Glass powder firing temperature control method and system

Abstract

The application discloses a glass powder sintering temperature control method and system, and belongs to the field of glass powder sintering temperature control. In order to solve the problem that the kiln partition temperature, material residence time, tail gas heat recovery and powder phase change window are difficult to be cooperatively adjusted, the application realizes the cooperative control of the sintering process through partition digital twinning, causal constraint multi-step prediction, phase change soft measurement and rolling optimization.

Description

Technical Field

[0001] This invention relates to the field of temperature control in glass powder firing, and more particularly to a method and system for temperature control in glass powder firing. Background Technology

[0002] Glass powder, microcrystalline glass powder, and other powder materials are widely used in electronic packaging, ceramic glazes, low-temperature co-fired materials, functional coatings, and composite materials. The firing process of these powders typically involves stages such as heating, softening, crystallization, holding, and cooling. The firing quality is closely related to factors such as kiln zone temperatures, heating rates, material residence time in each temperature zone, raw material formulation, particle size distribution, and atmospheric conditions. With the increasing application of continuous kilns, roller kilns, and pusher kilns in glass powder firing, the firing process has gradually evolved from traditional manual experience-based control and fixed temperature curve control to automated control methods based on sensor data acquisition, zoned temperature control, conveyor speed adjustment, and waste heat recovery from exhaust gas.

[0003] In existing technologies, kiln temperature control systems typically collect temperature data using thermocouples or infrared thermometers placed in different temperature zones, and adjust the heating power of each zone according to a preset temperature curve. Some systems also combine the operating speed of the conveyor device to control the material residence time, or adjust waste heat recovery valves based on the exhaust gas temperature and flow rate to improve energy utilization efficiency. In recent years, some technologies have also attempted to introduce model predictive control, digital twins, and data-driven modeling to predict changes in the kiln thermal field, thereby improving the response speed and stability of temperature control.

[0004] However, during the glass powder sintering process, the phase transformation properties of the powder, such as the softening point, crystallization initiation temperature, crystallization termination temperature, and coefficient of thermal expansion, are affected by the formulation composition, particle size distribution, batch differences, and historical sintering conditions. If control is based solely on a fixed temperature curve or single-zone temperature feedback, it is difficult to accurately reflect the actual phase transformation state of the powder during sintering. Therefore, the existing technology still has the following shortcomings:

[0005] First, existing kiln control mainly focuses on zone temperature, usually adjusting the heating power or conveying speed of each zone separately. It lacks coordinated prediction of the heat transfer relationship between upstream and downstream zones of the kiln, the material residence time lag relationship, and the change of tail gas enthalpy value. As a result, the control strategy is prone to lag behind the actual thermal field changes and it is difficult to correct the fluctuations of the firing curve in a timely manner.

[0006] Second, existing control methods typically do not incorporate phase change window parameters such as softening point, crystallization start temperature, and crystallization end temperature into online control constraints, or rely solely on offline detection results to set fixed process parameters. They cannot dynamically determine the duration of the material within the phase change window based on the real-time temperature field and batch status, which can easily lead to uneven crystal precipitation, inconsistent sintering degree, or increased deviation in the coefficient of expansion.

[0007] Third, existing exhaust gas waste heat recovery control systems primarily focus on energy saving by independently adjusting bypass valves, recovery valves, and other actuators, failing to integrate them with zoned temperature control, air-fuel ratio control, material residence time control, and phase change stability requirements. When batch fluctuations occur in raw material formulation or particle size distribution, fixed-weight or static control strategies struggle to balance calcination quality and energy consumption control, easily leading to decreased batch consistency or unstable waste heat recovery efficiency.

[0008] Therefore, there is an urgent need for a glass powder sintering temperature control method and system that can comprehensively consider zoned thermal field prediction, material residence time, exhaust gas heat recovery and powder phase change window, so as to achieve coordinated regulation of the sintering process, improve crystal phase consistency and reduce energy consumption. Summary of the Invention

[0009] One objective of this invention is to provide a method and system for temperature control in glass powder firing. Addressing the problem in existing technologies where it is difficult to coordinate the adjustment of kiln zone temperature, material residence time, exhaust gas heat recovery, and powder phase change window, this method uses zoned digital twin data acquisition to collect data on each zone's temperature, heating rate, material residence time, raw material formulation, particle size distribution, exhaust gas temperature, exhaust gas flow rate, waste heat recovery status, and historical firing results. It combines a causal constraint multi-step predictor with a phase change soft measurement model to output predicted thermal field, exhaust gas enthalpy, softening point, and crystallization window boundary. Within the rolling time domain, it dynamically corrects the phase change stability weight, energy-saving weight, and waste heat recovery weight, and jointly solves for the control variables. This allows the firing process, waste heat recovery, and phase change control to converge synchronously, reducing energy consumption and improving crystal phase consistency.

[0010] This invention provides a method for temperature control in glass powder firing, comprising: S1, collecting temperature, heating rate, material residence time, raw material formula, particle size distribution, tail gas temperature, tail gas flow rate, waste heat recovery status, and historical firing results of each zone of the kiln, and forming a batch state vector; S2, based on the state vector, calling a causal constraint multi-step predictor and a zone digital twin model, and outputting the predicted thermal field and tail gas enthalpy value of each zone; S3, based on the state vector and real-time temperature field, calling a phase change soft measurement model, outputting the softening point, crystallization initiation temperature, crystallization end temperature, and expansion coefficient deviation, and determining the window boundary according to the softening point, crystallization initiation temperature, and crystallization end temperature, and calculating the duration within the window and the batch deviation index; S4, in the rolling... In the dynamic time domain, based on the predicted thermal field, exhaust gas enthalpy, window boundary, duration within the window, and batch deviation index, the phase change stability weight, energy saving weight, and waste heat recovery weight are corrected. The constraint relaxation amount formed by the window boundary is input into the rolling optimizer to solve for the partition heating power, conveying speed, air-fuel ratio, bypass valve opening, and recovery valve opening, generating control commands. S5, the control commands are applied to the kiln and the actual partition temperature, actual residence time, exhaust gas temperature, exhaust gas flow rate, and actual firing results are collected. The actual partition temperature, actual residence time, exhaust gas temperature, exhaust gas flow rate, and actual firing results are written into the historical firing result database and fed back to the partition digital twin model and phase change soft measurement model to update the prediction parameters for the next cycle.

[0011] Optionally, S1 includes:

[0012] Data from sensors in each partition are collected according to a preset sampling period. The collected data is then processed by time alignment, outlier removal, and normalization to form the batch state vector.

[0013] Optionally, S2 includes:

[0014] A partition coupling matrix is ​​established based on the transmission relationship between the upstream and downstream partitions of the kiln, and a material residence time delay matrix is ​​established based on the conveying speed and partition length.

[0015] When the batch state vector is input into the causal constraint multi-step predictor, the prediction is restricted to using only the historical state of the current partition and the upstream partition, and the temperature and exhaust gas enthalpy of each partition are predicted step by step, and the prediction deviation is written back to the state parameters of the partition digital twin model.

[0016] Furthermore, the causal constraint multi-step predictor and the partitioned digital twin model use the current batch state vector, actual partition temperature trajectory, actual exhaust gas temperature trajectory, actual exhaust gas flow trajectory, and exhaust gas enthalpy value labeling obtained by converting actual exhaust gas temperature and actual exhaust gas flow from the historical firing sample library as training samples.

[0017] The phase change soft measurement model uses the current batch state vector, actual zone temperature trajectory, and softening point, crystallization start temperature, crystallization end temperature, and expansion coefficient labels obtained through offline thermal analysis as training samples. The model parameters are updated with the joint loss of zone temperature error, tail gas enthalpy error, softening point error, crystallization start temperature error, crystallization end temperature error, and expansion coefficient error.

[0018] Optionally, S3 includes:

[0019] By inputting the historical firing results, the current batch state vector, and the real-time temperature field into the phase change soft measurement model, the softening point, crystallization initiation temperature, crystallization termination temperature, and expansion coefficient deviation are obtained.

[0020] The lower and upper boundaries of the phase transition window are formed based on the softening point, crystallization initiation temperature, and crystallization end temperature, and the duration of time during which the measured kiln temperature is continuously located between the lower and upper boundaries is taken as the duration within the window.

[0021] Furthermore, the batch deviation index is determined by normalizing and weighting the corresponding parameters of the current batch formula, particle size distribution, historical softening point, and historical expansion coefficient with those of historical samples of the same batch, and is used as input to correct the update magnitude of the phase change stability weight, energy saving weight, and waste heat recovery weight.

[0022] Optionally, S4 includes:

[0023] The phase change stability weight, energy saving weight, and waste heat recovery weight are corrected online based on the duration within the window and the batch deviation index. The constraint relaxation amount formed by the window boundary is input into the rolling optimizer to generate control commands for zone heating power, conveying speed, air-fuel ratio, bypass valve opening degree, and recovery valve opening degree.

[0024] Furthermore, it is divided into two levels: feasible interval screening and control quantity solution. The feasible interval screening determines the feasible intervals of bypass valve opening and recovery valve opening based on exhaust gas temperature, exhaust gas flow rate and waste heat recovery status. The control quantity solution jointly solves the zone heating power, conveying speed and air-fuel ratio within the feasible intervals.

[0025] Optionally, S5 includes:

[0026] The actual zone temperature is compared with the predicted temperature trajectory in the predicted thermal field of each zone, and the deviation is determined by combining the actual residence time, actual exhaust gas temperature, actual exhaust gas flow rate and actual firing results; when the deviation exceeds the preset correction threshold, the zone digital twin model and phase change soft measurement model are corrected online using the actual firing results.

[0027] On the other hand, the present invention also provides a glass powder firing temperature control system, comprising:

[0028] The data acquisition module is used to read historical firing results from the historical firing result database and collect data on the temperature, heating rate, material residence time, raw material formula, particle size distribution, tail gas temperature, tail gas flow rate, and waste heat recovery status of each zone of the kiln, forming a batch state vector. The causal prediction module is used to call the causal constraint multi-step predictor and the zone digital twin model to output the predicted thermal field and tail gas enthalpy value of each zone. The phase change analysis module receives the batch state vector and real-time temperature field output by the data acquisition module, calls the phase change soft measurement model, outputs the softening point, crystallization initiation temperature, crystallization end temperature, and expansion coefficient deviation, and analyzes the softening point, crystallization initiation temperature, and crystallization end temperature based on the deviation. The system uses a beam temperature to determine the window boundary, calculates the duration within the window, and the batch deviation index. A rolling optimization module generates control commands based on the predicted thermal field, exhaust gas enthalpy, window boundary, duration within the window, and batch deviation index. An execution feedback module applies the control commands to the kiln and collects actual zone temperatures, actual residence times, exhaust gas temperatures, exhaust gas flow rates, and actual firing results. These data are then written into a historical firing result database and fed back to the causal prediction module, the zone digital twin model, and the phase transition analysis module to update the prediction parameters for the next rolling cycle.

[0029] The beneficial effects of this invention are:

[0030] 1. This invention generates a batch state vector by collecting data on the temperature, heating rate, material residence time, raw material formula, particle size distribution, tail gas temperature, tail gas flow rate, waste heat recovery status, and historical firing results of each zone of the kiln. Combined with a causal constraint multi-step predictor and a zone digital twin model, it outputs the predicted thermal field and tail gas enthalpy of each zone. This enables the control system to perceive in advance the heat transfer effect of the upstream zone on the downstream zone and the effect of material residence time lag, thereby reducing the kiln temperature control lag and improving the accuracy of predicting zone temperature trajectories and tail gas heat changes.

[0031] 2. This invention outputs the softening point, crystallization initiation temperature, crystallization end temperature, and expansion coefficient deviation online through a phase change soft measurement model. Based on this, the phase change window boundary is determined, the duration within the window is calculated, and the batch deviation index is determined. This allows the kiln control to no longer rely solely on a fixed temperature curve, but to dynamically constrain the phase change process of the material according to the current batch's formula, particle size distribution, and real-time temperature field. This reduces uneven crystal precipitation, inconsistent firing degree, and expansion coefficient deviation, thereby improving the firing quality and batch consistency of glass powder.

[0032] 3. In the rolling time domain, this invention dynamically corrects the phase change stability weight, energy saving weight, and waste heat recovery weight based on the predicted thermal field, exhaust gas enthalpy, window boundary, duration within the window, and batch deviation index. It also jointly solves the partitioned heating power, conveying speed, air-fuel ratio, bypass valve opening, and recovery valve opening, enabling the partitioned temperature control, material residence time control, and exhaust gas waste heat recovery to be optimized in a coordinated manner. This improves waste heat utilization efficiency and reduces firing energy consumption while ensuring phase change stability. Attached Figure Description

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0034] Figure 1 This is a control flowchart of the glass powder firing temperature control method of the present invention.

[0035] Figure 2 This is a flowchart of the S3 phase transition soft measurement of the present invention. Detailed Implementation

[0036] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0037] refer to Figures 1-2 The glass powder firing temperature control method includes: S1, collecting data on the temperature, heating rate, material residence time, raw material formula, particle size distribution, tail gas temperature, tail gas flow rate, waste heat recovery status, and historical firing results of each zone of the kiln, and forming a batch state vector; S2, based on the state vector, calling a causal constraint multi-step predictor and a zone digital twin model, outputting the predicted thermal field and tail gas enthalpy value of each zone; S3, based on the state vector and real-time temperature field, calling a phase change soft measurement model, outputting the softening point, crystallization initiation temperature, crystallization end temperature, and expansion coefficient deviation, and determining the window boundary based on the softening point, crystallization initiation temperature, and crystallization end temperature, calculating the duration within the window and the batch deviation index; S4, within the rolling time domain... Based on the predicted thermal field, exhaust gas enthalpy, window boundary, duration within the window, and batch deviation index, the phase change stability weight, energy saving weight, and waste heat recovery weight are corrected. The constraint relaxation amount formed by the window boundary is input into the rolling optimizer to solve for the zone heating power, conveying speed, air-fuel ratio, bypass valve opening, and recovery valve opening, generating control commands. S5. The control commands are applied to the kiln and the actual zone temperature, actual residence time, exhaust gas temperature, exhaust gas flow rate, and actual firing results are collected. The actual zone temperature, actual residence time, exhaust gas temperature, exhaust gas flow rate, and actual firing results are written into the historical firing result database and fed back to the zone digital twin model and phase change soft measurement model to update the prediction parameters for the next cycle.

[0038] In this specific embodiment, S1 includes:

[0039] This specific implementation takes a continuous roller kiln with six independent temperature control zones as the execution object. The preheating zone, heating zone 1, heating zone 2, crystallization heat preservation zone, slow cooling zone and discharge zone are sequentially labeled as zone 1 to zone 6. When the batch glass powder tray enters zone 1, the data acquisition module generates a batch identifier and records the raw material formula version, particle size test report number, tray mass and kiln entry time. Subsequent temperature, conveying and exhaust gas records are all associated with the batch identifier, zone number and unified timestamp.

[0040] The data acquisition module reads the temperature of three thermocouples arranged along the conveying direction in each zone, the power feedback of the heaters in each zone, the roller encoder pulse, the mass flow rate of the gas branch, the mass flow rate of the combustion air branch, the temperature of the air entering the furnace, the temperature of the thermocouple in the tail gas main pipe, the flow rate of the tail gas thermal mass flow meter, the position feedback of the bypass valve, the position feedback of the recovery valve, and the operating status of the waste heat exchanger at a sampling period of 2 seconds. The zone temperature is expressed as the median of the three thermocouples after cold junction compensation, in degrees Celsius. The flow rates of gas, combustion air, and tail gas are uniformly converted to kilograms per second. The temperature of the air entering the furnace is collected by the thermal resistance at the inlet of the combustion air main pipe and is in degrees Celsius. The valve position is uniformly converted to a dimensionless opening degree from 0 to 1.

[0041] The combustion controller divides the ratio of combustion air mass flow rate to fuel mass flow rate at the same sampling moment by the theoretical air-fuel mass ratio corresponding to the current fuel composition to obtain the dimensionless actual air-fuel ratio. The inlet thermal resistance reading of the combustion air main pipe is recorded as the furnace air temperature. Both are aligned with the heating power record according to the PLC monotonically increasing cycle number and timestamp. When the sampling interval is no more than 6s, the most recent valid value is used. When it exceeds 6s but does not exceed 30s, the most recent valid value is used and a rollback flag is set. When there is no valid gas flow, air flow or furnace air temperature for 30 consecutive seconds, a combustion input invalid flag is set, the current status update of the digital twin is frozen, and S4 is prohibited from issuing new incremental commands. The original flow, theoretical ratio version, calculated value, temperature value, quality flag and rollback source are all written to the batch status cache.

[0042] The controller calculates the heating rate based on the current zone temperature. ,in The partition numbers are 1 to 6. This is the sampling time sequence number. For the first Partition at time The median temperature, in degrees Celsius. The sampling period is 2 seconds. The unit is degrees Celsius per second, and the calculation result is written to the heating rate field of the batch time sequence record;

[0043] The estimated residence time of materials passing through the complete partition at the current speed is based on... Calculation, where For the first The effective length of the partition along the conveying direction, in meters, is measured and recorded in the kiln geometry parameter table during equipment acceptance. The conveying speed is calculated from the encoder pulses using the roller diameter coefficient, and the unit is meters per second. The estimated complete transit time for the corresponding partition is in seconds, and is used only as an estimate of the future time lag of S2 and the target dwell time of S4; when When the speed drops below the equipment's allowable lower limit by 0.002 m / s, the batch will be marked as suspended and the most recent valid estimated dwell time will be used without performing division.

[0044] When a batch is established, the data acquisition module creates online cumulative dwell time registers for the batch identifier and the six partition numbers. Registers that have not yet entered a partition are deterministically initialized to 0 and the status enumeration "not entered" is written. When an inlet photoelectric event arrives, the initial value of the partition register is kept at 0 and the status is changed to "occupied". Before the outlet photoelectric event arrives, the value is accumulated by 2 seconds for every valid 2-second sampling cycle. During pauses, reversals, or slippage, as long as the material tray is still determined to occupy the partition by the partition boundary event, the accumulation continues. After the outlet event arrives, the cumulative value of the partition is sealed and the status is changed to "left". It cannot be overwritten by the residual value of other batches or the previous production task. When the boundary photoelectric event is temporarily missing, the encoder cumulative position and the kiln geometric boundary are cross-checked. When both the photoelectric event and the encoder are invalid, the register is frozen and a cumulative dwell time invalid flag is set. Accumulation does not cross batches or partitions. The online cumulative value, register status, expected complete passage time, and expected remaining time are written to the batch status cache as fixed fields. S3 and S4 read the online cumulative value and register status. After the material tray leaves the area, S5 measures the entry and exit events to correct the final dwell time.

[0045] The raw material formulation record includes the mass fractions of silicon dioxide, boron oxide, aluminum oxide, alkali metal oxides, and nucleating components. The sum of all mass fractions is verified to be 1. The particle size record includes measurements taken by a laser particle size analyzer on three samples from the same batch before they enter the kiln. , and The unit is micrometers. The median of the three results is written into the batch static field. Historical firing results are calculated according to the same formula version and... The phase change detection values, finished product expansion coefficient, and temperature trajectory summary of the most recent 120 valid batches within the particle size range were retrieved.

[0046] Time alignment based on sampling time Based on the baseline, linear interpolation is performed on asynchronous sensor records with sampling intervals not exceeding 6s. Outlier removal is performed on readings that exceed the device range or deviate from the median of the most recent 15 valid samples of the same sensor by more than three times the median absolute deviation. When a single zone has a continuous missing measurement of no more than 10s, the one-step prediction value of the digital twin model of that zone is used to make up the missing measurement. When the continuous missing measurement exceeds 10s, a data invalidation flag is set and S4 is prohibited from issuing new incremental control commands.

[0047] Aligned Continuous variables according to Normalization, in which For the original or padded variable values, and These are the mean and standard deviation of this variable in the most recent 120 valid batches of the same formulation category, respectively, and compared with... For the same unit, the lower limit of the standard deviation is taken as 0.5% of the sensor's range. This means limiting the result to -4 to 4. The dimensionless normalized values, mean, standard deviation, sample batch list, and statistical update time are stored in the normalization parameter table;

[0048] The normalization parameter table clearly defines the average of similar historical batches. As a normalization baseline, the variable value is subtracted from this baseline and then divided by the standard deviation of similar historical batches. The interval between -4 and 4 is used as the cutoff interval. In this specific embodiment, a partition temperature record stores the variable name, partition number, baseline temperature, standard deviation, sensor range, number of sample batches, version number and effective time. When the formula version changes or the mean of the most recent 30 batches deviates from the baseline by more than one standard deviation, the parameters are recalculated. If less than 30 valid batches are reached, the initial parameter version verified by the process is used.

[0049] Each sensor calibration record uses the device number and channel number as keys. The fields include zero point, range coefficient, calibration temperature or flow point, allowable error, calibration date and validity period. Thermocouples are calibrated at three points: 400 degrees Celsius, 650 degrees Celsius and 850 degrees Celsius. Mass flow meters are calibrated at three points: 20%, 60% and 90% of the rated flow. If any calibration point exceeds the tolerance, the channel shall not be entered into the current batch state vector and shall be replaced by a redundant channel or a one-step predicted value of the digital twin.

[0050] The historical firing results summary is constructed based on valid batches with the same formula and particle size range that have completed kiln exit, quality inspection, and metering settlement. The retrieval cutoff time is fixed at the current batch's state vector generation time; records later than this time are not included in the statistics. Fields include the median historical softening point, median crystallization start and end temperatures, median expansion coefficient, mean zone temperature trajectory, mean tail gas enthalpy flow, median target crystalline phase mass fraction, median power consumption per unit batch, and median waste heat recovery. The target crystalline phase mass fraction is read from the corresponding valid process card and stored as a dimensionless mass fraction from 0 to 1. The power consumption per unit batch is obtained by summing the positive active power of the six energy meters during the period from kiln entry to kiln exit for that batch, and the unit is kilowatt-hours. Waste heat recovery is obtained by integrating the opening of the recovery valve, recoverable heat power and sampling interval during the same period, and the unit is kilowatt-hour. Each item simultaneously saves the median absolute deviation, number of valid samples, statistical cutoff time, metering or process card version and quality mark. The default is to use the most recent 60 valid batches. If there are fewer than 20 batches, the range is expanded to the adjacent particle size range but the formula version is kept consistent. If there are still fewer than 5 batches after expansion, the target crystal phase mass fraction is returned to the current valid process card value, the power consumption per batch is returned to the process card initial benchmark that has been recently replayed and is not less than 0.1 kilowatt-hour, and the waste heat recovery is returned to the positive benchmark given by the rated heat balance test of the heat exchanger. The process card return, power consumption benchmark return or recovery benchmark return mark are recorded respectively.

[0051] The data acquisition module uses a fixed field order to assemble the current batch state vector into a single vector. This vector includes the temperatures and heating rates of the six zones, the cumulative online residence time and estimated complete transit time of the six zones, the mass fractions of the five formulations, the three particle size quantiles, the actual air-fuel ratio, the furnace air temperature, the exhaust gas temperature, the exhaust gas flow rate, the bypass valve opening, the recovery valve opening, the waste heat exchanger status, and a summary of historical firing results containing the historical target crystal phase mass fraction, the unit batch power consumption, and the waste heat recovery amount. The module also writes the original values, normalized values, data validity flags, rollback sources, number of valid samples, statistical cutoff time, sampling time, and parameter table version of each field into the batch state cache, which serves as the unified input for S2 causal prediction and S3 phase softening measurement.

[0052] In this specific embodiment, S2 includes:

[0053] The causal prediction module reads the current batch state vector and kiln geometric parameter table written by S1, and performs mean, maximum, and last value aggregation on the 2-second sampled records with a prediction step size of 30 seconds to form the input sequence of the most recent 40 prediction steps. The input sequence retains the batch identifier, partition number, temperature, heating rate, online cumulative residence time, expected complete passage time, heating power, conveying speed, actual air-fuel ratio, furnace air temperature, exhaust gas temperature, exhaust gas flow rate, and valve position fields. The digital twin history training and current state update read the aligned data from S1. and Actual value; when there is a missing measurement rollback, the most recent valid value with the rollback flag is read. In future prediction, the S4 candidate air-fuel ratio is read in control steps 1 to 20 and the current valid value or the predicted value of the combustion air preheating model is used for the furnace air temperature. The value taken in step 20 is used in steps 21 to 120. Any invalid combustion input flag is passed to the prediction result cache.

[0054] Before the input sequence enters the predictor, the data in the batch state cache is cleaned according to the valid flags. Records without batch identifiers or in reverse time are deleted. Windows with fewer than 12 valid sampling points in a single 30s window are marked as sparse windows and filled with digital twin one-step state. If two consecutive sparse windows are opened, the multi-step prediction for this cycle is terminated. Only records with newer calibration versions and valid quality flags are retained for repeated timestamps.

[0055] Partition Coupling Matrix Established based on the heat transfer relationship from upstream to downstream, where and All are partition numbers from 1 to 6. Indicates the first The effect of zone temperature and heating power changes on the first The dimensionless influence coefficient of the next prediction step temperature in the partition, when Time fixed To block the reverse effect of the downstream future state on the upstream prediction, the remaining coefficients are calibrated by the least squares method with non-negative constraints based on the temperature step response of the most recent 60 valid batches, and the sum of the coefficients in each row is limited to 0 to 1.

[0056] Material residence time delay matrix according to Establish, among which , For the first The effective length of the partition is in meters. This represents the average conveying speed within the current scrolling window, expressed in meters per second. For a 30s prediction step size, For the upstream Partition status is passed to the first Integer prediction steps required for partitioning The corresponding elements are not included in the calculation. The matrix is ​​recalculated when the conveying speed changes by more than 5%.

[0057] The causal-constrained multi-step predictor employs a two-layer causal temporal convolutional network, with each layer having a kernel length of 3 and 64 channels. The temporal dimension of the input tensor only allows reading of the current prediction step and previous records, and the partition dimension is then multiplied by... and according to Extracting upstream features with corresponding time delays, the model outputs the equivalent temperature residuals of materials, exhaust gas temperature residuals, and absolute exhaust gas mass flow rates for the six zones over the next 120 prediction steps. The mass flow rate output is non-negatively truncated and limited to the flow metering range of 0 to 8 kg / s, with 120 prediction steps corresponding to a 60-minute rolling time domain;

[0058] The partitioned digital twin model uses discrete state updates. The state vector is arranged according to Fixed as a 13x1 column vector, where This indicates that the components within the parentheses are arranged in the order listed. to The equivalent temperature of materials in the six zones. to The heat storage state of the furnace lining is expressed in terms of equivalent temperature. This refers to the exhaust gas temperature; all components are in degrees Celsius. It is a 13x13 dimensionless state transition matrix, and its material temperature sub-block is in Press at time Write, in When the time is set to 0, the diagonal attenuation coefficients of the remaining furnace lining and tail gas sub-blocks are identified by the 30s step response; This is a 6x1 heating power vector in kilowatts. The conveying speed is expressed in meters per second. Air-fuel ratio, The temperature of the air entering the furnace is in degrees Celsius. It is 13 by 6, with each row representing degrees Celsius per kilowatt. It is 13 times 1, and the unit of each row is degrees Celsius divided by meters per second. It is 13 by 1 and each row is in degrees Celsius. The matrix is ​​a 13x1 dimensionless matrix. All five types of matrices are identified by historical step tests with the same 30s sampling period, making each term on the right side of the equation a 13x1 degree Celsius vector.

[0059] The residuals from the causal-constrained multi-step predictor are added to the state trajectories of the digital twin model to obtain the predicted thermal fields for each region. ,in The prediction step numbers are 1 to 120. For the first Partition in the future The material equivalent temperature is set in degrees Celsius. For predicted values ​​that are below room temperature or above the equipment's maximum permissible temperature of 950 degrees Celsius, boundary truncation is performed and model out-of-bounds flags are recorded.

[0060] Exhaust gas enthalpy flow according to Conversion, among which To predict exhaust gas mass flow rate in kilograms per second, This is the specific heat of exhaust gas at constant pressure, expressed in kilojoules per kilogram of Celsius. Its value is read from the exhaust gas property table based on the current formulation and air-fuel ratio and is limited to between 0.95 and 1.20. To predict exhaust gas temperature in degrees Celsius, The reference temperature is 25 degrees Celsius. The unit is kilowatt and it is written into the predicted tail gas enthalpy value sequence;

[0061] The exhaust gas property table uses formulation category and air-fuel ratio range as search keys. Fields include specific heat at constant pressure, effective temperature range, calibration batch, and version number. In this specific embodiment, a record stores information using borosilicate glass formulation and air-fuel ratio of 1.05 to 1.15 as keys. kJ / (kg·°C) and an effective range of 100 to 850°C. This record is obtained by fitting the component analysis and calorimetric calibration results of exhaust gas samples with the same formula. Recalibration is triggered after formula version or burner overhaul. If the result is not hit, the conservative upper limit of the adjacent air-fuel ratio range is used and a check mark is set.

[0062] Before release, the partition coupling matrix and time delay matrix were replayed and verified in 10 consecutive batches that were not calibrated. The predicted temperature was compared one step to 120 steps in each partition. The root mean square error of temperature in each partition was required to be no more than 8 degrees Celsius within 30 minutes and no more than 15 degrees Celsius within 60 minutes. The median error of exhaust gas enthalpy flow was no more than 8%. If the verification failed, the previous matrix version was retained and the step test sample was expanded before re-identification.

[0063] During the training of causal temporal convolutional networks, an AND is applied to the input sequence. A consistent lower triangular partition mask is used, the gradient of the downstream feature of the mask is fixed to 0, and a downstream feature perturbation experiment is performed on the validation set. When any downstream future feature perturbation causes the upstream prediction change to exceed the numerical error limit, the model version is rejected, thereby verifying that the prediction only uses the historical state of the current partition and the upstream partition.

[0064] The offline training of the model uses the current batch state vector, actual partition temperature trajectory, actual exhaust gas temperature trajectory, actual exhaust gas flow trajectory and exhaust gas enthalpy value converted according to the actual exhaust gas temperature and flow rate from the historical burn sample library. The model is divided into 70% training set, 15% validation set and 15% test set according to the batch. The temperature and enthalpy errors in the training loss are first divided by the standard deviation of their respective training sets to become dimensionless quantities, thereby avoiding the direct addition of different physical units.

[0065] The prediction results cache uses the batch identifier and prediction start time as a joint key to store 120-step six-zone prediction temperature matrix, 120-step exhaust gas temperature, exhaust gas flow and enthalpy sequence, coupling matrix version, time delay matrix version, causal predictor version, digital twin parameter version, and model out-of-bounds flag. S4 only uses records with the same version and whose generation time has not exceeded one control cycle. Expired records trigger the retention of the previous control instruction.

[0066] Calculate the predicted deviation for each control cycle. and ,in and All are six-partition Celsius vectors. and All are kilowatts; the digital twin model in this cycle only corrects the instantaneous state bias of the published parameters without changing them using recursive least squares gain, and writes the regression, residual and covariance required for the heat transfer coefficient update into the candidate statistical cache, without directly writing back the heat transfer coefficient or generating a new parameter version; the corrected instantaneous state estimate, the currently published digital twin parameter version, the predicted thermal field and exhaust gas enthalpy sequence of each region are written into the prediction result cache and transmitted to S4 according to a unified batch identifier; the candidate update of the heat transfer coefficient, the 2% change limit, the replay verification, the release and rollback are only executed after the trigger threshold is met for three consecutive cycles in S5.

[0067] In this specific embodiment, S3 includes:

[0068] The phase change analysis module reads the current batch state vector of S1, the 2-second real-time temperature field of the six zones, and the material equivalent temperature state of S2. It aggregates the most recent 20-minute records into a 40x47 input tensor using a 30-second window. The fixed channel table specifies that channels 1 to 5 represent the mass fractions of five formulation components: silica, boron oxide, alumina, alkali metal oxides, and nucleating components, respectively. Channels 6 to 8 represent... , and Channels 9 to 14 represent the average temperature of the six-zone window; channels 15 to 20 represent the heating rate of the six zones; channels 21 to 26 represent the cumulative online residence time of the six zones generated by S1 in batches; channels 27 and 28 represent the exhaust gas temperature and exhaust gas mass flow rate; channels 29 to 34 represent the standard deviation of the six-zone window temperature; channels 35 to 40 represent the temperature deviation of the six zones relative to the qualified trajectory of the same formula; channels 41 to 47 represent the historical softening point, crystallization start temperature, crystallization end temperature, expansion coefficient, target crystalline phase mass fraction, power consumption per unit batch, and waste heat recovery, respectively; channels 45 to 47 directly read the dimensionless mass fraction, kilowatt-hour power consumption, and kilowatt-hour recovery fields with the same names from the historical sintering result summary of S1. Simultaneously verify the number of valid samples, statistical cutoff time, rollback source, measurement or process card version, and quality mark. It is forbidden to fill in the quality or measurement results that have not yet been completed in the current batch. Each channel is processed in this order using the S1 normalized parameter table. Missing values ​​in the window are first filled with the previous valid window value in the same batch. If there is no previous value, the corresponding rollback value generated by S1 according to fixed rules is used and no different statistical measures are allowed. Independent quality marks are set. The 47-channel table, the statistical version of the three historical fields, and the soft measurement model version are bound together. Historical training data are reconstructed according to the same five-item formula mapping and the same cutoff time rule and verified batch by batch. The dimension, field order, statistical caliber, and version of the training end and the inference end are completely consistent before calculation can be performed.

[0069] The phase change soft measurement model employs a two-layer gated recurrent network and four regression output heads. The number of hidden units in the first and second layers are 96 and 48, respectively, and the four output heads respectively provide the softening point. Crystallization initiation temperature Crystallization end temperature and predicted values ​​of expansion coefficient All three temperature measurements are in degrees Celsius. The unit is per degree Celsius, and the model version is stored together with the training sample cutoff batch and the normalized parameter table version;

[0070] The training samples are taken as input from the current batch state vector and the actual zone temperature trajectory in the historical firing sample library, and obtained from the same batch retained samples through differential scanning calorimetry. , and and the actual coefficient of thermal expansion obtained through thermomechanical analysis For labeling, thermal analysis was performed in triplicate for each batch, and temperature results deviating from the median by more than 3 degrees Celsius or exceeding [a certain threshold] were discarded. The expansion coefficient results are given in degrees Celsius, and the average of the remaining results is used as the label.

[0071] The samples were stratified according to production date and formula version, divided into a 70% training set, a 15% validation set, and a 15% test set. Time series segments from the same batch were included in only one set. The model was released under the following conditions: the mean absolute error of the softening point and crystallization boundary in the test set did not exceed 6 degrees Celsius, and the mean absolute error of the expansion coefficient did not exceed [a certain value]. / degrees Celsius, if any indicator fails to meet the standard, the previous valid soft measurement model version will be retained;

[0072] The joint training loss of the causal predictor, the partitioned digital twin model, and the phase transition soft measurement model adopts ,in The errors correspond to six categories in sequence: zone temperature, exhaust gas enthalpy, softening point, crystallization initiation temperature, crystallization termination temperature, and expansion coefficient. The difference between the predicted value and the measured value, while retaining the corresponding physical units, The standard deviation of the same type of labels in the training set and its correlation with Same unit, The loss terms are dimensionless and sum to 1, and are adjusted inversely by the validation set error, so each loss term is a dimensionless quantity.

[0073] After model inference Calculate the deviation of the expansion coefficient, where The predicted inflation coefficient for the current batch. The target coefficient of thermal expansion is given by the formulation process card, and both are in units of degree Celsius. Retain the positive and negative directions and write them into the phase transition result record; when the invalid input data flag exists or any of the four outputs of the model exceeds the range of 10% expansion of the 1st to 99th percentile of the training samples, only the previous valid phase transition output of the same batch, the same formula version and whose generation time does not exceed 10 control cycles is allowed to be used, and the cycle is marked as soft measurement low confidence; when there is no previous valid output that meets the conditions, the results of other batches shall not be referenced, and the output available flags of softening point, crystallization start temperature, crystallization end temperature and expansion coefficient deviation shall be set to 0, the corresponding values ​​shall not be written, and the process shall be transferred to the conservative process card boundary or the whole kiln safety gate branch described later.

[0074] The lower and upper boundaries of the phase transition window are respectively set according to and Calculation, where and All units are degrees Celsius. and This is a dimensionless window coefficient, its value determined by the temperature range coverage of qualified crystalline phase samples from the most recent 60 batches of the same formulation and limited to 0.5 to 0.9; when or If the output order is deemed invalid, the previous valid model version is used for inference, and the process is repeated after retries. as well as If the temperature is strictly greater than 0 degrees Celsius and still does not meet the requirement, mark this cycle as low confidence for soft measurement and prohibit the use of the model boundary. Prioritize using the conservative process card boundary that is currently in effect and has the current formulation. After passing the strict positive window width verification, write it into the boundary source enumeration "conservative process card boundary" and set the boundary availability flag to 1. If there is no valid process card boundary, use the most recently valid adaptation boundary from the same batch that has not exceeded 10 control cycles. After passing the same verification, write it into the boundary source enumeration "most recently valid adaptation boundary from the same batch" and set the boundary availability flag to 1. If neither boundary exists or the verification fails, do not write the boundary boundary. and Write the boundary source enumeration "boundary unavailable" and set the boundary available flag to 0. S4 must first check the flag. If it is 0, keep the most recent security instruction, stop incremental optimization and generate a manual review alarm.

[0075] The window coefficient calibration record uses the formula version as the key. The fields include qualified batch number, thermal analysis temperature point, measured temperature trajectory of the heat preservation zone, target crystalline phase mass fraction, candidate coefficient, window coverage and verification error. In this specific implementation, the coefficients are fitted with the most recent 40 batches and verified with the next 20 batches. It is required that the rank correlation coefficient between the duration within the window and the target crystalline phase mass fraction is not less than 0.75. Otherwise, the previous window coefficient version is used.

[0076] The controller uses the median of the three effective thermocouples in the crystallization insulation zone as the measured kiln temperature. , The unit is Celsius and is consistent with , Using the same temperature scale, the system establishes an independent window counting record for each material tray based on the batch identifier and the position of the roller encoder, and writes the window status. Duration within the window And duration can be marked ,in For categorical window states, since enumerated values ​​are used and dimensions are not applicable, the states are sequentially "not entered," "occupied," and "left." The aforementioned duration can be marked with a dimensionless quantity of 0 or 1. When the batch has not yet entered the crystallization insulation zone, the dimensionless non-negative integer count is determined. , s and This zero value only indicates that there is no window duration observation yet and it does not participate in the S4 duration error aggregation; when the tray first enters the crystallization insulation zone... , For "occupation" and Press only while the tray is in the insulation zone when ,otherwise Form a dimensionless binary indicator and recursively deduce Duration within the window The unit is seconds; the counter for this tray is reset to zero when the temperature leaves the window, and resets when the tray leaves the insulation zone. Set as "left", keep The final duration is sealed, and the counting status of multiple overlapping trays is saved separately through batch identifiers without accumulating across batches;

[0077] Batch deviation calculation first involves the formula mass fraction, , , Establish a benchmark for similar batches using historical softening points and historical expansion coefficients, for each batch... Item deviation according to Normalization, in which This is the value corresponding to the current batch. This is the median of the most recent 120 qualified batches in the same formulation category. It is 1.4826 times the median absolute deviation and is consistent with Same unit, For dimensionless deviation components ranging from 0 to 1, when the denominator is lower than the gauge repeatability, gauge repeatability is used as the lower limit for the same unit.

[0078] Batch deviation index according to Calculation, where This represents the total number of included formulations, particle sizes, historical softening points, and historical expansion coefficients. The dimensionless weights obtained from the sensitivity calibration of firing failure rate based on the most recent 60 batches are... , The values ​​are between 0 and 1, and the larger the value, the greater the deviation of the current batch from the historical batches of the same type. When a certain item is missing, the item is removed from the summation and the remaining weights are renormalized.

[0079] The deviation weights are obtained by normalizing the absolute values ​​of logistic regression coefficients constrained by non-negativity and a sum of 1. The weights are verified on the most recent 20 reserved batches, with the firing test exceeding the allowable range of the process card as the standard. The weights are not updated when the area under the subject operating characteristic curve of the reserved batch is less than 0.75. In this specific implementation, a weight record simultaneously saves the weights of the formula component group, particle size group, historical softening point, historical expansion coefficient, training batch range, and effective time.

[0080] When a soft measurement output is marked as low confidence, S3 still retains the current cycle's input tensor, the previous valid phase transition output that meets the aforementioned batch and version conditions, or the absence of a usable rollback flag, the reason for the low confidence, and the original batch identifier verification result of the rollback output, as well as the window boundary source, boundary availability flag, and... , Only the results of the aforementioned actual rollback branches should be used and should not be uniformly overwritten; when the boundary source is a "conservative process card boundary" or a "most recently valid same batch adaptation boundary" and the boundary availability flag is 1, then... The lower limit is raised to 0.6, which increases the phase transition stability weight of S4 and prevents the use of low-confidence outputs to expand the control range for energy saving or waste heat recovery. When the boundary source is "boundary unavailable," no data is provided to S4. and This directly triggers the keep-the-most-safe-instructions, stops incremental optimization, and prompts manual review.

[0081] The phase change analysis module will be set according to the output availability flag. , , and Available values ​​and , , , The soft measurement confidence flag, the availability flags of each output, the rollback source, the model version, and the calculation time are written into the batch phase transition state record, and the batch identifier of the state record is verified to be consistent with the S1 state vector. When the four phase transition outputs are unavailable, null values ​​and reasons are recorded but not written into other batch values. The verified window boundary and the duration within the window are directly used as the S4 rolling optimization constraint, and the batch deviation index is used as the input of the S4 three-category target weight update magnitude.

[0082] In this specific embodiment, S4 includes:

[0083] The rolling optimization module reads the predicted thermal field and exhaust gas enthalpy sequence of each zone of S2 for the next 60 minutes, as well as the phase transition state record of S3, with a control cycle of 30 seconds. The rolling time domain includes 120 prediction steps and the first 20 control steps. The module establishes the current set of active material trays in the kiln based on the batch identifier of S1, inlet and outlet boundary events, and encoder position. And batch to partition occupancy mapping, to Indicates the active batch index, with Indicates partition numbers from 1 to 6, in order to Indicates the prediction step number from 1 to 120, and uses... Indicates batch In the prediction step Does it occupy the first The status of each active batch includes the equivalent temperature of the material in six zones, the heating rate of each zone, the online cumulative residence time, the enthalpy flow of the exhaust gas, the duration within the window, the batch deviation index, the source of the window boundary, and the boundary availability flag. The control quantities of the 21st to 120th prediction steps are fixed to the value of the 20th control step to close the long prediction time domain.

[0084] After the module establishes the active material tray set, it first executes the empty set access control; when the active material tray set... When no batches are included, the maximum batch value operation is not performed on the empty set, and phase change window, dwell time, or energy saving benefit rolling optimization is not performed, with deterministic conventions. , , and The four quantities are only written into the empty kiln cycle record and are not included in the optimization objective; among them, The dimensionless worst-case window duration error is between 0 and 1. A dimensionless worst-case batch deviation index ranging from 0 to 1. For the first Predict the most unfavorable non-negative dimensionless quality violation for all activity batches. Within 120 prediction steps The non-negative dimensionless phase transition stability cost formed by the sum of squares is calculated according to the maximum value and summation formula described later when the moving material tray set is not empty. The empty kiln command sets the conveying speed to 0m / s, the recovery valve opening to 0, and the bypass valve opening to 1. The insulation power and air-fuel ratio of the six zones are read from the empty kiln safety parameter table with the kiln identifier and empty kiln insulation mode as the joint key. The fields of this table include the insulation power of the six zones in kilowatts, the air-fuel ratio in dimensionless form, the upper and lower limits of the equipment permit, the version number, the check code, and the effective time. The parameters are calibrated by the empty kiln heat balance test during the equipment acceptance phase within all nameplate and temperature hard constraints, and are recalibrated and released after the burner, furnace lining, or conveying system is overhauled. When the parameter table verification fails, the unverified value is not used, and the S5 safety shutdown interlock is directly triggered. The empty kiln command records the empty collection flag, the parameter table version, and the access control reason. Only when the moving material tray set The following batch-by-batch calculation, weight adjustment, and multi-batch arbitration are performed only if at least one batch is included.

[0085] The heating power of the six zones, the overall kiln conveying speed, the air-fuel ratio, the bypass valve, and the recovery valve are treated as a single set of kiln-level shared control variables and are included in the same optimization problem. The equipment safety boundaries of all active batches are treated as non-relaxable hard constraints. The phase change window and dwell constraints of each batch are retained and the most unfavorable normalized violation is included in the phase change stability cost. The zone heating power constraint only aggregates the batches currently occupying the zone and those that will enter the zone in the predicted time domain. The overall kiln conveying speed constraint aggregates all active batches. When the constraint requirements of multiple batches are inconsistent, the equipment safety constraints are satisfied first. Then, the quality constraint arbitration order is determined according to the earliest expected kiln exit time, kiln entry time, and batch identifier in ascending order. The optimization output is always a kiln-level instruction version and records the set of covered batches, the zone occupancy mapping, and the constraint status of each batch.

[0086] Collection of materials in the activity tray When at least one batch is included, the module counts and filters records from the S3 window. "Occupied" or "Left" and The batches, forming duration errors, participate in the set. For each According to the target window duration specified in its process card Calculate the dimensionless duration error The "occupied" batch uses the current cumulative value, while the "left" batch uses the S3 archived value. and The units are all seconds. The timeframe is calibrated and limited to 60 to 1800 seconds and strictly greater than 0 based on the crystal phase detection results of qualified batches of the same formula; when the process card value is 0, exceeds the limit, or fails verification, only the most recent valid target duration that is consistent with the current formula version, is within the validity period, and strictly falls within 60 to 1800 seconds can be used, and an invalid target data and rollback flag is set; division must not be performed when there is no qualified rollback value, the target window duration of that batch is marked as unavailable, and the full kiln data access control described later is triggered; when Non-empty time order ,when Determine command when empty ; Another order The index represents the worst-case batch deviation in a non-empty active batch set. and All are dimensionless quantities between 0 and 1. When the active tray set is empty, the aforementioned neutral value is used directly without calculating the maximum batch value; the module will... , , and Write the phase change state record and the current cycle control record;

[0087] Before any batch weight, violation quantity, or objective function calculation, execute the full kiln data gate: If the phase change window boundary of any active batch is unavailable, the window width is not strictly positive, the target window duration is unavailable, the target dwell time of any partition that should be enabled is missing or not strictly positive, or the current position of the batch still in the kiln cannot form a unique occupation mapping with the effective partition geometry table, then stop the full kiln incremental optimization of this control cycle. Do not continue to change the shared control quantity after only excluding abnormal batches. Keep the most recently executed safety instruction confirmed by S5 and generate a manual review alarm containing batch identifier, original failure value, rollback retrieval result, failure field, and data version. Only when all active batches pass this gate will the subsequent basic weight selection, candidate trajectory calculation, and objective function solution proceed.

[0088] The unnormalized values ​​of phase change stability weight, energy saving weight, and waste heat recovery weight are respectively calculated as follows: , and Correction, among which , and The dimensionless basic weights in the process card are summed to 1. When multiple batches coexist, the valid process card of the highest priority active batch is selected as the sole source of the three basic weights, based on the earliest expected kiln exit time, kiln entry time, and batch identifier in ascending order. Each weight must be between 0 and 1, and the sum of the three must be 1. If the original process card weights are invalid, only the most recent process card with the same formula version, within the validity period, and verified by playback is allowed. The original version and the reverted version are recorded. If the three valid weights still cannot be obtained, the aforementioned kiln-wide data access control is executed, the incremental optimization for this cycle is stopped, and the most recent safety instruction is maintained. The lower limit of the three unnormalized values ​​is 0.05. Let the positive normalization denominator be 0.05. According to , and We obtain the three final dimensionless weights and When the most unfavorable risk in the activity batch increases, the weight of phase transition stability increases, while energy saving and recycling targets give way to quality constraints;

[0089] For each activity batch and prediction steps Set non-negative optimization relaxation variables and Only when And the batch boundary can be applied when the flag is 1. ,in For batch Predicted equivalent temperature of materials in the crystallization insulation zone and The boundary for writing the batch status record to S3 and passing the positive window width check is the unit of the four temperature quantities and the two types of relaxation variables. When the relaxation variable is not occupied and does not enter the crystallization insulation zone in the prediction time domain, the corresponding relaxation variable is fixed at 0. The initial upper limit of the two types of relaxation variables is 0 degrees Celsius. When recovery is feasible, it can only be relaxed in 2 degrees Celsius increments to 10 degrees Celsius, and the equipment safety boundary must not be relaxed.

[0090] When evaluating a set of candidate conveyor speed sequences, the initial values ​​are the current batch coordinates obtained from the cumulative displacement of the S1 encoder and the correction of inlet and outlet boundary events, and then... Candidate positions are calculated step by step; among which and Batch Current and the The predicted axial coordinates within the kiln are given in meters. For 30 seconds, sum the index. From 1 to The prediction step number, For the candidate sequence The conveying speed is measured in meters per second; the six-zone geometric table records the lower and upper boundary coordinates of each zone, which are non-overlapping and strictly increasing along the conveying direction, using the kiln identifier and geometric version as keys. All coordinates are in meters. When a candidate position falls within the left-closed, right-open coordinate interval of a zone, the corresponding... Set the value to 1 for all partitions and 0 for the rest, and record the first candidate step that reaches or crosses the upper boundary of that partition as . Candidate locations are allowed to cross shared boundaries of adjacent zones during forward transport. Upon crossing, the original zone departure step and the downstream adjacent zone entry step are recorded sequentially, and the unique occupancy mapping is updated. If a single 30s prediction step crosses multiple adjacent boundaries, the timing of the crossing is determined based on the candidate velocity and the coordinates of each boundary, and the sub-steps within that prediction step are used for recursion. When a batch first arrives at or crosses the upper boundary of the sixth zone during forward transport, this step is recorded as a predicted kiln exit step. This step is not counted in the occupancy of the sixth zone. From this step onwards, the six occupancy steps of the batch are... All values ​​are set to 0 and marked "predicted to have exited the kiln". Subsequent prediction steps no longer require it to uniquely fall into the six zones, nor do they apply temperature windows or zone dwell constraints. However, the batch index and the pre-kiln exit occupancy sequence are retained for calculating the estimated remaining kiln time, aggregation weights, and energy consumption baselines. Other zones will be set to zero if they do not reach the upper boundary within the prediction time domain. The value is truncated to 120 and a non-leaving zone flag is written. For each group of candidate control quantities, the batch position, occupancy mapping, leaving zone step, predicted kiln exit step, and dwell violation are recalculated. The aforementioned whole kiln data gate is executed without evaluating the candidate solution only when the candidate speed is paused or reversed, the geometric boundary is missing, overlap or not strictly increasing, the batch position falls outside the effective axial range of the whole kiln and does not leave the kiln through the designed outlet of the sixth zone, or a definite adjacent positive cross-zone sequence cannot be formed. Normal crossing of the shared boundary of adjacent zones does not trigger the gate.

[0091] For batch In the Cumulative online dwell time in each zone Using the second value and register status generated by S1 according to a unified clock: when the status is "not entered", 0 is used as the initial value of the cumulative stay for the candidate partition, and the value is accumulated from the first entry step of the candidate; when the status is "occupied", the current online cumulative value is read; when the status is "left", the sealed value is read, but the future stay constraint for the partition is no longer enabled; according to Predicting up to the The cumulative dwell time of each step; The batch leaving the aforementioned candidate positions was recursively determined. The step number of the partition, with a cutoff value of 120 used when the region has not left the partition in the prediction time domain, then the expected completion dwell time is calculated as follows: Calculate, where the summation index is used. For the prediction step number, The duration is 30 seconds; the unit for all three types of dwell time is seconds. Target dwell time in the zone Read batch The valid process card must be strictly greater than 0, and only if the batch is currently occupied or will enter the predicted time domain. The corresponding dwell constraint is only activated when partitioning; the aforementioned full-kiln data access control is executed when the target is invalid.

[0092] The feasible range screening first determines the recoverable heat based on the predicted exhaust gas temperature, predicted exhaust gas flow rate, and the operating status of the waste heat exchanger. ,in The dimensionless effective efficiency obtained from the most recent thermal balance test of the heat exchanger is limited to 0.45 to 0.78. and The units are all kilowatts. The bypass valve opening is then determined based on the heat exchanger's permissible heat load, flue gas pressure differential, and the minimum stable valve opening. and recovery valve opening The upper and lower bounds;

[0093] Valve position feasible range specification , and When the heat exchanger is shut down, or the flue gas pressure difference exceeds 2.5 kPa, or Fixed when below 20kW And order In other states, the load table of the recycling equipment will be used as the basis for calculation. The load is limited to a range that ensures the recovered heat load does not exceed the rated value. The load meter uses the exhaust gas temperature range, mass flow rate range, and equipment status as keys and is updated after the heat exchanger is cleaned or repaired.

[0094] Control vector of the rolling optimizer ,in For the first Zoned heating power, measured in kilowatts. The conveying speed is expressed in meters per second. Air-fuel ratio, and For dimensionless valve positions, the state transition adopts the S2 digital twin model. The upper and lower bounds of the control variables are derived from the heater nameplate, the allowable speed of the roller conveyor from 0.002 to 0.012 m / s, the air-fuel ratio of the burner from 0.95 to 1.25, and the aforementioned feasible valve position range, respectively.

[0095] For each activity batch and prediction steps Calculate the overtemperature amount and undertemperature Among them, strictly positive For batch The window width is given in degrees Celsius; the excessive dwell time is calculated for each enabled partition dwell constraint. Insufficient stay All four types of violations are non-negative dimensionless quantities; for batches For partitions that are currently unoccupied and will not be entered within the 120-step prediction time domain of the candidate sequence, the corresponding dwell constraints will not be enabled, and the excessive dwell time and insufficient dwell time of the batch corresponding to the partition will be deterministically set to 0 before the calculation of the maximum value of the six partitions.

[0096] Collection of materials in the activity tray When it contains at least one batch, let For batch In the Predict the worst-case quality violation of the step and let Aggregate all activity batches; phase transition stabilization costs within the rolling window are calculated as follows: When calculating and finding the maximum value, the batch, partition, violation type, and direction that generated the value are simultaneously saved, and the numerical difference does not exceed [a certain value]. The kiln exit time, kiln entry time, and batch identifier should be sorted in ascending order to eliminate ties. A smaller value indicates that the temperature window and dwell constraints of each activity batch under shared control are closer to being satisfied simultaneously; when the set is empty, the sum of the two maximum values ​​in this section is not performed, and the aforementioned value is used directly. and Record;

[0097] Energy saving cost Calculation, where The control step numbers are 1 to 20. The partition numbers are 1 to 6. The unit is kilowatt. The value is 30 seconds, converted to hours by dividing by 3600, with the numerator in kilowatt-hours; for each batch in the non-empty active batch set... Read the formula version and tray mass from the S1 batch status record. Convert all tray masses with mass unit enumeration and that have passed weighing verification to kilograms. Determine a unique tray mass interval starting from 0 kg and with a width of 5 kg (left-closed, right-open). Masses exactly at the upper boundary of the interval are assigned to the next mass interval. Based on the formula version and the tray mass interval, retrieve the most recent (but no more than 60) qualified batches with the same formula version and tray mass interval up to the current status time. Record the arithmetic average of their unit batch power consumption as... And the unit is kilowatt-hour; if the material tray mass is missing, negative, unit enumeration or weighing verification is invalid, or historical retrieval is empty, the recorded quality mark is invalid, or the average value is less than 0.1 kilowatt-hour, use the positive initial benchmark of not less than 0.1 kilowatt-hour from the valid process card of this batch that has been verified by playback, and record the batch rollback mark, quality range, benchmark source, benchmark version, and sample cutoff time; according to Batch obtained The positive value in the prediction time domain is the estimated remaining time in the kiln, in seconds, and then... The dimensionless aggregate weights are obtained, where The index is the summation index for the activity batch, with each weight strictly greater than 0 and the sum equal to 1; the unique energy consumption benchmark for kiln level is based on... The calculation is in kilowatt-hours, since the benchmark for all batches is no less than 0.1 kilowatt-hours. Maintain a strictly positive value; when multiple formulas or multiple material tray quality ranges coexist, they should be uniquely determined according to the same rules. It is a dimensionless quantity, and the smaller it is, the lower the energy consumption in the shared control time domain. It is not calculated when the set is empty. or And implement the aforementioned empty kiln access control;

[0098] Waste heat recovery revenue is calculated based on Calculation, where The control step numbers are 1 to 20. For the dimensionless opening of the recovery valve, For kilowatts, Convert seconds to hours; The baseline recovered heat for the most recent 60 valid batches, in kilowatt-hours, is required. Not less than 0.1 kWh. If the value is below this lower limit or there is no effective recovery sample, use the initial positive reference determined by the heat exchanger's rated test and record the retreat status. When the heat exchanger is shut down... thereby , It is a dimensionless quantity, and the larger the value, the higher the recovery benefit;

[0099] The optimization objective is to minimize ,in , and For the weights after online correction, This is the sum of squared changes of each control variable relative to its equipment range in adjacent steps, after normalization. To suppress the dimensionless coefficients controlling jitter, the waste heat gain term is negative to increase the recovery benefit and reduce the total cost. The rolling optimizer uses sequential quadratic programming and limits the solution time to 5 seconds in each 30-second cycle.

[0100] Before the optimization parameters are released, replay verification is performed using the 20 most recent qualified batches and 10 boundary batches. The verification indicators include window boundary violation duration, unit batch power consumption, waste heat recovery, control variable change rate, and number of solution timeouts. It is required that the number of hard constraint violations of all equipment is 0 and the phase change window violation duration is not higher than the previous parameter version. If the verification fails, the update of basic weight, heat exchange efficiency, or control jitter coefficient will not be released.

[0101] For each activity batch and prediction step, four types of dimensionless quality violations are calculated: overheating is measured by dividing the positive difference between the predicted temperature and the upper boundary by the strictly positive window width; underheating is measured by dividing the positive difference between the lower boundary and the predicted temperature by the window width; excessive dwell time is measured by dividing the positive difference between the cumulative online dwell time and the target value by the strictly positive target dwell time; and insufficient dwell time is measured by dividing the positive difference between the target value and the expected completion dwell time by the target value. For each type, the maximum value is first taken within the prediction step, and then a ranking is determined by prioritizing equipment safety, from largest to smallest violation, from earliest to latest expected kiln exit time, from earliest to latest kiln entry time, and in ascending order of batch identifier. The first item in the ranking is defined as the dominant violation. The power direction within the same zone is determined by the highest-ranked batch occupying that zone, and the kiln speed direction is determined by the highest-ranked item across all activity batches. Values ​​equal must not exceed a certain threshold. The dimensionless differences are considered parallel and eliminated according to the above time and batch identifier order;

[0102] When the solver obtains a solution that satisfies the hard constraints within 5 seconds, it only issues the first control step's partition heating power, conveyor speed, air-fuel ratio, bypass valve opening, and recovery valve opening; if the original problem is infeasible, it first relaxes the constraints according to the aforementioned levels. and If it is still not feasible to maintain the upper limit but not relax the equipment safety limits, the dominant violation will cover the instructions of the previous cycle: if overheating or excessive residence is predicted, the power of the relevant zone will be reduced by 10% and the conveying speed will be increased by 10%; if underheating or insufficient residence is predicted, the power of the relevant zone will be increased by 5% and the conveying speed will be reduced by 10% within the nameplate and temperature rise constraints; if flue pressure difference or heat exchanger failure occurs, the most recent safe power and speed will be maintained, the recovery valve will be closed and the bypass valve will be fully opened; if the opposite direction violations of different activity batches for the same shared speed or the same zone power all exceed 1 and any action in one direction will cause the violation in the other direction to continue to increase, it will be judged as an unresolvable same-level conflict, the most recent safe parameters will be maintained, incremental issuance will be prohibited and it will be transferred to manual review; the above power, speed and valve position are all truncated to the upper and lower limits of the control quantity defined in S4, and the control quantity not covered by this branch will retain the value of the previous cycle. The equipment safety interlock in S5 always has the highest priority, and the dominant constraint, sorting basis, coverage field, conflict status and optimization infeasibility alarm will be recorded at the same time.

[0103] When a numerical solution has been obtained but a violation of any hard constraint is detected, such as power, conveying speed, air-fuel ratio, valve position, or flue pressure difference, the rolling optimization module determines the solution as invalid and prohibits its issuance. It records the name, direction, amount of violation, batch, and prediction step number of the violation constraint, and then generates a unique rollback instruction according to the same dimensionless dominant violation. When an unresolvable conflict of the same level occurs, the most recent safe parameter is maintained, incremental issuance is prohibited, and manual review is required. No branch may exceed the higher priority of the S5 device interlock.

[0104] Control commands are written to the control command queue using the kiln identifier and effective start time as a joint key. Fields include a unique kiln-level version number, the set of covered batches, batch-to-zone occupancy mapping, the geometry version used for the occupancy mapping, the target power for the six zones, the target conveying speed, the target air-fuel ratio, the target opening degree of the two valves, the effective start time, the 30-second validity period, the optimizer version, the phase change weight, the energy saving weight, the waste heat recovery weight, the source batch and process card version of the base weight, the batch-by-batch energy consumption baseline, the batch-by-batch material tray quality range, the estimated remaining time in the kiln, the aggregation weight, the kiln-level energy consumption baseline, the source of the batch-by-batch baseline, the baseline version, the sample cutoff time, the constraint status of each batch, and the overall kiln data. Access control status, empty set flag, and infeasibility status flag; when generating queue records, the kiln identifier, kiln-level version number, the covered batch set sorted in ascending order by batch identifier, the target power of zone six, the target conveying speed, the target air-fuel ratio, the target opening degree of the bypass valve and recovery valve, the effective start time, and the validity period are serialized in the above fixed field order and UTF-8 encoding. The instruction content hash is calculated using the SHA-256 algorithm, and the hash algorithm identifier, the normalized serialization version, and the expected hash value are written into the queue. Only one kiln-level version is allowed to be in the executable state within the same validity period, and it is read by the S5 execution feedback module in the next PLC scan cycle.

[0105] In this specific embodiment, S5 includes:

[0106] The execution feedback module reads the unique kiln-level instruction currently valid from the control instruction queue. It first recalculates the instruction content hash according to the normalized serialized version written in S4, fixed field order, UTF-8 encoding, and SHA-256 algorithm, and compares it with the expected hash value in the queue. Then, it verifies that the batch set it covers is consistent with the current active material tray set in S1, and that there is no second executable version within the same validity period. After all verifications pass, the PLC converts the target power of the six zones into the heater power regulator setpoint, the target conveyor speed into the roller conveyor frequency converter frequency, and the target air-fuel ratio into the setpoint for the gas valve and combustion fan linkage. It also sends the target opening degree of the bypass valve and recovery valve to the corresponding electric actuator. If the set is inconsistent, the version is duplicated, or the hash verification fails, the incremental instruction is rejected while maintaining the most recent safety parameters. Simultaneously, the recalculated hash value, expected hash value, algorithm identifier, serialized version, and rejection reason are recorded.

[0107] Before the instruction is executed, the equipment interlock check is performed and the priority is determined in the order of combustion safety, roller conveyor stall, zone overheating, flue high pressure, and heat exchanger failure: when flame detection fails, the corresponding gas valve is closed and the safe purge air volume is maintained; when the roller conveyor stalls, the conveyor drive is stopped, the heating of the affected and upstream zones is cut off, and the material protection shutdown is initiated; when a zone overheats and there is no stall, the heating of that zone is cut off and the conveyor speed is increased within the allowable range; when the flue pressure difference exceeds 2.5 kPa or the heat exchanger fails, the recovery valve is closed, the bypass valve is fully opened, and the combustion load is limited; when concurrent faults occur, the highest priority action is used and the non-conflicting low priority actions are merged; the fault can only be reset after it disappears for three consecutive PLC scan cycles and is confirmed by the operator. At the same time, the reason for rejection, the interlock time, the final safety parameter set, and the reset status are written in the control instruction record.

[0108] Under normal operating conditions, the equipment sets limits on the rate of change of heating power, conveying speed, and valve position. The heating power change in a single cycle shall not exceed 8% of the rated power of the corresponding zone, the conveying speed change in a single cycle shall not exceed 0.001 m / s, and the valve position change in a single cycle shall not exceed 0.12. The PLC collects the actual zone temperature, encoder speed, actual exhaust gas temperature, actual exhaust gas flow rate, and actuator position feedback every 2 seconds, and calculates the cumulative displacement when batches enter the kiln. Subsequently, the integral encoder displacement was identified by batch. ,in For batch identification, For sampling sequence number, The unit is meters. The encoder speed is expressed in meters per second. The sampling period is 2 seconds; the photoelectric switches at the entrance and exit of each zone provide... and ,in The partition numbers are 1 to 6, and the unit for both times is seconds. The actual dwell time is calculated as follows: The calculation is in seconds, and the pause time is included in the dwell time. When there is a reversal, slippage, or encoder failure, the boundary photoelectric event is used as the standard and an invalid trajectory flag is set. Batches that lack complete entry and exit events are not included in the training set.

[0109] Feedback acquisition uses a PLC monotonically increasing cycle counter for alignment. The original sampled values, engineering units, quality marks, and calibration versions of the actual zone temperature, actual residence time, actual exhaust gas temperature, and actual exhaust gas flow are all saved. When the effective time of the actuator feedback and the control command differs by more than 4 seconds, the cycle is marked as an execution delay and excluded from the real-time model update sample.

[0110] The current rolling period has ended and the number of valid sampling points has been reached. Calculate the root mean square deviation of the temperature trajectory in time ,in It is a dimensionless positive integer. and The first The actual temperature and previously predicted temperature of the zone at the same moment, both in degrees Celsius. The unit is Celsius; when If the formula is not executed, the deviation is marked as invalid and the model update is prohibited for that cycle. At the same time, the difference between the actual residence time and the predicted residence time of the effective batch and the difference between the actual exhaust gas enthalpy flow and the predicted exhaust gas enthalpy flow are calculated.

[0111] The actual firing results are written after the batch leaves the kiln and completes the inspection. The fields include the softening point, crystallization start temperature and crystallization end temperature measured by differential scanning calorimetry, the expansion coefficient measured by thermomechanical analysis, the target crystalline phase mass fraction measured by X-ray diffraction, the product moisture content, the inspection time and the inspection equipment number. Each result retains three parallel test values, the effective mean and the anomaly rejection flag.

[0112] The historical firing results database stores the following data using the batch identifier as the primary key: S1 current batch state vector, S2 predicted thermal field and exhaust gas enthalpy, S3 phase change output, S4 control command, actual zone temperature, actual residence time, actual exhaust gas temperature, actual exhaust gas flow rate, actual actuator position, and actual firing results. It also stores the data mode version, model version, process card version, and sensor calibration version, so that subsequent training samples can be reproduced according to the same time reference and version relationship.

[0113] Historical firing results are recorded by batch end events. The system first verifies that the state vector, prediction results, control commands, actual trajectories and test results have the same batch identifier. Then, the actual trajectories are sorted by a unified timestamp and a summary of the number of valid sampling points, missing test ratio, trajectory start and end times and data completeness is generated. When the test results have not yet been returned, the record status is set to pending inspection and it is not included in the training set. The record status is updated to valid only after the inspection is completed and the completeness is not less than 95%.

[0114] Online trajectory correction criteria Digital twin online correction is triggered when any of the following conditions occur consecutively for three valid rolling cycles: exceeding 12 degrees Celsius, the absolute value of the actual residence time deviation exceeding 5% of the target value, or the absolute value of the exhaust gas enthalpy deviation exceeding 8% of the reference enthalpy. Offline quality correction is independently determined when the batch leaves the kiln and the thermomechanical analysis results arrive at the event. It is triggered when the actual expansion coefficient deviation of at least two of the three most recently inspected and valid batches of the same formula exceeds [a certain threshold]. The soft measurement model is corrected when the temperature reaches / degrees Celsius. The offline results are only counted as one batch event and are not counted repeatedly in the rolling cycle. Each threshold is derived from the 95th percentile of the deviation of the most recent 60 qualified batches and is updated with the process card version.

[0115] After online trajectory correction is triggered, the partitioned digital twin model performs a recursive least squares update with a forgetting factor of 0.995 based on the actual state and control variables of the current cycle. The single change of heat transfer and heat storage parameters is limited to 2% of the original value. After offline quality correction is triggered, the phase change soft measurement model adds the batches that have obtained offline thermal analysis annotations and passed the integrity verification to the incremental training cache and updates the last layer regression parameters with one-tenth of the original learning rate. The new parameters will take effect no earlier than the next effective control cycle of the next batch. Before obtaining the actual firing results, only the digital twin state bias is allowed to be updated, and the soft measurement model is not updated.

[0116] The next cycle prediction parameter package includes the updated digital twin state bias, heat transfer coefficient, soft measurement regression layer parameters, normalization parameter version, training sample cutoff batch, validation metrics, and rollback version number. The parameter package is verified by content hash and effective control cycle. S2 and S3 will only switch versions after both read the same effective cycle and the hash verification passes; otherwise, the previous valid parameter package will continue to be used.

[0117] The online correction process retains the prediction results, error statistics and control simulation results of the model before and after correction for the same verification batch. The digital twin temperature error, tail gas enthalpy flow error, softening point error, crystallization boundary error and expansion coefficient error are compared according to the original physical units. The errors of different units are not directly summed. The verification record is saved with the parameter package for at least three model versions to support deterministic rollback.

[0118] The updated model is first validated by replaying on the 20 most recent valid batches that did not participate in this incremental training. The next cycle model version is only released when the root mean square deviation of temperature, exhaust gas enthalpy deviation, and the four phase change output errors are all no higher than the previous model and at least one of them is reduced. Otherwise, the previous valid version is restored and the reason for the rollback is recorded. The execution feedback module finally writes the new version number, correction status, parameter package hash, effective period, and next cycle prediction parameters back to the parameter repository of the partitioned digital twin model and the phase change soft measurement model. The release or rollback event is recorded in the historical burn-in result database. The record also saves the threshold item that triggers the correction, the number of consecutive out-of-tolerance periods, the list of validation batches, the error before and after release, the data integrity, and the operator's confirmation status. This allows the model update, command execution, and batch results to be fully traced and audited cycle by cycle, completing the closed-loop update.

[0119] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0120] This invention forms a continuous control chain around the technical problem of "synchronous control of zoned thermal field changes, material residence time lag, exhaust gas waste heat recovery, and powder phase change window" during glass powder sintering. Specifically, a zoned digital twin model is used to characterize the heat transfer relationship and operating status between different zones of the kiln. A causal constraint multi-step predictor restricts the use of the historical states of the current zone and upstream zones during the prediction process, and recursively predicts the temperature of subsequent zones and exhaust gas enthalpy by combining the material residence time lag. The phase change soft measurement model further outputs the softening point, crystallization start temperature, crystallization end temperature, and expansion coefficient deviation based on the current batch status and real-time temperature field, converting the powder phase change process into a window boundary and duration within the window that can participate in online optimization. The rolling optimizer then jointly solves the zoned heating power, conveying speed, air-fuel ratio, bypass valve opening, and recovery valve opening based on the predicted thermal field, exhaust gas enthalpy, phase change window constraints, and batch deviation index. Therefore, the above algorithm combination can transform the originally dispersed temperature control, residence time control and waste heat recovery control into a coordinated control process within the same rolling time domain, so that the control commands can directly act on the kiln thermal field, material phase change state and tail gas energy flow, thereby reducing control lag, improving the consistency of crystal phase precipitation and reducing firing energy consumption.

[0121] This invention provides adaptive improvements to address issues such as upstream and downstream thermal coupling, material transport lag, batch formulation and particle size fluctuations, and phase change window drift during glass powder calcination. Firstly, it introduces a partition coupling matrix and a material residence time lag matrix into the multi-step prediction process, and sets causal constraints to ensure that the prediction process conforms to the physical relationship of heat and material transfer along the transport direction in a continuous kiln, avoiding the impact of future downstream states on upstream control that is inconsistent with process logic. Secondly, in phase change soft measurement, the softening point, crystallization initiation temperature, crystallization end temperature, and expansion coefficient deviation are used as model outputs, and phase change window boundaries and constraint relaxation amounts are formed accordingly. This allows the algorithm to transform phase change states that are difficult to measure directly online into control constraints that can be handled by a rolling optimizer. Thirdly, it uses a batch deviation index to correct the phase change stability weight, energy-saving weight, and waste heat recovery weight online, enabling the control strategy to adjust according to changes in raw material formulation, particle size distribution, and historical calcination results, avoiding imbalances in quality and energy consumption targets caused by fixed weights under batch fluctuations. Through the above structural improvements, the algorithm of this invention is not a simple application of a general prediction model or optimization model, but is matched with the thermal field coupling, phase change window and waste heat recovery characteristics in the glass powder firing temperature control scenario, thereby enabling more stable synergistic optimization among zoned temperature, material residence time, phase change quality and energy utilization.

Claims

1. A method for controlling the firing temperature of glass powder, characterized in that, include: S1. Collect the temperature, heating rate, material residence time, raw material formula, particle size distribution, tail gas temperature, tail gas flow rate, waste heat recovery status and historical firing results of each zone of the kiln, and form a batch status vector. S2. Based on the state vector, call the causal constraint multi-step predictor and the partitioned digital twin model to output the predicted thermal field and exhaust gas enthalpy value of each partition. S3. Based on the state vector and real-time temperature field, call the phase change soft measurement model to output the softening point, crystallization start temperature, crystallization end temperature and expansion coefficient deviation. Determine the window boundary based on the softening point, crystallization start temperature and crystallization end temperature, and calculate the duration within the window and the batch deviation index. S4. In the rolling time domain, based on the predicted thermal field, exhaust gas enthalpy, window boundary, duration within the window, and batch deviation index, the phase change stability weight, energy saving weight, and waste heat recovery weight are corrected, and the constraint relaxation amount formed by the window boundary is input into the rolling optimizer to solve for the partition heating power, conveying speed, air-fuel ratio, bypass valve opening degree, and recovery valve opening degree, and generate control commands. S5. Apply control commands to the kiln and collect actual zone temperature, actual residence time, exhaust gas temperature, exhaust gas flow rate and actual firing results. Write the actual zone temperature, actual residence time, exhaust gas temperature, exhaust gas flow rate and actual firing results into the historical firing result database and feed them back to the zone digital twin model and phase change soft measurement model to update the prediction parameters for the next cycle.

2. The glass powder firing temperature control method according to claim 1, characterized in that, Step S1 includes: Data from sensors in each partition are collected according to a preset sampling period. The collected data is then processed by time alignment, outlier removal, and normalization to form the batch state vector.

3. The glass powder firing temperature control method according to claim 1, characterized in that, Step S2 includes: A partition coupling matrix is ​​established based on the transmission relationship between the upstream and downstream partitions of the kiln, and a material residence time delay matrix is ​​established based on the conveying speed and partition length. When the batch state vector is input into the causal constraint multi-step predictor, the prediction is restricted to using only the historical state of the current partition and the upstream partition, and the temperature and exhaust gas enthalpy of each partition are predicted step by step. The prediction deviation is then written back to the state parameters of the partition digital twin model.

4. The glass powder firing temperature control method according to claim 1, characterized in that, Step S3 includes: By inputting the historical firing results, the current batch state vector, and the real-time temperature field into the phase change soft measurement model, the softening point, crystallization initiation temperature, crystallization termination temperature, and expansion coefficient deviation are obtained. The lower and upper boundaries of the phase transformation window are formed based on the softening point, crystallization initiation temperature, and crystallization end temperature, and the duration of time during which the measured kiln temperature is continuously located between the lower and upper boundaries is taken as the duration within the window.

5. The glass powder firing temperature control method according to claim 1, characterized in that, Step S4 includes: The phase change stability weight, energy saving weight, and waste heat recovery weight are corrected online based on the duration within the window and the batch deviation index. The constraint relaxation amount formed by the window boundary is input into the rolling optimizer to generate control commands for zone heating power, conveying speed, air-fuel ratio, bypass valve opening degree, and recovery valve opening degree.

6. The glass powder firing temperature control method according to claim 1, characterized in that, Step S5 includes: The actual zone temperature is compared with the predicted temperature trajectory in the predicted thermal field of each zone, and the deviation is determined by combining the actual residence time, actual exhaust gas temperature, actual exhaust gas flow rate and actual firing results; when the deviation exceeds the preset correction threshold, the zone digital twin model and phase change soft measurement model are corrected online using the actual firing results.

7. The glass powder firing temperature control method according to claim 4, characterized in that, The batch deviation index is determined by normalizing and weighting the corresponding parameters of the current batch formula, particle size distribution, historical softening point, and historical expansion coefficient with those of historical samples of the same batch, and is used as input to correct the update magnitude of the phase change stability weight, energy saving weight, and waste heat recovery weight.

8. The glass powder firing temperature control method according to claim 5, characterized in that, Step S4 is divided into two levels: feasible interval screening and control quantity solution. The feasible interval screening determines the feasible intervals of the bypass valve opening and the recovery valve opening based on the exhaust gas temperature, exhaust gas flow rate and waste heat recovery status. The control quantity solution jointly solves the zone heating power, conveying speed and air-fuel ratio within the feasible intervals.

9. The glass powder firing temperature control method according to claim 3, characterized in that, The causal constraint multi-step predictor and the partitioned digital twin model use the current batch state vector, actual partition temperature trajectory, actual exhaust gas temperature trajectory, actual exhaust gas flow trajectory, and exhaust gas enthalpy value labeling obtained by converting actual exhaust gas temperature and actual exhaust gas flow in the historical firing sample library as training samples, respectively. The phase change soft measurement model uses the current batch state vector, actual zone temperature trajectory, and softening point, crystallization start temperature, crystallization end temperature, and expansion coefficient labels obtained through offline thermal analysis as training samples. The model parameters are updated with the joint loss of zone temperature error, tail gas enthalpy error, softening point error, crystallization start temperature error, crystallization end temperature error, and expansion coefficient error.

10. A glass powder firing temperature control system for performing the method according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to read historical firing results from the historical firing result database and collect data on the temperature, heating rate, material residence time, raw material formula, particle size distribution, tail gas temperature, tail gas flow rate and waste heat recovery status of each zone of the kiln, forming a batch status vector. The causal prediction module is used to call the causal constraint multi-step predictor and the partitioned digital twin model to output the predicted thermal field and exhaust gas enthalpy value for each partition. The phase change analysis module is used to receive the batch state vector and real-time temperature field output by the data acquisition module, call the phase change soft measurement model, output the softening point, crystallization start temperature, crystallization end temperature and expansion coefficient deviation, and determine the window boundary based on the softening point, crystallization start temperature and crystallization end temperature, and calculate the duration within the window and the batch deviation index. The rolling optimization module is used to generate control commands based on the predicted thermal field, exhaust gas enthalpy, window boundary, duration within the window, and batch deviation index. The execution feedback module is used to apply the control commands to the kiln and collect the actual zone temperature, actual residence time, exhaust gas temperature, exhaust gas flow rate and actual firing results. The actual zone temperature, actual residence time, exhaust gas temperature, exhaust gas flow rate and actual firing results are written into the historical firing result database and fed back to the causal prediction module, the zone digital twin model and the phase change analysis module respectively to update the prediction parameters for the next rolling cycle.

Images