooth small fluctuations in the acquired load signal, remove high-frequency interference and occasional anomalies during the acquisition process, and output high-confidence load time-series data of the physical sensing layer.

[0165] Furthermore, a dynamic load fluctuation analysis algorithm (parameters: sliding window length, fluctuation intensity criterion, threshold sliding mechanism) is adopted to calculate the short-cycle mean, variance, kurtosis and other statistical characteristics of the load level of the local process unit, and to extract fluctuation characteristic indicators representing various working condition change modes such as stable load, sudden and violent fluctuation.

[0166] A load change scenario recognition algorithm (parameters: multivariate time series feature set, number of cluster categories, scenario label dictionary) is adopted. The load fluctuation features within different time windows are input into the scenario clustering and label discrimination model to realize the automatic judgment of typical working condition change scenarios such as stable production, load shock, load overload, and standby, and to label the results with scenario category labels.

[0167] Furthermore, by using a dynamic load anomaly capture algorithm (parameters: prediction model type, anomaly threshold, real-time residual), a residual prediction relationship is established based on historical load models and actual sampled values. This enables timely alarms and type attribution for sudden load anomalies, improving the adaptive response capability of process units to extreme production fluctuations.

[0168] Through the aforementioned multi-level data acquisition and dynamic feature extraction, automatic scene recognition and anomaly capture algorithms, the load level change scenario of local process units at any given moment is dynamically captured, forming a traceable and quantifiable autonomous scheduling perception load data stream, realizing high-resolution dynamic perception of production status changes.

[0169] For example, in the press unit of the copper clad laminate production line, the autonomous scheduling intelligent agent is configured with a high-precision load current sensor with a cycle of 1 second. The load acquisition buffer length is set to 1800 (corresponding to 30 minutes of data). After acquisition, a sliding window of 10 seconds and a third-order median filter are used to denoise the signal. The outlier discrimination threshold is the mean ± 3 times the standard deviation. A fluctuation analysis algorithm is applied to extract the mean, variance, kurtosis, and other features of each 1-minute window. These features are input into the scene recognition model, and the number of clusters is set to 4. The operating data is automatically classified into "stable production", "impact pressurization", "load overload", and "...". In the "standby" scenario, dynamic anomaly capture is triggered when there are strong load changes. Based on the prediction results of the sliding window VAR model, if the residual between the observed load and the predicted value is greater than 5% of the rated value, an overload alarm is initiated. This process realizes real-time monitoring of the compressor load fluctuation and automatic identification of multiple scenarios. During the test period, the scenario discrimination accuracy reached 98%, and the anomaly capture response time was less than 2 seconds. The output autonomous scheduling perception load data stream is linked with the production line anomaly early warning system, which effectively improves the time domain response of scenario prediction, resource pre-allocation and flexible switching, and realizes high-precision dynamic availability of scheduling perception under critical load scenarios.

[0170] S4.3 utilizes an autonomous scheduling intelligent agent to instantiate process status parameters of the input multi-level digital twin modeling structure to determine equipment health status. Through embedded health assessment algorithms, it models and predicts trends of key health indicators such as equipment aging and failure probability, thereby improving the equipment health status dimension in the autonomous scheduling perception features.

[0171] The input consists of instantiated process state parameters in a multi-level digital twin modeling structure, covering dynamic operating condition variables at the equipment and process levels, including status data such as load, current, voltage, temperature rise, and cumulative running time during the equipment's operating cycle, as well as synchronized historical maintenance records.

[0172] An embedded health assessment algorithm (parameters: equipment type, list of key operating condition variables, age and operating cycle thresholds, historical fault labels) is adopted. The autonomous scheduling intelligent agent automatically reads and parses the status parameters of the target equipment to realize the original health model of the equipment's operating status.

[0173] Furthermore, the weighted exponential moving average (EWMA) method (parameters: operating condition variable time series window, weighted decay coefficient) is used to smooth the trend of operating condition variables (such as temperature rise, current fluctuation, short-term load), eliminate occasional anomalies within the calculation period, and output a stable basic sequence of equipment health.

[0174] Using an equipment aging modeling algorithm, based on the theoretical lifespan model and historical operating behavior (parameters: cumulative uptime, load utilization, number of failures), various time-series variables are input into a health estimation formula based on physical mechanisms, and a preliminary health score H is output.Sinit (t):

[0175]

[0176] Among them, T op (t) represents the cumulative runtime up to time t, where T is the runtime of the previous time. max For the rated life of the equipment, N fd (t) represents the cumulative number of failures, N fa To allow for a maximum number of faults, L avg (t) represents the average load intensity, L max The nominal maximum load of the equipment is denoted as α, β, and γ, which are the weights of each influencing factor.

[0177] Furthermore, a Bayesian estimation of the potential failure probability of key components is performed using a probabilistic statistical modeling algorithm based on historical operating conditions and anomaly labels (parameters: historical anomaly operating condition sequence, prior probability distribution, and posterior failure window). The following failure probability prediction model is established:

[0178]

[0179] in, This represents the number of fault labels counted within the nearest Δt window. denoted as the total number of samples, and k0, k1 as the anti-zeroing smoothing constants.

[0180] Furthermore, a multivariate time-series health prediction algorithm combined with an LSTM recurrent neural network (parameters: input window length, number of network layers, number of hidden units) is adopted. The synchronized process state time-series sequence is input to dynamically predict the future trend range of equipment health and output the health prediction trajectory HS. pred (t+τ).

[0181] Preliminary Health Score (HS) init (t), Real-time failure probability P fault (t) and health trend prediction HS pred (t+τ) is normalized and weighted to generate a multi-component autonomous scheduling perceived health index H. final (t):

[0182] H final (t)=w1·HS init (t)+w2·[1-P fault (t)]+w3·HS pred (t+τ)

[0183] Among them, w1, w2, and w3 are the fusion weights, which are adaptively adjusted based on historical backtesting data.

[0184] Through autonomous scheduling intelligent agents, Hfinal (t) Store and synchronize to the scheduling awareness feature set to realize multi-level, dynamic and predictable quantification of device health status, providing a highly reliable criterion for subsequent cross-device scheduling weight allocation and resource collaborative optimization.

[0185] For example, for the laminating machine equipment in a copper clad laminate production line, the input operating parameters are set as follows: current cumulative running time T op (t) = 27840 hours, theoretical lifespan T max = 35,000 hours, cumulative failures N in the past year fd (t) = 2, with an upper limit of N. fa =5, average load L over the past month avg (t) = 7.1kW, maximum rated L max =10kW, parameter weight configuration α=0.6, β=0.2, γ=0.2, substitute into the formula:

[0186] HS init (t)=1-[0.0×(27840 / 35000)+0.2X(2 / 5)+0.2×(7.1 / 10)]

[0187] =1 - (0.477 + 0.08 + 0.142) = 0.301

[0188] Fault probability statistics sampled over the past 6 months:

[0189] According to the formula:

[0190]

[0191] Health prediction HS approximately 3 months later using LSTM pred (t+τ)=0.28.

[0192] Let the fusion weights be w1 = 0.5, w2 = 0.2, and w3 = 0.3, then the final health perception will be:

[0193] H final (t)=0.5×0.301+0.2×(1-0.00185)+0.3×0.28≈0.1505+0.1996+0.084

[0194] =0.4341

[0195] The current health normalization index of this press is 0.43, which is lower than the set health threshold of 0.6. This is fed back to the autonomous scheduling perception feature set to realize the intelligent generation of flexible workstation scheduling and maintenance suggestions. In the production line scheduling simulation backtest, the health model improved the early warning rate of predicted faults by 37%.

[0196] S4.4 employs an autonomous scheduling intelligent agent to perform real-time energy efficiency analysis and calculation on energy consumption parameters from a multi-level digital twin modeling structure. Based on the correlation between process load and equipment health status, it outputs energy consumption levels and energy efficiency risk indicators from the autonomous scheduling perception features.

[0197] S4.5 combines the autonomous scheduling perception features obtained by the autonomous scheduling intelligent agent from the local process unit with the autonomous scheduling perception features exchanged between adjacent units through a distributed message mechanism to form a joint autonomous scheduling perception feature covering both local and adjacent units, laying a multi-dimensional decision-making foundation for dynamically optimizing cross-unit resource scheduling.

[0198] S4.6, based on the joint features of autonomous scheduling perception, summarizes the key correlation indicators involving load level, equipment health status and energy consumption parameters under specific operating conditions, and generates the final autonomous scheduling perception feature set through an automatic feature importance evaluation algorithm, which serves as the core input for subsequent target weight allocation and multi-objective scheduling game.

[0199] like Figures 1-2 As shown, S5 calculates the global target weight vector of the current production line based on the perception characteristics of the autonomous scheduling intelligent agent, combined with order priority, real-time electricity price changes and equipment aging trends, to achieve periodic adaptive generation of multi-objective optimization weights.

[0200] Further, in step S5,

[0201] S5.1 performs data preprocessing on autonomous scheduling sensing features, including time-series alignment and discretization, to solve the problem of feature information distortion caused by inconsistent sampling frequencies of autonomous scheduling sensing features from different devices, and obtain a time-series dataset of autonomous scheduling sensing features with a unified structure.

[0202] S5.2 uses an autonomous scheduling-aware feature time-series dataset and a process rule reasoning algorithm to dynamically calculate the order priority of each order within a time window, thereby obtaining a dynamic order priority weight vector.

[0203] S5.3 integrates autonomous scheduling sensing feature time-series data with real-time energy consumption information, uses time-period weighted operators to extract energy consumption curves for each process in the production line, and introduces a real-time electricity price acquisition interface to calibrate the energy consumption target weight in real time based on the energy consumption curves.

[0204] S5.4 uses time-series data of autonomous scheduling perception features and predictive modeling algorithms (such as the equipment remaining life prediction RUL model) to quantify the current aging degree of each device, generate equipment aging trend indicators, and calibrate the target weight components of equipment aging accordingly.

[0205] S5.5 performs normalized weighted fusion processing on the dynamic order priority weight vector, energy consumption target weight, and equipment aging target weight components, and calls the multi-objective weight adaptive adjustment algorithm to periodically generate a global target weight vector as input parameters for subsequent swarm intelligence optimization algorithms.

[0206] like Figures 1-2 As shown in step S6, the global objective weight vector is input into the evolutionary game and swarm intelligence optimization algorithm to perform resource scheduling collaborative game and scheduling scheme evolution among multiple intelligent agents, generating the optimal solution set for cross-device scheduling under the current multi-objective constraints.

[0207] Further, in step S6,

[0208] S6.1 takes the input global target weight vector and performs weight normalization transformation on multi-objective constraint indicators such as process priority, equipment energy consumption index, equipment health parameters, and capacity completion rate based on the multi-objective weighted mapping mechanism to obtain a multi-objective weighted evaluation matrix, which provides a weight basis for subsequent resource allocation modeling.

[0209] S6.2 takes the normalized multi-objective weighted evaluation matrix as input and adopts a distributed knowledge sharing mechanism to fuse multi-objective constraint information with the local state characteristics of each autonomous scheduling intelligent agent, thereby realizing local state synchronization and target collaboration information update across intelligent agents and providing a complete state space for the collaborative game algorithm.

[0210] Based on the synchronized local state space and global objective constraint information, S6.3 initializes the resource allocation strategy encoding for each autonomous scheduling intelligent agent and establishes a resource scheduling collaborative game model using a game strategy generation mechanism to simulate the interaction relationship of scheduling schemes among multiple intelligent agents.

[0211] S6.4 implements a multi-round evolutionary game for the resource scheduling collaborative game model. Through game state transition rules and multi-objective fitness functions, it dynamically adjusts the resource allocation code of the autonomous scheduling intelligent agent and selects a group of collaborative scheduling strategies with global advantages across multiple generations.

[0212] S6.5 employs swarm intelligence optimization algorithms (such as multi-objective particle swarm optimization, genetic evolution, or ant colony optimization) to use the advantageous scheduling strategy group output by collaborative game as the population base, and performs multi-objective adaptive parameter mutation and population evolution iteration to dynamically find the optimal resource scheduling solution set across equipment and processes.

[0213] S6.6 uses the optimal resource scheduling solution set output from the evolutionary game and swarm intelligence optimization stages as evaluation criteria, and conducts multiple subjective and objective comprehensive evaluations. Finally, it selects and outputs the optimal cross-equipment scheduling solution set that satisfies the multi-objective constraints, as the basis for production line scheduling decisions.

[0214] like Figures 1-2 As shown in Figure S7, based on the generated cross-device scheduling optimal solution set, multiple schemes are simulated and verified in parallel on the digital twin production line. The final scheduling control parameter set that achieves the optimal overall goal and meets the requirements for flexible process switching is selected first.

[0215] Further, in step S7,

[0216] S7.1 performs structured loading of the optimal solution set for cross-device scheduling and introduces it into the digital twin production line simulation platform to ensure that each scheduling control parameter group fully maps its resource allocation relationship and scheduling logic in the virtual environment, providing accurate input for the subsequent parallel simulation process.

[0217] S7.2 uses a digital twin simulation engine to conduct parallel simulations of multiple scenarios based on the structured loading of the scheduling and control parameter group. The simulation process covers key scenarios such as equipment load dynamics, process switching response, and production line energy consumption fluctuations, so as to systematically reproduce the real production line operation mechanism.

[0218] S7.3 uses a multi-objective performance evaluation algorithm to calculate the response output of each group of scheduling control parameters on multi-dimensional indicators such as production line capacity achievement rate, energy efficiency utilization rate and equipment health, in order to normalize the indicators and construct a feature evaluation matrix.

[0219] Based on the feature evaluation matrix, S7.4 further adopts a flexible switching adaptability judgment method to test the response delay, resource scheduling flexibility and process adaptability of the scheduling scheme corresponding to each group of control parameters under different process switching scenarios with specific indicators, and selects the parameter set that supports dynamic process switching.

[0220] S7.5 applies a global optimal target weight comprehensive evaluation algorithm to select the final scheduling control parameter group that has the highest overall target achievement score and can achieve the best balance under multi-target conflict. The final scheduling control parameter group is then output in a standardized professional terminology for efficient execution by downstream production lines.

[0221] like Figures 1-2 As shown, S8 sends the final scheduling control parameter set to the physical production line for scheduling execution and records the actual production response status online. Through the digital twin modeling structure and feedback data, the evolution model is continuously calibrated to achieve adaptive parameter correction.

[0222] Further, in step S8,

[0223] S8.1 serializes and encapsulates the final generated scheduling control parameter group, and sends the scheduling control parameter group to the distributed device controller of the physical production line based on the industrial control communication protocol, so as to realize the synchronous deployment of cross-device scheduling instructions.

[0224] S8.2, based on the scheduling control parameter group received by the distributed device controller, controls each production line device to execute production tasks according to the predetermined scheduling priority, timing and energy consumption constraints, monitors the real-time execution status of scheduling instructions, and generates device response data streams.

[0225] S8.3 performs cascaded acquisition and real-time multi-dimensional archiving of equipment response data streams, obtaining key production response characteristics such as equipment operating status, energy consumption curves, capacity achievement, and load fluctuations, providing high-fidelity data support for the subsequent calibration of digital twin modeling structures.

[0226] S8.4 performs state alignment and error quantification processing on the collected production response features and the existing digital twin modeling structure. Based on the data alignment error analysis algorithm, it outputs a dynamic accuracy correction factor as the basic parameter for adjusting the evolution model.

[0227] S8.5 uses a dynamic accuracy correction factor to adaptively correct the mapping relationship between decision parameters and targets in the evolutionary scheduling model, optimizes the behavior mapping and feedback mechanism of the digital twin in real time, and corrects the virtual-real deviation of the model from the actual production line conditions.

[0228] S8.6 utilizes the modified evolutionary scheduling model to re-evaluate the execution effect of the current scheduling strategy, generate a performance fitness feedback report, and provide performance traceability and data-driven basis for continuous iteration and the generation of future scheduling schemes, thereby realizing closed-loop adaptive optimization of the evolutionary model.

[0229] like Figures 1-2 As shown, S9 monitors key process indicators after scheduling execution, including equipment health, energy consumption curve and capacity achievement rate, to determine whether there are abnormal fluctuations. If performance is detected to deviate from the set threshold, the scheduling parameter adaptive adjustment mechanism is automatically triggered to achieve closed-loop optimization under abnormal conditions.

[0230] Further, in step S9,

[0231] Based on the digital twin modeling structure, S9.1 performs real-time data collection and historical trajectory synchronization mapping of key process indicators such as equipment health, energy consumption curves and capacity achievement rate of physical production lines to obtain multi-dimensional process monitoring data covering the entire scheduling and execution process.

[0232] S9.2 utilizes a high-frequency dynamic data detection algorithm to identify abnormal fluctuations in multi-dimensional process monitoring data. This includes employing a multi-objective threshold determination method and a time-series trend analysis model to automatically identify and classify abnormal states such as significant declines in equipment health, unexpected jumps in energy consumption curves, and production capacity achievement rates falling below preset standards.

[0233] S9.3 Based on the anomaly discrimination results, extract the abnormal fluctuation feature parameters, and associate these feature parameters with the response behavior of the historical digital twin model to analyze the inherent process causal mechanism of the anomaly, so as to generate an anomaly state identification report based on the relationship between the cause and effect.

[0234] S9.4 inputs the generated abnormal state identification report into the adaptive parameter adjustment algorithm, adopts self-learning rules under multi-objective optimization (such as reinforcement learning type weight adjustment, parameter gradient backpropagation, etc.), and automatically generates scheduling parameter correction suggestions that are consistent with the abnormal features.

[0235] S9.5 feeds back the scheduling parameter correction suggestions output by the adaptive parameter adjustment algorithm to the digital twin modeling structure, and drives the evolutionary scheduling model to update parameters, thereby closing the loop to correct the existing scheduling strategy and realizing the adaptive optimization process under abnormal conditions.

[0236] S9.6 uses a digital twin modeling structure with updated closed-loop parameters to further monitor the recovery trend of key process indicators after targeted adjustments in real time, verify the positive impact of adaptive optimization measures on equipment health, energy consumption curves and capacity achievement rate, so as to continuously improve the robustness and self-learning capability of the scheduling system.

[0237] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this application.

Claims

1. A copper-clad plate production line collaborative regulation method based on digital twinning, characterized in that, The application relates to a multi-level digital twin modeling method for a copper-clad plate production line. S1, real-time operation data of each device and process in the copper-clad plate production line is acquired to establish a data input basis for a multi-level digital twin; S2, the acquired real-time operation data is normalized, and label data sets are obtained by labeling the data features of the normalized data based on different device types and process attributes, so that the working condition variables of different devices and process differences are distinguished; S3, based on the normalized and labeled label data sets, a multi-level digital twin modeling structure is constructed for each device type and process link, specifically including: Based on the normalized and labeled label data sets as input conditions, the functional unit splitting processing is carried out on each device type and process link to generate the working condition state sub-module description parameters of each level of device and process; For the generated working condition state sub-module description parameters, a multi-objective factor mapping method is applied to build a one-to-one correspondence relationship between the multi-objective indexes of load, energy consumption, productivity and device health degree and the working condition state variables; Based on the working condition state sub-module structure description after the multi-objective factor mapping, the dynamic behavior of each device type and process link is modeled by using a time sequence modeling algorithm, and time sequence working condition modeling parameters reflecting the characteristics of real-time production process changes are outputted; Fusion of the above time sequence working condition modeling parameters and working condition state sub-module structure, a distributed virtual-real mapping mechanism is adopted to realize real-time synchronization of each level of working condition data of the physical production line to the digital twin structure, and a data dynamic mirror with local process details and global collaborative relationship is obtained; On the basis of the obtained multi-level digital twin structure, an index consistency calibration algorithm is executed to align the multi-objective indexes of the physical production line working condition data and the digital twin prediction output, and an index mapping consistency parameter is outputted; S4, the multi-level digital twin modeling structure is inputted into the autonomous scheduling intelligent agent, the load level, device health state and energy consumption parameters of the local and adjacent units are continuously sensed and analyzed by the autonomous scheduling intelligent agent to generate autonomous scheduling sensing features under specific working conditions, specifically including: The instantiated process state parameters in the constructed multi-level digital twin modeling structure are interface encapsulated; Based on the instantiated process state parameters of the multi-level digital twin modeling structure, an autonomous scheduling sensing activation process is executed, the load level data of the local process unit is collected and fluctuated in real time by the autonomous scheduling intelligent agent, and the production load change scene is dynamically captured; The device health degree is determined by using the autonomous scheduling intelligent agent to the instantiated process state parameters of the input multi-level digital twin modeling structure, and the key health indexes of device aging and fault probability are modeled and trend predicted by the embedded health evaluation algorithm; The energy consumption parameters from the multi-level digital twin modeling structure are analyzed in real time by using the autonomous scheduling intelligent agent, and the energy consumption level and energy efficiency risk indexes in the autonomous scheduling sensing features are outputted based on the correlation between process load and device health state. The autonomous scheduling perception features obtained from the local process unit are synthesized with the autonomous scheduling perception features exchanged between adjacent units through a distributed message mechanism to obtain autonomous scheduling perception joint features covering the local and adjacent units; Based on the autonomous scheduling perception joint features, key correlation indicators related to load level, equipment health status and energy consumption parameters under specific working conditions are summarized, and the final autonomous scheduling perception feature set is generated through an automatic feature importance evaluation algorithm; S5, based on the autonomous scheduling intelligent agent perception features, the global target weight vector of the current production line is calculated combined with the order priority, real-time electricity price change and equipment aging trend; S6, the global target weight vector is input into the evolutionary game and swarm intelligence optimization algorithm to perform resource scheduling collaborative game and scheduling scheme evolution among multiple intelligent agents to generate a set of optimal solutions for cross-device scheduling under current multi-objective constraints; S7, based on the generated set of optimal solutions for cross-device scheduling, multiple scheme parallel simulation verification is performed on the digital twin production line, and the final scheduling control parameter group that achieves the optimal total target and meets the process flexible switching demand is preferentially selected; S8, the final scheduling control parameter group is sent to the physical production line for scheduling execution, and the actual production response state is recorded online, and the evolutionary model is continuously calibrated through digital twin modeling structure and feedback data.

2. The copper-clad plate production line collaborative regulation method based on digital twinning according to claim 1, wherein The S8 further comprises: S9, monitor the key process indicators after scheduling execution, including equipment health degree, energy consumption curve and production capacity achievement rate, judge whether there is abnormal fluctuation, if the performance deviates from the set threshold, automatically trigger the scheduling parameter adaptive adjustment mechanism to realize closed-loop optimization under abnormal state.

3. The copper-clad plate production line collaborative regulation method based on digital twinning according to claim 1, characterized in that, Real-time operation data of each device and process in the copper-clad plate production line is obtained, and a multi-level intelligent sensor networking method is used to configure multi-modal intelligent sensors including current, voltage, temperature, vibration, pressure, power physical quantities and operating state parameters.

4. The copper-clad plate production line collaborative regulation method based on digital twinning according to claim 1, characterized in that, The step S2 specifically comprises: Based on the industrial data standardization criteria, the normalized process parameter set is generated to generate the standardized process parameter set of the device and process working condition; Based on the standardized process parameter set, the device and process attribute dictionary is constructed according to the production line device type and process attribute to form a unified device type label and process link label; The normalized process parameter set is assembled into a feature vector according to the corresponding device type label and process link label, and each data sample is assigned a corresponding multi-property label through feature splicing and attribute mapping to generate a structured feature labeled sample; The feature consistency verification algorithm is used on the structured feature labeled sample to verify the standardized consistency of the feature vector dimension and value range of different devices and processes in the series data, eliminate the distribution deviation caused by potential heterogeneous data, and generate a consistent feature labeled sample set; The consistent feature labeled sample set is subjected to process difference identification processing, and the key variable distribution across devices and processes is further screened based on the process fluctuation identification model to extract the working condition variable set for multi-objective scheduling optimization.

Images