AI-based method and system for compounding magnesia-calcia sand with electrically fused magnesia-calcia sand

Abstract

The application discloses an AI-based magnesium-calcium sand and electrically fused magnesium-calcium sand compounding method and system, relates to the technical field of magnesium-calcium refractory ceramic preparation and intelligent manufacturing, and solves the technical problem that the raw material state changes of different batches of magnesium-calcium sand and electrically fused magnesium-calcium sand are difficult to form closed-loop control. The application obtains a raw material state vector, determines target compounding parameters and initial process parameters in combination with quality constraints and production line process boundaries, collects process time sequence data in compounding processing and predicts deviation trends to form dynamic control constraints, performs quality detection on a compounding product, takes quality results as feedback, and stores the quality results according to batches for control correction and model updating. The method is beneficial to improving the adaptability of a production process to raw material batch fluctuations and is suitable for the production of green special refractory ceramics.

Description

Technical Field

[0001] This invention relates to the field of magnesium-calcium refractory ceramics preparation and intelligent manufacturing technology, specifically to an AI-based composite method and system for magnesium-calcium sand and fused magnesium-calcium sand. Background Technology

[0002] Magnesia-calcium refractories are alkaline refractories with magnesium oxide and calcium oxide as the main refractory components. They can be used in working linings, functional refractory components, or related refractory products in steel smelting and other high-temperature industrial equipment. Magnesia-calcium sand can be produced by sintering, while fused magnesia-calcium sand is produced by electrofusion. The two types of raw materials have different material properties in terms of grain state, density, chemical composition uniformity, and particle size. Therefore, in the specific production process of refractory products, the combination of magnesia-calcium sand and fused magnesia-calcium sand, according to the product performance requirements, is a crucial factor affecting the performance of magnesia-calcium refractories.

[0003] The production of refractory materials typically involves multiple processes, including raw material inspection, batching, mixing, molding, and heat treatment. Even for raw materials of the same nominal category, different production batches may still exhibit differences in chemical composition, impurity composition, moisture content, and particle size distribution. When production formulas and process settings are primarily executed according to fixed formulas or predetermined process parameters, changes in the state of the raw materials may render the original compounding relationships and process parameters incompatible with the current state of the raw materials.

[0004] The blending relationship between magnesia-calcium sand and fused magnesia-calcium sand is strongly correlated with subsequent process control. Changes in the mass distribution or particle size distribution between magnesia-calcium sand and fused magnesia-calcium sand may alter the appropriate process parameters for mixing, molding, and heat treatment. If only the raw material formulation is optimized while subsequent processes continue to use fixed process parameters, the blending result cannot be applied to the actual production process. Furthermore, if only process sensor data is used to adjust parameters such as temperature, pressure, or time without considering the current batch's raw material state and the actual blending result, it is difficult to accurately distinguish between process deviations caused by changes in raw material state and those caused by disturbances in production equipment.

[0005] Current production processes are typically controlled based on the deviation between current detected values ​​and setpoints. When the process exhibits thermal inertia, material residence time, or process response delays, compensation is implemented only after the actual detected value reaches or exceeds the process boundary, reducing the available adjustment range. While timing prediction methods can obtain trends in subsequent process states, if the predicted results are only used for alarms and do not contribute to the formation of the actual control action range, a direct data dependency remains between the prediction process and the equipment control process.

[0006] Furthermore, product surface defect detection, chemical composition testing, and production process data typically originate from different production stages. If product quality inspection results are only used for product conformity determination and are not fed back to the compounding and process control processes that produced the product, it is difficult to establish a complete correlation between control actions, production process status, and final product quality, and it is also not conducive to using actual product quality results to correct subsequent production processes. Summary of the Invention

[0007] The purpose of this invention is to provide an AI-based method and system for compounding magnesium calcium sand and fused magnesium calcium sand. This invention solves the technical problems in the existing production process of compounding magnesium calcium sand and fused magnesium calcium sand, such as the inability to timely adjust the compounding parameters and process parameters in response to batch fluctuations in raw materials, the disconnect between process state prediction results and actual control actions, and the inability to effectively feed product quality inspection results back to the production control process. These problems make it difficult to form a closed-loop production system covering raw material state, compounding parameters, process control, and quality inspection.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand includes:

[0010] S1: Obtain the chemical composition, particle size distribution, impurity content and moisture data of the current batch of magnesium calcium sand and fused magnesium calcium sand respectively. Perform source consistency verification, anomaly handling and normalization on the obtained data to form a raw material state vector, and obtain the quality constraints of the target product and the process boundary of the production line.

[0011] S2, input the raw material state vector and candidate compounding parameters into the material performance prediction and feasibility judgment model, determine the target compounding parameters within the quality constraints and production line process boundaries, and input the target compounding parameters and raw material state vector into the process parameter prediction model to obtain the initial process parameters for composite production;

[0012] S3, according to the target compounding parameters, measure and mix magnesium calcium sand and fused magnesium calcium sand, perform at least one of the composite treatments of molding and heat treatment according to the initial process parameters, and collect at least one of the process timing data of temperature, time, cooling rate and pressure during the composite treatment process;

[0013] S4, predict the deviation trend of subsequent process states relative to the process boundary of the production line based on process timing data, generate dynamic control constraints based on the deviation trend, and adjust the corresponding process parameters within the action range limited by the dynamic control constraints.

[0014] S5 performs surface defect detection and / or composition uniformity detection on the composite product to generate quality results. The quality results are used as feedback for the control strategy. The quality results are stored in batch association with the actual compounding parameters and process time data for subsequent control correction of the current batch and / or model update of subsequent batches. The output is magnesium-calcium special refractory ceramic composite product and corresponding batch traceability data.

[0015] As a preferred technical solution of the present invention, the source consistency verification includes: obtaining data from at least two chemical component detection sources for the same batch of raw materials, and determining the source difference based on the detection results of the corresponding chemical components; when the source difference exceeds the consistency threshold, marking the corresponding raw material status as an abnormal state that cannot be directly entered into the compounding parameter calculation, and triggering at least one of the following operations: resampling, re-detection, and detection equipment status check; the consistency threshold is determined based on the repeatability data of the detection equipment and the source difference distribution of historical valid samples.

[0016] As a preferred technical solution of the present invention, the material performance prediction and feasibility discrimination model includes a continuous performance prediction model, a feasibility discrimination model, and a candidate parameter search model; the continuous performance prediction model obtains the corresponding material performance prediction results based on the raw material state vector and candidate compounding parameters; the feasibility discrimination model filters out candidate compounding parameters that do not meet the quality constraints based on the material performance prediction results and the quality constraints of the target product; the candidate parameter search model determines the subsequent parameters to be evaluated based on the evaluated candidate compounding parameters and their evaluation results, and determines the target compounding parameter among the feasible candidate compounding parameters.

[0017] As a preferred technical solution of the present invention, the candidate compounding parameters include at least the mass distribution parameters of magnesium calcium sand and fused magnesium calcium sand, and at least one of the gradation parameters of raw materials of different particle sizes; when generating candidate compounding parameters, a total amount conservation constraint is applied to the mass distribution parameters of each raw material, and the value boundary of the candidate compounding parameters is determined according to the executable range of the feeding equipment, the current raw material inventory status and the quality constraint of the target product.

[0018] As a preferred technical solution of the present invention, the input of the process parameter prediction model includes at least the target compounding parameter, the current batch raw material state vector and the current state of the production line, and the output includes at least one of the mixing parameter, heat treatment parameter, heat preservation parameter, cooling parameter and molding pressure parameter; the output of the process parameter prediction model is used as the initial set value of the production controller after being verified by the production line process boundary, and the production controller performs basic closed-loop control according to the deviation between the actual measured value and the initial set value.

[0019] As a preferred embodiment of the present invention, step S4 includes: using a time-series prediction model to predict the process state within a subsequent time range based on process time-series data within a preset time window; comparing the predicted process state with the corresponding production line process boundary to determine the degree of prediction deviation; when the predicted process state tends towards or exceeds the corresponding production line process boundary, narrowing the allowable action range of the control strategy for the corresponding process parameter and / or changing the allowable adjustment direction of the corresponding process parameter; when the degree of prediction deviation reaches the abnormal triggering condition, switching to the preset safe process parameter and / or pausing to enter the next process.

[0020] As a preferred technical solution of the present invention, the control strategy adopts a reinforcement learning control model. The reinforcement learning control model generates control actions based on the production process status. The control actions include at least increasing, maintaining, and decreasing the corresponding process setpoints. Process deviations, energy consumption, and abnormal states during the production process form stage feedback, and the quality results obtained after the batch is completed form terminal feedback, so as to establish a correlation between the quality inspection results and the process control actions that generated the quality inspection results.

[0021] As a preferred technical solution of the present invention, surface defect detection includes: acquiring a surface image of the composite product under controlled illumination conditions, performing standardized preprocessing on the surface image, and inputting it into a convolutional neural network to obtain a detection result including at least one of the following: qualified state, crack state, inclusion state, and porosity state; composition uniformity detection includes performing chemical composition detection at at least one location on the same batch of composite products, and when the detection result does not meet the batch composition consistency condition, marking the batch quality result as abnormal and generating a corresponding process feedback identifier.

[0022] As a preferred technical solution of the present invention, the raw material state vector, target compounding parameters, actual executed formula, initial process parameters, process timing data, control actions and quality results of each production batch are associated with the same batch identifier to form a batch sample; when a new batch sample meets the preset update conditions and / or a change in the distribution of input data is detected, at least one of the material performance prediction model, process parameter prediction model, timing prediction model and control strategy is updated, and the model versions before and after the update and the corresponding applicable data range are saved.

[0023] This invention also discloses an AI-based composite system for magnesia-calcium sand and fused magnesia-calcium sand, used to execute the AI-based composite method for magnesia-calcium sand and fused magnesia-calcium sand described above. The system includes a raw material data acquisition unit, an intelligent decision-making unit, a production execution unit, a process monitoring unit, a quality inspection unit, and a closed-loop control unit. The raw material data acquisition unit includes a data interface connected to at least one of a chemical composition detection device, a particle size detection device, and a moisture detection device, used to acquire raw material detection data of magnesia-calcium sand and fused magnesia-calcium sand and form a raw material state vector. The intelligent decision-making unit includes a processor and a memory, used to receive the raw material state vector, target product quality constraints, and production line process boundaries, perform material performance prediction, feasibility assessment, and compounding parameter optimization, output target compounding parameters, and determine initial process parameters based on the target compounding parameters and the raw material state vector. The production execution unit includes a raw material feeding unit. The system comprises at least one of equipment, mixing equipment, molding equipment, and heat treatment equipment, and includes a PLC controller and / or DCS controller connected to the equipment for performing composite production according to target compounding parameters and initial process parameters; a process monitoring unit includes process sensors installed in the production execution unit and data acquisition interfaces connected to the process sensors for collecting process time-series data and forming dynamic control constraints sent to the production execution unit based on the predicted deviation trend of the process time-series data; a quality inspection unit includes a machine vision acquisition device and / or a chemical composition detection device for generating quality results of the composite product; and a closed-loop control unit is communicatively connected to the intelligent decision-making unit, the production execution unit, the process monitoring unit, and the quality inspection unit for feeding back the quality results to the control strategy and storing the raw material status, compounding parameters, production process data, and quality results of the same batch in association.

[0024] As a preferred embodiment of the present invention, it further includes an industrial communication gateway, a batch data storage device, and a management server; the industrial communication gateway is connected to the PLC system and / or DCS system of the production line; the intelligent decision-making unit and the closed-loop control unit are set in a distributed computing environment consisting of field edge computing devices, production control servers, or both; when the communication between the intelligent decision-making unit and the production execution unit is interrupted and / or the model output exceeds the process boundary of the production line, the PLC controller and / or DCS controller maintain the most recent valid safety setting value or switch to the preset safety control logic.

[0025] In addition, the present invention also discloses an electronic device, including a processor, a memory, and a data interface for communicating with refractory material production equipment. The memory stores a computer program, and when the computer program is executed by the processor, the electronic device performs the above method.

[0026] In addition, the present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above method performed by a computing device.

[0027] As a preferred technical solution of the present invention, in step S2, before inputting the raw material state vector and candidate compounding parameters into the material performance prediction and feasibility judgment model and determining the target compounding parameters within the quality constraints and production line process boundaries, the method further includes: extracting the mass fraction of first calcium oxide, the mass fraction of second calcium oxide, the average particle size of magnesia-calcium sand, and the average particle size of fused magnesia-calcium sand from the raw material state vector; calculating the interface thermal stress mismatch factor based on the mass fraction of first calcium oxide, the mass fraction of second calcium oxide, the average particle size of magnesia-calcium sand, and the average particle size of fused magnesia-calcium sand; the calculation formula for the interface thermal stress mismatch factor is:

[0028]

[0029] In the formula, The interfacial thermal stress mismatch factor is expressed in units of 1. ; The mass fraction of calcium oxide in magnesium calcium sand is dimensionless. The mass fraction of second calcium oxide in fused magnesium calcium sand is dimensionless. The average particle size of fused magnesia-calcium sand is given in units of 1. ; The average particle size of magnesium-calcium sand is given in units of 1. ; The preset basic coefficient of thermal expansion, in units of The interfacial thermal stress mismatch factor is spliced ​​into the raw material state vector. When the interfacial thermal stress mismatch factor is greater than the preset mismatch threshold, the pore-forming agent mass distribution parameter for water treatment filtration application is added to the candidate compounding parameters, and the material performance prediction and feasibility judgment model is controlled to reduce the predicted interfacial strength value corresponding to the candidate compounding parameter by a preset ratio.

[0030] As a preferred technical solution of the present invention, in step S3, at least one composite treatment in molding and heat treatment is performed according to the initial process parameters, including performing heat treatment; moisture data is read from the raw material state vector, and the moisture data is input into the hydration vapor pressure prediction model built into the intelligent decision unit to calculate the critical heating rate during the heat treatment process; the critical heating rate is obtained by dividing the product of the maximum allowable internal tensile strength of the billet and the square of the average pore radius by the product of the water vapor dynamic viscosity, the water content per unit volume, and the temperature-related volume expansion term; wherein, the temperature-related volume expansion term is the sum of 1 and the product of the volume expansion coefficient and the current absolute temperature; when the control strategy generates dynamic control constraints, the critical heating rate is used as the upper limit boundary of the heat treatment equipment when it reaches the hydration reaction sensitive temperature range; when the subsequent process state indicated by the time sequence prediction model is close to the critical heating rate, the heating power of the heat treatment equipment is forcibly reduced to limit the expansion stress of hydration vapor inside the composite product.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] This invention uses the raw material state vectors of the current batch of magnesia-calcium sand and fused magnesia-calcium sand as input sources, driving the generation of compounding parameters and initial process parameters. Based on the time-series data during the composite processing, it predicts the deviation trend of future process states, thereby forming dynamic control constraints to adjust process parameters in real time. Finally, it uses the quality results of the composite product to correct the control strategy and update the model. This constructs a complete data closed loop covering "raw material state - compounding parameters - process control - quality feedback", significantly improving the adaptability of the magnesia-calcium special refractory ceramic production process to raw material batch fluctuations and ensuring the consistency of product quality between batches.

[0033] This invention effectively prevents the propagation of abnormal detection data to downstream decision chains by performing source consistency verification of raw material testing data and marking and re-acquiring abnormal data that exceeds the consistency threshold, thus ensuring the reliability of compounding parameter calculation from the source. Before determining the target compounding parameters, this invention first calculates the interfacial thermal stress mismatch factor based on the mass fraction of calcium oxide and particle size difference in the raw materials, and then concatenates this factor into the raw material state vector. When the mismatch factor exceeds the preset mismatch threshold, it automatically introduces the pore-forming agent mass allocation parameter into the candidate compounding parameters and performs interface strength prediction value attenuation on the scheme without the introduction of pore-forming agent. This proactively compensates for the risk of micro-cracking caused by the difference in thermal expansion when different raw materials are compounded during the batching stage, thereby improving the structural integrity of the composite product.

[0034] Furthermore, when heat treatment is involved in composite processing, this invention calculates the critical heating rate based on the moisture data in the raw material state vector and uses the critical heating rate as the upper limit boundary of the heat treatment equipment within the sensitive temperature range. When the subsequent process state output by the time-series prediction model indicates that the actual heating rate approaches the critical rate, the system forcibly reduces the heating power, effectively preventing the billet from cracking due to the accumulation of hydration vapor pressure and ensuring production safety. This invention employs a reinforcement learning control model, using process deviations, energy consumption, and abnormal states during production as stage feedback, and the quality results after the batch is completed as terminal feedback. This allows for the evaluation and optimization of the energy consumption state of different control actions while meeting product quality and safety constraints, effectively reducing production energy consumption. Attached Figure Description

[0035] 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.

[0036] Figure 1 This is a flowchart of the method of the present invention.

[0037] Figure 2 This is a flowchart of the raw material state vector preprocessing subprocess of the present invention.

[0038] Figure 3 This is a flowchart of the dynamic safety control branch of the heat treatment in this invention.

[0039] Figure 4 This is a flowchart of the batch data storage and model update branch of the present invention. Detailed Implementation

[0040] 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.

[0041] The following is in conjunction with the appendix Figures 1-4 The embodiments of the present invention will be described in detail below.

[0042] The magnesium calcium sand in this invention is made of - The system comprises magnesia-calcium refractory raw materials with the main refractory components obtained through a non-electro-fusion process. Fused magnesia-calcium sand is a magnesia-calcium refractory raw material formed via an electrofusion process. The raw material state vector is formed from the test data of the current actual production batch, including at least two of the following: chemical composition data, particle size distribution data, impurity content data, and moisture data. The compounding parameters characterize the actual mass distribution relationship between magnesia-calcium sand, fused magnesia-calcium sand, and corresponding raw materials of different particle sizes. The production line process boundary is the allowable parameter range determined by at least one of the following: the production equipment safety range, the product process design range, and the historical qualified batch process interval. Dynamic control constraints are the real-time allowable action ranges further formed within the production line process boundary based on the current process state and the predicted process state.

[0043] Example 1: This example uses a magnesium-calcium refractory ceramic production line with raw material detection, automatic feeding, mixing, molding and heat treatment capabilities.

[0044] After magnesia-calcium sand and fused magnesia-calcium sand enter the production area, separate raw material batch identification systems are established. Chemical composition analysis equipment is used to obtain... , In addition, other chemical component data required for current product quality control are obtained, including particle size distribution data of raw materials through particle size detection equipment and moisture content data of raw materials through moisture detection equipment.

[0045] Data obtained from different testing devices are linked through corresponding raw material batch identifiers. When the same chemical component in the same batch has two testing sources, the consistency of the corresponding test results is judged. When the test difference exceeds the consistency threshold, raw material sampling and / or testing are repeated; when the test difference is within the consistency threshold range, the corresponding test results are used as valid raw material status data.

[0046] The effective continuous features are normalized, for example, by subtracting the mean of the training samples from each feature value and then dividing by the standard deviation, to form dimensionless features, which constitute the raw material state vector of the current batch of magnesium calcium sand and fused magnesium calcium sand.

[0047] The production control system reads the quality constraints of the product to be prepared and the corresponding production line process boundaries of the feeding equipment, forming equipment, and heat treatment equipment. Based on the current raw material inventory status and the executable range of the feeding equipment, multiple candidate compounding parameters are generated.

[0048] The material performance prediction model reads the current raw material state vector and candidate compounding parameters to obtain the material performance prediction results corresponding to each candidate compounding parameter. The feasibility discrimination model compares the material performance prediction results with the target product quality constraints and eliminates candidate compounding parameters that do not meet the target product quality constraints.

[0049] A parameter search is performed within the remaining candidate compounding parameter range to determine the target compounding parameter. The target compounding parameter includes at least the mass distribution relationship between magnesia-calcium sand and fused magnesia-calcium sand, and when raw materials of different particle sizes are stored and fed separately, it also includes the mass distribution relationship between corresponding particle sizes.

[0050] After determining the target compounding parameters, input the target compounding parameters, the current raw material state vector, and the current state of the production line into the process parameter prediction model to obtain at least one initial process parameter among the mixing parameters, molding parameters, and heat treatment parameters.

[0051] Initial process parameters are sent to the corresponding production controller after being verified against the production line process boundaries. For parameters within the production line process boundaries, the production controller uses them as the initial setpoints for the current production batch; for parameters outside the production line process boundaries, they are not sent directly to the equipment actuators.

[0052] The feeding equipment measures the magnesia-calcium sand and fused magnesia-calcium sand separately according to the target compounding parameters. The weighing equipment collects the actual feed amount and compares it with the target feed amount. When the feed deviation is within the allowable range, the raw materials enter the mixing equipment; when the feed deviation exceeds the allowable range, at least one of the following operations is performed: compensating feed, re-measuring, or suspending subsequent processes.

[0053] Metered magnesia-calcium sand and fused magnesia-calcium sand enter a mixing device to form a composite material. The composite material then enters a molding device and / or a heat treatment device according to the product process.

[0054] During the molding process, the actual pressure is collected by a pressure sensor; during the heat treatment process, the actual temperature is collected by a temperature detection device; during the heat preservation process, the actual heat preservation start time and end time are recorded; during the controlled cooling process, the actual cooling state is obtained based on the continuous temperature detection results.

[0055] The process monitoring unit generates process time series data by arranging the above data in chronological order and inputs it into the time series prediction model to obtain the prediction results of subsequent process states.

[0056] When the predicted process state is within the production line process boundary, the production controller executes control according to the current allowed action range; when the predicted process state tends towards the production line process boundary, the dynamic control constraint module reduces the action range for further adjustment in the corresponding boundary direction; when the predicted process state crosses the production line process boundary, further adjustment in the out-of-bounds direction is prohibited; when the safety control trigger condition is reached, the system switches to the preset safety control state.

[0057] After the composite product is manufactured, it enters the quality inspection process. Surface images of the product are acquired under controlled lighting conditions, preprocessed, and then input into a convolutional neural network to obtain surface condition detection results. For production batches requiring component analysis, samples are taken from the composite product for chemical composition analysis, and the results are compared with the corresponding component control conditions.

[0058] The surface condition test results and chemical composition test results are used to form the quality results of the current production batch, and are correlated with the raw material state vector, target compounding parameters, actual feeding data, initial process parameters, process timing data and control actions of the current batch.

[0059] When the quality results are generated, if there are still adjustable subsequent processes in the production process, the allowable control range of the corresponding subsequent processes is adjusted according to the quality results; when the current production batch has ended, the quality results and corresponding production data are used to form subsequent model update data.

[0060] Example 2: This example is used for heat treatment processes with significant process response delays. Initial process parameters for the heat treatment process are determined based on the target compounding parameters and the current batch of raw materials. These heat treatment process parameters, verified by the production line process boundary, are then sent to the heat treatment controller.

[0061] The temperature detection device continuously collects the actual temperature during the heat treatment process and generates temperature time-series data according to the sampling time. The time-series prediction model obtains the temperature prediction results for the subsequent time range based on the temperature time-series data within the current time window and the corresponding control status.

[0062] When both the actual temperature and the predicted temperature are within the production line process boundary, and there is an allowable distance between the predicted temperature and the corresponding boundary, the control strategy maintains the current complete operating range.

[0063] When the predicted temperature has not yet exceeded the production line process boundary, but the prediction results show that the temperature continues to approach the upper process boundary, the dynamic control constraint is reduced to further increase the allowable adjustment range of the heat treatment input.

[0064] When the prediction results indicate that the temperature will exceed the upper limit of the process, the dynamic control constraint prohibits the control strategy from continuing to increase the heat treatment input, and only allows the current control state to be maintained and / or the corresponding heat treatment input to be reduced.

[0065] When the predicted result reaches the safety control trigger condition of the heat treatment equipment, the heat treatment equipment will be switched to the preset safety control state.

[0066] During the production process, the current actual temperature, temperature prediction results, dynamic control constraints, actual control actions, and corresponding energy consumption data are stored together according to the same time base.

[0067] After the composite product completes quality testing, the actual product quality results are correlated with the process timing data and control actions of the heat treatment stage for that batch. This allows the timing prediction results not only to generate anomaly alerts but also to directly constrain the permissible control actions of the heat treatment equipment, and enables the final product quality results to be correlated with the corresponding heat treatment control process.

[0068] When using a reinforcement learning control model, process deviations, energy consumption, and abnormal states during the production process form stage feedback, while the final product quality result forms the terminal feedback. The feedback consists of negative process deviations, negative energy consumption, and negative abnormal state penalties, plus the quality evaluation value at the end of the batch. Each item is multiplied by a corresponding weighting coefficient, for example, deviation weight 0.5, energy consumption weight 0.3, abnormal state weight 0.8, and quality weight 1.0. When multiple control actions meet both product quality constraints and production line safety constraints, different control actions are evaluated using corresponding equipment energy consumption data to reduce unnecessary production energy consumption.

[0069] Example 3: This example is used for the intelligent transformation of an existing magnesia-calcium refractory material production line with a PLC system and / or DCS system. The existing PLC system and / or DCS system continues to be responsible for the basic control of the feeding equipment, mixing equipment, molding equipment, and heat treatment equipment. An industrial communication gateway is set up between the existing production control system and the intelligent decision-making unit.

[0070] The industrial communication gateway reads the actual equipment status, process sensor data, and production execution status from the PLC system and / or DCS system, and sends the read data to the process monitoring unit and intelligent decision-making unit.

[0071] The chemical composition data, particle size distribution data, and moisture data obtained by the raw material testing equipment are sent to the raw material data acquisition unit through the data interface and associated with the corresponding production batch according to the raw material batch identifier.

[0072] The intelligent decision-making unit determines the target blending parameters based on the current batch of raw material state vectors, and determines the initial process parameters based on the target blending parameters and the current state of the production line. Both the target blending parameters and the initial process parameters are verified against the production line process boundaries before being sent by the industrial communication gateway to the corresponding control interfaces of the PLC system and / or DCS system.

[0073] During production, the process monitoring unit continuously acquires actual process time-series data and uses a time-series prediction model to obtain predictions of subsequent process states. Based on the relationship between the predicted process states and the production line process boundaries, dynamic control constraints are formed, and the corresponding allowable adjustment ranges are sent to the PLC system and / or DCS system.

[0074] The PLC system and / or DCS system perform corresponding equipment adjustments within the limits of dynamic control constraints.

[0075] After production is completed, quality inspection data is obtained through machine vision acquisition devices and / or chemical composition detection devices. The quality inspection data is stored in association with production batch identifiers, target compounding parameters, actual feeding data, process timing data, and equipment control actions.

[0076] When communication between the intelligent decision-making unit and the production execution unit is interrupted, the industrial communication gateway stops sending new model control commands to the PLC system and / or DCS system, and the PLC system and / or DCS system maintain the most recent valid safety setpoint and / or execute the preset safety control logic.

[0077] Once communication is restored, the current equipment status, current production stage, and current effective process boundary are reacquired. Intelligent control data interaction resumes once the acquired data meets the control conditions.

[0078] Furthermore, a device health status detection function can be set up in the system. Device health status data is generated by collecting at least one of the following: production equipment running time, load status, equipment temperature, and vibration data. When the device health status indicates a change in the executable capability of the corresponding device, the corresponding production line process boundary is tightened according to the actual allowable operating state of the device, and the adjusted production line process boundary is then used to form dynamic control constraints.

[0079] Furthermore, a unified data management server can be set up. The data management server stores the correspondence between raw material batch identifiers, production batch identifiers, target compounding parameters, actual feeding data, production process data, quality results, and model versions, and stores them using data storage structures suitable for batch data and continuous process data, respectively.

[0080] Example 4: This example specifically applies the AI-based composite method of magnesia-calcium sand and fused magnesia-calcium sand to the preparation of magnesia-calcium porous ceramic filters for treating industrial phosphorus-containing wastewater and heavy metal wastewater. The calcium oxide component in the magnesia-calcium material exhibits extremely high free activity after high-temperature calcination, demonstrating strong chemical adsorption and precipitation retention capabilities for phosphate ions and heavy metal cations in water. This application scenario requires the composite product to possess both a well-developed interconnected pore network to ensure water flow capacity and sufficient micro-interface strength to resist erosion and long-term hydration in complex aquatic environments.

[0081] The raw material data acquisition unit reads the basic test data of the current batch of magnesia-calcium sand and fused magnesia-calcium sand using chemical composition and particle size analysis equipment. The chemical composition analysis equipment uses an X-ray fluorescence spectrometer, and the particle size analysis equipment uses a laser particle size analyzer. The acquired data is converted into a structured raw material state vector within the computing device. The processor of the intelligent decision-making unit extracts key feature dimensions from the raw material state vector, specifically including the mass fraction of first calcium oxide in magnesia-calcium sand, the mass fraction of second calcium oxide in fused magnesia-calcium sand, the average particle size of magnesia-calcium sand, and the average particle size of fused magnesia-calcium sand.

[0082] Due to differences in the thermal history of their preparation processes, fused magnesia-calcium sand and electrofused magnesia-calcium sand exhibit inconsistent degrees of calcium oxide grain development. During high-temperature composite heat treatment, their coefficients of linear expansion differ, making microscopic tearing highly likely at grain boundaries. The intelligent decision-making unit calculates the interfacial thermal stress mismatch factor based on the extracted data, with the computational logic following a mapping between physical geometry and thermodynamics.

[0083]

[0084] In the formula, The interfacial thermal stress mismatch factor characterizes the theoretical shear stress limit at the interface between two phase materials, with units of . ; The mass fraction of calcium oxide in magnesium calcium sand is dimensionless. The mass fraction of second calcium oxide in fused magnesium calcium sand is dimensionless. The average particle size of fused magnesia-calcium sand is given in units of 1. ; The average particle size of magnesium-calcium sand is given in units of 1. ; The preset basic coefficient of thermal expansion is obtained by fitting the magnesium-calcium binary phase diagram with historical calorimetric data, and its value is set to [value missing]. Exponential term The wedge-shaped compression effect of coarse-grained fused magnesia-calcium sand on fine-grained magnesia-calcium sand matrix was characterized. The greater the difference in particle size, the stronger the nonlinear characteristics of local stress amplification.

[0085] The calculated interfacial thermal stress mismatch factor is concatenated as a new feature dimension into the raw material state vector. The numerical comparison module within the intelligent decision-making unit performs a threshold determination on the interfacial thermal stress mismatch factor. When the interfacial thermal stress mismatch factor exceeds a preset mismatch threshold (set to [value] in this embodiment), [the threshold is determined]. When the system determines that direct mixing and sintering will lead to the collapse of the filter's microstructure, the material performance prediction and feasibility assessment model triggers a compensation mechanism, forcibly allocating polymethyl methacrylate microspheres as a pore-forming agent mass allocation parameter among the candidate compounding parameters. The pore-forming agent volatilizes during the heat treatment stage, forming buffer pores that absorb and release the accumulated thermal stress at the interface between the two phases. The model algorithm applies a coefficient of 1 to the predicted interface strength values ​​corresponding to the candidate compounding parameters without added pore-forming agent. The attenuation process forces the candidate parameter search model to converge to the parameter solution containing the optimal pore-forming agent ratio in the multi-objective particle swarm optimization space, generating the target composite parameters.

[0086] The raw materials are accurately weighed by the feeding equipment and dispersed and homogenized by the mixing equipment. The composite processing process then enters the critical heat treatment stage. Magnesium-calcium materials readily absorb moisture and form calcium hydroxide when exposed to air. In the early stages of heat treatment, the rapid vaporization of free water and the removal of crystal water will accumulate huge hydration vapor pressure inside the green body, directly causing macroscopic cracking or even explosion of the composite product.

[0087] The process monitoring unit retrieves moisture data from the raw material state vector. The hydration vapor pressure prediction model within the intelligent decision-making unit calculates the critical heating rate during heat treatment based on the physicochemical gas law and the principles of porous media fluid mechanics. The mathematical expression for the dynamic equilibrium between gas permeation and internal tension within the micropores is:

[0088]

[0089] In the formula, The critical heating rate represents the maximum heating gradient that the billet can withstand without structural damage. Unit; The maximum allowable internal tensile strength of the composite product preform was determined through previous finite element analysis. ; The average pore radius of the composite product preform, in units of ; The dynamic viscosity of water vapor at the current temperature is calculated and updated in real time according to the Sutherland formula, with units of [missing value]. ; Water content per unit volume, as characterized by moisture data, is expressed in units of... It is calculated by multiplying the mass fraction of the raw material's moisture content by the bulk density of the formed green body. For example, when the mass fraction of moisture is 0.42% and the green body density is 2500 kg / m³, Approximately 10.5 kg / m³; The volumetric expansion coefficient is dimensionless and takes values ​​of [value missing]. ; The current absolute temperature of the heat treatment is collected and fed back in real time by thermocouples arranged inside the heat treatment equipment, and the unit is 1. This formula is an empirical formula based on the stated units. Substituting the numerical values ​​of each physical quantity yields the result. The critical heating rate is expressed in units. The moisture concentration, steam viscosity, and volume expansion effect in the denominator together determine the resistance to steam overflow. The greater the resistance, the more strictly the allowable heating rate is compressed.

[0090] The heat treatment controller receives dynamic control constraints including the critical heating rate. to Within the sensitive temperature range of hydration reaction and violent steam release, the time-series prediction model frequently collects current heating power and kiln temperature rise slope. The core of the prediction calculation extrapolates future trends. The temperature rise curve over several minutes. The difference between the predicted heating rate and the critical heating rate is fed into the stage feedback function of the reinforcement learning control model. The stage feedback function includes a deep penalty term for out-of-bounds behavior.

[0091]

[0092] In the formula, This is the penalty value for exceeding the boundary. As a penalty weighting coefficient, To predict the heating rate, the cubic exponent amplifies the risk gradient when approaching the critical boundary. When the predicted process condition indicates that the actual heating rate has reached the critical heating rate... At this time, the closed-loop control unit blocks the conventional PID regulation commands and directly issues a forced reduction command to the natural gas flow regulating valve of the heat treatment equipment. The heating power of the equipment is physically limited, and the temperature rise curve exhibits an adaptive and gradual transition, precisely controlling the internal transient vapor pressure below the tensile strength limit.

[0093] The quality inspection unit performs surface defect detection and water treatment performance evaluation on the magnesium-calcium porous ceramic filter bodies that come off the production line. A machine vision acquisition device, combined with a multi-angle slicing light source, scans the product surface; a convolutional neural network extracts crack morphology features and quantifies the surface crack rate. Filter bodies from the same batch are extracted from the chemical experimental platform and placed in an environment with an initial phosphorus concentration of... Dynamic cross-flow filtration and adsorption tests were performed on simulated industrial wastewater to determine... The residual phosphate content in the wastewater after 10 minutes was converted into the phosphorus removal rate. Batch production data and final quality results are summarized in Table 1.

[0094] Table 1. Comparison of production batches and water treatment performance of magnesium-calcium porous ceramic filter media.

[0095] B2026-1 0.42 4.6 15.2 14.8 0.3 92.4 B2026-2 0.85 6.8 7.5 7.1 0.5 95.8 B2026-3 1.25 9.4 4.8 4.5 0.4 98.6 B2026-4 1.80 11.2 3.2 3.0 0.8 99.1

[0096] The table data clearly demonstrates the ability of first-principles-based dynamic control constraints to manage highly sensitive raw materials. Batches B2026-3 and B2026-4 utilized raw material storage capacity with extremely high moisture content and significant particle size variations. The interfacial thermal stress mismatch factor significantly exceeded... The threshold, the intelligent decision-making unit decisively introduced a pore-forming agent, altering the micromechanical conduction path of the matrix. The harsh high moisture conditions caused the calculated critical heating rate to drop to extremely low levels. And #imgpt61#. The closed-loop control unit forces the heat treatment equipment to operate at a near-constant, extremely slow temperature rise. The surface crack rate of the finished product is controlled below #imgpt62#, effectively reducing the risk of kiln explosions or large-scale scrap accidents that conventional refractory material plants may face under such harsh raw material conditions.

[0097] The introduction of pore-forming agents and the suppression of hydration cracks jointly created a complete and highly interconnected porous network. Batch B2026-4, while maintaining structural integrity, achieved a wastewater phosphorus removal rate of #imgpt63#. Free calcium oxide microcrystals were fully exposed on the pore wall surface, undergoing a high-speed interfacial precipitation reaction with phosphate ions in the wastewater to generate a stable hydroxyapatite adsorption layer. Data from each batch was packaged and uploaded to the batch data storage device. The material performance prediction model retrieved this data to perform backpropagation iterative optimization, continuously narrowing the error boundary of process control, thus achieving highly consistent intelligent manufacturing of special water treatment ceramic materials under extremely unstable raw material environments.

[0098] The initial training process for the AI ​​model is as follows: At least 100 historical batches of production data are collected, including raw material state vectors, compounding parameters, process parameters, process time series data, and quality results. These data are then used to train material performance prediction models, feasibility judgment models, process parameter prediction models, and time series prediction models, respectively. Taking the random forest regression model as an example, the raw material state vectors and compounding parameters are used as input features during training, and the corresponding measured material performance values ​​are used as outputs. A mean squared error loss function is employed, and a grid search is used to determine the number and maximum depth of trees. Cross-validation is then performed to prevent overfitting. The support vector machine model uses qualified / unqualified labels for binary classification. The training methods are as follows: The process parameter prediction model can use a neural network, taking the target compound parameters and raw material state vector as input, and the optimized process parameters as output, and using the root mean square error loss function for training; the time series prediction model uses a long short-term memory network, taking process time series data and control actions as input, and outputting process variable values ​​for multiple future time steps, using the root mean square error as the training loss function, and terminating training using an early stop method; the reinforcement learning control model is pre-trained in a simulation environment constructed from the aforementioned historical data, and then fine-tuned in online production. Its Q-network or policy network adopts a fully connected layer structure, with a discount factor of 0.95 and a learning rate of 0.001 during training.

[0099] In some preferred embodiments, the electronic device includes a processor, a memory, and a data interface for communicating with the refractory material production equipment. The memory stores computer programs for performing raw material data processing, compounding parameter determination, process parameter determination, timing state prediction, dynamic control constraint formation, quality feedback, and batch data association. After executing the computer programs, the processor completes the corresponding data processing and interacts with the corresponding raw material testing equipment, production controller, process data acquisition equipment, and quality testing equipment through the data interface.

[0100] In some preferred embodiments, a computer-readable storage medium may store a computer program for performing the above-described data processing procedures. When executed by a processor, the computer program implements at least one of the following data processing procedures: raw material state vector formation, target compounding parameter determination, initial process parameter determination, predicted process state acquisition, dynamic control constraint formation, and quality result feedback.

[0101] Through the above implementation methods, the current raw material state of magnesia-calcium sand and fused magnesia-calcium sand can continuously influence the determination of target compounding parameters and production process parameters. The time-series prediction results during the production process can influence the allowable control range of the actual production equipment. The product quality inspection results can return the control data of the corresponding production batch, thereby realizing closed-loop control of raw materials, compounding, process and quality suitable for the production of green special refractory ceramics.

[0102] 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 the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0103] 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. An AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand, characterized in that, include: S1: Obtain the chemical composition, particle size distribution, impurity content and moisture data of the current batch of magnesium calcium sand and fused magnesium calcium sand respectively. Perform source consistency verification, anomaly handling and normalization on the obtained data to form a raw material state vector, and obtain the quality constraints of the target product and the process boundary of the production line. S2, input the raw material state vector and candidate compounding parameters into the material performance prediction and feasibility judgment model, determine the target compounding parameters within the quality constraints and production line process boundaries, and input the target compounding parameters and raw material state vector into the process parameter prediction model to obtain the initial process parameters for composite production; S3, according to the target compounding parameters, measure and mix magnesium calcium sand and fused magnesium calcium sand, perform at least one of the composite treatments of molding and heat treatment according to the initial process parameters, and collect at least one of the process timing data of temperature, time, cooling rate and pressure during the composite treatment process; S4, predict the deviation trend of subsequent process states relative to the process boundary of the production line based on process timing data, generate dynamic control constraints based on the deviation trend, and adjust the corresponding process parameters within the action range limited by the dynamic control constraints. S5 performs surface defect detection and / or composition uniformity detection on the composite product to generate quality results. The quality results are used as feedback for the control strategy. The quality results are stored in batch association with the actual compounding parameters and process time data for subsequent control correction of the current batch and / or model update of subsequent batches. The output is magnesium-calcium special refractory ceramic composite product and corresponding batch traceability data.

2. The AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand according to claim 1, characterized in that: Source consistency verification includes: obtaining data from at least two chemical component detection sources for the same batch of raw materials, determining source differences based on the detection results of the corresponding chemical components; when the source difference exceeds the consistency threshold, marking the corresponding raw material status as an abnormal state that cannot be directly entered into the calculation of compounding parameters, and triggering at least one of the following operations: resampling, retesting, and testing equipment status check; the consistency threshold is determined based on the repeatability data of the testing equipment and the distribution of source differences of historical valid samples.

3. The AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand according to claim 1, characterized in that: The material performance prediction and feasibility judgment model includes a continuous performance prediction model, a feasibility judgment model, and a candidate parameter search model; the continuous performance prediction model obtains the corresponding material performance prediction results based on the raw material state vector and candidate compounding parameters; The feasibility discrimination model filters out candidate compounding parameters that do not meet the quality constraints based on the predicted material properties and the quality constraints of the target product. The candidate parameter search model determines the subsequent parameters to be evaluated based on the evaluated candidate composite parameters and their evaluation results, and identifies the target composite parameter from the feasible candidate composite parameters.

4. The AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand according to claim 3, characterized in that: The candidate compounding parameters include at least the mass distribution parameters of magnesia-calcium sand and fused magnesia-calcium sand, and at least one of the gradation parameters of raw materials of different particle sizes. When generating candidate compounding parameters, a total amount conservation constraint is applied to the mass distribution parameters of each raw material, and the value boundaries of the candidate compounding parameters are determined according to the executable range of the feeding equipment, the current raw material inventory status, and the quality constraints of the target product.

5. The AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand according to claim 1, characterized in that: The inputs to the process parameter prediction model include at least the target compounding parameters, the current batch raw material state vector, and the current state of the production line. The outputs include at least one of the following: mixing parameters, heat treatment parameters, heat preservation parameters, cooling parameters, and molding pressure parameters. The outputs of the process parameter prediction model are used as the initial setpoints of the production controller after being verified by the production line process boundary. The production controller performs basic closed-loop control based on the deviation between the actual measured values ​​and the initial setpoints.

6. The AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand according to claim 1, characterized in that: Step S4 includes: using a time-series prediction model to predict the process state within a subsequent time range based on process time-series data within a preset time window; comparing the predicted process state with the corresponding production line process boundary to determine the degree of prediction deviation; when the predicted process state tends towards or exceeds the corresponding production line process boundary, narrowing the allowable action range of the control strategy for the corresponding process parameter and / or changing the allowable adjustment direction of the corresponding process parameter; when the degree of prediction deviation reaches the abnormal triggering condition, switching to the preset safe process parameter and / or pausing to enter the next process.

7. The AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand according to claim 1, characterized in that: The control strategy adopts a reinforcement learning control model, which generates control actions based on the production process status. The control actions include at least increasing, maintaining, and decreasing the corresponding process setpoints. Process deviations, energy consumption, and abnormal states during the production process form stage feedback, and the quality results obtained after the batch is completed form terminal feedback, so as to establish a correlation between the quality inspection results and the process control actions that generated the quality inspection results.

8. The AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand according to claim 1, characterized in that: Surface defect detection includes: acquiring surface images of the composite product under controlled illumination conditions, performing standardized preprocessing on the surface images, and inputting them into a convolutional neural network to obtain a detection result that includes at least one of the following: qualified state, crack state, inclusion state, and porosity state; composition uniformity detection includes detecting the chemical composition at at least one location of the same batch of composite products, and when the detection result does not meet the batch composition consistency condition, the batch quality result is marked as abnormal, and a corresponding process feedback identifier is generated.

9. The AI-based composite method for magnesium-calcium sand and fused magnesium-calcium sand according to claim 1, characterized in that: The raw material state vector, target compounding parameters, actual executed formula, initial process parameters, process timing data, control actions and quality results of each production batch are associated with the same batch identifier to form a batch sample; When a new batch of samples meets the preset update conditions and / or a change in the distribution of input data is detected, at least one of the material performance prediction model, process parameter prediction model, time series prediction model, and control strategy is updated, and the model versions before and after the update and the corresponding applicable data range are saved.

10. An AI-based composite system for magnesium-calcium sand and fused magnesium-calcium sand, used to perform the AI-based composite system for magnesium-calcium sand and fused magnesium-calcium sand as described in any one of claims 1-9, characterized in that: The system comprises a raw material data acquisition unit, an intelligent decision-making unit, a production execution unit, a process monitoring unit, a quality inspection unit, and a closed-loop control unit. The raw material data acquisition unit includes a data interface connected to at least one of a chemical composition detection device, a particle size detection device, and a moisture detection device, used to acquire raw material detection data of magnesia-calcium sand and fused magnesia-calcium sand and form a raw material state vector. The intelligent decision-making unit includes a processor and a memory, used to receive the raw material state vector, target product quality constraints, and production line process boundaries, perform material performance prediction, feasibility assessment, and compounding parameter optimization, output target compounding parameters, and determine initial process parameters based on the target compounding parameters and the raw material state vector. The production execution unit includes at least one of a raw material feeding device, a mixing device, a molding device, and a heat treatment device, and includes a PLC controller and / or a DCS controller connected to the device, used to execute composite production according to the target compounding parameters and initial process parameters. The process monitoring unit includes process sensors installed in the production execution unit and a data acquisition interface connected to the process sensors, used to collect process time-series data and form dynamic control constraints sent to the production execution unit based on the predicted deviation trend of the process time-series data. The quality inspection unit includes a machine vision acquisition device and / or a chemical composition detection device, used to generate quality results for the composite product; the closed-loop control unit is communicatively connected to the intelligent decision-making unit, the production execution unit, the process monitoring unit and the quality inspection unit, used to feed back the quality results to the control strategy, and to associate and store the raw material status, compounding parameters, production process data and quality results of the same batch.

Images