Intelligent preparation and performance prediction method and system for large-crystal magnesia

Abstract

The present application relates to inorganic non-metallic refractory intelligent manufacturing and arc smelting process control technology, discloses a method and system for intelligent preparation and performance prediction of large crystalline magnesia. Obtain the composition data of magnesite raw materials and the target performance vector, generate the batching vector and the batch smelting control boundary; fusion time alignment temperature, furnace atmosphere, electric power and furnace pressure time series data form smelting state vector; adopt neural network to build adaptive control model, generate adjusting amount in batch smelting control boundary and execute in arc smelting equipment; adopt neural network to build performance prediction model, according to the reacquired smelting state time series, predict bulk density, apparent porosity, compression strength and thermal expansion coefficient, and the performance deviation is fed back to the subsequent control cycle; after batch ends, form batch optimization sample to update next batch parameters, form batch performance feedback closed loop and batch process update closed loop.

Description

Technical Field

[0001] This invention relates to intelligent manufacturing and arc melting process control technology for large crystalline magnesia, and discloses intelligent preparation and performance prediction methods and systems for large crystalline magnesia. Background Technology

[0002] Macrocrystalline magnesia is a refractory raw material formed by high-temperature melting and cooling crystallization of magnesium oxide as the main component. It is used as the lining material for smelting equipment in iron and steel enterprises, smelting equipment in non-ferrous metal enterprises, furnace bodies in coking enterprises, and related high-temperature kilns in coal enterprises. These kilns experience harsh conditions such as high temperatures, slag, and corrosive gases during smelting, coking, or high-temperature processing, placing high demands on the bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion of macrocrystalline magnesia. The performance of macrocrystalline magnesia materials is affected by the composition of the main component and associated impurities of magnesite raw materials, the smelting temperature history, energy input, furnace atmosphere, furnace pressure, and the disturbance state of materials or melt. Magnesite is a natural mineral, and the magnesium oxide content and associated components such as calcium, silicon, iron, and aluminum vary in different mining areas, different mining layers, or different pre-treated batches of raw materials. It is difficult to achieve a consistent melting and crystallization state for different batches of raw materials by using fixed batching ratios and fixed smelting control parameters. While existing automated production methods can improve the accuracy of batching by using weighing devices, the accuracy of weighing mainly reduces the execution error of the actual feed amount relative to the set feed amount, and cannot make a fixed batching scheme automatically adapt to changes in the chemical composition of raw materials.

[0003] The electric arc melting process is characterized by high temperature, variable coupling, and varying response speeds for different state variables. Changes in electrical power alter the energy input and temperature rise process, oxygen supply affects the furnace atmosphere, and the state of materials and melt influences the arc load. The sampling frequencies and response speeds of temperature, atmosphere, electrical power, and furnace pressure detection devices are typically inconsistent. Directly combining outputs from different sensors without a unified time reference may result in state vectors containing detection data that do not correspond to the same control stage. The final material properties of large-crystalline magnesia, such as bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion, are usually obtained after melting, cooling, and finished product testing. Actual test results cannot be directly obtained during the melting stage. If performance prediction results are only used for production monitoring or early display without participating in the control of the actual actuators, they cannot alter the subsequent melting process before the current material properties are finalized.

[0004] If adaptive control or learning-based control models are only limited by the fixed equipment safety range, different batches of magnesite raw materials will share the same feasible action area. However, the equipment safety range represents the limit to which the equipment can operate. The suitable process area corresponding to the current raw material composition and target material properties is a material preparation constraint, and the two are not equivalent. In the final stage of smelting or during the initial stage of heat preservation and crystallization while still implementing melt disturbance control, if the stirring control quantity continues to change after the material properties have approached the target, it may cause flow near the solid-liquid interface and continuous fluctuations in the temperature boundary. The thermal risk of arc melting equipment is related to the temperature distribution inside the molten pool. Discrete temperature measurement points cannot directly provide the spatial relationship between the highest temperature position inside the furnace and the safety boundary at the bottom of the furnace. If a digital twin model only displays a virtual temperature state without changing the feasible range of subsequent control actions, it cannot directly form control constraints. If the finished product test results obtained after production are only used for product grading and are not re-correlated with the raw material composition, actual batching, process status, control actions, and predicted results, the actual results formed by the previous production batch cannot be directly used to correct the parameter generation process of the next production batch. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for intelligent preparation and performance prediction of large-crystal magnesia. It solves the technical problem that under the conditions of batch differences in magnesite raw materials, disturbances in the electric arc melting process, and lag in the final material performance testing, there is a lack of continuous data dependence and physical control correlation between raw material composition, actual melting state, control actions, performance prediction, finished product testing, and parameters of the next batch. This results in fixed batching and unified control boundaries being unable to adapt to batch changes, and performance prediction results being unable to participate in the melting control of the current batch.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] The magnesite raw material undergoes pretreatment, and the composition data of the current production batch is obtained using raw material composition analysis equipment. This raw material composition data is used to characterize at least the main magnesium oxide components and the major associated components actually detected in the current production task. The raw material composition data is associated with the production batch identifier.

[0008] Obtain the target performance vector corresponding to the current production task. The target performance vector is derived from the planned performance grade of large crystalline magnesia, the production task, or the corresponding quality control requirements. In some preferred embodiments, it includes at least the target bulk density, target apparent porosity, target compressive strength, and target coefficient of thermal expansion.

[0009] Retrieves historical process records from the database of large-crystal magnesia preparation processes. These historical records are linked by production batch, including raw material composition, actual batching results, smelting process timing data, smelting control actions, performance prediction results, and product performance test results.

[0010] The batch parameter generation module generates a batching vector and a batch melting control boundary based on the current raw material composition data, target performance vector, and historical process records. The batching vector determines the addition relationship of each raw material component that the automatic batching equipment needs to execute for this batch; the batch melting control boundary determines the available range of electrical power, oxygen supply, and stirring amount during the adaptive control process of this batch, as well as the variation range constraints set according to production needs.

[0011] The power regulation amount represents the power control command generated by the action strategy network in the corresponding control cycle and sent to the power regulation device after being mapped by the batch smelting control boundary; the oxygen supply regulation amount represents the oxygen supply control command sent to the oxygen supply regulation device; and the stirring amount regulation amount represents the stirring control command sent to the stirring actuator. The change in action between adjacent control cycles is represented by the difference between the corresponding control commands of two adjacent control cycles.

[0012] In some preferred embodiments, adjacent historical batches are first screened from historical process records according to the current raw material composition. During screening, different component variables can be standardized first, and then historical batches can be sorted or weighted according to the similarity of their compositions. Batches whose product performance does not meet the current target constraints, whose key data is missing, or whose process records are abnormal are not directly used as valid reference samples for generating the current parameters.

[0013] When historical samples can support surrogate optimization, batching parameters and smelting control parameters are jointly set in the candidate space. The actual product performance of historical batches is used as one of the evaluation criteria for candidate parameters, and target performance constraints and equipment operation constraints are added. The surrogate model is used to estimate the relationship between the combination of candidate parameters and the target performance, and the acquisition strategy is used to select subsequent candidate points from the candidate space.

[0014] The surrogate model can be a Gaussian process model, or other surrogate models capable of establishing the relationship between input parameters and product performance based on limited historical samples. The data acquisition strategy can balance utilizing existing optimal regions with exploring unknown regions. The batch parameter generation process simultaneously determines the batching vector and the corresponding feasible melting process region.

[0015] The batching parameters for the current batch are determined from the candidate parameter combinations that satisfy the target performance constraints and equipment operation constraints, and the batch smelting control boundary is determined based on the feasible distribution of the corresponding smelting control parameters.

[0016] When the number of historical batches is insufficient to establish a stable proxy model, historical batches with raw material composition similar to the current batch and whose final product performance meets the target requirements can be selected. An initial process region is formed based on the actual distribution of the corresponding control parameters. This initial process region is then combined with the allowable range of the arc melting equipment to form the initial batch melting control boundary. As new batch optimization samples are written to the database, the batch melting control boundary is updated.

[0017] The automated batching equipment performs material metering and feeding according to the batching vector. After batching is completed, the weighing or feeding record is generated to form the actual batching result. Subsequent control states preferably use the actual batching result confirmed by actual execution, so that the control model can receive the material composition status that has entered the current production batch.

[0018] The batched materials enter the electric arc melting equipment. During the melting process, at least the timing data for temperature, furnace atmosphere, electrical power, and furnace pressure are collected. Since different sensors may have different sampling frequencies, a unified time reference is established based on the control cycle used by the adaptive control. The control cycle is determined based on the rate of change of the electric arc melting equipment's state, the actuator's response capability, the sensor's sampling capability, and the control computing equipment's processing capability.

[0019] For high-frequency process variables with multiple sampling points within a control cycle, one or more of the following can be extracted: the effective value at the end of the cycle, the cycle mean, the change, the trend of change, and the degree of fluctuation. For process variables with a sampling frequency lower than the control frequency, the most recent effective measurement value from the current control cycle is obtained according to the timestamp, and effectiveness information is generated.

[0020] The time condition used to determine whether a low-frequency measurement value is still valid is determined based on the normal sampling period of the corresponding sensor, the equipment response time, and the technical requirements of the production line's testing equipment. Historical measurement values ​​that have exceeded the valid time condition are not used as new normal measurement values.

[0021] After time alignment, process data characteristics, data validity information, and actual batching results are combined to form a melting state vector. The actual batching results constitute a batch-level state within the same production batch, with temperature, furnace atmosphere, electrical power, and furnace pressure forming a dynamic state that changes with the control cycle. This melting state vector is then input into an adaptive control model constrained by batch melting control boundaries.

[0022] In some preferred embodiments, the adaptive control model includes a state feature extraction network, an action policy network, and a value evaluation network.

[0023] The state feature extraction network generates state features based on the smelting state vector. In subsequent control cycles where the performance prediction model has already generated the performance deviation vector, the performance deviation vector, as additional state information, participates in the generation of state features along with the current smelting state. The action strategy network generates normalized control actions based on the state features. After boundary mapping, the normalized control actions are converted into electrical power adjustment, oxygen supply adjustment, and stirring adjustment.

[0024] Boundary mapping ensures that the actual output control actions are within the current batch melting control boundary. Boundary mapping can employ continuous mapping, trimming, action masking, feasible action filtering, or other processing methods that keep the action output within the feasible process region of the current batch.

[0025] Furthermore, boundary proximity evaluation information is generated based on the distance between the control action and the batch melting control boundary, enabling the action strategy to reduce control behaviors that continuously approach the batch process boundary while meeting performance targets. The electrical power adjustment amount actually changes the electrical energy input through the energy input regulation mechanism of the electric arc melting equipment. The oxygen supply adjustment amount changes the oxygen flow rate or corresponding oxygen supply status entering the electric arc melting equipment through the oxygen supply regulation device. The stirring amount adjustment amount changes the material or melt disturbance state through the stirring actuator configured on the production line. Temperature is a detected melting state variable and is not used as a directly executed action.

[0026] In some preferred embodiments, the value evaluation network includes a first value evaluation branch and a second value evaluation branch. The two value evaluation branches each generate an action value evaluation based on the same state characteristics and control actions. The evaluation result with the lower action value from the two evaluation results is used in the action policy network update to reduce the possibility of a single value evaluation result being consistently too high, leading to overly aggressive control actions.

[0027] After the motion adjustment is applied to the electric arc melting equipment through the actual actuator, the predicted state generated by the motion strategy network, value assessment network or other internal models is not used to replace the actual state of the next control cycle.

[0028] Upon reaching the next control cycle, the outputs of the temperature detection device, atmosphere detection device, power monitoring device, and furnace pressure detection device are reread, time alignment is re-executed, and a new melting state vector is formed. Therefore, the consequences of the control actions are determined by the actual response of the electric arc melting equipment.

[0029] The melting state vectors re-acquired over multiple consecutive control cycles are arranged in chronological order to form a melting state time series, which is then input into the performance prediction model. The performance prediction model predicts the final properties of the large-crystal magnesia material based on the melting state time series already generated for the current batch and raw material batch information.

[0030] In some preferred embodiments, the performance prediction model includes a temporal feature weighted layer, a long short-term memory network, a shared feature layer, and multiple task output layers.

[0031] A time-series feature weighting layer forms a weighted state sequence based on the correlation between different process features, different time positions, and the final material properties. This ensures that the smelting state segments, which have a significant impact on material properties, receive corresponding weights in subsequent time-series feature extraction. A Long Short-Term Memory (LSTM) network receives this weighted state sequence and extracts the time-dependent features formed between different control cycles. A shared feature layer fuses the time-dependent features output by the LTM network with the current raw material composition information. Multiple task output layers output predicted values ​​for bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion based on the shared features. These predicted values ​​form a predicted performance vector.

[0032] The performance prediction model is trained using historical batches that have completed production and yielded finished product testing results. The training data is associated with raw material composition, actual batching results, melting state time series, and actual product performance according to production batch identifiers, avoiding incorrect combinations of melting process data and product performance labels from different production batches.

[0033] When the model is used for online prediction of the production process, the input data includes data acquired before the current prediction point in time and process settings determined before the current prediction point in time. Actual cooling data or other future detection data occurring after the current prediction point in time are not used as input for the current prediction.

[0034] For batches with different melting times, a time series can be established according to the corresponding process stages, using effective sequence length markers or sequence masks. This ensures that data from different batches enters the performance prediction model while maintaining the actual process sequence. After the performance prediction model outputs the predicted performance vector, each predicted performance component is compared with the corresponding target value in the target performance vector.

[0035] For the Each performance component in the control cycle The normalized directed bias is obtained by:

[0036] ;

[0037] in, Indicates the first Each performance component in the control cycle Normalization bias; Indicates the performance component number; Indicates the control cycle number; Indicates the first Each performance component in the control cycle The predicted value; Indicates the first Target values ​​for each performance component; Indicates the first The normalized scale corresponds to each performance component; the upper cap indicates the predicted quantity, and the superscript asterisk indicates the target quantity.

[0038] The normalization scale is determined based on the allowable range of the corresponding product performance level, the performance statistics of historical qualified batches, or the process scale verified in production. Each performance component uses a normalization scale corresponding to its own physical quantity; therefore, the above-mentioned normalization directional deviation is a dimensionless quantity.

[0039] The performance deviation vector is formed by the normalized directed deviations of each performance component. This performance deviation vector serves as additional state information for subsequent control cycles, enabling the adaptive control model to not only receive the current furnace condition but also identify the direction and degree of deviation of the predicted final material properties from the current production target. The performance deviation is also used to generate control evaluation information.

[0040] In some preferred embodiments, the following is adopted:

[0041] ;

[0042] in, Indicates control cycle The initial control evaluation quantity; Indicates the weight of the performance deviation evaluation item; Indicates the first The component weights of each performance component; Indicates the first Each performance component in the control cycle Normalization bias; Indicates the weight of energy consumption evaluation items; Indicates control cycle The corresponding normalized energy consumption; Indicates the weight of the evaluation item that is close to the boundary; Indicates control cycle The control action is the boundary proximity evaluation quantity relative to the current batch smelting control boundary.

[0043] Normalized energy consumption, boundary proximity evaluation quantity, and the above-mentioned normalized directional deviation are all dimensionless quantities. The weights of performance deviation evaluation items, performance component weights, energy consumption evaluation items, and boundary proximity evaluation items are dimensionless weights that match the corresponding evaluation items. Therefore, the initial control evaluation quantity is a dimensionless control evaluation quantity.

[0044] The boundary proximity evaluation quantity is determined based on the normalized distance between each control action and the corresponding batch smelting control boundary. The closer the control action is to the corresponding boundary, the larger the boundary proximity evaluation quantity. The weights of the performance deviation evaluation item, energy consumption evaluation item, and boundary proximity evaluation item are determined according to the importance of the target performance and the control requirements of energy and process boundaries.

[0045] Furthermore, to reduce the high-frequency fluctuations of the stirring control command between adjacent control cycles after the predicted bulk density enters the target neighborhood, a stirring adjustment penalty term activated based on the bulk density deviation condition is added to the initial control evaluation quantity. The stirring quantity adjustment of the current control cycle and the stirring quantity adjustment of the previous control cycle are normalized according to the currently effective batch melting control boundary to obtain the corresponding normalized stirring control action. The absolute value of the difference between the normalized stirring control action of the current control cycle and the normalized stirring control action of the previous control cycle is defined as the stirring action fluctuation. Since the fluctuation is formed by the normalized control action, it is a dimensionless quantity. Let the normalized directional deviation of the performance component corresponding to the bulk density be... subscript Represents the bulk density performance component. Indicates the control cycle number.

[0046] Set a step activation function that depends on the volume density normalization bias. When the absolute value of the normalized deviation of the volume density is not greater than the set density deviation threshold, the step activation function is set to 1; when the absolute value of the normalized deviation of the volume density is greater than the set density deviation threshold, the step activation function is set to 0.

[0047] The set density deviation threshold is determined by normalizing the allowable volumetric density deviation range corresponding to the current product grade, or by determining it based on the volumetric density deviation distribution in historical qualified batches that have been confirmed to meet the requirements of the current product grade. A corrected control evaluation quantity is formed based on the step activation function and the agitation fluctuation.

[0048] ;

[0049] in, Indicates control cycle Correction control evaluation quantity; Indicates control cycle The initial control evaluation quantity; This represents the dimensionless stirring penalty coefficient; This represents a step activation function that depends on the volume density normalization bias. Indicates control cycle With control cycle The difference in normalized stirring control action; Indicates control cycle The fluctuation of the stirring action.

[0050] The step activation function, stirring penalty coefficient, and stirring motion variability are all dimensionless quantities. Therefore, the stirring adjustment penalty term has the same dimension as the initial control evaluation quantity and the modified control evaluation quantity, and the modified control evaluation quantity is a dimensionless control evaluation quantity. The stirring penalty coefficient is used to limit the relative role of stirring motion variability in control evaluation after entering the target neighborhood of volume density. Its parameter is calibrated based on the control model training results or the variability range of stirring control motion during the stable phase in historical qualified batches.

[0051] When the volumetric density prediction deviation has not yet entered the target neighborhood defined by the set density deviation threshold, the step activation function is set to 0, the corrected control evaluation value is equal to the initial control evaluation value, and the stirring adjustment penalty term does not participate in the evaluation of the current control cycle. When the volumetric density prediction deviation enters the target neighborhood, the step activation function is set to 1, and the greater the difference in normalized stirring control actions between two adjacent control cycles, the greater the decrease in the corrected control evaluation value relative to the initial control evaluation value.

[0052] The value assessment network uses the corrected control evaluation quantity to evaluate the state-action combination and forms value assessment information for updating the action policy network parameters accordingly. When using time-difference-based value learning, a time-difference error is formed based on the corrected control evaluation quantity, the value estimate corresponding to the state in the next control cycle, and the value estimate of the current state action, and the time-difference error is used to update the value assessment network.

[0053] The action strategy network updates parameters using a strategy gradient approach that matches the continuous action strategy, based on the evaluation results of the control actions from the value assessment network. The stirring adjustment penalty does not directly limit the stirring actuator to stopping or maintaining a fixed physical value. Instead, it adds an evaluation cost when the stirring control action in adjacent cycles undergoes significant changes after the predicted volume density enters the target neighborhood, thus biasing the update direction of the action strategy towards smaller stirring action fluctuations. The adaptive control model calculates subsequent control actions based on the new actual melting state, performance deviation vector, and control evaluation information. These control actions are then applied to the actual arc melting equipment, and the actual state is re-acquired and performance is re-predicted, thus forming an intra-batch performance feedback closed loop.

[0054] Intra-batch performance feedback sets joint termination conditions. Target performance constraints are derived from the current production task, and target allowable ranges can be set for each performance component in the predicted performance vector. Melting state stability conditions are established based on the current melting process stage. Stability conditions can be determined based on the statistical range of temperature changes, furnace atmosphere changes, electrical power changes, and furnace pressure changes in historical qualified production batches at the same process stage, or based on the process control range that has been verified on the production line.

[0055] The corresponding batch performance feedback phase ends only when the predicted performance meets the target performance constraints, and the temperature, furnace atmosphere, electrical power, and furnace pressure meet the stability conditions of the corresponding smelting stage for multiple consecutive control cycles. If the predicted performance enters the target allowable range but the smelting state has not yet stabilized, state acquisition and control continue.

[0056] If the melting state has stabilized but the predicted performance still does not meet the target performance constraints, adaptive control continues to be executed within the batch melting control boundaries. When sensor data is invalid or exceeds the allowable valid time, the corresponding data will not participate in the performance convergence judgment. The system maintains the most recently verified safety control state, switches to the preset equipment control mode, or triggers manual inspection according to the production line safety control requirements.

[0057] Overheating, abnormal furnace pressure, electrical faults, or other safety interlocks of the electric arc melting equipment exist independently of the performance prediction closed loop and have a higher execution priority than the adaptive control output. After the production batch completes melting and cooling, the finished product testing equipment obtains the product performance test results. The product performance test results corresponding to the performance prediction model include at least bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion.

[0058] When product quality grading is required based on grain size, the grain size is obtained from the corresponding testing equipment during the finished product inspection stage and used as information for product grade determination. Grain size is not included as a component of the four-dimensional prediction performance vector when no corresponding prediction task output layer is set.

[0059] Based on the production batch identifier, the product performance test results are matched with the corresponding effective predicted performance vector for this batch to form the prediction residual. The batch optimization sample is associated with at least the magnesite raw material composition data, actual batching results, batch smelting control boundaries, smelting state time series, smelting control action sequence, predicted performance vector, product performance test results, and prediction residual.

[0060] The optimized batch sample is written into the large-crystal magnesia preparation process database. After the raw material composition detection is completed for the next production batch, the batch parameter generation module reads the historical process records that have added the optimized sample for the current batch, and regenerates the batching vector and batch melting control boundary for the next batch.

[0061] The prediction residuals are also used for performance prediction model calibration. When the prediction residuals meet the model calibration trigger conditions, a model calibration sample will be generated for the corresponding completed production batch. The model calibration trigger conditions are determined based on the acceptance tolerance error of the performance prediction model or the statistical range of prediction residuals from historical qualified batches.

[0062] This results in two data-dependent feedback processes. The first feedback process uses the control cycle as its time scale. The actual smelting state enters the adaptive control model, control actions are applied to the equipment, and the actual equipment response enters the performance prediction model. Predicted performance deviations then influence subsequent control cycles. The second feedback process uses the production batch as its time scale. The finished product testing results, along with the batch's raw materials, actual batching, control boundaries, state time series, and action records, form a batch optimization sample. Newly added batch optimization samples participate in the generation of the next batch's batching vector and batch smelting control boundaries.

[0063] When setting digital twin safety constraints, the furnace lining thickness, furnace bottom boundary position, and other structural parameters used to build the spatial model are obtained from the structural parameters of the electric arc melting equipment. The digital twin model updates its internal thermal state based on the actual collected time-series data of temperature, furnace atmosphere, electrical power, and furnace pressure.

[0064] In some preferred embodiments, a discrete computational grid is established vertically within the arc melting equipment. Vertical depth coordinates are then established. The coordinate orientation remains fixed within the same digital twin model, and the furnace bottom depth coordinates are marked as follows: The computational domain and furnace lining boundary conditions are formed using the structural parameters of the electric arc melting equipment. Actual electric power data is used to form the heat source input corresponding to the energy input location; actual temperature data is used for model state correction or temperature boundary correction; furnace atmosphere data and furnace pressure data are used to correct the boundary parameters related to the heat transfer state of the molten pool surface.

[0065] The correspondence between the furnace atmosphere and pressure and the boundary heat transfer parameters is determined by the thermal balance calibration data of the corresponding equipment, historical production data, or the offline calibration process of the digital twin model, and is updated based on the actual collected data during online calculation. The digital twin model generates temperature estimates at discrete depth positions along the vertical axis in the current control cycle through unsteady-state heat conduction calculations, a calibrated reduced-order thermal model, or a thermal state model that can maintain the same input-output relationship, forming a vertical temperature distribution vector.

[0066] Determine the highest temperature value and its corresponding spatial location in the vertical temperature distribution vector. When multiple discrete nodes have the same highest temperature value, select the node with the smallest spatial distance to the furnace bottom depth coordinate as the node corresponding to the highest temperature.

[0067] The absolute spatial distance between the depth coordinate corresponding to the highest temperature and the furnace bottom depth coordinate is defined as the molten pool safety margin.

[0068] ;

[0069] in, Indicates control cycle The safety margin of the molten pool, expressed in meters (m); This indicates the furnace bottom depth coordinates used for safety assessment, in meters. Indicates control cycle The depth coordinates corresponding to the highest temperature in the vertical temperature distribution vector are given in meters. The molten pool safety margin characterizes the degree to which the highest temperature region in the digital twin model approaches the safe judgment position at the furnace bottom. The margin threshold is determined based on the structural design of the electric arc melting equipment, the permissible condition of the furnace bottom refractory lining, equipment maintenance requirements, or the safe operating distance of the equipment.

[0070] When the molten pool safety margin is not less than the set margin threshold, the upper limit of the electrical power limit is not triggered, and the original allowable range of electrical power for the current batch remains in effect. When the molten pool safety margin is less than the set margin threshold, the boundary contraction condition is triggered, and the upper limit of the initial allowable range of electrical power for the current batch is reduced according to the ratio of the molten pool safety margin to the set margin threshold.

[0071] ;

[0072] in, Indicates control cycle The upper limit of the allowable range of electrical power is calculated based on the safety margin of the molten pool, in kW. This indicates the upper limit of the initial allowable range of electrical power generated by the batch parameter generation module for the current raw material batch, in kW; Indicates control cycle The safety margin of the molten pool, expressed in meters (m); This indicates the setting of a margin threshold, in meters (m).

[0073] Since the safety margin of the molten pool and the set margin threshold have the same dimension, and their ratio is a dimensionless quantity, the upper limit of the allowable power range for the current control cycle and the upper limit of the initial allowable power range have the same power dimension. When the upper limit of the allowable power range calculated according to the above ratio is not lower than the lower limit of the current allowable power range, the calculated value is used as the upper limit of the allowable power range for the current control cycle.

[0074] If the proportional reduction result is lower than the lower limit of the current allowable power range, then the proportional result is no longer used to form a normal adaptive control boundary, and the system is processed according to the preset equipment safety control logic. The preset equipment safety control logic includes entering the equipment safety power reduction state, executing equipment-level protection control, or being taken over by equipment safety interlocking, and its execution priority is higher than performance optimization control.

[0075] After generating the contracted batch melting control boundary, it is sent to the adaptive control model. When the action policy network outputs a normalized electric power control action, the physical quantity mapping is re-executed based on the currently effective contracted electric power allowable range. If the mapping result exceeds the upper limit of the currently effective electric power allowable range, the final electric power control command is limited to within the upper limit.

[0076] The power control command is sent to the power regulation device through the execution interface to reduce or limit the power entering the electric arc melting equipment. After execution, the actual equipment status is re-acquired, and the digital twin model and performance prediction model are updated. The digital twin status does not replace the actual sensor status, but is used to form physical constraint information for dynamically adjusting the batch melting control boundary.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] This invention establishes a continuous data dependency and physical control relationship between magnesite raw material composition data, actual batching results, smelting state time series, adaptive control actions, performance prediction results, finished product inspection results, and parameters for the next batch. This ensures that raw material batch information continues to define the smelting control boundary of the current production batch after automatic batching, avoiding the insufficient adaptability of fixed batching schemes and uniform control ranges to different raw material batches. The power, oxygen supply, and stirring adjustments output by the adaptive control model are actually applied to the arc melting equipment by the actuators, forming the smelting state vector for the next control cycle based on the newly acquired actual process data. This avoids replacing the actual equipment response with the predicted state from the internal model, thereby improving control reliability. The performance prediction model outputs predicted values ​​for bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion based on the actual smelting state time series. The performance deviation vector is returned to subsequent control cycles, allowing the final material performance information to participate in the current batch smelting control before finished product inspection, improving product consistency. After the batch is completed, the product performance test results are used to form the predicted residual and batch optimization sample, which are written into the historical process record to update the batching vector and batch melting control boundary for the next batch, forming a closed loop of cross-batch process update.

[0079] By using a proxy optimization model to screen historical batches with adjacent raw material compositions and product performance meeting target constraints, parameters for the current batch are generated from a joint candidate space of batching parameters and smelting control parameters, improving parameter rationality. Multi-sampling rate process data are aligned according to a unified control cycle to generate validity information, reducing the likelihood of different time-location detection values ​​being combined into the same furnace condition at the same control moment. The adaptive control model employs a dual-value evaluation branch and a strategy of updating with lower action values, reducing the risk of overly aggressive control actions due to excessively high single-value evaluations.

[0080] After the predicted bulk density enters the target neighborhood, a conditionally activated stirring adjustment penalty is applied based on the normalized stirring control action changes in adjacent control cycles. This reduces unnecessary fluctuations in stirring control, stabilizing melt flow and crystallization boundaries. The digital twin model reconstructs the vertical temperature distribution based on equipment structural parameters and actual process data, forming a molten pool safety margin. When the molten pool safety margin falls below a set margin threshold, the allowable upper limit of electrical power is dynamically reduced, allowing the estimation of the internal thermal state of the equipment to directly alter the feasible domain of the action strategy, thus improving equipment thermal safety. This invention enhances the batch adaptability of raw materials, the accuracy of process control, the performance prediction guidance capability, and the safety of equipment operation. Attached Figure Description

[0081] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0082] Figure 1 This is a flowchart of the method of the present invention.

[0083] Figure 2 This is a flowchart of the sub-process for time alignment of process data and construction of melting state in the multi-sampling rate process of the present invention.

[0084] Figure 3 This is a flowchart of the intra-batch performance feedback closed-loop process based on the adaptive control model and performance prediction model of the present invention. Detailed Implementation

[0085] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0086] The following is in conjunction with the appendix Figures 1-3 The embodiments of the present invention will be described in detail below.

[0087] In this invention, magnesite raw material composition data refers to the composition data obtained by testing the pretreated magnesite raw material of the current production batch using raw material composition testing equipment. The specific composition detected is determined based on the production line's testing capacity and the production task of large-crystal magnesia, and at least includes the associated components used to describe the main magnesium oxide component and the components actually monitored in the current production task.

[0088] The batching vector refers to the batching target generated by the batch parameter generation module based on the current raw material composition, target performance, and historical process records, and sent to the automatic batching equipment. The actual batching result refers to the actual addition relationship of this batch obtained after the automatic batching equipment completes metering and feeding, based on weighing, metering, or actual feeding records. The batch melting control boundary refers to the process control feasible domain generated for the current production batch and provided to the adaptive control model. It lies within the safe operating range of the equipment and is determined based on the current raw material composition, target material performance, and historical production data. The equipment operating boundary refers to the equipment-level operating limits determined by the structure of the electric arc melting equipment, the electrical system, the furnace condition, and the safety requirements of the production equipment.

[0089] The electrical power regulation quantity refers to the electrical power control command generated by the action strategy network for a control cycle and sent to the electrical power regulation device after mapping to the control boundary of the currently effective batch of smelting. The oxygen supply regulation quantity refers to the oxygen supply control command generated by the action strategy network for a control cycle and sent to the oxygen supply regulation device after mapping to the control boundary of the currently effective batch of smelting. The stirring quantity regulation quantity refers to the stirring control command generated by the action strategy network for a control cycle and sent to the stirring actuator after mapping to the control boundary of the currently effective batch of smelting. The normalized stirring control action refers to the control quantity obtained by converting the stirring quantity regulation quantity to a predetermined dimensionless scale according to the stirring control range corresponding to the control boundary of the currently effective batch of smelting.

[0090] The agitation fluctuation refers to the absolute value of the difference between the normalized agitation control action of the current control cycle and the previous control cycle. The set density deviation threshold is a dimensionless threshold used to determine whether the predicted volumetric density falls within the target neighborhood; it is formed by normalizing the volumetric density target allowable range or based on the normalized deviation distribution of volumetric density from historical qualified batches. The initial control evaluation quantity refers to the control evaluation quantity formed based on performance deviation, energy consumption, and boundary proximity before the addition of agitation adjustment penalties for the volumetric density target neighborhood. The corrected control evaluation quantity refers to the control evaluation quantity formed after deducting the agitation adjustment penalty term based on the volumetric density deviation condition from the initial control evaluation quantity.

[0091] A smelting state vector refers to a data object composed of the time-aligned dynamic process state and the actual batching results within a unified control cycle. A smelting state time series refers to multiple smelting state vectors arranged in chronological order. A target performance vector refers to the target material performance data obtained from the current production task or corresponding quality requirements. A predicted performance vector refers to the predicted results of the performance prediction model, outputting bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion based on the process information already obtained for the current batch. A performance deviation vector refers to the scaled deviation between each performance component in the predicted performance vector and its corresponding target performance component.

[0092] Digital twin state refers to the internal state estimate of the equipment formed by updating the digital twin model of the smelting process based on the structural parameters of the electric arc melting equipment and actual process sensor data. The furnace bottom depth coordinate refers to the furnace bottom boundary position used for equipment thermal risk safety assessment in the unified vertical spatial coordinate system of the digital twin model. The depth coordinate corresponding to the highest temperature refers to the vertical depth coordinate corresponding to the highest temperature value in the vertical temperature distribution vector output by the digital twin model; when multiple highest temperature nodes exist, the node coordinate with the smallest distance from the furnace bottom depth coordinate is used. The molten pool safety margin refers to the absolute spatial distance between the depth coordinate corresponding to the highest temperature and the furnace bottom depth coordinate, used to characterize the degree to which the high-temperature region in the digital twin model approaches the furnace bottom safety assessment position. The set margin threshold refers to the safety margin judgment value determined based on the structural design of the electric arc melting equipment, the permissible state of the furnace bottom refractory lining, maintenance requirements, or the safe operating distance of the equipment.

[0093] Product performance test results refer to the material performance results obtained by finished product testing equipment after the current batch has completed melting and cooling. Prediction residuals refer to the difference between the predicted performance of the current batch and the corresponding product performance test results. Batch optimization samples refer to the data set formed by associating the current batch's raw material composition, actual batching, control boundaries, process status, control actions, predicted performance, product performance test results, and prediction residuals according to the production batch identifier.

[0094] The production system includes magnesite raw material pretreatment equipment, raw material composition testing equipment, automatic batching equipment, automatic material distribution equipment, electric arc melting equipment, process sensing equipment, control and computing equipment, finished product testing equipment, and a database for the preparation process of large-crystal magnesite. The raw material pretreatment equipment is located at the front end of the production process and is used to perform at least one of the following treatments based on the state of the magnesite raw material entering the production line: crushing, screening, and impurity removal, so that the raw material reaches the material state required for subsequent composition testing, metering and batching, and smelting. The raw material composition testing equipment is located after pretreatment and before automatic batching to obtain the composition data of the magnesite raw material for the current production batch.

[0095] The automatic batching equipment receives the batching vector generated by the batch parameter generation module, performs metering and feeding of each material source according to the batching vector, and feeds back the actual batching results to the control computing equipment. The automatic material distribution equipment delivers the batched materials into the electric arc melting equipment and forms a corresponding furnace feed distribution according to the production line layout requirements. The process sensing equipment includes at least a temperature detection device, an atmosphere detection device, a power monitoring device, and a furnace pressure detection device.

[0096] The control computing device establishes data connections with raw material composition detection equipment, automatic batching equipment, process sensing equipment, finished product detection equipment, and the large-crystal magnesia preparation process database. It also connects to the power regulation device, oxygen supply regulation device, and stirring device of the electric arc melting equipment via an execution interface. The control computing device has an internal state buffer for storing control actions from adjacent control cycles. This state buffer stores at least the normalized stirring control actions from the previous control cycle for use in calculating stirring action fluctuations.

[0097] When employing digital twin constraints, the control computing device runs a digital twin model of the melting process, or communicates with the computing device running the digital twin model. The digital twin model receives structural parameters of the arc melting equipment and actual sensor data, and uses the resulting safety margin information to correct the batch melting control boundaries.

[0098] After magnesite raw materials enter the production line, they undergo crushing, screening, and impurity removal based on their initial particle size and impurity status. The output from the crushing stage proceeds to the screening stage. Materials exceeding the target particle size range of the current production line can be returned to the corresponding crushing stage for reprocessing. Materials meeting the requirements of subsequent processes enter the impurity removal stage or the component testing preparation stage. Impurity removal can employ magnetic impurity separation or other mineral sorting methods suitable for magnesite production. When using wet processing, the material is ensured to enter a state where it can be normally processed by the corresponding testing and feeding equipment before subsequent component testing and batching.

[0099] The raw material composition detection equipment acquires the composition data of the current production batch and associates the detected values ​​with the batch identifier, detection time, and raw material source identifier. It performs unit-wide and validity checks on the raw material composition data. If the detection equipment reports an anomaly or the data validity does not meet requirements, re-detection is performed or the anomaly handling process is initiated according to the production line detection rules. Valid raw material composition data is sent to both the batch parameter generation module and the large-crystal magnesia preparation process database.

[0100] The batch parameter generation module reads the current magnesite raw material composition data, target performance vector, and historical process records. First, it filters historical samples suitable for generating parameters for the current batch. Historical sample filtering considers at least the similarity of raw material composition, whether product performance meets current production targets, and the completeness of production records. For cases where different component variables have different numerical scales, scale unification is performed first, and then the compositional similarity between the current batch and historical batches is calculated. Historical batches with significant compositional differences have their reference weight reduced or are excluded from direct candidate samples for the current batch.

[0101] After forming a valid historical sample set, candidate parameter combinations are generated in the joint candidate space composed of batching parameters and smelting control parameters. The batching parameters in the candidate parameter combinations satisfy the material constraints that the actual batching equipment can execute; the candidate smelting parameters satisfy the allowable range of the arc melting equipment. The surrogate model establishes an estimation relationship between candidate parameters and final performance based on historical batch parameters and corresponding product performance. After being filtered by target performance constraints and equipment constraints, the batching vector is determined from the feasible candidate parameters. The candidate smelting parameters that, together with the batching vector, satisfy the performance target form the feasible region of the current batch control variables, and further form the batch smelting control boundary.

[0102] Batch melting control boundaries may include permissible power ranges, permissible power variation ranges between adjacent control cycles, permissible oxygen supply ranges, oxygen supply variation ranges, permissible agitation ranges, and other control action boundaries that the production line actually needs to restrict. Batch melting control boundaries do not replace equipment safety interlocks. Equipment safety interlocks represent safety restrictions that equipment is prohibited from crossing; batch melting control boundaries are process control areas further established within these safety restrictions for the current raw materials and current product objectives.

[0103] The batching vector is sent to the automatic batching equipment. The automatic batching equipment executes the feeding according to the target addition relationship corresponding to each raw material source, and records the actual feeding result through weighing or metering devices. When there is an execution error between the actual feeding result and the target corresponding to the batching vector, the batching equipment corrects it through a weighing control loop. After batching is completed, the actual batching result is generated. The actual batching result is written into the current production batch record and also serves as the batch-level input for subsequent smelting state construction.

[0104] After the material enters the electric arc melting equipment, process sensors begin to collect melting status data. The temperature detection device outputs melting thermal status data. The atmosphere detection device outputs oxygen content in the furnace or other atmosphere status data detected by the production line. The power monitoring device outputs energy input and load-related data for the electric arc melting equipment. The furnace pressure detection device outputs furnace pressure status data. The control calculation equipment sets a unified control cycle.

[0105] For data with a sampling frequency higher than the control frequency, corresponding time-series features are extracted within the current control cycle time window, including one or more of the following: effective value at the end of the cycle, mean, change, trend, and fluctuation characteristics. For data with a sampling frequency lower than the control frequency, the most recent valid measurement value is read in the current control cycle, and the time status of that measurement value from the end of the current control cycle is calculated or recorded to form validity information. If the measurement value has exceeded the allowable valid time of the corresponding sensor, the measurement value is not marked as the latest normal measurement. The source of the data validity time condition is at least one of the following: the normal sampling cycle of the sensor, the sensor response characteristics, and the production line detection and control requirements.

[0106] The time-aligned temperature, furnace atmosphere, electrical power, and furnace pressure states are combined with the actual batching results to form a smelting state vector. The smelting state vectors of adjacent control cycles are stored chronologically, forming a smelting state time series. The smelting state vector of the current control cycle is then fed into the state feature extraction network. If a valid performance deviation vector has already been formed in the previous control cycle, this performance deviation vector, along with the current smelting state, participates in state feature generation.

[0107] In the initial control phase, before performance prediction feedback is generated, the adaptive control model executes control based on the current smelting state and batch smelting control boundaries. After the first set of effective predicted performances is formed, performance deviations are entered into subsequent control cycles. The action strategy network outputs normalized control actions based on state characteristics. These normalized actions are mapped to physical quantities corresponding to the batch smelting control boundaries for electrical power, oxygen supply, and stirring amount. The mapping simultaneously applies action amplitude limits and adjacent control cycle rate limits. The value assessment network evaluates the corresponding control actions based on current state characteristics, actions, and control evaluation information. When using two value assessment branches, each branch uses independent model parameters to form action value evaluations, and the branch with the relatively lower action value evaluation participates in the action strategy update.

[0108] The adaptive control model is trained based on the recorded states from historical batches, actual control actions, the actual states after the actions are executed, and corresponding control evaluation information to form state transition samples. State transitions occurring during online control are written to the database after the batch ends for subsequent offline training or calibration. The strategy generated by the model training is limited by the equipment's allowable control range and the batch smelting control boundaries before being put into production. The adjustments for electrical power, oxygen supply, and stirring are sent to the corresponding actuators via the execution interface. After execution is completed, the actual sensor data is re-acquired in the next control cycle.

[0109] Training samples for the performance prediction model are constructed from historical process records of completed production and finished product testing. Each training sample must contain at least the raw material composition data, actual batching results, melting state time series, and product performance test results for the same production batch. Production batch identifiers serve as the basis for data association. When constructing the supervisory training data, the consistency between the production batch identifiers of various process data and product testing labels is checked.

[0110] The performance prediction model first uses a time-series feature weighting layer to form a weighted sequence of features from different time positions and process states. A Long Short-Term Memory (LSTM) network processes this weighted sequence to generate time-dependent features characterizing the historical changes in the smelting process. A shared feature layer fuses these time-dependent features with the current raw material composition. The volumetric density, apparent porosity, compressive strength, and coefficient of thermal expansion task output layers each output their corresponding performance prediction values ​​based on the shared features.

[0111] During training, each performance component undergoes data standardization or normalization according to its own numerical scale, and the shared feature part is updated jointly by the prediction errors corresponding to multiple tasks. In online applications, performance prediction is triggered at a set prediction cycle or process stage. The performance prediction input uses the state time series already obtained at the current prediction time point and batch-level information determined before the current prediction time point. Subsequent cooling states that have not yet actually occurred are not used as the actual process input at the current time point. If the cooling control program has been determined before production, the pre-determined program parameters are used as known process conditions; the actual cooling process data is written into the corresponding batch history record after it actually occurs.

[0112] After obtaining the predicted performance vector, the target performance vector is read according to the current production task. The directed deviations corresponding to bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion are calculated according to the aforementioned normalized deviation relationship. Each component forms the performance deviation vector. The performance deviation vector continues into the next control cycle. The initial control evaluation further incorporates normalized energy consumption and boundary proximity evaluation quantities. Normalized energy consumption is formed based on the measured electrical energy input of the current control cycle and the reference scale of the corresponding production stage. Boundary proximity evaluation quantities are formed based on the distance between the action and the current batch smelting control boundary. When the performance deviation increases, the performance evaluation item reduces the corresponding action evaluation. When the control action continues to approach the current batch process boundary, the boundary proximity evaluation item reduces the corresponding action evaluation. When the performance deviation state and the process boundary state are similar, the energy consumption evaluation item is used to distinguish actions with different energy inputs.

[0113] During the final stages of smelting or the initial stages of heat preservation and crystallization while still subject to stirring control of the melt, frequent adjustments to the stirring control quantity will alter the melt flow state as the predicted bulk density enters the target allowable neighborhood. To differentiate between the adjustment phase when material properties are far from the target and the maintenance phase after material properties have entered the target neighborhood, a conditionally activated stirring adjustment penalty is set based on the initial control evaluation quantity. The control calculation device stores the stirring quantity adjustment amount of the previous control cycle in the state buffer. After the stirring quantity adjustment amount is generated by the action strategy network for the current control cycle, the stirring quantity adjustment amount of the current control cycle and the stirring quantity adjustment amount of the previous control cycle in the buffer are converted into a unified dimensionless normalized stirring control action according to the currently effective stirring control range. The difference between the normalized stirring control actions of adjacent control cycles is obtained by subtracting the normalized stirring control action of the previous control cycle from the current normalized stirring control action, and its absolute value is taken as the stirring action fluctuation of the current control cycle.

[0114] After the volumetric density task output layer generates the predicted volumetric density for the current control cycle, it calculates the volumetric density normalization deviation based on the target volumetric density value and the corresponding normalization scale. The absolute value of the volumetric density normalization deviation is compared to a set density deviation threshold. If the absolute value of the volumetric density normalization deviation is greater than the set density deviation threshold, the step activation function is set to 0, and no stirring adjustment penalty is added to the initial control evaluation quantity in the current control cycle. If the absolute value of the volumetric density normalization deviation is not greater than the set density deviation threshold, the step activation function is set to 1, and the product of the stirring penalty coefficient and the stirring action fluctuation is used as the stirring adjustment penalty term, which is deducted from the initial control evaluation quantity. The stirring penalty coefficient is initially determined based on the distribution of stirring action changes in adjacent control cycles after the volumetric density of historical qualified production batches enters the target neighborhood, and then calibrated through offline strategy verification or production verification within the equipment's allowable range.

[0115] After utilizing the corrected control evaluation value, the value assessment network is updated based on the state transition data formed by the actual collected state. When two value assessment branches are used, both branches receive training information corresponding to the corrected control evaluation value and participate in the action policy update with the relatively lower action value result. During the policy gradient update process, the greater the fluctuation of the stirring action and the more the volume density is already in the target neighborhood, the lower the corrected control evaluation value obtained by the corresponding action, thereby reducing the tendency to select actions with high stirring fluctuations in the parameter update direction. The control mechanism does not require the stirring amount adjustment to become zero, nor does it require the stirring actuator to stop operating. Its control objective is to reduce unnecessary changes in control commands in adjacent cycles after entering the volume density target neighborhood. If other performance components still deviate significantly from the target, the performance deviation information continues to participate in the initial control evaluation and the state input of the next cycle. Therefore, the stirring stability constraint does not eliminate the influence of other performance objectives on the action policy.

[0116] The control computing equipment determines the target performance constraints and melting state stability conditions in each performance prediction cycle. The target performance constraints are derived from the allowable performance ranges corresponding to the current production task. Melting state stability conditions are determined separately for each current process stage. The allowable temperature and power variations may differ between the initial melting stage and the melt stabilization stage; therefore, the stability conditions correspond to the current process stage. Only when both the target performance constraints and the melting state stability conditions are met simultaneously within the required continuous control cycle is the batch performance feedback considered to have reached the current stage's termination condition. If the joint termination condition is not met, state acquisition, adaptive control, actual execution, state updates, and performance prediction continue.

[0117] When a process sensor fails to output a new measurement value as expected, the system first checks if the most recent valid measurement value is still within the allowed valid time range. If it is, the most recent valid value is used, and the corresponding validity status is retained. If it exceeds the allowed valid time range, the corresponding status is marked as invalid or low-confidence. When the input status is insufficient to support normal model control, a pre-defined degradation control strategy for the production line is executed. This degradation control strategy includes maintaining the most recently verified equipment control state, switching to the equipment's existing conventional control mode, or requesting manual intervention. Equipment safety interlocks act directly on the execution layer. When an equipment safety interlock is triggered, the adaptive control output does not prevent the safety interlock from executing.

[0118] After the current batch completes arc melting, it enters the cooling process. After the specified cooling is completed, the actual product performance is obtained by the finished product testing equipment. When the actual test items correspond to the output of the performance prediction model, at least the bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion are obtained. The product performance test results are matched with the effective predicted performance of the corresponding production batch. The final effective predicted performance of this batch can be used, or the predicted performance corresponding to multiple key process stages can be saved according to the production process needs, and corresponding relationships can be established separately. The prediction residual between the predicted performance and the product performance test results is calculated. The prediction residual is correlated with the raw material composition, actual batching results, batch melting control boundaries, melting state time series, melting control action sequence, predicted performance, and product performance test results of this batch. The resulting optimized batch sample is written into the large-crystal magnesia preparation process database.

[0119] After the raw material composition testing for the next production batch is completed, the batch parameter generation module rereads the historical process records. These records include newly generated batch optimization samples from the previous production batch. If the raw material composition of the new batch is similar to that of the next batch, and the product performance test results meet the corresponding production targets, the actual batching and effective melting control area of ​​the new batch become part of the historical basis for generating parameters for the next batch. If a certain control area consistently corresponds to a large performance deviation across multiple batches, the candidate performance for that area is re-estimated based on the newly added product performance test results during subsequent proxy optimization. If the prediction residual meets the performance prediction model calibration trigger condition, the newly added test samples are used to perform parameter calibration on the performance prediction model.

[0120] In the implementation using digital twin constraints, a digital twin model of the melting process corresponding to the actual electric arc melting equipment is established. The digital twin model receives structural parameters of the electric arc melting equipment and real-time sensor data. The structural parameters of the electric arc melting equipment include at least the furnace geometry, furnace lining thickness, furnace bottom safety determination position, and equipment parameters related to thermal state calculations. The digital twin model employs a physical model, a reduced-order physical model, or a combined physical and data model that reflects the equipment's thermal state, flow state, or safe operating margin. The output of the digital twin model is used to determine the degree of proximity between the current state of the actual equipment and the equipment's operating boundary. The digital twin model does not replace the actual process sensors in generating the actual state of the control closed loop.

[0121] A discrete axial computational grid is established along the vertical direction of the electric arc melting equipment. Each discrete node has a unique depth coordinate. Actual electrical power data is used to update the internal heat source state related to the electric arc energy input. Actual temperature measurements are used to correct the temperature state of the corresponding space or boundary. Furnace atmosphere and furnace pressure information are used to correct the heat exchange boundary parameters corresponding to the molten pool surface or the gas region inside the furnace. The vertical temperature distribution vector for the current control cycle is obtained by solving through time stepping. The spatial depth coordinates of the node corresponding to the highest temperature value are determined by traversing the temperature distribution vector. If there are multiple nodes with the same highest temperature value, the node closest to the safe judgment position at the furnace bottom is selected to ensure a unique result in subsequent safety margin calculations. The absolute distance between the depth coordinates corresponding to the highest temperature and the depth coordinates at the furnace bottom is calculated to obtain the molten pool safety margin.

[0122] When the molten pool safety margin is within the normal range, the adaptive control model operates according to the original smelting control boundary of the current batch. When the molten pool safety margin is less than the set margin threshold, the boundary correction module reduces the upper limit of the allowable power range for subsequent control cycles. The reduced upper limit of the allowable power range is sent to the action strategy network through the batch smelting control boundary input. The normalized power action output by the action strategy network is mapped to physical quantities using the new power range. Control results exceeding the reduced upper limit of the power range are limited to within the allowable range. If the upper limit reduction result formed based on the safety margin is lower than the lower limit of the current batch's allowable power range, the adaptive performance control will not continue to generate action intervals with invalid upper and lower limit relationships, but will instead enter the equipment's preset safety control process.

[0123] When the source of magnesite raw materials changes, the new raw material composition distribution may deviate from the training sample range of the existing performance prediction model. When the production task switches to a new large-crystal magnesia performance grade, the target performance distribution may also change. If the number of batches corresponding to the new scenario that have completed production and obtained product performance test results is small, model calibration can be used first. The performance prediction model maintains the existing structure of the temporal feature weighted layer, long short-term memory network, shared feature layer, and task output layer. New scenario batches that have completed production are used to form model calibration samples. When the sample size is small, the shared temporal feature parameters can be kept unchanged, and the four task output layers are mainly updated. Alternatively, the shared temporal feature part can be adjusted with a parameter update magnitude lower than that of the task output layers. The update of the task output layer uses the new scenario product performance test results as supervision labels. As the number of new scenario batches increases, the decision to update the shared feature part is made based on independent validation samples. The new scenario product performance data is simultaneously written into the batch optimization samples and used to gradually update the batch smelting control boundaries corresponding to the new raw material or new product grade.

[0124] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further illustrated below with reference to specific implementation examples.

[0125] Example 1: A batch of pretreated magnesite raw material is selected. The raw material composition detection equipment obtains the main magnesium oxide components of the current batch and the associated components monitored by the production line, and associates this data with the production batch identifier. The control calculation equipment reads the target performance information corresponding to the current production task, including bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion. The batch parameter generation module reads historical batch data from the large-crystal magnesite preparation process database, and filters historical batches with similar raw material composition, complete records, and product performance test results that meet the corresponding target requirements. Based on the filtered samples, the current batch batching vector and the batch smelting control boundaries corresponding to electrical power, oxygen supply, and stirring amount are formed.

[0126] The automatic batching equipment executes the batching vector. After batching, the actual batching result is formed based on the actual metering records. The material enters the electric arc melting equipment. The process sensing equipment continuously acquires data on temperature, furnace atmosphere, electric power, and furnace pressure. The control computing equipment performs time alignment on the process data according to a unified control cycle and forms a melting state vector by combining the actual batching result with the aligned process state. The melting state vector is input into the state feature extraction network. The action strategy network generates control actions based on the state features and forms electric power adjustment, oxygen supply adjustment, and stirring adjustment based on the current batch melting control boundary. The three adjustment quantities are sent to the corresponding actuators. After the actuators complete their actions, the process sensing data is re-acquired in subsequent control cycles. The newly acquired data is re-time aligned to obtain a new melting state vector. The continuously formed melting states are arranged in chronological order to form a melting state time series. The performance prediction model outputs predicted values ​​for bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion based on the melting state time series and the current batch raw material composition state.

[0127] In one embodiment, the target bulk density is 3.45 g / cm³, the target apparent porosity is 3.0%, the target compressive strength is 100 MPa, and the target coefficient of thermal expansion is 14.0 × 10⁻⁻⁻⁶. 6 / °C. The corresponding normalization scales were 0.10 g / cm³, 1.0 percentage point, 20 MPa, and 1.0 × 10⁻⁻⁻⁶. 6 / °C. During one control cycle, the performance prediction model yielded a bulk density of 3.40 g / cm³, apparent porosity of 3.6%, compressive strength of 92 MPa, and coefficient of thermal expansion of 14.5 × 10⁻⁻⁻⁶. 6 / °C.

[0128] According to the normalized deviation relationship, the magnitudes of the four deviations are 0.50, 0.60, 0.40, and 0.50, respectively. When the component weights of the four performance items are all set to 0.25, the weighted magnitude of the performance deviation is 0.50. Currently, the normalized energy consumption is 0.40, the boundary proximity evaluation quantity is 0.10, the performance deviation item evaluation weight is 0.60, the energy consumption item evaluation weight is 0.25, and the boundary proximity evaluation item weight is 0.15. Calculated based on the initial control evaluation quantity, the performance deviation component is -0.30, the energy consumption component is -0.10, the boundary proximity component is -0.015, and the current initial control evaluation quantity is -0.415. The performance deviation vector and the corresponding control evaluation information are returned to the action policy network and the value assessment network.

[0129] In the next control cycle, the action strategy network selects a new action within the batch melting control boundary. After the action is executed, the actual state is re-acquired. The new state time series corresponds to the predicted performance values ​​of a bulk density of 3.44 g / cm³, apparent porosity of 3.1%, compressive strength of 99 MPa, and coefficient of thermal expansion of 14.1 × 10⁻⁻⁻⁴.6 At / °C, the corresponding normalized deviation amplitudes for the four items are 0.10, 0.10, 0.05, and 0.10, respectively, with a weighted performance deviation of 0.0875. The current normalized energy consumption is 0.42, and since the control action is within the batch melting control boundary, resulting in the boundary approaching evaluation value being 0, calculated using the same evaluation weights, the performance deviation is -0.0525, the energy consumption is -0.105, and the boundary approaching part is 0, resulting in an initial control evaluation value of -0.1575. The system continues to execute the control cycle.

[0130] The batch feedback ends when all four predicted performance parameters are within the current product target allowable range, and the temperature, furnace atmosphere, electrical power, and furnace pressure meet the stable conditions of the current process stage for multiple consecutive control cycles. Product performance test results are obtained after melting and cooling. The system compares the product performance test results with the last effective predicted performance to form a prediction residual. The prediction residual, current raw material composition, actual batching results, batch melting control boundaries, melting state time series, and complete control action records together form the batch optimization sample. The batch optimization sample is written to the large-crystal magnesia preparation process database. When the next batch of raw materials completes composition testing, the batch parameter generation module reads historical records containing data from this batch and redetermines the next batch batching vector and batch melting control boundaries.

[0131] Example 2: This example adds multi-sampling rate state validity control and digital twin boundary correction to Example 1. In this example, magnesite raw material enters the particle size processing flow, and after crushing, screening, and impurity removal, it enters the automatic batching system. The raw material detection equipment obtains the main chemical composition of the current batch. The batching system executes the actual batching based on the output of the batch parameter generation module. The material enters the electric arc melting equipment through the automatic material distribution system. The electric arc melting equipment is equipped with temperature detection devices, oxygen content detection devices, power monitoring devices, and furnace pressure detection devices. Different sensors are configured with sampling frequencies according to the response capabilities of the corresponding equipment. The control calculation equipment adopts a fixed control cycle.

[0132] Each control cycle extracts at least one of the following from the power data sampled at higher frequencies: the last value of the window, the average value of the window, and the change in the window value. Simultaneously, it reads the latest valid values ​​for temperature, oxygen content, and furnace pressure. When a low-frequency sensor does not generate a new detection value in the current control cycle, the system reads the most recent detection data within the allowed valid time range and adds a valid status to the corresponding data. Old data exceeding the valid time range is not treated as the latest normal measurement. The synchronized data, together with the actual batching results, forms a melting state vector.

[0133] A digital twin model of the smelting process is further established. The digital twin model updates the internal thermal state of the equipment based on the structural parameters of the arc melting equipment and real-time sensor status. The digital twin model employs a physical-thermal model, a reduced-order model, or a hybrid model calibrated with equipment structure and historical data. The digital twin output is used to determine the proximity of the current furnace thermal state to the equipment's permissible state boundary. When the digital twin model determines that the internal thermal state is within the normal range, the action strategy network outputs control actions according to the original smelting control boundaries for the current batch. When the digital twin model determines that a certain thermal state is close to the equipment's operating boundary, the boundary correction module reduces the permissible range of subsequent control variables related to the risk state. The control boundary corrected by the digital twin model is sent to the action strategy network. The action strategy network outputs physical control commands within the corrected boundary range.

[0134] In the automated production line, the power regulation device, oxygen supply regulation device, and stirring actuator each receive outputs from the adaptive control model. After each action, the state is retrieved from the actual sensors. The performance prediction model uses the actual state time series for prediction. Performance deviations are then fed back to the adaptive control model. During the finished product inspection stage, testing equipment adapted to the corresponding performance indicators is used to obtain product performance test results. The test results for bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion are used for performance prediction model error assessment. Grain size is obtained through a separate testing process when it is needed for final grade determination. After the production batch is completed, the system writes the product performance test results, quality grade, prediction residuals, and complete process records back to the large-crystal magnesia preparation process database.

[0135] Example 3: This example is a further optimization based on Example 2. In this example, after the new magnesite raw material enters the production line, the raw material pretreatment and component detection are completed in the aforementioned manner. The batch parameter generation module compares the composition status of the new raw material with the raw material composition distribution in the historical process records. For new raw materials with sufficient adjacent historical batch support, the batching vector and batch smelting control boundary are obtained according to the normal batch parameter generation method. For new raw materials with significant differences from the historical main raw material composition and few samples of the new scenario that have completed product performance testing, a relatively limited initial batch smelting control boundary is first formed based on the existing relatively close qualified historical batches and the allowable operating range of the equipment.

[0136] The production process operates using a method of real-time data acquisition, adaptive control, actual execution, and performance prediction. After each new raw material batch is completed, the finished product testing equipment obtains the product performance test results. Model calibration samples are formed using the melting state time series of the corresponding new raw material batch and the product performance test results. In the initial stage with few samples, the shared temporal feature parameters of the performance prediction model are kept unchanged or updated with a parameter update magnitude lower than that of the task output layer. The four task output layers are updated primarily based on the product performance labels of the new scenario. After the model is updated, the prediction results are verified using data from completed batches that did not participate in the corresponding parameter updates. Simultaneously, the batch optimization samples formed by each new raw material production batch are written into the large-crystal magnesia preparation process database. As the number of new raw material production batches increases, the batch parameter generation module gradually increases the reference role of the new raw material product performance samples, and updates the corresponding batching vector and batch melting control boundary for the new raw material scenario accordingly. New target performance level scenarios are processed according to the same data relationships, where the new target performance vector comes from the corresponding production task, and the product performance test results come from the completed new-level batches.

[0137] Example 4: This example is a further optimization of Example 3. It further explains the stable control of stirring action after the predicted bulk density enters the target neighborhood, and the digital twin power boundary contraction process for furnace bottom thermal risk. In the late melting stage or during the initial stages of holding and crystallization while stirring control is still in effect, continuously changing the stirring control command alters the melt flow state and further changes the flow and heat transfer conditions near the solid-liquid interface. When the predicted bulk density has entered the target neighborhood, the changes in stirring control commands from adjacent control cycles are treated as additional evaluation costs, reducing frequent disturbances to the target state in the control strategy.

[0138] The control computing device establishes a continuous state buffer in memory to store the normalized stirring control action time sequence output by the action strategy network. The normalized deviation of the volumetric density in the current control cycle is calculated according to the aforementioned normalized directed deviation relationship. The set density deviation threshold is formed by normalizing the volumetric density target allowable range corresponding to the current product grade. When the absolute value of the normalized volumetric density deviation is greater than the set density deviation threshold, the step activation function is set to 0. When the absolute value of the normalized volumetric density deviation is not greater than the set density deviation threshold, the step activation function is set to 1. The stirring amount adjustment formed by the action strategy network in the current control cycle is read, and the stirring amount adjustment corresponding to the previous control cycle is extracted from the continuous state buffer. The stirring amount adjustment values ​​of the two control cycles are normalized according to the stirring control range effective for the current batch. The difference between the normalized stirring control actions of the current control cycle and the previous control cycle is the difference between the normalized stirring control actions of adjacent control cycles, and its absolute value is the stirring action fluctuation. The control computing device obtains the initial control evaluation quantity formed by the performance feedback module. The step activation function output, the stirring motion fluctuation, and the stirring penalty coefficient are multiplied to form a stirring adjustment penalty term. The adjusted control evaluation term is obtained by subtracting this stirring adjustment penalty term from the initial control evaluation value.

[0139] For example, when the predicted bulk density has not yet entered the target neighborhood defined by the set density deviation threshold, the step activation function is 0, the corrected control evaluation quantity is equal to the initial control evaluation quantity, and the action strategy is based on the evaluation of basic performance deviation, energy consumption, and batch smelting control boundaries. Once the predicted bulk density enters the target neighborhood, the step activation function switches to 1. At this point, the greater the change in normalized stirring control action in adjacent control cycles, the greater the stirring adjustment penalty. The value assessment network updates the value assessment based on the corrected control evaluation quantity and actual state transition information. When using time-difference updates, the time-difference error is formed based on the corrected control evaluation quantity, the value estimate of the next actual state, and the action value of the current state. The action strategy network updates parameters through strategy gradients based on the value assessment results, ensuring that actions with smaller stirring changes in adjacent cycles after the predicted bulk density has entered the target neighborhood obtain a relatively favorable strategy update direction when other evaluation conditions are comparable.

[0140] Furthermore, a one-dimensional discrete spatial grid is established within the digital twin model along the vertical axis of the arc melting equipment. In one embodiment, the grid is set to 100 discrete nodes, arranged sequentially from the selected upper reference plane of the furnace body towards the furnace bottom. The actual number of nodes is determined based on the equipment geometry, discrete error requirements, and the computational load allowed per control cycle. A unified vertical depth coordinate system is established, with the depth direction increasing towards the furnace bottom. The furnace lining thickness, furnace bottom depth coordinates, and geometric parameters required for the digital twin thermal model are obtained from the structural parameters of the arc melting equipment. The actual collected temperature, furnace atmosphere, electrical power, and furnace pressure time-series data are input.

[0141] Electrical power data serves as the basis for updating the state of the heat source related to energy input. Temperature detection data serves as the basis for correcting the state of the digital twin model. Furnace atmosphere and pressure data are used to correct the boundary state of heat exchange between the furnace gas and the molten pool surface. The heat transfer parameter mapping relationship is pre-established based on the corresponding equipment's heat balance data, historical production data, or offline thermal model calibration data. Time-step calculations are performed using an unsteady-state heat conduction model or a reduced-order thermal model calibrated with actual data to obtain the temperature estimates of each vertical discrete node in the current control cycle, and a vertical temperature distribution vector is formed according to the spatial order of the nodes.

[0142] A maximum value search is performed on the vertical temperature distribution vector. If a unique highest temperature node exists, its corresponding depth coordinates are obtained. If multiple nodes with the same highest temperature exist, the node with the smallest distance to the furnace bottom depth coordinates is selected to ensure that subsequent safety margins use the highest temperature region location closer to the furnace bottom. The molten pool safety margin is formed based on the absolute distance between the depth coordinates corresponding to the highest temperature and the furnace bottom depth coordinates. The set margin threshold is determined based on the furnace bottom structure design, the allowable state of the refractory lining, or the equipment's safe operation requirements. When the molten pool safety margin is not less than the set margin threshold, the upper limit of electrical power is not triggered. When the molten pool safety margin is less than the set margin threshold, the boundary contraction condition is triggered. The boundary management module reads the initial upper limit of the allowable electrical power range for the current batch. A new upper limit of the allowable electrical power range is formed based on the ratio of the molten pool safety margin to the set margin threshold. The contracted allowable electrical power range is written into the currently effective batch smelting control boundary. The adaptive control model uses the new allowable electrical power range in the next normalized action mapping. If the power control command obtained from the normalized power action mapping of the action strategy network exceeds the upper limit of the current allowable power range, it is restricted to within the upper limit. The power control command is sent to the power regulation device through the execution interface, so that the actual energy input is constrained by the contracted control boundary. After execution, temperature, furnace atmosphere, power, and furnace pressure data are re-acquired, and the digital twin state is updated again in subsequent control cycles.

[0143] In one embodiment, the target bulk density is not less than 3.30 g / cm³, the density deviation threshold is set to 0.05, the stirring penalty coefficient is set to 2.0, the furnace bottom depth coordinate is set to 2.50 m, the margin threshold is set to 0.40 m, the upper limit of the allowable range of the initial electric power of the current batch is set to 6500 kW, and the unified control cycle is set to 10 s.

[0144] Table 1. Timing table of control parameters for smelting large-crystal magnesia;

[0145] Control cycle number Volume density normalized deviation absolute value Step activation function value Normalized stirring action fluctuation degree Control evaluation correction amount Maximum temperature depth coordinate (m) Melt pool safety margin (m) Current effective electric power upper limit (kW) Electric power control command after boundary restriction (kW) 1450 0.18 0 0.12 0 1.85 0.65 6500 6250 1451 0.15 0 0.15 0 1.90 0.60 6500 6380 1452 0.08 0 0.18 0 2.05 0.45 6500 6490 1453 0.04 1 0.15 -0.30 2.12 0.38 6175 6175 1454 0.02 1 0.02 -0.04 2.15 0.35 5687.5 5687.5 1455 0.01 1 0.00 0 2.15 0.35 5687.5 5520

[0146] During control cycles 1450 to 1452, the absolute value of the normalized deviation of the volumetric density was greater than 0.05, the step activation function was set to 0, the stirring regulation penalty term was 0, and the corrected control evaluation value was equal to the initial control evaluation value. The maximum temperature depth coordinate gradually changed from 1.85m to 2.05m, and the corresponding molten pool safety margin changed from 0.65m to 0.45m, but remained no less than the set margin threshold of 0.40m. Therefore, the upper limit of the allowable power range remained at 6500kW.

[0147] Upon entering control cycle 1453, the absolute value of the normalized deviation of the volume density is 0.04, not exceeding 0.05, and the step activation function is set to 1. The normalized stirring action fluctuation is 0.15, the stirring penalty coefficient is 2.0, resulting in a deduction of 0.30 for the initial control evaluation quantity. In the same control cycle, the highest temperature depth coordinate is 2.12m, the furnace bottom depth coordinate is 2.50m, and the molten pool safety margin is 0.38m. 0.38m is less than the set margin threshold of 0.40m, thus triggering power boundary contraction. The current upper limit of the allowable power range is 6175kW. When the normalized power action of the action strategy network corresponds to a power control command higher than 6175kW under the original initial boundary mapping, after remapping and limiting the boundary after contraction, the final control command does not exceed 6175kW.

[0148] In control cycle 1454, the absolute value of the normalized deviation of the volumetric density is 0.02, and the step activation function remains at 1. The normalized stirring action fluctuation is 0.02, resulting in a deduction of 0.04 to the initial control evaluation value. The highest temperature corresponds to a depth coordinate of 2.15m, the molten pool safety margin is 0.35m, and the current upper limit of the allowable power range is 5687.5kW. In control cycle 1455, the normalized stirring action fluctuation is 0, therefore the stirring adjustment penalty term is 0. This state indicates that the normalized stirring control action of the current control cycle is the same as the previous control cycle. The molten pool safety margin remains at 0.35m, therefore the current upper limit of the allowable power range continues to be 5687.5kW. When the power control command generated by the action strategy network after mapping is 5520kW, this control command is lower than the current allowable upper limit and is executed by the power regulation device.

[0149] A mechanism for evaluating changes in stirring action triggered by the target neighborhood of volume density, and a mechanism for adjusting the feasible domain of electric power action triggered by the thermal state of the furnace bottom. The former changes the control evaluation information used by the value assessment network for updating the action strategy, while the latter changes the physical boundary used by the electric power action mapping in the action strategy network. After the control action is executed, the actual sensor data is reread, and based on this, the melting state, performance prediction, and digital twin state for the next control cycle are formed.

[0150] Furthermore, the intelligent preparation and performance prediction system for large-crystal magnesia includes raw material composition detection equipment, automatic batching equipment, process sensing equipment, electric arc melting equipment, finished product detection equipment, control computing equipment, a large-crystal magnesia preparation process database, and execution interfaces. The control computing equipment includes at least one processor, a memory connected to the processor, and an industrial communication interface. The batch parameter generation module, melting state construction module, adaptive control module, performance prediction module, performance feedback module, and batch update module are jointly implemented by the program instructions and processor in the control computing equipment, or executed separately by distributed control computing equipment and exchanged data through a production network.

[0151] The data output interface of the raw material composition detection equipment is connected to the batch parameter generation module. The automatic batching equipment receives the batching vector and outputs the actual batching result to the smelting state construction module. The data output interface of the process sensing equipment is connected to the smelting state construction module. The output of the adaptive control module is connected to the power regulation device, oxygen supply regulation device, and stirring device through the execution interface. The adaptive control module stores the action information of adjacent control cycles in its internal storage or connected storage area. When the stirring action fluctuation penalty is enabled, the storage area stores at least the normalized stirring control action of the previous control cycle. The performance feedback module provides the adaptive control module with a performance deviation vector and data for constructing the initial control evaluation quantity. The value assessment network evaluates the state action based on the initial control evaluation quantity or based on the corrected control evaluation quantity after the volume density enters the target neighborhood.

[0152] The performance prediction module receives data from the melting state time series re-acquired after the actual execution of actions. The performance prediction module outputs data to the performance feedback module. Performance deviation information from the performance feedback module is then fed into the adaptive control module for subsequent control cycles. In systems employing digital twin boundary constraints, the control computing device also includes or is connected to a digital twin computing module. The digital twin computing module receives structural parameters of the arc melting equipment, temperature data, furnace atmosphere data, electric power data, and furnace pressure data, forming a temperature distribution along the vertical axis and the corresponding molten pool safety margin. The boundary correction module updates the upper limit of the allowable electric power range based on a comparison between the molten pool safety margin and a set margin threshold. The updated upper limit of the allowable electric power range is written into the currently effective batch melting control boundary and serves as the physical boundary for the normalized electric power action mapping of the adaptive control module.

[0153] The data output of the finished product testing equipment is connected to the batch update module. The batch update module writes the optimized batch samples into the large-crystal magnesia preparation process database. The batch parameter generation module reads the updated large-crystal magnesia preparation process database again when the next production batch starts. The memory of the control computing device stores program instructions for performing data time alignment, batch parameter generation, adaptive control, performance prediction, performance deviation feedback, stirring action fluctuation evaluation, digital twin status update, batch smelting control boundary correction, and batch sample update. After the processor executes the program instructions, it generates corresponding calculation results based on the actual input data and receives equipment testing data or outputs control signals through the industrial communication interface.

[0154] A state buffer for evaluating stirring fluctuations is stored in the memory. At the end of each control cycle, the normalized stirring control action for that cycle is written into the state buffer and read as data from the previous control cycle in the next control cycle. Equipment structural parameters for digital twin temperature distribution calculations are stored in the equipment parameter area. These equipment structural parameters correspond to specific arc melting equipment, and are updated when the equipment structure, furnace lining condition, or furnace bottom boundary parameters used for safety assessment change. The program instructions are stored in a non-transitory computer-readable storage medium. When the program instructions are executed by the processor, the processor completes data processing, model calculation, control signal generation, and database update steps; raw material detection, batching, melting, process detection, control action execution, and finished product performance detection are respectively implemented by the corresponding physical devices connected to the control and calculation equipment.

[0155] 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.

[0156] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent preparation and performance prediction of large-crystalline magnesia, characterized in that, include: Acquire pre-processed magnesite raw material composition data, target performance vector, and historical process records. Generate the batching vector and batch smelting control boundary for the current batch based on the magnesite raw material composition data, target performance vector, and historical process records. Perform batching according to the batching vector to obtain the actual batching result. The time sequence data of temperature, furnace atmosphere, electric power and furnace pressure in the smelting process are time-aligned according to a unified control cycle, and the time-aligned process data is fused with the actual batching results to form a smelting state vector. The melting state vector is input into the adaptive control model constrained by the batch melting control boundary to generate the adjustment of electric power, oxygen supply and stirring amount. After being acted on by the corresponding actuators in the electric arc melting equipment, the actual process data is re-acquired to form an updated melting state time series. The updated smelting state time series is input into the performance prediction model to obtain the predicted performance vector. The predicted performance vector is compared with the target performance vector to form a performance deviation vector. The performance deviation vector is used as additional state information and control evaluation information for subsequent control cycles. The actual state acquisition, adaptive control, physical execution and performance prediction are executed cyclically until the predicted performance meets the target performance constraint and the smelting state meets the stability condition. After the batch is completed, the product performance test results obtained from the finished product test are used to form a prediction residual with the corresponding predicted performance vector. The prediction residual, actual batching results, batch melting control boundary, melting state time series and melting control record are associated to form a batch optimization sample and written into the historical process record for use in the generation of batching vector and batch melting control boundary for the next batch.

2. The intelligent preparation and performance prediction method for large-crystalline magnesia according to claim 1, characterized in that: Based on the current magnesite raw material composition, historical batch samples with adjacent raw material compositions and product performance meeting the corresponding target constraints are selected from historical process records. In the joint candidate space composed of batching parameters and smelting control parameters, a surrogate optimization model is used to generate candidate parameter combinations based on the historical batch samples. The batching parameters of the current batch are determined from the candidate parameter combinations that meet the target performance constraints and equipment operation constraints, and the batch smelting control boundary is formed based on the feasible region of the corresponding candidate smelting control parameters.

3. The intelligent preparation and performance prediction method for large-crystal magnesia according to claim 1, characterized in that: The time alignment includes: extracting at least one of the following from process data with multiple sampling points within a control cycle: the effective value at the end of the cycle, statistical features, and change features; obtaining the most recent effective measurement value for process data with a sampling frequency lower than the control frequency according to the timestamp, and generating validity information indicating the timeliness of the measurement value; and combining the process features with a common control cycle identifier, the validity information, and the actual batching results to form the smelting state vector.

4. The intelligent preparation and performance prediction method for large-crystal magnesia according to claim 1, characterized in that: The adaptive control model includes a state feature extraction network, an action strategy network, and a value evaluation network. The state feature extraction network generates state features based on the smelting state vector and the performance deviation vector received in subsequent control cycles. The action strategy network generates normalized control actions based on the state features and maps the normalized control actions to the control range defined by the batch smelting control boundary, forming the adjustment quantities of electric power, oxygen supply, and stirring amount. The value evaluation network updates the action strategy network based on the state features, control actions, and control evaluation information corresponding to the performance deviations.

5. The intelligent preparation and performance prediction method for large-crystal magnesia according to claim 4, characterized in that: The value assessment network includes a first value assessment branch and a second value assessment branch that are independent of each other. The first value assessment branch and the second value assessment branch evaluate the same state characteristics and control actions respectively, and update the action strategy network using the evaluation result corresponding to the lower action value among the two evaluation results. After the electric power, oxygen supply and stirring amount are adjusted by the actual actuator, the smelting state of the next control cycle is formed by the re-acquired actual process data, instead of replacing the actual process data with the predicted state generated by the adaptive control model or process simulation model.

6. The intelligent preparation and performance prediction method for large-crystalline magnesia according to claim 4, characterized in that: When the value assessment network updates the action strategy network based on state characteristics, control actions, and control evaluation information corresponding to performance deviations, it normalizes the stirring amount adjustment in the current control cycle and the previous control cycle according to the currently effective batch smelting control boundary, and forms the stirring action fluctuation based on the absolute value of the difference between the two normalized stirring amount adjustment values. When the absolute value of the normalized volume density deviation in the performance deviation vector is not greater than the set density deviation threshold, the stirring adjustment penalty term is activated, and the stirring adjustment penalty term is positively correlated with the stirring action fluctuation and the stirring penalty coefficient. When the absolute value of the normalized volume density deviation is greater than the set density deviation threshold, the stirring adjustment penalty term is not activated. The modified control evaluation quantity, obtained by subtracting the stirring adjustment penalty term from the initial control evaluation quantity, is used to replace the initial control evaluation quantity, and the action policy network is updated based on the policy gradient to suppress the fluctuation of the stirring adjustment quantity in adjacent control cycles after the predicted volume density enters the target neighborhood defined by the set density deviation threshold.

7. The intelligent preparation and performance prediction method for large-crystal magnesia according to claim 1, characterized in that: The performance prediction model includes a time-series feature weighting layer, a long short-term memory network, a shared feature layer, and multiple task output layers. The time-series feature weighting layer generates a weighted state sequence based on the correlation between different time positions and different process features and the material properties. The long short-term memory network extracts the time-dependent features from the weighted state sequence. The shared feature layer integrates the time-dependent features and magnesite raw material composition information. The multiple task output layers output predicted values ​​of bulk density, apparent porosity, compressive strength, and coefficient of thermal expansion, respectively, and form the predicted performance vector.

8. The intelligent preparation and performance prediction method for large-crystal magnesia according to claim 1, characterized in that: The difference between each performance component in the predicted performance vector and the target performance vector is processed according to the corresponding normalized scale to form a performance deviation vector. The normalized scale is determined based on the target allowable range of the corresponding performance component, the statistical scale of historical qualified batches, or the process scale verified by production. The control evaluation information is formed based on the normalized deviation of each performance component, the energy consumption corresponding to the current control cycle, and the degree to which the control action approaches the batch smelting control boundary.

9. The intelligent preparation and performance prediction method for large-crystal magnesia according to claim 1, characterized in that: The stability conditions are determined by the range of temperature, furnace atmosphere, electric power, and furnace pressure changes corresponding to the current smelting process stage. In-batch performance feedback ends only when all performance components in the predicted performance vector meet the target allowable range, and the temperature, furnace atmosphere, electric power, and furnace pressure meet the stability conditions for multiple consecutive control cycles. When the validity information of the process data does not meet the input requirements, performance convergence judgment based on the corresponding invalid process data is suspended, and a preset safety control strategy is executed. The equipment safety interlock of the electric arc melting equipment has a higher execution priority than the control output of the adaptive control model.

10. A smart preparation and performance prediction system for large crystalline magnesia, used to execute the smart preparation and performance prediction method for large crystalline magnesia as described in any one of claims 1-9, characterized in that: It includes a control computing device, a database for the preparation process of large crystal magnesia, a raw material composition detection device, an automatic batching device, a process sensing device, an electric arc melting device, a finished product detection device, and an execution interface connecting the control computing device and the actuator of the electric arc melting device; The control computing device includes a batch parameter generation module, a smelting state construction module, an adaptive control module, a performance prediction module, a performance feedback module, and a batch update module; The batch parameter generation module receives magnesite raw material composition data output by the raw material composition detection device and reads historical process records from the large crystal magnesia preparation process database to generate batching vectors and batch smelting control boundaries. The batching vectors are sent to the automatic batching equipment for execution. The smelting state construction module receives the actual batching results fed back by the automatic batching equipment and the time-series data of temperature, furnace atmosphere, electric power and furnace pressure collected by the process sensing equipment, which are used to form a smelting state vector and a smelting state time series. The adaptive control module receives the melting state vector, the batch melting control boundary, and the performance deviation information formed by the performance feedback module. It sends the adjustment amount to the electric power adjustment device, oxygen supply adjustment device, and stirring device of the electric arc melting equipment through the execution interface, and receives the re-collected melting state after the adjustment amount is executed. The performance prediction module outputs a predicted performance vector based on the melting state time series formed after the adjustment is executed, and the performance feedback module forms performance deviation information based on the predicted performance vector and the target performance vector and sends it to the adaptive control module. The batch update module receives the product performance test results output by the finished product testing equipment and the smelting control record of the current batch, forms a batch optimization sample including the predicted residual, and writes it into the large crystal magnesia preparation process database for use by the batch parameter generation module of subsequent production batches.

Images