Model prediction based finishing temperature optimization system

Abstract

The present application relates to hot continuous rolling temperature control, and discloses a model prediction-based finishing temperature optimization system and method. The system establishes material logical segments by length and compensates calculation, communication and valve delay, uses a thickness-direction heat transfer model and a residual model trained by reliable samples to predict surface, core and finishing temperatures; constructs spectral consistency residual, energy balance residual and valve flow residual for the same material segment, determines the contribution of measurement pollution, heat transfer drift and actuator abnormality through a sensitivity matrix, and adjusts measurement covariance, process covariance and actuator covariance respectively to obtain the mean value and confidence interval of the finishing temperature; under water quantity, speed and equipment constraints, probability constraint rolling optimization with upper and lower limit risk allocation is used to execute the first control quantity, and sample learning according to abnormal causes, heat transfer parameter updating or spray capacity tightening are performed. The present application can reduce the risk of incorrect adjustment caused by temperature measurement distortion and cooling capacity drift.

Description

Technical Field

[0001] This invention belongs to the field of hot strip rolling production process control technology, specifically, it relates to a model-based prediction-based final rolling temperature optimization system. Background Technology

[0002] The final rolling temperature is a key control variable in the hot continuous rolling temperature regime, typically measured by a radiation temperature measuring device located at the exit of the last stand in the finishing mill. In the finishing zone, the strip sequentially undergoes air convection, radiative heat dissipation, roll contact heat exchange, plastic deformation heat generation, and inter-stand water spray cooling. Furthermore, the strip speed, reduction, thickness, thermal properties of the steel grade, oxide scale condition, water temperature, water pressure, and nozzle operating status all change with the production process. Therefore, the final rolling temperature control object exhibits characteristics such as nonlinearity, control delay, time-varying parameters, and multi-actuator coupling.

[0003] Current methods for controlling the final rolling temperature typically employ heat transfer mechanism models or statistical models to predict the final rolling temperature. The water volume between stands and the strip threading speed are then adjusted based on the inlet temperature difference, outlet temperature difference, or predicted temperature difference. Alternatively, the strip is divided into logical segments, and control commands are issued in advance based on the speed integral. These methods can improve the accuracy of final rolling temperature measurement when equipment conditions and temperature measurement conditions are stable.

[0004] However, existing controls typically treat the deviation between the predicted and measured final rolling temperature as the same type of deviation. In reality, the same temperature deviation may be caused by different physical reasons: water mist, water film, oxide scale shedding, or changes in emissivity in the outlet temperature measurement field can cause measurement offsets, while the actual thermal state of the strip may not change synchronously; nozzle blockage, water temperature changes, or alterations in the boiling heat transfer zone can cause the effective heat transfer capacity to drift, resulting in changes in the actual strip temperature; valve dead zones, changes in water supply differential pressure, or flow meter deviations can cause the actual spraying action to deviate from the control command.

[0005] When the controller cannot distinguish between the above causes, it may produce erroneous control in the opposite direction. Mistaking measured contamination for actual temperature drop will cause the controller to reduce cooling water or increase speed, thus raising the actual final rolling temperature. Mistaking actual heat transfer attenuation for measured noise will maintain the original control settings, causing the temperature deviation to continue to expand along the strip length. If contamination samples are used for online learning, measurement deviations may also be solidified into model parameters, affecting subsequent material sections.

[0006] Furthermore, existing point predictive control only provides the optimizer with a single temperature value, without explicitly indicating the reliability of the prediction under the current operating conditions. When steel grade or specification changes, the temperature measurement field of view is disturbed, or the spraying capacity decreases, even if the point predicted value is close to the target temperature, the prediction variance may still increase significantly. In this case, if optimization continues to be performed close to the upper and lower temperature limits, the risk of the actual final rolling temperature exceeding the limits cannot be explicitly constrained.

[0007] In view of this, the present invention is proposed. Summary of the Invention

[0008] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by the present invention is as follows: The model-predictive final rolling temperature optimization system is deployed on the process control server of the hot strip finishing mill and communicates with the finishing mill's temperature measurement devices, speed and position detection devices, rolling condition detection devices, inter-stand spraying devices, and basic automation system, including: The process data acquisition module is used to collect signals from at least two radiation temperature measurement channels corresponding to the same material segment, material running speed, rolling condition parameters, spray control commands, actual spray flow rate, and spray branch pressure difference; The material segment synchronization module is used to divide the strip steel into material logic segments along the length direction, update the position of each material logic segment according to the integral result of the material running speed, and determine the control command issuance time based on the estimated time of the material logic segment arriving at the spray zone, as well as the calculation delay, communication delay and spray valve response delay. The thermal state prediction module is used to obtain the a priori predicted values ​​of surface temperature, core temperature, effective heat transfer coefficient, and final rolling temperature for each material logic segment using a thickness-oriented unsteady-state heat transfer model and a data residual model. The residual cause decomposition module is used to construct, for the same material logic segment, a spectral consistency residual characterizing the consistency of the at least two radiation temperature measurement channels, an energy balance residual characterizing the consistency of the enthalpy change of the material segment with the heat balance of the rolling process, and a valve flow residual characterizing the consistency of the measured flow rate of the spray and the expected flow rate determined based on the valve opening and the pressure difference of the spray branch. It also determines the abnormal contribution weights of measurement contamination, heat transfer drift, and actuator abnormality based on a pre-calibrated residual sensitivity matrix. The uncertainty propagation module is used to adjust the measurement covariance based on the abnormal contribution weight of measurement contamination, adjust the thermal state process covariance based on the abnormal contribution weight of heat transfer drift, adjust the actuator covariance based on the abnormal contribution weight of actuator anomaly, and propagate the adjusted covariances to the finishing mill exit to obtain the predicted mean and predicted variance of the final rolling temperature. The confidence constraint optimization module is used to solve a rolling optimization problem containing a final rolling temperature probability constraint under equipment constraints, using the inter-stand spray flow rate and rolling speed correction amount of each material logic segment in the prediction time domain as decision variables. The final rolling temperature probability constraint is determined based on the predicted mean of the final rolling temperature, the predicted variance of the final rolling temperature, and the risk amount allocated to the lower and upper temperature limits. The execution and safety interlocking module is used to execute the first control quantity in the control sequence obtained from the rolling optimization, and to stop online learning and switch to the preset safety control setting when there is no feasible solution that satisfies the equipment constraints or the abnormal contribution weight does not meet the preset confidence condition.

[0009] In a preferred embodiment of the present invention, the temperature measuring device includes a dual-band pyrometer, or includes a first pyrometer and a second pyrometer with overlapping fields of view and different spectral responses; the spectral consistency residual is determined by the temperature difference obtained by calibrating the two radiation temperature measuring channels respectively, the reference deviation between channels, and the standard deviation of the channel difference under stable operating conditions.

[0010] In a preferred embodiment of the present invention, the thickness-oriented unsteady-state heat transfer model discretizes the strip thickness direction into 3-7 temperature nodes, including plastic deformation heat, spray water cooling heat transfer, radiation heat transfer, air convection heat transfer and roll contact heat transfer; each temperature node is remapped according to volume conservation after the material section is pressed down by the stand, the core temperature is obtained from the temperature node located in the middle of the thickness, and the thickness-oriented temperature gradient is obtained from the difference between the core temperature and the average values ​​of the upper and lower surface temperatures.

[0011] In a preferred embodiment of the present invention, the residual cause decomposition module standardizes the spectral consistency residual, energy balance residual, and valve flow residual to form a residual feature vector. The residual sensitivity matrix is ​​obtained by applying finite perturbations to measurement contamination, heat transfer drift, and actuator anomalies, or by using calibration tests. The non-negative contribution values ​​of each anomaly source are solved using optimization with sparse constraints, and then the non-negative contribution values ​​are normalized to obtain the corresponding anomaly contribution weights. When the residual sensitivity matrix does not meet the preset identifiable condition, or the residual feature vector exceeds the limit and the weights of each anomaly contribution are all lower than the dominant threshold, the cause is output as unclear.

[0012] In a preferred embodiment of the present invention, the uncertainty propagation module uses an extended Kalman filter, an unscented Kalman filter, or a Bayesian state estimator to maintain the thermal state mean and covariance; when the abnormal contribution weight of measurement contamination increases, the measurement covariance is increased; when the abnormal contribution weight of heat transfer drift increases, the thermal state process covariance is increased; when the abnormal contribution weight of actuator malfunction increases, the actuator covariance is increased; and the actuator covariance is propagated to the thermal state prediction covariance via the control input sensitivity matrix.

[0013] In a preferred embodiment of the present invention, the allowable range of final rolling temperature is set as follows: The total risk limit for exceeding the boundary is The risk amounts allocated to the lower temperature limit and the upper temperature limit are respectively and ,and When the final rolling temperature prediction error approximately follows a normal distribution, the final rolling temperature probability constraint includes... as well as ,in, This is the predicted average final rolling temperature. The standard deviation of the final rolling temperature prediction. These are the standard normal quantiles.

[0014] In a preferred embodiment of the present invention, when the abnormal contribution weight of measured pollution exceeds the pollution threshold, the data residual model and effective heat transfer coefficient are prohibited from being updated with the outlet temperature of the corresponding material logic segment; when the abnormal contribution weight of heat transfer drift exceeds the heat transfer threshold, the effective heat transfer coefficient is updated within the preset physical boundary and single-cycle update range; when the abnormal contribution weight of actuator malfunction exceeds the actuator threshold, the upper limit of the available flow rate of the corresponding spray zone is reset according to the actual spray flow rate, or the corresponding spray zone is removed from the decision variables of rolling optimization.

[0015] In a preferred embodiment of the present invention, the material segment synchronization module divides the material into logical segments according to a set length of 0.5-2.0m, and increases the set length when the material running speed increases, and decreases the set length when the inlet temperature change rate or outlet temperature change rate exceeds the change rate threshold; when the execution and safety interlock module has no feasible solution to the rolling optimization problem, it first converts the final rolling temperature target from a hard constraint to a soft constraint with a penalty term, and when there is still no feasible solution, it maintains the current rolling speed and restores the verified inter-stand spray safety setting.

[0016] Compared with the prior art, the present invention has the following advantages: This invention utilizes the combined relationship of spectral consistency residuals, energy balance residuals, and valve flow residuals to distinguish between measurement contamination, heat transfer drift, and actuator anomalies. The distinction results directly determine whether temperature samples participate in learning, whether the effective heat transfer coefficient is updated, and whether abnormal spray zones continue to participate in optimization, thereby reducing the risk of measurement spurious deviations triggering real regulation and abnormal samples contaminating the model.

[0017] This invention maps different causes of anomalies to measurement covariance, thermal process covariance, and actuator covariance, respectively, and propagates each covariance into a confidence interval for the final rolling temperature. Therefore, under conditions where the mean predicted final rolling temperature is the same but the prediction confidence level differs, the optimizer can employ different temperature safety margins.

[0018] This invention implements rolling optimization based on material segment synchronization and allocates the total out-of-bounds risk between the upper and lower limits of the final rolling temperature, so that the probability constraint is consistent with the confidence level of the two-sided interval required by reality; at the same time, it retains the equipment safety hard constraint and the graded backoff mechanism, which is conducive to maintaining feasible safety control of the system when the temperature measurement is disturbed, the spraying capacity is reduced, or the optimization has no feasible solution.

[0019] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0020] In the attached diagram: Figure 1 This is a schematic diagram of the overall structure and closed-loop data flow of the final rolling temperature optimization system of the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the multi-residual causal decomposition and causal conditional model update relationship of the present invention.

[0022] Figure 3 This is a schematic diagram of the execution flow of a single model prediction control cycle in this invention.

[0023] Figure reference numerals: 100, Process data acquisition module; 200, Material segment synchronization module; 300, Thermal state prediction module; 400, Residual cause decomposition module; 500, Uncertainty propagation module; 600, Confidence constraint optimization module; 700, Execution and safety interlock module. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention.

[0025] Example 1: like Figure 1 As shown, this embodiment is applied to a seven-stand hot strip finishing mill. The process data acquisition module 100 reads inlet and outlet radiation temperature measurement signals, stand speed and position encoder signals, reduction, rolling force or motor power, strip thickness and width, steel grade identification, spray valve opening, measured spray flow rate, and branch pressure difference from the primary basic automation system via an industrial communication interface, and also reads the water temperature and ambient temperature. All data are accompanied by timestamps and quality marks.

[0026] At least a dual-band pyrometer should be installed at the finishing mill exit. The two radiation channels of the dual-band pyrometer have different center wavelengths and observe the same strip area. If two independent pyrometers are used on site, their fields of view should be overlapped after calibration, and their sampling times should be aligned using a unified clock. The process data acquisition module 100 performs range verification, mutation rate verification, and data quality marking for each channel, but does not directly replace subsequent causal decomposition based on whether a single channel exceeds its limit.

[0027] The material segment synchronization module 200 assigns a unique number i to each material logic segment. The thermal state prediction module 300, residual cause decomposition module 400, uncertainty propagation module 500, and confidence constraint optimization module 600 all index data using this number. The execution and safety interlock module 700 writes the control results and their actual execution feedback back to the same number so that closed-loop verification can be performed after the corresponding material logic segment reaches the finishing mill exit.

[0028] Example 2: Material logic segment synchronization and delay compensation divide the strip entering the finishing rolling zone into material logic segments of a set length of 0.5-2.0m. The nominal segment length can be set to 1.0m; when the material running speed increases, the segment length is appropriately increased to limit the amount of calculation; when the inlet temperature change rate or the outlet temperature change rate exceeds the change rate threshold, the segment length is reduced to improve the control resolution of the temperature change region.

[0029] In the k-th control cycle, the position of material logic segment i relative to the reference temperature measurement point is updated according to the following formula:

[0030] In the formula, This refers to the position of material logic segment i. The material travel speed is obtained by combining the frame encoder speed with the forward slip correction. For the first... Spray area between racks, material logic segment The estimated arrival time and the control command issuance time are as follows:

[0031]

[0032] In the formula, Let n be the position of the nth spray zone relative to the reference temperature measurement point. , Material logic segment Reaching the The expected average speed before each spray zone , The response delay obtained after identification of the nth spray valve To calculate the delay, Communication delay. Control commands in i , The timing is set so that the spraying action corresponds to the time when material logic segment i arrives at the spraying area.

[0033] The response delay of the spray valve can be identified online through the cross-correlation between the valve step command and the measured flow response, or through a first-order inertial pure delay model, and limited to the allowable range of the equipment. When the delay identification result exceeds the limit or the flow feedback is unavailable, a verified conservative delay value is adopted and the actuator uncertainty is increased.

[0034] Example 3: The physical-data fusion thermal state prediction module 300 establishes the thickness-oriented thermal state for each material logical segment. The strip thickness direction is discretized into N temperature nodes, where N ranges from 3 to 7. The state vector can be represented as:

[0035] In the formula, and Corresponding to the upper surface temperature and lower surface temperature The core temperature Tc is obtained by interpolation from the mid-thickness node or adjacent nodes; heff is the effective heat transfer coefficient under the current spray and surface conditions. is the equivalent radiative emissivity or dual-band emissivity ratio correction; bT is the temperature measurement bias state. The thickness-oriented temperature gradient characterization quantity G is not treated as a repeating state independent of each temperature node, but is determined by... Calculated.

[0036] Thick-section heat transfer employs a one-dimensional unsteady-state heat conduction equation, and uses the heat converted from plastic deformation work as the volume heat source:

[0037] In the formula, ρ is the density. Here, k is the specific heat capacity, k is the thermal conductivity, and z is the thickness direction coordinate. It serves as a heat source for plastic deformation. The boundary heat flow of the water spray surface includes at least spray water cooling, radiation, and air convection, and in the roll contact area, it also includes roll contact heat transfer. The boundary of the water spray surface can be represented as:

[0038] In the formula, and These are the measured flow rate of the spray system, the branch pressure difference, and the water temperature, respectively. These are surface state parameters such as oxide scale, roughness, and boiling heat transfer zone. Here, is the Stefan-Boltzmann constant, Ta is the ambient temperature, and hconv is the air convection heat transfer coefficient. All temperatures should be converted to absolute temperatures when calculating the radiation terms.

[0039] After each press down of the frame, the thickness node position is remapped according to the thickness and volume conservation relationship after pressing, and the node temperature is interpolated using energy conservation. The heat exchange from roll contact, deformation, and spraying is integrated according to the residence time of the material logic segment in the corresponding region, thereby avoiding the heat calculation deviation caused by using only instantaneous speed.

[0040] To compensate for the spray distribution, localized oxide scale, and equipment thermal state that are difficult to accurately describe by the compensation mechanism model, a data residual model is set up. This model only predicts the residual error of the prior temperature of the mechanism and does not directly replace the thickness-direction heat transfer equation.

[0041] In the formula, u includes the spray flow rate and speed correction between each rack, and d includes the steel type, thickness, width, inlet temperature, reduction regime, water temperature, and ambient temperature. Here, z represents the mechanistic parameters, and z is the input to the residual model. The data residual model can be a set of five light quantum models, with each sub-model outputting a temperature correction. The dispersion of the sub-model outputs constitutes the uncertainty of the data model. .

[0042] Example 4: Multi-residual structure and anomaly genetic decomposition, such as Figure 2 As shown, the residual causation decomposition module 400 constructs three types of residuals with different sensitivities to different causes of anomalies. The first type is the spectral consistency residual. This is obtained after the two radiation channels are calibrated. and The spectral consistency residual is:

[0043] In the formula, This represents the reference difference between the two radiation channels under a clean field of view. This represents the standard deviation of the channel difference under stable operating conditions. When water film, smoke, oxide scale abrupt changes, or lens contamination cause different attenuations in different bands, rλ varies accordingly. If contamination causes approximately proportional attenuation in two bands, the system also considers energy balance residuals, single-channel quality indicators, and temporal continuity for judgment.

[0044] The second category is energy balance residuals. For the same material logic segment, between two adjacent thermal state verification locations, the enthalpy change of the material segment is calculated based on the state estimation. And compared with the heat transfer of rolling deformation (Edef), radiative heat transfer (Qrad), roll contact heat transfer (Qroll), water cooling heat transfer (Qwater), and air convection heat transfer (Qconv):

[0045] The third category is valve flow residual, which is used to reflect valve dead zone, nozzle blockage, changes in water supply pressure differential, or abnormal flow feedback.

[0046] In the formula, This is the actual measured flow rate of the spray system. To determine based on valve opening The calibrated flow coefficient, For the pressure difference of the spray branch, This represents the standard deviation of the flow model error under stable operating conditions. It includes conversion factors determined by fluid density and pipeline structure.

[0047] To ensure that the non-negative outlier contribution values ​​are compatible with the residual direction, this embodiment constructs the residual feature vector from the magnitudes of the standardized residuals. When it is necessary to retain the direction of deviation, sensitivity baselines can be set separately for the positive and negative directions. The sensitivity matrix S of the residual eigenvector to the measured contamination cm, heat transfer drift ch, and actuator anomaly ca is obtained through historical calibration experiments or finite perturbation of the mechanistic model.

[0048] Within each control cycle, the anomaly contribution vector a is solved using an optimization problem with non-negativity constraints and sparse penalties:

[0049] In the formula, W is the residual weight matrix, λ is the sparsity penalty coefficient, and a = [am, ah, aa]^T. The normalized contribution weight of each anomaly source is calculated according to the following formula:

[0050] In the formula, To prevent positive numbers with a denominator of zero, pm, ph, and pa represent the relative contribution weights of measurement contamination, heat transfer drift, and actuator malfunction, respectively. These are used for confidence assessment in the control logic and are not required to be interpreted as statistically calibrated posterior probabilities. If calibrated using fault samples, these contribution weights can be further mapped to posterior probabilities.

[0051] Water mist or lens contamination typically increases the spectral consistency residual significantly while the valve flow residual remains near normal; nozzle blockage usually increases both the valve flow residual and the energy balance residual simultaneously; when the flow rate is normal but the boiling heat transfer range changes, the energy balance residual increases while the valve flow residual remains normal. To ensure identifiability, the column correlation or condition number of the sensitivity matrix is ​​checked after calibration. When the sensitivity matrix does not meet the preset identifiability conditions, or when the residual eigenvector exceeds the limit but the maximum anomaly contribution weight is lower than the dominant threshold, the current state is marked as having an unclear cause.

[0052] Example 5: Uncertainty propagation based on causal conditions: The uncertainty propagation module 500 uses an extended Kalman filter, an unscented Kalman filter, or an equivalent Bayesian state estimator to maintain the thermal state mean and covariance. In this embodiment, the measurement covariance R, the thermal state process covariance Q, and the actuator covariance are adjusted according to the cause of the anomaly. :

[0053] In the formula, and This is the amplification factor determined through calibration. Measurement contamination primarily reduces the reliability of observations, heat transfer drift primarily reduces the reliability of the thermal state transition model, and actuator anomalies primarily reduce the reliability of translating control commands into actual flow rates. The three types of covariance are adjusted separately to avoid different anomalies sharing the same error amplification rule.

[0054] With linearized state transition matrix A k Control input sensitivity matrix B k and observation matrix H k The local model representing the current operating point has the following prediction covariance, filter gain, and correction covariance:

[0055]

[0056]

[0057] By mapping the predicted state of the material logic segment to the finishing rolling temperature at the finish mill exit using the output mapping CF, we obtain:

[0058] In the formula, and These represent the predicted mean and standard deviation of the final rolling temperature, respectively. For errors exhibiting significant non-Gaussian characteristics, a scenario set can be constructed based on the data residual model set, spray flow error, and valve response delay disturbance, and the scenario quantile can be used to approximate the normal standard deviation.

[0059] Example 6: Confidence-constrained rolling optimization of final rolling temperature; the confidence-constrained optimization module 600 targets several consecutive material logic segments within the prediction time domain, using the adjustable flow rate of the spray zone. and the allowable rolling speed correction amount These are the decision variables. The objective function weights the final rolling temperature deviation, prediction variance, thickness temperature gradient, spray flow rate variation, and velocity variation:

[0060] In the formula, The target final rolling temperature is defined by wT, wσ, wG, wq, and wv, which are non-negative weights. The prediction variance term enables the optimizer to select the option with lower uncertainty among options with similar point prediction temperatures; the thickness temperature gradient term is used to suppress the surface and core temperature difference caused by excessive instantaneous water cooling.

[0061] Let the allowable range of final rolling temperature be... The total permissible risk of crossing the boundary is The total risk of exceeding the boundary is allocated as follows: [Assignment based on temperature below...] Risk and temperatures higher Risk ,satisfy When the final rolling temperature prediction error approximately follows a normal distribution, the two-sided probability constraint is transformed into:

[0062]

[0063] In the formula, and This represents the standard normal quantile. When the probability of the final rolling temperature falling within the allowable range is not less than 95%, and the risk is equally distributed between the upper and lower limits, The corresponding quantile is approximately 1.96; when only one direction is constrained and the one-sided confidence level is 95%, the quantile is 1.645.

[0064] The optimization problem also satisfies the minimum and maximum available flow rates for each spray zone, single-cycle valve opening or flow rate change, maximum capacity of the main water supply pipe, rolling speed and acceleration range, upper limits of motor power and stand load, allowable range of looper quantity, and speed correction amplitude for limiting thickness disturbances. If pa indicates that the capacity of a certain spray zone is insufficient, the maximum available flow rate of that zone is tightened according to the actual measured flow rate of the spray. When the capacity is lower than the minimum requirement for participation in control, the control variable of that spray zone is fixed to a safe value or removed from the decision variables.

[0065] like Figure 3 As shown, after obtaining the control sequence, the system only sends the first set of spray flow and speed correction values ​​corresponding to the current material logic segment to the basic automation system. In the next control cycle, the material segment status and actuator status are reread and the solution is solved continuously. Before issuing control commands, the execution and safety interlock module 700 verifies the basic automation permission, rack load, looper quantity, speed change rate, and water supply capacity.

[0066] When the optimization problem has no feasible solution, first convert the temperature target into a soft constraint with a penalty term, but do not relax the hard constraints on equipment safety; if it still has no feasible solution, maintain the current rolling speed, restore the verified inter-stand spray safety settings, and stop online learning. In cases where the cause is unclear, simultaneously increase R, Q, and And limit the single-cycle control variation until the residual returns to a reliable range.

[0067] Example 7: When the material logic segment reaches the finishing mill exit, the material segment synchronization module 200 uses number i to align the measured exit temperature with the previous predicted state, anomaly contribution weight, spray effect, and speed correction. Only when pm, ph, and pa are all below the corresponding confidence thresholds, the two radiation temperature measurement channels are consistent, the data quality flag is valid, and the basic automation system has no fault flags, will the corresponding sample enter the data residual model update queue.

[0068] When PM exceeds the pollution threshold, the learning of the corresponding temperature sample is frozen, and the effective heat transfer coefficient is not updated using that temperature sample. When PH exceeds the heat transfer threshold while PM and PA are low, the heff and the heat transfer correction coefficients corresponding to the steel grade and thickness are updated using a recursive least squares method with upper and lower bounds or a state estimator. The update range per cycle is limited to 5%-10% of the current value. When PA exceeds the actuator threshold, the available capacity model of the valve or spray area is updated, but the anomaly is not written into the heat transfer parameters of the steel grade.

[0069] The data residual model is updated using a sliding window or an incremental window. Before the update, material segment number consistency verification, timestamp verification, input range verification, and abnormal contribution weight verification are performed on the candidate samples. After the update, the new model is checked for error on the retained reliable sample set. Only when the verification error is not higher than the current model and the output change rate is within the limit is it put into online use; otherwise, the original model is retained.

[0070] Example 8: Taking a seven-stand hot strip mill finishing unit as an example, the strip width is 1250mm, the target thickness is 3.0mm, the finishing inlet temperature is 1040-1060℃, the target finishing temperature is 860℃, and the allowable range is 850-870℃. The control cycle is 0.2s, and the nominal length of the material logic segment is 1.0m. The spray zones between stands F4-F5, F5-F6, and F6-F7 are adjustable spray zones, with an allowable flow rate of 0-120m³ / h per zone. 3 / h, single-cycle flow rate variation not exceeding 10m³ 3 / h; The permissible speed range for the last frame is 7-12 m / s, and the speed correction acceleration limit is ±0.15 m / s. 2 .

[0071] Under stable operating conditions, the reliable material logic segment is used to determine and Finite perturbations are applied to the temperature measurement optical path attenuation, effective heat transfer coefficient, and available valve flow rate to form a sensitivity matrix S. The sparsity penalty coefficient λ for online solution can be taken as 0.05. If any anomaly contribution weight is greater than 0.65, the corresponding anomaly source is determined to be dominant; if the residual eigenvector significantly exceeds the limit but the maximum anomaly contribution weight is less than 0.55, it is determined to be of unclear cause. The above thresholds should be adjusted according to the unit calibration data.

[0072] The following three scenarios illustrate how the results of residual cause decomposition are incorporated into model updates and control constraints. The values ​​in the table are examples of the calculation process and are not considered as fixed parameters for all units.

[0073]

[0074] Given a permissible temperature range of 850-870℃, a total out-of-bounds risk α of 0.05, and equal risk allocation between the upper and lower limits, if the prediction standard deviation of a certain material logic segment... For a temperature of 3℃, its predicted mean must satisfy approximately 855.88-864.12℃. When anomalies contribute... As the range increases, the feasible mean range narrows accordingly. The optimizer will prioritize redistributing the cooling effect using the remaining available spray zones and the allowed speed correction amount. When this cannot be met, it will enter a soft constraint and safety rollback.

[0075] Disturbances in the temperature measurement field of view will not directly trigger real adjustments opposite to spurious temperature differences; decreased nozzle or valve capacity will not be learned as a material heat transfer property; and real heat transfer drift will not be ignored as ordinary measurement noise. The anomaly identification results simultaneously affect model learning switches, covariance updates, and rolling optimization constraints, forming a closed-loop control chain.

[0076] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A model-based final rolling temperature optimization system, deployed on the finishing mill process control server and communicating with on-site detection, spraying, and basic automation equipment, characterized in that, include: The data acquisition unit is used to collect multi-source process data on strip temperature measurement, rolling operation, and spraying conditions. The material segment synchronization unit is used to divide the strip into logical segments, update the segment positions in real time, and compensate for the control delay of the entire link. The thermal state prediction unit, relying on the thickness heat transfer mechanism model and the data residual model, predicts the a priori values ​​of the hot state and final rolling temperature of the strip. The residual cause analysis unit constructs multiple types of physical residuals and decouples and identifies three types of anomalies and their contribution weights: temperature measurement, heat transfer and spray actuators. The uncertainty propagation unit corrects the multi-dimensional covariance based on the weights of various anomalies, and derives the mean and variance of the final rolling temperature prediction. The confidence-constrained optimization unit uses the spray flow rate and rolling speed correction as optimization variables to solve the rolling control optimization problem with temperature probability constraints. The safety interlocking unit issues the first-step control quantity for optimization and switches the safety control logic when the operating condition is abnormal or the optimization is unsuccessful.

2. The model-predictive final rolling temperature optimization system according to claim 1, characterized in that, The data acquisition unit is configured with at least two spectrally differentiated radiation temperature measurement channels to construct spectral residuals for characterizing temperature measurement consistency; the multiple physical residuals also include energy balance residuals characterizing the matching degree of rolling heat balance and valve flow residuals characterizing the matching degree of spray command and actual flow.

3. The model-predictive final rolling temperature optimization system according to claim 1, characterized in that, The residual cause analysis unit uses a pre-calibrated residual sensitivity matrix and non-negative optimization with sparse constraints to solve for the contribution value of each anomaly source. After normalization, the anomaly contribution weight is obtained. If the residual exceeds the limit but there is no dominant anomaly source, the fault cause is output as unknown.

4. The model-predictive final rolling temperature optimization system according to claim 1, characterized in that, The thickness-oriented heat transfer model is discretized into 3-7 temperature nodes along the strip thickness, uniformly taking into account multiple heat transfer boundaries such as plastic deformation heat, water cooling, radiation, convection, and roll contact. After the stand is pressed down, the temperature nodes are remapped according to the volume conservation principle to solve for the core temperature and thickness-oriented temperature gradient of the strip.

5. The model-predictive final rolling temperature optimization system according to claim 1, characterized in that, The uncertainty propagation unit uses Kalman filtering or Bayesian estimation to maintain the thermal state covariance; increasing the temperature measurement contamination weight amplifies the measurement covariance, increasing the heat transfer drift weight amplifies the thermal state process covariance, increasing the spray actuator abnormality weight amplifies the actuator covariance, and coupling the actuator uncertainty to the thermal state prediction output.

6. The model-predictive final rolling temperature optimization system according to claim 1, characterized in that, The confidence constraint optimization unit presets the upper and lower limit ranges of the final rolling temperature and the total temperature out-of-bounds risk threshold, and distributes the total risk to the upper and lower temperature boundaries; it constructs a two-sided temperature probability constraint based on the normal distribution quantile, and simultaneously superimposes hardware boundary constraints such as spray flow rate, rolling speed, and equipment load to construct the optimization problem.

7. The model-predictive final rolling temperature optimization system according to claim 1, characterized in that, When rolling optimization fails to find a feasible solution, the hard temperature constraint is first converted into a soft constraint with a penalty term; if there is still no solution, the current rolling speed is maintained, the preset spray safety setting is switched, and online model learning is paused.

8. The model-predictive final rolling temperature optimization system according to claim 1, characterized in that, The system divides the strip steel material into logical segments with a variable length of 0.5-2.0m; the segment length is increased when the rolling speed is increased and decreased when the strip steel temperature fluctuates drastically; the synchronization unit combines the time when the segment arrives at the spray zone with multiple delays in calculation, communication or valve response to accurately match the timing of the control command issuance.

9. The model-predictive final rolling temperature optimization system according to claim 1, characterized in that, The system updates the control model differently based on the weight of different anomalies: when temperature measurement pollution is dominant, the temperature measurement sample learning is frozen; when heat transfer drift is dominant, the heat transfer coefficient is slightly updated within the physical limits.

10. The model-predictive final rolling temperature optimization system according to claim 9, characterized in that, When the spray actuator malfunction is dominant, only the upper limit of the spray flow rate is corrected, without changing the parameters of the strip heat transfer model; only when all malfunction weights are below the confidence threshold is the residual model updated with the measured outlet temperature used.

Images