A heat treatment furnace group energy consumption collaborative optimization and scheduling system

Abstract

The present application relates to a kind of heat treatment furnace group energy consumption collaborative optimization and scheduling system, specifically relates to the field of industrial internet of things heat treatment furnace group control, by process feasible domain pre-screening module, parameter space is evaluated using knowledge graph jointly confidence, output high potential process parameter feasible domain after pre-screening, lay high-quality foundation for subsequent optimization, simulation optimization module is based on heat treatment process digital twin model and embedded bayesian optimization algorithm of safety barrier function, in the feasible domain, find the optimal process parameters that meet all constraints by simulation optimization, virtual-real closed-loop execution and fine-tuning module is through real-time digital thread synchronization and online correction optimization, ensure that actual process path always remains in safety envelope execution, knowledge-model collaborative evolution module then uses the verification sample driving data-driven agent model of each batch and the confidence weight of associated rule in knowledge graph collaborative evolution, so that the optimization and execution capability of the whole scheme obtains sustained self-enhancement.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) control of heat treatment furnace groups, and more specifically, to a collaborative optimization and scheduling system for energy consumption of heat treatment furnace groups. Background Technology

[0002] In the field of heat treatment processing services, the precise setting of process parameters is the cornerstone for achieving the expected metallographic structure, hardness, wear resistance, and other key performance indicators of workpieces, while simultaneously ensuring the safety and economy of the production process. Faced with the continuous emergence of new materials, new components, and highly personalized non-standard heat treatment requirements from customers, companies often lack mature process specification databases that can be directly accessed. In actual production, operators usually perform operations based on relatively conservative process cards formed from historical experience. While this fixed set of parameters can basically guarantee the bottom line of quality and safety, it may not be the optimal solution that balances quality, safety, and energy consumption under the current equipment status and incoming material conditions. Especially for stable batch production, subtle optimization of process parameters can often bring significant energy savings and efficiency improvements. However, this optimization process is highly dependent on the personal experience of process experts and is carried out through costly, lengthy, and risky "trial and error" experiments. A large amount of valuable production process data, such as multi-dimensional process parameter time-series curves, real-time energy consumption flow, and final quality inspection results, have not been systematically linked, analyzed, and fed back under the existing technical architecture, forming "data silos" that cannot effectively drive the continuous iteration and optimization of process knowledge.

[0003] In existing technologies, the process parameters for heat treatment generally exhibit solidification or semi-solidification characteristics, and their optimization and adjustment heavily rely on human experience and judgment. Although distributed control systems or programmable logic controllers can accurately execute preset process curves at the automation control level, they themselves lack the ability to automatically find the optimal solution in a complex multidimensional constraint space. From a technical perspective, heat treatment process parameter optimization is a typical multi-objective, high-dimensional, nonlinear engineering optimization problem with complex constraints. Optimization objectives typically include at least the pursuit of minimum energy consumption and optimal comprehensive quality indicators, while constraints cover hard quality constraints ensuring acceptable metallographic structure and hard hard constraints such as furnace pressure, combustible gas concentration, and temperature exceeding limits. These objectives and constraints are mutually coupled, mutually restraining, and may even conflict. Traditional experience-based trial-and-error methods or solutions based on classical optimization algorithms struggle to efficiently and reliably find the global or near-global optimal solution in this high-dimensional, non-convex solution space. Meanwhile, the microscopic differences between batches of incoming workpieces, the slow drift in the performance of heat treatment equipment, and the fluctuations in production environment all require process parameters to have a certain degree of adaptive adjustment capability, which existing fixed parameter patterns cannot meet. Therefore, current technology lacks an intelligent method and system that can deeply integrate process mechanism knowledge, historical production data, and real-time operating status, and can automatically and efficiently search for and verify the optimal or better combination of process parameters while strictly meeting all quality and safety constraints. This has become a key bottleneck restricting the heat treatment processing industry from further improving energy efficiency, ensuring quality stability, and realizing process digitalization. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a collaborative optimization and scheduling system for energy consumption of heat treatment furnace groups. It solves the problems mentioned in the background art by using a process feasible domain pre-screening module, a simulation optimization module, a virtual-real closed-loop execution and fine-tuning module, and a knowledge-model collaborative evolution module.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: a process feasible domain pre-screening module, a simulation optimization module, a virtual-real closed-loop execution and fine-tuning module, and a knowledge-model co-evolution module connected in sequence, wherein; The process feasibility domain pre-screening module responds to process optimization requests for target workpiece materials by acquiring and integrating multi-source data, including historical process documents, quality inspection reports, and production process time-series data related to the target workpiece materials. Based on the multi-source data, a knowledge graph containing entities, attributes, and their inter-attribute rules is constructed. This knowledge graph is used to characterize the relationships between five dimensions: materials, process parameters, process status, quality indicators, and energy consumption. Based on the explicit rule constraints extracted from the knowledge graph and the parameter association strength function obtained through joint distribution learning of the acquired multi-source data, the module performs a joint confidence assessment of the target process parameter space and outputs a pre-screened high-potential process parameter feasibility domain with a confidence level higher than a preset threshold. Simulation optimization module: Receives the pre-screened feasible domain of high-potential process parameters, constructs an optimization simulation environment based on a digital twin model of the heat treatment process that integrates physicochemical mechanisms and data-driven proxy models, and adopts a Bayesian optimization algorithm with embedded safety barrier functions. In the optimization simulation environment, with the minimum energy consumption as the optimization objective, it performs iterative simulation optimization within the feasible domain of high-potential process parameters, and outputs the optimal process parameters obtained through simulation optimization that satisfy all preset quality constraints and safety constraints. Virtual-real closed-loop execution and fine-tuning module: The optimal process parameters are sent to the industrial control system for execution. During the execution process, the actual furnace sensor data is synchronized to the digital twin model of the heat treatment process through real-time digital thread to perform parallel simulation prediction. When it is predicted that the process path will violate either the quality constraint or the safety constraint, online correction and optimization are initiated, and the process parameter fine-tuning amount is dynamically calculated and sent to keep the actual process path within the safety envelope. Knowledge-Model Co-evolution Module: After each batch of actual production is completed, a verification sample consisting of the actual process parameters, actual energy consumption data and actual quality inspection data is collected. The verification sample is used to learn and update the data-driven proxy model part in the digital twin model of the heat treatment process online. Based on the degree of conformity between the verification sample and the association rules stored in the knowledge graph, the confidence weight of the corresponding association rules stored in the knowledge graph is dynamically adjusted. In a preferred embodiment, the process feasibility domain pre-screening module uses natural language processing technology to parse historical process documents and quality inspection reports, extracting material grade entities representing material information, process name entities representing process type, process parameter entities representing specific values, and quality conclusion entities representing the results, and identifying the applicability, cause, and correlation relationships between these entities. Simultaneously, extract from the production process time series data the actual process parameter curve sequence, actual energy consumption curve sequence, and final quality index quantitative data corresponding to historical process documents and quality inspection reports, indexed by production batch and with timestamps.

[0006] In a preferred embodiment, the process of constructing a knowledge graph containing entities, attributes, and rules of association between them based on multi-source data is as follows: The extracted entities, identified relationships, actual process parameter curve sequences, actual energy consumption curve sequences, and final quality index quantification data are aligned and integrated using production batches as the association key to form structured associated data records. Using material grade entities representing material information, process parameter entities representing specific values, key process state entities extracted from actual process parameter curve sequences, quality indicator entities extracted from final quality indicator quantification data, and energy consumption entities extracted from actual energy consumption curve sequences as nodes, and using the identified relationships and data-driven relationships statistically derived from structured associated data records as edges, a knowledge graph is constructed to represent the relationships between five dimensions: materials, process parameters, process states, quality indicators, and energy consumption. The attributes of the edges include the frequency of the relationship based on multi-source data statistics, and a confidence weight representing the credibility of the relationship.

[0007] In a preferred embodiment, the specific operation of jointly evaluating the confidence level of the target process parameter space based on explicit rule constraints extracted from the knowledge graph and the parameter correlation strength function obtained by joint distribution learning of the acquired multi-source data is as follows: Define the combination of process parameters selected within the target process parameter space as a candidate process parameter vector; extract all deterministic process rules that exist in the form of conditional logic from the knowledge graph as a set of explicit rule constraints; For each explicit rule constraint in the set of explicit rule constraints, a satisfaction function is defined. When the candidate process parameter vector satisfies the explicit rule constraint, the value of its corresponding satisfaction function is 1, otherwise it is 0. At the same time, from the structured associated data records, a number of preset key process parameter pairs are selected. For each key process parameter pair, based on all historical data combinations corresponding to the key process parameter pair in the structured associated data records, a parameter correlation strength function describing the probability of the occurrence of the numerical combination of the key process parameter pair is learned using the kernel density estimation method. The joint confidence assessment is achieved by calculating a comprehensive score for each candidate process parameter vector. The comprehensive score is obtained by multiplying the first calculation term and the second calculation term. The first calculation term is the product of the satisfaction function values ​​corresponding to each explicit rule constraint in the set of explicit rule constraints. The second calculation term is the weighted sum of the correlation strength function values ​​between the corresponding parameters for all key process parameters. Finally, the set of all candidate process parameter vectors whose comprehensive scores exceed the preset threshold is output as the pre-screened high-potential process parameter feasible region.

[0008] In a preferred embodiment, the specific operation of constructing an optimization simulation environment based on a digital twin model of the heat treatment process that integrates physicochemical mechanisms and data-driven proxy models in the simulation optimization module is as follows: A digital twin model of the heat treatment process is established and initialized. The digital twin model of the heat treatment process is composed of a mechanism calculation module based on the physicochemical mechanism of heat treatment and a data-driven proxy model trained based on historical process documents, quality inspection reports and production process time series data. The digital twin model of the heat treatment process receives a candidate process parameter vector from the pre-screened high-potential process parameter feasible domain as input. The mechanism calculation module simulates the phase change, heat transfer and mass diffusion laws in the heat treatment process. The data-driven proxy model compensates for the unmodeled dynamic characteristics and deviations. The model comprehensively outputs the predicted quality index, predicted safety index and predicted total energy consumption corresponding to the candidate process parameter vector. Based on this digital twin model of the heat treatment process, an optimization simulation environment is constructed with the goal of minimizing the predicted total energy consumption and with all preset quality and safety constraints as conditions. The feasible domain of the pre-screened high-potential process parameters is set as the entire parameter space that can be searched in the optimization simulation environment. In the optimized simulation environment, surrogate models based on Gaussian processes are established for the predicted total energy consumption and all preset safety constraints. The surrogate models output the predicted mean and prediction uncertainty of the predicted total energy consumption, as well as the predicted mean and prediction uncertainty of each safety constraint for the candidate process parameter vector. Based on the output of the proxy model, perform the following calculations: Calculate the expected improvement value of the predicted total energy consumption: Obtain the predicted mean of the predicted total energy consumption at the candidate process parameter vector and the currently observed optimal predicted total energy consumption value, and calculate the difference between the two; multiply the difference by an improvement scaling factor between zero and one, and add the result of the multiplication to the currently observed optimal predicted total energy consumption value to obtain the expected improvement value. Calculate the overall safety probability of all preset safety constraints: For each safety constraint, obtain its predicted mean and prediction uncertainty at the candidate process parameter vector; divide the predicted mean by the prediction uncertainty to obtain the standardized safety margin index corresponding to the safety constraint; calculate the cumulative probability of the standardized safety margin index under the standard normal distribution; multiply the cumulative probabilities corresponding to all safety constraints together to obtain the overall safety probability.

[0009] In a preferred embodiment, the specific operation of iterative simulation optimization using a Bayesian optimization algorithm with an embedded security barrier function is as follows: In the optimized simulation environment, a Bayesian optimized acquisition function with embedded safety barrier function is constructed. The Bayesian optimized acquisition function is composed of the improved expected value, a safety probability multiplicative factor of an adjustable parameter power of the overall safety probability, and a deterministic violation penalty term multiplied together. The optimization process iteratively maximizes the Bayesian optimization acquisition function to select the next simulation evaluation point. It uses a digital twin model of the heat treatment process to simulate and obtain the predicted total energy consumption and safety constraint values ​​of the point and updates the surrogate model until the termination condition is met. Finally, from all simulation evaluation points that meet the preset quality constraints and safety constraints, the point with the lowest predicted total energy consumption is selected as the optimal process parameters obtained through simulation optimization.

[0010] In a preferred embodiment, the specific operation of synchronizing actual furnace sensor data to the digital twin model of the heat treatment process and performing parallel simulation prediction in the virtual-real closed-loop execution and fine-tuning module is as follows: The system receives the optimal process parameters obtained through simulation optimization from the simulation optimization module and sends these optimal process parameters as the baseline process setting to the industrial control system to start actual production. At the same time, a real-time digital thread is started, which periodically collects real-time temperature, pressure and atmosphere concentration data from the sensors of the actual heat treatment furnace at a preset high frequency to form a real-time state vector. The real-time digital thread performs a synchronous prediction operation on the acquired real-time state vector, specifically as follows: The process is synchronously injected into the digital twin model of the heat treatment process. The digital twin model of the heat treatment process takes the real-time state vector injected at the current moment as the initial state of the simulation, and takes the process settings of the future time period that have not yet been executed in the optimal process parameters as the input conditions. It performs fast forward simulation calculations to predict the process path state trajectory within a fixed time window in the future, and derives the corresponding predicted quality indicators and predicted safety indicators based on the predicted state trajectory.

[0011] In a preferred embodiment, when it is predicted that the process path will violate either the quality constraints or the safety constraints, online correction optimization is initiated, and the specific operation of dynamically calculating and issuing the fine-tuning amount of the process parameters is as follows: A dynamic safety envelope is defined for each quality constraint and safety constraint. Each dynamic safety envelope consists of an inner warning boundary located inside and an outer hard boundary located outside. Based on the process path state trajectory predicted by the digital twin model of the heat treatment process, the safety margin of each point on the process path state trajectory relative to the warning boundary and hard boundary of the corresponding constraint is calculated. When it is predicted that the safety margin of any constraint will be lower than a preset positive trigger threshold within a future time window, online correction optimization is immediately triggered. Online calibration optimization constructs and solves a rolling time-domain optimization problem. The decision variable of the rolling time-domain optimization problem is the sequence of process parameters to be adjusted within a finite number of future control cycles starting from the current moment. The objective function of the rolling time-domain optimization problem consists of three weighted sums. The first part is a parameter adjustment penalty term, which is used to minimize the weighted deviation between the sequence of process parameters to be adjusted and the corresponding sequence of optimal process parameters. The second part is an energy consumption deviation penalty term, which is used to penalize the increase in the predicted total energy consumption relative to the original process caused by the adjustment of process parameters. A non-negative energy consumption weight coefficient is used to adjust the importance of the energy consumption deviation penalty term in the objective. The third part is a safety risk barrier term, which consists of an exponential safety barrier function. The exponential safety barrier function takes the predicted future path safety margin under the sequence of process parameters to be adjusted as input. Solving this rolling time-domain optimization problem yields the optimal process parameter adjustment sequence. The fine-tuning amount of the process parameters corresponding to the next control cycle in the optimal process parameter adjustment sequence is then sent to the industrial control system for execution. After that, the system returns to perform synchronous prediction operations through a real-time digital thread, forming a closed-loop operation until production ends.

[0012] In a preferred embodiment, the specific operation of using validation samples to perform online learning and updating of the data-driven proxy model part of the digital twin model of the heat treatment process in the knowledge-model co-evolution module is as follows: After each batch of actual production is completed, a verification sample is collected, consisting of the actual process parameters, actual energy consumption data and actual quality inspection data. The actual process parameters in the verification sample are input into the digital twin model of the heat treatment process, and a forward simulation is performed to obtain the predicted quality index and predicted total energy consumption of the current batch based on the current state of the digital twin model of the heat treatment process. Calculate the quality prediction deviation between the predicted quality index and the actual quality inspection data, and calculate the energy consumption prediction deviation between the predicted total energy consumption and the actual energy consumption data; based on the quality prediction deviation and the energy consumption prediction deviation, construct an integrated loss function to guide the parameter update of the data-driven proxy model. The integration loss function is composed of a weighted sum of three parts: quality prediction loss, energy consumption prediction loss, and knowledge consistency loss. By minimizing the ensemble loss function, the internal parameters of the data-driven agent model are iteratively optimized, thereby enabling its online learning and updating.

[0013] In a preferred embodiment, the specific operation of dynamically adjusting the confidence weight of the corresponding association rule stored in the knowledge graph based on the degree of conformity between the verification sample and the association rule stored in the knowledge graph is as follows: While updating the agent model using validation samples, the confidence weights of the association rules stored in the knowledge graph are dynamically adjusted. For each relevant stored association rule, the degree of conformity with the current validation sample is determined: if the actual quality inspection data supports the result inferred by the association rule, the evidence is marked as a validation of the association rule; otherwise, it is marked as a violation. If the actual quality inspection data deviates significantly from the expected range set by the association rule, the association rule is marked as having generated an observation conflict in this event. Subsequently, a Bayesian update process incorporating a memory decay factor is used to adjust the confidence weight of the association rule. This process treats the current confidence weight of the association rule as the probability of its validity and the observed validation or violation evidence as a conditional event. The specific calculation process for the updated confidence weight is as follows: Obtain the likelihood value of the current evidence when the association rule is true, and multiply the likelihood value by the memory decay factor raised to the power of the current confidence weight; at the same time, obtain the likelihood value of the current evidence when the association rule is false, and multiply the likelihood value by the memory decay factor raised to the power of the difference between the current confidence weight and the previous one; divide the product of the former by the sum of the product of the latter and the former, and the quotient is the updated confidence weight. After updating the confidence weights of all relevant association rules, the graph propagation process of the confidence weights is executed: For each entity node in the knowledge graph, based on the updated confidence weights of the association rules of all edges directly connected to it, and the current knowledge reliability scores of the neighboring entity nodes connected to these edges, a new knowledge reliability score for that entity node is calculated; the calculation process for the new knowledge reliability score is as follows: Multiply the old knowledge reliability score of the entity node by one and subtract the propagation coefficient, then add the propagation coefficient multiplied by a weighted average. The weighted average is the weighted sum of the old knowledge reliability scores of all neighboring entity nodes, with the updated confidence weights of the connected edges as the weights.

[0014] The beneficial effects of this invention are as follows: the process feasible region pre-screening module uses a knowledge graph to jointly evaluate the confidence of the parameter space, outputting a pre-screened high-potential process parameter feasible region, laying a high-quality foundation for subsequent optimization; the simulation optimization module, based on a digital twin model of the heat treatment process and a Bayesian optimization algorithm with embedded safety barrier functions, finds the optimal process parameters obtained through simulation optimization that satisfy all constraints within the feasible region; the virtual-real closed-loop execution and fine-tuning module ensures that the actual process path is always executed within the safety envelope through real-time digital thread synchronization and online correction optimization; and the knowledge-model co-evolution module uses the validation samples of each batch to drive the co-evolution of the proxy model and the confidence weights of the association rules in the knowledge graph, enabling the optimization and execution capabilities of the entire scheme to continuously self-enhance. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1-2 The heat treatment furnace group energy consumption collaborative optimization and scheduling system shown includes: a process feasible domain pre-screening module, a simulation optimization module, a virtual-real closed-loop execution and fine-tuning module, and a knowledge-model collaborative evolution module connected in sequence. The process feasibility domain pre-screening module responds to process optimization requests for target workpiece materials by acquiring and integrating multi-source data, including historical process documents, quality inspection reports, and production process time-series data related to the target workpiece materials. Based on the multi-source data, a knowledge graph containing entities, attributes, and their inter-attribute rules is constructed. This knowledge graph is used to characterize the relationships between five dimensions: materials, process parameters, process status, quality indicators, and energy consumption. Based on the explicit rule constraints extracted from the knowledge graph and the parameter association strength function obtained through joint distribution learning of the acquired multi-source data, the module performs a joint confidence assessment of the target process parameter space and outputs a pre-screened high-potential process parameter feasibility domain with a confidence level higher than a preset threshold. Simulation optimization module: Receives the pre-screened feasible domain of high-potential process parameters, constructs an optimization simulation environment based on a digital twin model of the heat treatment process that integrates physicochemical mechanisms and data-driven proxy models, and adopts a Bayesian optimization algorithm with embedded safety barrier functions. In the optimization simulation environment, with the minimum energy consumption as the optimization objective, it performs iterative simulation optimization within the feasible domain of high-potential process parameters, and outputs the optimal process parameters obtained through simulation optimization that satisfy all preset quality constraints and safety constraints. Virtual-real closed-loop execution and fine-tuning module: The optimal process parameters are sent to the industrial control system for execution. During the execution process, the actual furnace sensor data is synchronized to the digital twin model of the heat treatment process through real-time digital thread to perform parallel simulation prediction. When it is predicted that the process path will violate either the quality constraint or the safety constraint, online correction and optimization are initiated, and the process parameter fine-tuning amount is dynamically calculated and sent to keep the actual process path within the safety envelope. Knowledge-Model Co-evolution Module: After each batch of actual production is completed, a verification sample consisting of the actual process parameters, actual energy consumption data and actual quality inspection data is collected. The verification sample is used to learn and update the data-driven proxy model part in the digital twin model of the heat treatment process online. Based on the degree of conformity between the verification sample and the association rules stored in the knowledge graph, the confidence weight of the corresponding association rules stored in the knowledge graph is dynamically adjusted.

[0020] In this embodiment, it is specifically necessary to explain that in the process feasibility domain pre-screening module, historical process documents and quality inspection reports are parsed using natural language processing technology to extract material grade entities that represent material information, process name entities that represent process type, process parameter entities that represent specific values, and quality conclusion entities that represent the results, and identify the applicability, cause, and correlation relationships between these entities. Simultaneously, extract from the production process time series data the actual process parameter curve sequence, actual energy consumption curve sequence, and final quality index quantitative data corresponding to historical process documents and quality inspection reports, indexed by production batch and with timestamps; Natural language processing technology specifically includes named entity recognition and relation extraction steps. Through pre-trained language models or entity recognition models customized for professional texts in the heat treatment field, it automatically identifies and classifies process parameters such as material grade, process name, temperature, time, atmosphere, and carbon potential from unstructured text, as well as their values ​​and quality conclusions such as "qualified" and "decarburization layer depth exceeds the standard". The actual process parameter curve sequence refers to the continuous change data of heating temperature, furnace pressure, protective gas flow rate, and carbon potential recorded at a fixed sampling frequency (e.g., once per second) throughout the entire heat treatment cycle. The actual energy consumption curve sequence refers to the total power consumption of the equipment or natural gas flow data recorded with the same or compatible timestamps. The final quality index quantification data refers to the numerical or graded results of hardness value, diffusion layer depth, metallographic structure grade, etc., obtained after testing the finished workpiece. All data are associated through a unified production batch number to ensure that the process settings, process records, and final results of the same batch of workpieces are traceable. The process of constructing a knowledge graph containing entities, attributes, and the rules of association between them based on multi-source data is as follows: The extracted entities, identified relationships, actual process parameter curve sequences, actual energy consumption curve sequences, and final quality index quantification data are aligned and integrated using production batches as the association key to form structured associated data records. Using material grade entities representing material information, process parameter entities representing specific values, key process state entities extracted from actual process parameter curve sequences, quality indicator entities extracted from final quality indicator quantification data, and energy consumption entities extracted from actual energy consumption curve sequences as nodes, and using the identified relationships and data-driven relationships statistically derived from structured associated data records as edges, a knowledge graph is constructed to represent the relationships between five dimensions: materials, process parameters, process states, quality indicators, and energy consumption. The attributes of the edges include the frequency of the relationship based on multi-source data statistics, and a confidence weight representing the credibility of the relationship. Alignment and fusion refer to, for a production batch, associating the process setting entities extracted from its corresponding process documentation (e.g., "quenching temperature: 850℃") and the quality conclusion entities extracted from the quality inspection report (e.g., "hardness: HRC58") with the actual process parameter curve sequence (e.g., temperature curve), actual energy consumption curve sequence (e.g., power curve), and final quality index quantitative data (e.g., "core hardness: 58.2 HRC") in the time or logical dimension to form a complete data record containing "setting parameters - actual process - final result - total energy consumption". The key process state entities extracted from the actual process parameter curve sequence refer to the data obtained by performing time-series data analysis. Indicators obtained through feature engineering that characterize process stages or states, such as "average temperature duration," "peak temperature," "average heating rate," and "average carbon potential during carburizing period," are energy consumption entities extracted from actual energy consumption curve sequences. These refer to statistical quantities such as "total power consumption" and "energy consumption per unit mass of workpiece." Data-driven relationships are strong associations between entities that are not explicitly stated in the text, obtained through association rule mining or statistical analysis of structured associated data records. For example, "when the average heating rate is within a certain range, it co-occurs frequently with lower energy consumption entity values." The initial confidence weight can be assigned based on the statistical frequency of this relationship in historical data (such as support and confidence). Based on the explicit rule constraints extracted from the knowledge graph and the parameter correlation strength function obtained through joint distribution learning of the acquired multi-source data, the specific operation for joint confidence evaluation of the target process parameter space is as follows: Define the combination of process parameters selected within the target process parameter space as a candidate process parameter vector; extract all deterministic process rules that exist in the form of conditional logic from the knowledge graph as a set of explicit rule constraints; For each explicit rule constraint in the set of explicit rule constraints, a satisfaction function is defined. When the candidate process parameter vector satisfies the explicit rule constraint, the value of its corresponding satisfaction function is 1, otherwise it is 0. At the same time, from the structured associated data records, a number of preset key process parameter pairs are selected. For each key process parameter pair, based on all historical data combinations corresponding to the key process parameter pair in the structured associated data records, a parameter correlation strength function describing the probability of the occurrence of the numerical combination of the key process parameter pair is learned using the kernel density estimation method. The joint confidence assessment is achieved by calculating a comprehensive score for each candidate process parameter vector. The comprehensive score is obtained by multiplying a first calculation term by a second calculation term. The first calculation term is the product of the satisfaction function values ​​corresponding to each explicit rule constraint in the set of explicit rule constraints. The second calculation term is the weighted sum of the correlation strength function values ​​between all key process parameters and their corresponding parameters. The calculation process for the weighted summation is as follows: The inter-parameter correlation strength function value of each key process parameter pair is multiplied by its corresponding weight coefficient and then summed, and then divided by a normalization factor determined based on the weight coefficients of all key process parameter pairs and the range of inter-parameter correlation strength function values; the weight coefficient of each key process parameter pair is jointly determined by the confidence weight of the corresponding relationship of the key process parameter pair in the knowledge graph for subsequent dynamic adjustment and the preset business importance. Finally, the set of all candidate process parameter vectors whose comprehensive scores exceed the preset threshold is output as the feasible region of high-potential process parameters after pre-screening. Key process parameter pairs refer to pre-selected combinations of parameters whose interactions are considered to have a significant impact on the outcome, based on process knowledge. Examples include "quenching temperature and tempering temperature" and "carburizing temperature and carburizing time." In kernel density estimation methods, a Gaussian kernel function can be used, with its bandwidth parameter selected through cross-validation to smoothly estimate the joint probability density distribution of the parameter pair values ​​in historical success cases. This distribution function is the correlation strength function between parameters, with its value ranging from 0 to positive infinity. After subsequent normalization, it can be used to characterize relative probability. The weighting coefficients can be obtained by normalizing the product of confidence weights and business importance coefficients. The normalization factor can be taken as all weighting coefficients. The product of the sum of numbers and the typical range of the correlation strength function between parameters is used to map the weighted summation result to a reasonable scale, such as between 0 and 1, so that it can be multiplied with the first calculation term (whose value is 0 or 1). The preset threshold is an empirical value between 0 and 1, used to balance the strictness of the screening with the size of the feasibility domain. For example, in a specific embodiment, the preset threshold can be set to 0.7, which means that only candidate process parameter vectors with a comprehensive score higher than 0.7 are retained. This can effectively filter out most parameter combinations that violate hard rules or are extremely uncommon empirically, thereby reducing the size of the parameter space to be optimized by more than 70% and significantly improving the efficiency of the subsequent simulation optimization module.

[0021] In this embodiment, it is specifically necessary to explain the specific operation of constructing and optimizing the simulation environment based on the digital twin model of the heat treatment process that integrates physicochemical mechanisms and data-driven proxy models in the simulation optimization module: A digital twin model of the heat treatment process is established and initialized. This model is composed of a mechanism calculation module based on the physicochemical mechanisms of heat treatment and a data-driven proxy model trained from historical process documents, quality inspection reports, and production process time-series data. The digital twin model receives a vector of candidate process parameters from a pre-screened high-potential feasible domain as input. The mechanism calculation module simulates phase transitions, heat transfer, and mass diffusion during the heat treatment process, while the data-driven proxy model compensates for unmodeled dynamic characteristics and biases. It then comprehensively outputs predicted quality indicators, predicted safety indicators, and predicted total energy consumption corresponding to the candidate process parameter vectors. The mechanism calculation module can be established based on finite element analysis or computational fluid dynamics methods to solve for temperature fields, stress fields, and phase transition dynamics. The data-driven proxy model can employ a deep neural network, with training data derived from sample pairs generated by fusing historical process documents, quality inspection reports, and production process time-series data, containing "input process parameters - output quality, safety, and energy consumption results." The fusion method can be either using the output of the mechanism model as part of the input to the data-driven proxy model, or integrating the outputs of both through weighted averaging or other methods. Based on this digital twin model of the heat treatment process, an optimization simulation environment is constructed with the goal of minimizing the predicted total energy consumption and with all preset quality and safety constraints as conditions. The feasible domain of the pre-screened high-potential process parameters is set as the entire parameter space that can be searched in the optimization simulation environment. In the optimized simulation environment, surrogate models based on Gaussian processes are established for the predicted total energy consumption and all preset safety constraints. The surrogate models output the predicted mean and prediction uncertainty of the predicted total energy consumption, as well as the predicted mean and prediction uncertainty of each safety constraint for the candidate process parameter vector. The prediction uncertainty is usually represented by the square root of the diagonal of the posterior covariance matrix of the Gaussian process, and its value increases as the evaluation point moves further away from the existing observation point. Based on the output of the proxy model, perform the following calculations: Calculate the expected improvement value of the predicted total energy consumption: Obtain the predicted mean of the predicted total energy consumption at the candidate process parameter vector and the currently observed optimal predicted total energy consumption value, and calculate the difference between the two; multiply the difference by an improvement scaling factor between zero and one. The improvement scaling factor is dynamically adjusted according to the magnitude of the prediction uncertainty of the predicted total energy consumption. When the prediction uncertainty is zero, the value is one, and the value decreases when the prediction uncertainty increases; add the result of the multiplication to the currently observed optimal predicted total energy consumption value to obtain the expected improvement value. The dynamic adjustment relationship between the improvement scaling factor and the prediction uncertainty can be achieved through a monotonically decreasing function. For example, the improvement scaling factor is equal to one divided by the sum of the product of one and the prediction uncertainty, to ensure that the greater the prediction uncertainty, the greater the discount on the expected improvement value, thereby balancing exploration and utilization. Calculate the overall safety probability of all preset safety constraints: For each safety constraint, obtain its predicted mean and prediction uncertainty at the candidate process parameter vector; divide the predicted mean by the prediction uncertainty to obtain the standardized safety margin index corresponding to the safety constraint; calculate the cumulative probability of the standardized safety margin index under the standard normal distribution; multiply the cumulative probabilities corresponding to all safety constraints to obtain the overall safety probability. The standardized safety margin index is essentially the signal-to-noise ratio, and its larger value indicates a higher confidence level in satisfying the constraint; the multiplication operation of the cumulative probabilities is based on the assumption that each safety constraint is independent, in order to estimate the joint probability that all constraints are satisfied simultaneously. The specific steps for iterative simulation optimization using a Bayesian optimization algorithm with embedded safety barrier functions are as follows: In the optimized simulation environment, a Bayesian optimization acquisition function embedded with a safety barrier function is constructed. The Bayesian optimization acquisition function is composed of the improved expected value, a safety probability multiplicative factor of an adjustable parameter power of not less than zero for the overall safety probability, and a deterministic violation penalty term. This construction method ensures that the acquisition function value is positive only when the improved expected value is positive, the overall safety probability is not zero, and there is no deterministic violation, thereby guiding the search away from high-risk areas. Among them, the adjustable parameter is used to adjust the weight of the safety probability multiplicative factor. The value range is greater than or equal to 0.5 and less than or equal to 3. The specific value of the adjustable parameter can be preset according to the safety requirements of the optimization process. For example, 2.5 is used when high safety is required, and 1 is used when moderate exploration is allowed. The deterministic violation penalty term is an exponential decay function constructed with a negative exponent, using an adjustable parameter greater than zero as the base and the sum of the absolute values ​​of the predicted mean of all safety constraints violated by the predicted mean as the negative exponent. A safety constraint violated by the predicted mean is one whose predicted mean at the candidate process parameter vector is less than zero. The sum of the absolute values ​​of the predicted mean of all safety constraints violated by the predicted mean is obtained by iterating through all preset safety constraints, filtering out those with a predicted mean less than zero, and summing the absolute values ​​of their predicted means. The adjustable parameter greater than zero is the penalty strength coefficient, with a value range greater than or equal to one and less than or equal to ten. The penalty strength coefficient can be set to 2.5, so that once any deterministic violation occurs, the acquisition function value will rapidly decay to near zero, thus effectively shielding the region during the optimization search. The optimization process iteratively maximizes the Bayesian optimization acquisition function to select the next simulation evaluation point. It uses a digital twin model of the heat treatment process to simulate and obtain the predicted total energy consumption and safety constraint values ​​at that point and updates the surrogate model until the termination condition is met. Finally, from all simulation evaluation points that meet the preset quality constraints and safety constraints, the point with the lowest predicted total energy consumption is selected as the optimal process parameters obtained through simulation optimization. The termination condition can be set as the improvement of several consecutive iterations being less than a preset threshold, or the total number of simulations reaching a preset upper limit, such as one hundred times.

[0022] In this embodiment, it is specifically necessary to explain the following operation in the virtual-real closed-loop execution and fine-tuning module: Synchronizing actual furnace sensor data to the digital twin model of the heat treatment process via a real-time digital thread and performing parallel simulation prediction. The system receives the optimal process parameters obtained through simulation optimization from the simulation optimization module and sends these optimal process parameters as the baseline process setting to the industrial control system to start actual production. At the same time, a real-time digital thread is started. This real-time digital thread periodically collects real-time temperature, pressure and atmosphere concentration data from the sensors of the actual heat treatment furnace at a preset high frequency to form a real-time state vector. The preset high frequency can be ten times per second or higher to ensure that the simulation initial state of the digital twin model can closely follow the changes in the actual physical process. The real-time digital thread performs a synchronous prediction operation on the acquired real-time state vector, specifically as follows: The data is synchronously injected into the digital twin model of the heat treatment process. The digital twin model of the heat treatment process takes the real-time state vector injected at the current moment as the initial state of the simulation, and takes the process settings of the future time period that have not yet been executed in the optimal process parameters as the input conditions. It performs fast forward simulation calculations to predict the process path state trajectory within a fixed time window in the future, and derives the corresponding predicted quality indicators and predicted safety indicators based on the predicted state trajectory. The length of the fixed time window in the future can be set according to the dynamic characteristics of the process, such as thirty seconds, to balance the forward-looking nature of the prediction and the real-time requirements of the calculation. When it is predicted that the process path will violate either the quality constraints or the safety constraints, online correction and optimization are initiated. The specific operation of dynamically calculating and issuing the fine-tuning amount of the process parameters is as follows: A dynamic safety envelope is defined for each quality constraint and safety constraint. Each dynamic safety envelope consists of an inner warning boundary and an outer hard boundary. The inner warning boundary can be obtained by shrinking the outer hard boundary inward by a preset percentage, such as 10%, to provide a buffer zone. Based on the process path state trajectory predicted by the digital twin model of the heat treatment process, the safety margin of each point on the process path state trajectory relative to the warning boundary and hard boundary of the corresponding constraint is calculated. The safety margin reflects the relative distance of the predicted state from the warning boundary. The safety margin can be calculated by the formula: predicted state value minus inner warning boundary value, then divided by the difference between outer hard boundary value and inner warning boundary value. When it is predicted that the safety margin of any constraint will be lower than a preset positive trigger threshold within a future time window, online correction optimization is immediately triggered. The preset positive trigger threshold can be set to 0.2, which means that intervention is initiated when the predicted path enters the buffer zone 20% inside the warning boundary. Online calibration optimization constructs and solves a rolling time-domain optimization problem. The decision variables of the rolling time-domain optimization problem are the sequence of process parameters to be adjusted within a finite number of future control cycles starting from the current moment. The process parameter sequence refers to the set of process parameters corresponding to multiple consecutive future control cycles arranged in chronological order starting from the current moment. The finite number of future control cycles can be set to ten control cycles, and the length of each control cycle is consistent with the synchronization cycle of the real-time digital thread. The objective function of the rolling time-domain optimization problem consists of a weighted sum of three parts. The first part is a parameter adjustment penalty term, used to minimize the weighted deviation between the sequence of process parameters to be adjusted and the sequence corresponding to the optimal process parameters. The values ​​of the elements of the weight matrix are inversely proportional to the allowable adjustment range of each process parameter. For example, the adjustment penalty weight for the temperature parameter can be set to ten, and the adjustment penalty weight for the time parameter can be set to one. The second part is an energy consumption deviation penalty term, used to penalize the increase in the predicted total energy consumption relative to the original process caused by the adjustment of process parameters, and through... A non-negative energy consumption weighting coefficient is used to adjust the importance of the energy consumption deviation penalty term in the objective. The non-negative energy consumption weighting coefficient can be set to 0.05. The third part is the safety risk barrier term, which consists of an exponential safety barrier function. The exponential safety barrier function takes the predicted future path safety margin under the process parameter sequence to be adjusted as input. The construction process of the exponential safety barrier function is as follows: a barrier strength coefficient greater than zero is used as the base, and the product of the negative decay rate coefficient and the safety margin is used as the exponent. The barrier strength coefficient can be set to one hundred, and the decay rate coefficient can be set to ten. Under this construction, when the safety margin is large, the function output value approaches zero and has little impact on the optimization objective. When the safety margin decreases and approaches the warning boundary, the function output value increases exponentially and sharply, thereby forming a repulsive force in the optimization model, forcing the optimized process parameter sequence to make the future path move away from the constraint boundary. The barrier strength coefficient is used to adjust the overall magnitude of the repulsive force, and the decay rate coefficient is used to adjust the rate of increase of the repulsive force as the safety margin decreases. Solving this rolling time-domain optimization problem yields the optimal process parameter adjustment sequence. The fine-tuning amount of the process parameters corresponding to the next control cycle in the optimal process parameter adjustment sequence is then sent to the industrial control system for execution. Afterward, synchronous prediction operations are performed through a real-time digital thread. Based on the new actual state and the adjusted settings, the real-time digital thread continues to run, performing forward prediction and risk assessment. This closed loop of perception-prediction-optimization-execution continues to run at a speed higher than the dynamic changes in the process until the production of this batch is completed. The closed loop running until the production of this batch is completed means that the operation of this module is terminated when the actual process execution time reaches the total process time specified by the optimal process parameters obtained through simulation optimization, or when a production completion signal is received from the underlying industrial control system.

[0023] In this embodiment, it is specifically necessary to explain the following operation in the knowledge-model co-evolution module: The online learning and updating of the data-driven proxy model part of the digital twin model of the heat treatment process using validation samples is as follows: After each batch of actual production is completed, a verification sample is collected, consisting of the actual process parameters, actual energy consumption data, and actual quality inspection data. The actual process parameters in the verification sample are input into the digital twin model of the heat treatment process, and a forward simulation is performed to obtain the predicted quality index and predicted total energy consumption of the current batch based on the current state of the digital twin model of the heat treatment process. The forward simulation uses the complete sequence of actual process parameters of the current batch as input to simulate the entire heat treatment cycle in order to obtain the prediction results corresponding to the actual end time of production. Calculate the quality prediction deviation between the predicted quality index and the actual quality inspection data, and calculate the energy consumption prediction deviation between the predicted total energy consumption and the actual energy consumption data; based on the quality prediction deviation and the energy consumption prediction deviation, construct an integrated loss function to guide the parameter update of the data-driven proxy model. The integration loss function consists of a weighted sum of three parts: a quality prediction loss term, an energy consumption prediction loss term, and a knowledge consistency loss term. The quality prediction loss term is calculated from the quality prediction deviation, and the energy consumption prediction loss term is calculated from the energy consumption prediction deviation. Their proportions in the total loss are adjusted by quality loss weights and energy consumption loss weights, respectively. These weights can be preset according to the business priorities of quality compliance and energy saving; for example, the quality loss weight can be set to 1.0, and the energy consumption loss weight to 0.3. The construction process of the knowledge consistency loss term is as follows: First, all stored association rules matching the actual process parameters of the current verification sample are retrieved from the knowledge graph. For each matching association rule, a rule importance decay factor characterizing its temporal importance is calculated. This factor is the decay factor of the association rule's importance since its last verification. The negative exponential function of the number of batches experienced results in recently triggered rules having a higher weight in model training. The rule importance decay factor can be expressed as base 0.9 raised to the power of the number of batches since the last validation, meaning that for each new batch of unvalidated records, its importance decays by 10%. Subsequently, the difference between the prediction quality index output by the data-driven surrogate model and the result inferred by the association rule is calculated as the rule distance. The difference measure can be calculated using Euclidean distance or mean absolute error. Finally, the confidence weight of the current association rule, the rule importance decay factor, and the rule distance are multiplied together, and the calculated results of all matching association rules are summed and multiplied by a knowledge consistency loss weight to obtain the knowledge consistency loss term. The knowledge consistency loss weight can be set to 0.1 to achieve a balance between model fitting the data and following historical knowledge. By minimizing the ensemble loss function, the internal parameters of the data-driven agent model are iteratively optimized, thereby enabling its online learning and updating. The optimization algorithm can be stochastic gradient descent or Adam optimizer, iterating until the loss function converges or reaches the preset number of iterations. The specific operation of dynamically adjusting the confidence weight of the corresponding association rule stored in the knowledge graph based on the degree of conformity between the verification sample and the association rule stored in the knowledge graph is as follows: While updating the agent model using validation samples, the confidence weights of the association rules stored in the knowledge graph are dynamically adjusted. For each relevant stored association rule, its conformity with the current validation sample is determined: if the actual quality inspection data supports the result inferred by the association rule, the evidence is marked as a validation of the association rule; otherwise, it is marked as a violation. If the actual quality inspection data deviates significantly from the expected range set by the association rule, the association rule is marked as having generated an observation conflict in this event. "Significant deviation" can be determined by judging whether the actual quality inspection data falls outside the confidence interval of the result inferred by the association rule, for example, the deviation from the mean exceeds two standard deviations. Subsequently, a Bayesian update process incorporating a memory decay factor is used to adjust the confidence weight of the association rule. This process treats the current confidence weight of the association rule as the probability of its validity and the observed validation or violation evidence as a conditional event. The specific calculation process for the updated confidence weight is as follows: The algorithm obtains the likelihood value of the current evidence given the association rule is true, and multiplies this likelihood value by the memory decay factor raised to the power of the current confidence weight. Simultaneously, it obtains the likelihood value of the current evidence given the association rule is false, and multiplies this likelihood value by the memory decay factor raised to the power of the difference between the current confidence weight and the previous one. The product of the former and the latter is divided by the sum of their products, and the quotient is the updated confidence weight. The memory decay factor is a constant between zero and one, used to control the influence of historical confidence on the new weight. A value less than one means the system assigns a higher weight to new evidence. The memory decay factor can be set to 0.9, allowing the system to retain most of its historical experience while continuously learning. If an association rule is marked as having an observation conflict, an additional conflict penalty factor is applied when calculating its updated confidence weight. This penalty factor can be 0.8, meaning the calculated updated confidence weight is multiplied by 0.8. After updating the confidence weights of all relevant association rules, the graph propagation process of the confidence weights is executed: For each entity node in the knowledge graph, based on the updated confidence weights of the association rules of all edges directly connected to it, and the current knowledge reliability scores of the neighboring entity nodes connected to these edges, a new knowledge reliability score for that entity node is calculated; the calculation process for the new knowledge reliability score is as follows: The old knowledge reliability score of the entity node is multiplied by one and the propagation coefficient is subtracted. Then, the propagation coefficient is multiplied by a weighted average. The weighted average is the weighted sum of the old knowledge reliability scores of all neighboring entity nodes, with the updated confidence weights of the connected edges as weights. The propagation coefficient is a constant between zero and one, used to control the intensity of information absorption from neighboring nodes during the calculation process. The propagation coefficient can be set to 0.2, indicating that about 20% of the new reliability score of the entity node comes from the influence of its neighbors. After this graph propagation process, the entity nodes connected by high-confidence association rules will obtain higher knowledge reliability scores. The knowledge reliability scores can be used in the subsequent process feasibility domain pre-screening module as a basis for giving higher weight to association rules involving high-scoring entities when evaluating candidate process parameter vectors.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0030] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A system for coordinated optimization and scheduling of energy consumption in a heat treatment furnace group, characterized in that, Specifically, it includes: The sequentially connected modules are: a process feasible domain pre-screening module, a simulation optimization module, a virtual-real closed-loop execution and fine-tuning module, and a knowledge-model co-evolution module. The process feasibility domain pre-screening module responds to process optimization requests for target workpiece materials by acquiring and integrating multi-source data, including historical process documents, quality inspection reports, and production process time-series data related to the target workpiece materials. Based on the multi-source data, a knowledge graph containing entities, attributes, and their inter-attribute rules is constructed. This knowledge graph is used to characterize the relationships between five dimensions: materials, process parameters, process status, quality indicators, and energy consumption. Based on the explicit rule constraints extracted from the knowledge graph and the parameter association strength function obtained through joint distribution learning of the acquired multi-source data, the module performs a joint confidence assessment of the target process parameter space and outputs a pre-screened high-potential process parameter feasibility domain with a confidence level higher than a preset threshold. Simulation optimization module: Receives the pre-screened feasible domain of high-potential process parameters, constructs an optimization simulation environment based on a digital twin model of the heat treatment process that integrates physicochemical mechanisms and data-driven proxy models, and adopts a Bayesian optimization algorithm with embedded safety barrier functions. In the optimization simulation environment, with the minimum energy consumption as the optimization objective, it performs iterative simulation optimization within the feasible domain of high-potential process parameters, and outputs the optimal process parameters obtained through simulation optimization that satisfy all preset quality constraints and safety constraints. Virtual-real closed-loop execution and fine-tuning module: The optimal process parameters are sent to the industrial control system for execution. During the execution process, the actual furnace sensor data is synchronized to the digital twin model of the heat treatment process through real-time digital thread to perform parallel simulation prediction. When it is predicted that the process path will violate either the quality constraint or the safety constraint, online correction and optimization are initiated, and the process parameter fine-tuning amount is dynamically calculated and sent to keep the actual process path within the safety envelope. Knowledge-Model Co-evolution Module: After each batch of actual production is completed, a verification sample consisting of the actual process parameters, actual energy consumption data and actual quality inspection data is collected. The verification sample is used to learn and update the data-driven proxy model part in the digital twin model of the heat treatment process online. Based on the degree of conformity between the verification sample and the association rules stored in the knowledge graph, the confidence weight of the corresponding association rules stored in the knowledge graph is dynamically adjusted.

2. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 1, characterized in that: In the process feasibility domain pre-screening module, historical process documents and quality inspection reports are parsed using natural language processing technology. From these documents, material grade entities representing material information, process name entities representing process type, process parameter entities representing specific values, and quality conclusion entities representing the results are extracted. The applicability, cause, and correlation relationships between these entities are also identified. Simultaneously, extract from the production process time series data the actual process parameter curve sequence, actual energy consumption curve sequence, and final quality index quantitative data corresponding to historical process documents and quality inspection reports, indexed by production batch and with timestamps.

3. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 2, characterized in that: The process of constructing a knowledge graph containing entities, attributes, and rules of association between them based on multi-source data is as follows: The extracted entities, identified relationships, actual process parameter curve sequences, actual energy consumption curve sequences, and final quality index quantification data are aligned and integrated using production batches as the association key to form structured associated data records. Using material grade entities representing material information, process parameter entities representing specific values, key process state entities extracted from actual process parameter curve sequences, quality indicator entities extracted from final quality indicator quantification data, and energy consumption entities extracted from actual energy consumption curve sequences as nodes, and using the identified relationships and data-driven relationships statistically derived from structured associated data records as edges, a knowledge graph is constructed to represent the relationships between five dimensions: materials, process parameters, process states, quality indicators, and energy consumption. The attributes of the edges include the frequency of the relationship based on multi-source data statistics, and a confidence weight representing the credibility of the relationship.

4. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 3, characterized in that: The specific operation of jointly evaluating the confidence level of the target process parameter space based on explicit rule constraints extracted from knowledge graphs and the parameter correlation strength function obtained through joint distribution learning of acquired multi-source data is as follows: Define the combination of process parameters selected within the target process parameter space as a candidate process parameter vector; extract all deterministic process rules that exist in the form of conditional logic from the knowledge graph as a set of explicit rule constraints; For each explicit rule constraint in the set of explicit rule constraints, a satisfaction function is defined. When the candidate process parameter vector satisfies the explicit rule constraint, the value of its corresponding satisfaction function is 1, otherwise it is 0. At the same time, from the structured associated data records, a number of preset key process parameter pairs are selected. For each key process parameter pair, based on all historical data combinations corresponding to the key process parameter pair in the structured associated data records, a parameter correlation strength function describing the probability of the occurrence of the numerical combination of the key process parameter pair is learned using the kernel density estimation method. The joint confidence assessment is achieved by calculating a comprehensive score for each candidate process parameter vector. The comprehensive score is obtained by multiplying the first calculation term and the second calculation term. The first calculation term is the product of the satisfaction function values ​​corresponding to each explicit rule constraint in the set of explicit rule constraints. The second calculation term is the weighted sum of the correlation strength function values ​​between the corresponding parameters for all key process parameters. Finally, the set of all candidate process parameter vectors whose comprehensive scores exceed the preset threshold is output as the pre-screened high-potential process parameter feasible region.

5. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 4, characterized in that: In the simulation optimization module, the specific operation of constructing and optimizing the simulation environment based on the digital twin model of the heat treatment process that integrates physicochemical mechanisms and data-driven proxy models is as follows: A digital twin model of the heat treatment process is established and initialized. The digital twin model of the heat treatment process is composed of a mechanism calculation module based on the physicochemical mechanism of heat treatment and a data-driven proxy model trained based on historical process documents, quality inspection reports and production process time series data. The digital twin model of the heat treatment process receives a candidate process parameter vector from the pre-screened high-potential process parameter feasible domain as input. The mechanism calculation module simulates the phase change, heat transfer and mass diffusion laws in the heat treatment process. The data-driven proxy model compensates for the unmodeled dynamic characteristics and deviations. The model comprehensively outputs the predicted quality index, predicted safety index and predicted total energy consumption corresponding to the candidate process parameter vector. Based on this digital twin model of the heat treatment process, an optimization simulation environment is constructed with the goal of minimizing the predicted total energy consumption and with all preset quality and safety constraints as conditions. The feasible domain of the pre-screened high-potential process parameters is set as the entire parameter space that can be searched in the optimization simulation environment. In the optimized simulation environment, surrogate models based on Gaussian processes are established for the predicted total energy consumption and all preset safety constraints. The surrogate models output the predicted mean and prediction uncertainty of the predicted total energy consumption, as well as the predicted mean and prediction uncertainty of each safety constraint for the candidate process parameter vector. Based on the output of the proxy model, perform the following calculations: Calculate the expected improvement value of the predicted total energy consumption: Obtain the predicted mean of the predicted total energy consumption at the candidate process parameter vector and the currently observed optimal predicted total energy consumption value, and calculate the difference between the two; multiply the difference by an improvement scaling factor between zero and one, and add the result of the multiplication to the currently observed optimal predicted total energy consumption value to obtain the expected improvement value. Calculate the overall safety probability of all preset safety constraints: For each safety constraint, obtain its predicted mean and prediction uncertainty at the candidate process parameter vector; divide the predicted mean by the prediction uncertainty to obtain the standardized safety margin index corresponding to the safety constraint; calculate the cumulative probability of the standardized safety margin index under the standard normal distribution; multiply the cumulative probabilities corresponding to all safety constraints together to obtain the overall safety probability.

6. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 5, characterized in that: The specific operation of iterative simulation optimization using a Bayesian optimization algorithm with embedded security barrier function is as follows: In the optimized simulation environment, a Bayesian optimized acquisition function with embedded safety barrier function is constructed. The Bayesian optimized acquisition function is composed of the improved expected value, a safety probability multiplicative factor of an adjustable parameter power of the overall safety probability, and a deterministic violation penalty term multiplied together. The optimization process iteratively maximizes the Bayesian optimization acquisition function to select the next simulation evaluation point. It uses a digital twin model of the heat treatment process to simulate and obtain the predicted total energy consumption and safety constraint values ​​of the point and updates the surrogate model until the termination condition is met. Finally, from all simulation evaluation points that meet the preset quality constraints and safety constraints, the point with the lowest predicted total energy consumption is selected as the optimal process parameters obtained through simulation optimization.

7. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 6, characterized in that: In the virtual-real closed-loop execution and fine-tuning module, the specific operation of synchronizing actual furnace sensor data to the digital twin model of the heat treatment process through a real-time digital thread and performing parallel simulation prediction is as follows: The system receives the optimal process parameters obtained through simulation optimization from the simulation optimization module and sends these optimal process parameters as the baseline process setting to the industrial control system to start actual production. At the same time, a real-time digital thread is started, which periodically collects real-time temperature, pressure and atmosphere concentration data from the sensors of the actual heat treatment furnace at a preset high frequency to form a real-time state vector. The real-time digital thread performs a synchronous prediction operation on the acquired real-time state vector, specifically as follows: The process is synchronously injected into the digital twin model of the heat treatment process. The digital twin model of the heat treatment process takes the real-time state vector injected at the current moment as the initial state of the simulation, and takes the process settings of the future time period that have not yet been executed in the optimal process parameters as the input conditions. It performs fast forward simulation calculations to predict the process path state trajectory within a fixed time window in the future, and derives the corresponding predicted quality indicators and predicted safety indicators based on the predicted state trajectory.

8. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 7, characterized in that: The specific steps for initiating online correction and optimization, dynamically calculating and issuing fine-tuning amounts of process parameters when it is predicted that the process path will violate either the quality constraints or the safety constraints are as follows: A dynamic safety envelope is defined for each quality constraint and safety constraint. Each dynamic safety envelope consists of an inner warning boundary located inside and an outer hard boundary located outside. Based on the process path state trajectory predicted by the digital twin model of the heat treatment process, the safety margin of each point on the process path state trajectory relative to the warning boundary and hard boundary of the corresponding constraint is calculated. When it is predicted that the safety margin of any constraint will be lower than a preset positive trigger threshold within a future time window, online correction optimization is immediately triggered. Online calibration optimization constructs and solves a rolling time domain optimization problem. The decision variables of the rolling time domain optimization problem are the sequence of process parameters to be adjusted in the future finite number of control cycles starting from the current moment. The objective function of the rolling time-domain optimization problem consists of a weighted sum of three parts. The first part is the parameter adjustment penalty term, which is used to minimize the weighted deviation between the sequence of process parameters to be adjusted and the corresponding sequence of optimal process parameters. The second part is the energy consumption deviation penalty term, which is used to penalize the increase in the predicted total energy consumption relative to the original process caused by the adjustment of process parameters. A non-negative energy consumption weight coefficient is used to adjust the importance of the energy consumption deviation penalty term in the objective. The third part is the safety risk barrier term, which consists of an exponential safety barrier function. The exponential safety barrier function takes the predicted future path safety margin under the sequence of process parameters to be adjusted as input. Solving this rolling time-domain optimization problem yields the optimal process parameter adjustment sequence. The fine-tuning amount of the process parameters corresponding to the next control cycle in the optimal process parameter adjustment sequence is then sent to the industrial control system for execution. After that, the system returns to perform synchronous prediction operations through a real-time digital thread, forming a closed-loop operation until production ends.

9. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 8, characterized in that: In the knowledge-model co-evolution module, the specific operation of using validation samples to perform online learning and updating of the data-driven proxy model part in the digital twin model of the heat treatment process is as follows: After each batch of actual production is completed, a verification sample is collected, consisting of the actual process parameters, actual energy consumption data and actual quality inspection data. The actual process parameters in the verification sample are input into the digital twin model of the heat treatment process, and a forward simulation is performed to obtain the predicted quality index and predicted total energy consumption of the current batch based on the current state of the digital twin model of the heat treatment process. Calculate the quality prediction deviation between the predicted quality index and the actual quality inspection data, and calculate the energy consumption prediction deviation between the predicted total energy consumption and the actual energy consumption data; based on the quality prediction deviation and the energy consumption prediction deviation, construct an integrated loss function to guide the parameter update of the data-driven proxy model. The integration loss function is composed of a weighted sum of three parts: quality prediction loss, energy consumption prediction loss, and knowledge consistency loss. By minimizing the ensemble loss function, the internal parameters of the data-driven agent model are iteratively optimized, thereby enabling its online learning and updating.

10. The energy consumption collaborative optimization and scheduling system for a heat treatment furnace group according to claim 9, characterized in that: The specific operation of dynamically adjusting the confidence weight of the corresponding association rule stored in the knowledge graph based on the degree of conformity between the verification sample and the association rule stored in the knowledge graph is as follows: While updating the data-driven proxy model using validation samples, the confidence weights of the association rules stored in the knowledge graph are dynamically adjusted. For each related stored association rule, the degree of conformity with the current validation sample is determined: if the actual quality inspection data supports the result inferred by the association rule, the evidence is marked as a validation of the association rule; otherwise, it is marked as a violation. If the actual quality inspection data deviates significantly from the expected range set by the association rule, then the association rule is marked as having generated an observation conflict in this event. Subsequently, a Bayesian update process incorporating a memory decay factor is used to adjust the confidence weights of the association rule. This process treats the current confidence weight of the association rule as the probability of its validity, and the observed validation or violation evidence as a conditional event; the specific calculation process for the updated confidence weight is as follows: Obtain the likelihood value of the current evidence when the association rule is true, and multiply the likelihood value by the memory decay factor raised to the power of the current confidence weight; at the same time, obtain the likelihood value of the current evidence when the association rule is false, and multiply the likelihood value by the memory decay factor raised to the power of the difference between the current confidence weight and the previous one; divide the product of the former by the sum of the product of the latter and the former, and the quotient is the updated confidence weight. After updating the confidence weights of all relevant association rules, the graph propagation process of the confidence weights is executed: For each entity node in the knowledge graph, based on the updated confidence weights of the association rules of all edges directly connected to it, and the current knowledge reliability scores of the neighboring entity nodes connected to these edges, a new knowledge reliability score for that entity node is calculated; the calculation process for the new knowledge reliability score is as follows: Multiply the old knowledge reliability score of the entity node by one and subtract the propagation coefficient, then add the propagation coefficient multiplied by a weighted average. The weighted average is the weighted sum of the old knowledge reliability scores of all neighboring entity nodes, with the updated confidence weights of the connected edges as the weights.

Images