Physically Constrained Predictive Control Methods, Systems, and Programs for Continuous Casting Quality

By establishing a dual-source continuous casting time-series corpus and a sparse hybrid expert Transformer model, the problems of incomplete data and inaccurate predictions caused by process changes in continuous casting production were solved, achieving high consistency and reliability in quality prediction and control.

CN123116503APending Publication Date: 2026-11-17HUA DATA TECH (SHANGHAI) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611530260.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the existing continuous casting production process, the fixed time window data acquisition method leads to inaccurate defect detection results when the process changes or the data is incomplete. It is difficult to adapt to changes in operating conditions, and the model prediction results lack physical reliability, resulting in control recommendations that are not suitable for the current state.

Method used

By acquiring measured time series data of continuous casting and multi-physics simulation data, a dual-source continuous casting time series corpus is established. Physical consistency pre-training is performed using a sparse hybrid expert Transformer. Modulation latent states are embedded in the equipment, steel grade and cross-section to establish a basic time series model. Control quantities are determined by projection of the physical feasible set and confidence level to form a closed loop of prediction verification and control execution.

Benefits of technology

It improves the physical consistency, safety, and operational reliability of continuous casting quality prediction and control results, reduces the impact of process state misalignment and data loss, expands the sample coverage of extreme working conditions, and improves the consistency of predictions across equipment, steel grades, and cross sections.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116503A_ABST
    Figure CN123116503A_ABST
Patent Text Reader

Abstract

This invention provides a method, system, and program product for predictive control of continuous casting quality based on physical constraints. The method includes: acquiring measured time-series data of continuous casting, quality inspection results, and multiphysics simulation data; establishing a dual-source continuous casting time-series corpus based on the correlation between casting speed integral and defect location and billet slice process state; setting position anchor points along the metallurgical length, aggregating process data and multi-source features to form a sequence of material point trajectory aligned terms; performing physical consistency pre-training on a sparse mixing expert Transformer by combining solidification heat transfer, solute conservation, flow continuity, phase diagram consistency, and global conservation constraints to obtain a basic time-series model; adapting the model using target continuous casting scenario data, projecting the predicted physical state field onto a physically feasible set, and determining the confidence level; generating target control increments based on the projection results, and executing safety verification issuance, digital twin review issuance, or expert rule rollback according to the confidence level.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology in iron and steel metallurgy, and in particular to a method, system, and program product for predictive control of continuous casting quality based on physical constraints. Background Technology

[0002] In continuous casting production, basic automation systems and process automation systems typically collect data on molten steel temperature, casting speed, crystallizer cooling, secondary cooling water supply, electromagnetic stirring, roll gap, and equipment vibration, forming process time sequence data according to the collection time. Existing quality prediction methods mostly extract process data within a fixed time window, using the results of segregation detection, flaw detection, or surface inspection from historical heats to train classification or regression models to predict the risk of defects such as center segregation and internal cracks, as well as the solidification endpoint. When process adjustments are needed, secondary cooling water supply, casting speed, electromagnetic stirring, or light reduction parameters are then provided based on the prediction results or preset expert rules.

[0003] When the continuous casting process experiences sudden changes in casting speed, steel grade change, cross-section switching, secondary cooling circuit anomaly, nozzle blockage, missing sensor data, or equipment status drift, the data within a fixed time window can easily become misaligned with the actual process experienced by the cast billet slice. This makes the defect detection results unable to accurately correspond to the temperature, cooling, and stress states at the formation stage. For newly built casting machines, steel grades with few samples, and extreme operating conditions, insufficient historical data coverage can also cause the model to extrapolate based on the statistical correlation in the training data, resulting in contradictory predictions regarding temperature, solid fraction, solute concentration, or solidification endpoint. Such predictions may be numerically within a common range, but they may not conform to phase transformation laws, energy balance, solute balance, or solidification evolution sequence. Conventional probability values ​​are also insufficient to accurately reflect this physical distortion.

[0004] The aforementioned failures can further propagate to the process parameter calculation stage, resulting in a lack of verifiable consistency between defect risk assessment, control quantity generation, and equipment execution. Directly using model outputs may produce control suggestions unsuitable for the current steel grade, cross-section, or equipment status, while relying solely on rule rollback is ill-suited to continuously changing operating conditions. Therefore, maintaining the physical reliability of continuous casting quality prediction results under conditions of changing operating conditions and incomplete data, and generating safe and executable control quantities accordingly, is a technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method, system, and program product for predictive control of continuous casting quality based on physical constraints, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The method for predictive control of continuous casting quality based on physical constraints includes the following steps:

[0008] S1. Obtain the measured time series data, quality inspection results and multi-physics simulation data of continuous casting, associate the defect location with the process state of the corresponding billet slice based on the casting speed integral, unify the timestamp and metallurgical length coordinates of the two types of data, and establish a dual-source continuous casting time series corpus.

[0009] S2. Set position anchor points along the metallurgical length, collect process data, physical derivation features, vibration features, image features and validity masks of the aggregated billet slices when they pass through each position anchor point, generate position words and insert event words to form a material point trajectory aligned word sequence;

[0010] S3, based on lexical sequences and solidification heat transfer, solute conservation, flow continuity, and phase... Figure 1 To ensure consistency and global conservation loss, the sparse hybrid expert Transformer is pre-trained for physical consistency. By embedding the modulation hidden state using equipment, steel grade, and cross-section, a time-series basic model with defect risk prediction, solidification endpoint prediction, and process parameter inversion functions is established.

[0011] S4. Adapt the time series basic model using the massless time series data of the target continuous casting scenario to obtain the defect risk, solidification endpoint and physical state field. Project the physical state field to the physical feasible set defined by the phase diagram boundary, energy conservation, solute conservation and monotonic relationship of solid fraction, and determine the confidence level based on the prediction interval, projection residual, projection convergence state and input validity.

[0012] S5. Generate candidate control quantities based on the projected physical state field and the current control quantity. Determine the target control increment through time series basic model deduction and physical feasibility screening. After performing safety verification according to the confidence level, issue the control quantity, issue it after digital twin review, or roll back the control according to expert rules. Then, feed the execution response and quality inspection results back to the dual-source continuous casting time series corpus to incrementally update the time series basic model.

[0013] The continuous casting quality prediction and control system based on physical constraints includes: a dual-source corpus construction module, a trajectory lexical generation module, a basic model training module, a scene prediction projection module, and a control decision update module connected in sequence.

[0014] The beneficial effects of this invention are as follows:

[0015] This invention forms a closed loop of prediction verification and control execution by integrating physical consistency pre-training, physical feasibility set projection, confidence grading, and hierarchical control. This suppresses the transmission of prediction distortion to the control end when operating conditions change or data is incomplete, thereby improving the physical consistency, safety, and operational reliability of continuous casting quality prediction and control results. By correlating defect locations with billet slicing process states based on casting speed integrals and unifying the spatiotemporal coordinates of measured data and multiphysics simulation data, the corresponding error between quality inspection results and defect formation processes is reduced, the sample coverage of extreme operating conditions is expanded, and the data consistency and sample reliability of the dual-source continuous casting time-series corpus are improved.

[0016] This invention aggregates process data, physical derived features, vibration features, and image features through location anchors, and introduces event terms and validity masks to establish a material point trajectory aligned term sequence. This reduces process state misalignment caused by fixed time windows and improves the spatial alignment accuracy of multi-source features and the identifiability of missing data states. By constraining a sparse hybrid expert Transformer with physical consistency loss and embedding modulated latent states using equipment, steel grade, and cross-section, the time-series basic model learns the evolution law of continuous casting state and scene differences, reducing physical errors caused by statistical extrapolation under few sample conditions and improving the consistency of predictions across equipment, steel grades, and cross-sections.

[0017] This invention adapts the target continuous casting scenario data to a time-series basic model and projects the physical state field onto a physically feasible set. This corrects outputs that violate phase diagram limits, energy conservation, solute conservation, and the monotonic relationship of the solid fraction. The confidence level is then determined by combining the prediction interval and the projected state, improving the physical legitimacy of the prediction results and the reliability of the confidence judgment. The target control increment is determined through forward extrapolation and physical feasibility screening of candidate control quantities. Safety verification, digital twin review, or expert rule backoff are implemented according to the confidence level to reduce the risk of unreliable predictions directly driving the equipment. Incremental updates are performed based on the execution response and quality inspection results, improving control stability and the model's continuous adaptability. Attached Figure Description

[0018] Figure 1 This is a flowchart of the continuous casting quality prediction and control method based on physical constraints according to the present invention.

[0019] Figure 2 This is a framework diagram of the continuous casting quality prediction and control system based on physical constraints according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: As Figure 1 As shown, this embodiment provides a method for predictive control of continuous casting quality based on physical constraints, including the following steps:

[0022] S1. Obtain the measured time series data, quality inspection results and multi-physics simulation data of continuous casting, associate the defect location with the process state of the corresponding billet slice based on the casting speed integral, unify the timestamp and metallurgical length coordinates of the two types of data, and establish a dual-source continuous casting time series corpus.

[0023] S2. Set position anchor points along the metallurgical length, collect process data, physical derivation features, vibration features, image features and validity masks of the aggregated billet slices when they pass through each position anchor point, generate position words and insert event words to form a material point trajectory aligned word sequence;

[0024] S3, based on lexical sequences and solidification heat transfer, solute conservation, flow continuity, and phase... Figure 1 To ensure consistency and global conservation loss, the sparse hybrid expert Transformer is pre-trained for physical consistency. By embedding the modulation hidden state using equipment, steel grade, and cross-section, a time-series basic model with defect risk prediction, solidification endpoint prediction, and process parameter inversion functions is established.

[0025] S4. Adapt the time series basic model using the massless time series data of the target continuous casting scenario to obtain the defect risk, solidification endpoint and physical state field. Project the physical state field to the physical feasible set defined by the phase diagram boundary, energy conservation, solute conservation and monotonic relationship of solid fraction, and determine the confidence level based on the prediction interval, projection residual, projection convergence state and input validity.

[0026] S5. Generate candidate control quantities based on the projected physical state field and the current control quantity. Determine the target control increment through time series basic model deduction and physical feasibility screening. After performing safety verification according to the confidence level, issue the control quantity, issue it after digital twin review, or roll back the control according to expert rules. Then, feed the execution response and quality inspection results back to the dual-source continuous casting time series corpus to incrementally update the time series basic model.

[0027] Specifically, it includes the following:

[0028] S1. Production data for two consecutive years was collected from the company's three in-service continuous casting machines, covering more than 60 steel grades and approximately 120,000 heats. The collected data included: crystallizer liquid level height, crystallizer vibration frequency and amplitude, crystallizer thermocouple temperature array, crystallizer inlet and outlet water temperature and flow rate, secondary cooling circuit water flow rate and pressure, casting speed, roll gap, electromagnetic stirring current and frequency, drive roll vibration signal; temperature field, billet shell thickness and solidification endpoint position output by the solidification heat transfer model; steel grade, actual smelting composition, casting schedule, tundish temperature, ladle change events, steel grade change events and cross-section switching events; as well as carbon segregation values ​​of drill samples, low-magnification inspection results, ultrasonic flaw detection results and surface defect detection results.

[0029] Using "casting batch - furnace batch - cast flow - billet slice" as the four-level data primary key, data from different sources are associated with the corresponding billet slices. Data from the L1 basic automation system and the L2 process automation system are accessed through the OPC UA interface, while steel grade, furnace batch, and quality data from the manufacturing execution system are accessed through the data interface. A unified clock source is used to calibrate the timestamps of each acquisition device.

[0030] Data that exceeds the sensor's range or whose rate of change between adjacent sampling times exceeds the maximum rate of change specified in the equipment manual is marked as invalid data. For single-point anomalies or short-term missing data with a duration of less than two original sampling periods, linear interpolation is performed using adjacent valid sampling points. For continuously drifting data, correction is performed based on equipment calibration records and data from redundant sensors in the same loop.

[0031] If the normalized deviation between the corrected data and the redundant sensor data still exceeds the 95th percentile of the deviation of the corresponding channel training samples, the corrected result will no longer replace the original data. Instead, the missing marker will be retained, and a corresponding channel validity mask will be set. The channel validity mask and the corresponding process data will be synchronously input into the subsequent material point trajectory alignment word segmenter.

[0032] The location of defects detected at the billet exit is traced back based on the casting speed. For the i-th billet slice, the moment it enters the meniscus of the crystallizer is recorded. Its metallurgical length position at time t is determined by the following formula:

[0033] ;

[0034] in, This represents the metallurgical length of the i-th cast billet slice from the meniscus at time t, calculated cumulatively by the casting speed over the corresponding time interval. Indicates time The actual casting speed is collected by the casting speed channel of the L1 basic automation system, and the unit is m / s. Therefore, the physical coordinates of the defects detected at the exit are inversely correlated with the process states experienced by the corresponding billet slice in the crystallizer, secondary cooling zone, straightening zone and air cooling zone.

[0035] The casting speed integral is calculated in segments using the same sampling time as the L1 basic automation system. When the continuous missing time of casting speed data does not exceed 10s, the effective casting speeds before and after the missing interval are used for linear interpolation; when the continuous missing time exceeds 10s, the position backtracking of the corresponding billet slice is stopped and it is marked as an incomplete trajectory sample. It is not used as a defect location supervision sample, but the remaining effective process data is retained for unsupervised pre-training.

[0036] In addition to the measured data, a multiphysics simulation platform for continuous casting was established to couple and calculate solidification heat transfer, solute transport, molten steel flow in the crystallizer, and billet shell stress. The solidification heat transfer and solute transport processes were discretized using the finite volume method, the molten steel flow in the crystallizer was calculated using the pressure-velocity coupling method, and the thermal stress and strain of the billet shell were calculated using the finite element method.

[0037] The simulation area is meshed along the billet pulling direction, with local mesh refinement performed near the meniscus, crystallizer outlet, secondary cooling zone boundary, straightening zone, and solidification endpoint. After two consecutive mesh refinements, the current mesh is designated as the simulation mesh when the change in the solidification endpoint position does not exceed 1% and the change in the maximum billet shell thickness does not exceed 2%.

[0038] The initial molten steel temperature in the simulation is taken as the tundish temperature of the corresponding heat, the initial composition is taken as the actual smelting composition of the corresponding heat, and the initial solid fraction is set to 0. The heat flux at the surface boundary of the crystallizer is determined based on the cooling water flow rate, the temperature difference between the inlet and outlet water, and the effective heat exchange area; the heat exchange conditions at the boundary of the secondary cooling zone are determined based on the water flow density, nozzle arrangement, and cooling water temperature of each secondary cooling loop; the air-cooled zone adopts radiation heat transfer and natural convection heat transfer boundaries; and the central symmetry plane of the billet adopts an adiabatic boundary.

[0039] The density, specific heat capacity, thermal conductivity, latent heat of solidification, liquidus temperature, and solidus temperature of the steel were determined by CALPHAD calculations and the corresponding steel property data. The simulation time step was adaptively adjusted within the range of 0.01 to 0.1 s based on the convergence state of the temperature field and solidity field.

[0040] The simulation conditions include normal conditions, extreme conditions, and defect injection conditions. The normal conditions cover the commonly used casting speed, superheat, secondary cooling water distribution, and electromagnetic stirring range for various steel grades; the extreme conditions include superheat exceeding the conventional process upper limit by 10°C, a step change in casting speed of ±0.4 m / min, a reduction of more than 50% in the flow rate of a single secondary cooling circuit, cessation of electromagnetic stirring, partial blockage of the nozzle, and a decrease in the heat exchange performance of the protective slag.

[0041] For the injection of center segregation defects, a roll gap bulge displacement is set in the two-phase region with a solid fraction of 0.2 to 0.8, and the interdendritic feeding flow velocity coefficient is increased by 10% to 30%, causing the enriched solute to migrate towards the center of the billet. When the ratio of the central composition of the billet to the initial composition exceeds the segregation threshold specified in the quality standard of the corresponding steel grade, the corresponding position is marked as the true value of center segregation.

[0042] For the internal crack injection scenario, tensile strain caused by roll gap misalignment or bulging is set within the brittle temperature range of the steel grade. The cumulative tensile strain at the corresponding location is obtained by accumulating the strain increment at each simulation time. When the cumulative tensile strain exceeds the critical strain determined by high-temperature mechanical tests, the corresponding location is marked as the true value of the internal crack.

[0043] For the injection of surface longitudinal cracks, the local surface heat flux of the crystallizer is reduced by 15%–30%, creating a local difference in billet shell thickness. When the local thermal stress exceeds the critical strength of the steel grade at the corresponding temperature, the corresponding location is marked as the true value of the surface longitudinal crack. For the injection of corner cracks, the local heat transfer intensity in the corner region is increased, and tensile strain is superimposed in the straightening zone. When the corner temperature enters the brittle temperature range of the steel grade and the cumulative strain exceeds the critical strain, the corresponding location is marked as the true value of the corner crack.

[0044] For the slag entrainment defect injection condition, protective slag particles are released into the crystallizer flow field, and particle trajectory tracking is performed. When the protective slag particles enter the billet shell capture area and the local solid fraction reaches the particle capture threshold, the corresponding position is marked as the true value of slag entrainment. For the bulging defect injection condition, the inter-roller displacement is calculated based on the static pressure of molten steel, roll gap, billet shell thickness, and high-temperature elastic-plastic parameters of the steel grade. When the maximum inter-roller displacement exceeds the allowable bulging amount of the corresponding casting machine, the corresponding position is marked as the true value of bulging.

[0045] Each simulation case outputs a timestamp, metallurgical length coordinates, process parameters, temperature field, solidity field, solute concentration field, billet shell thickness, stress-strain state, and defect truth labels, generating a total of 12,000 sets of virtual casting data. The simulation data and the measured data use the same field names, physical units, timestamp format, and metallurgical length coordinates to ensure that the two types of data can be mixed and used in subsequent training batches.

[0046] In the initial training phase, a higher proportion of simulated data is used to allow the model to preferentially obtain state representations that conform to physical laws. As training progresses, the proportion of real-world data is gradually increased. The proportion of simulated data in the t-th training step is set according to the following formula:

[0047] ;

[0048] in, This represents the proportion of simulated language at the start of training, set to 0.8. This represents the minimum percentage of simulated corpus that can be retained during the later stages of training, set to 0.15. e is the natural constant and the base of the exponential decay operation. This is the decay coefficient of the simulated corpus proportion, which is a positive number. Before training, the total number of training steps T and the proportion of the target simulated corpus corresponding to the T-th training step are determined. ,in ; T and Substituting into the above formula and solving inversely, we get... Maintain during training Unchanged. At least 15% of the simulation data will still be used in the later stages of training to maintain coverage of extreme and hazardous conditions.

[0049] The dataset is divided into training, validation, and test sets according to the heat batch, with proportions of 80%, 10%, and 10%, respectively. All billet slices generated from the same heat batch are only included in one dataset, and data from the same heat batch are not simultaneously assigned to the training, validation, or test sets to avoid information leakage caused by data from the same heat batch being spread across sets.

[0050] S2. Divide the area from the meniscus to the cut position after the slab has solidified into multiple equidistant anchor points. For a slab continuous casting machine with a metallurgical length of 28m, set the anchor point spacing to 0.1m, forming a total of 281 anchor points. Define the area 0.05m before and after each anchor point as its data aggregation neighborhood. For continuous casting machines of other metallurgical lengths, maintain the 0.1m anchor point spacing, and merge any section less than 0.1m at the end into the previous anchor point.

[0051] For each slab slice, the time when it passes through each anchor point is determined based on the integral result of the casting speed obtained from S1, and the process parameters in the neighborhood of the corresponding anchor point are extracted from the dual-source continuous casting time sequence corpus.

[0052] For 1Hz temperature, pulling speed, water flow rate, and pressure data, the aggregation weight is determined according to the distance between the sampling location and the location anchor point, with sampling points closer to the location anchor point receiving a larger weight. The aggregation value of the location anchor point is obtained by multiplying each valid sample value in the neighborhood by its corresponding weight, summing the results, and then dividing by the sum of all valid weights.

[0053] When there is only one valid sampling point in the neighborhood, the sampling value is used directly; when there is no valid sampling point in the neighborhood, but the distance between adjacent valid positions does not exceed 0.3m, linear interpolation is performed; when the distance between adjacent valid positions exceeds 0.3m, no interpolation is performed, but the median value of the corresponding channel training set is filled in, and the channel validity mask is set to invalid.

[0054] For 10kHz vibration data, vibration signal segments are extracted according to the neighborhood of the location anchor point, and the root mean square value, peak value, kurtosis, and energy in the frequency bands of 0–200Hz, 200–1000Hz, and 1000–5000Hz are extracted from the signal segments. In this way, the high-frequency vibration signal is converted into vibration characteristics that can be mapped to the metallurgical length coordinate system together with the 1Hz process data.

[0055] Image data includes infrared thermograms and visible light images located at the crystallizer outlet, the observation position in the secondary cooling zone, and the detection position on the surface of the cast billet. Each frame of the image is associated with the corresponding slice of the cast billet and the position anchor point according to the acquisition timestamp and the integral result of the casting speed. For the infrared thermogram, the average temperature, maximum temperature, temperature standard deviation, lateral temperature gradient, and longitudinal temperature gradient are extracted within the cast billet area. For the visible light image, grayscale conversion, median filtering, and edge detection are performed sequentially to extract the grayscale mean, grayscale standard deviation, edge pixel ratio, and high gradient connected region area ratio. The above values ​​are arranged in a fixed field order to form image features. When the image is blurry, occluded, or there is no valid image at the corresponding position, the median value of the corresponding feature in the training set is filled into the corresponding field, and the validity mask of the corresponding channel is set to invalid.

[0056] In addition to the original acquisition channels, physical derived channels are calculated and spliced. These physical derived channels include: superheat obtained from the difference between the temperature of the molten steel in the ladle and the liquidus temperature; heat flux density of the crystallizer obtained from the flow rate of the cooling water in the crystallizer, the temperature difference between the inlet and outlet water, and the effective heat exchange area; the secondary cooling water volume per unit mass of molten steel; the cooling intensity of each section of the secondary cooling system; the estimated shell thickness of the billet obtained based on the solidification coefficient and solidification time; and dimensionless characteristics reflecting the heat transfer intensity of the cast billet, the relationship between convection and heat diffusion at casting speed, and the relationship between sensible and latent heat.

[0057] For channels with defined physical ranges, normalization is performed according to the lower and upper physical allowable limits. For vibration and image features without fixed physical ranges, robust normalization is performed using the median and interquartile range of the corresponding channels in the training set. The validation, test, and online data all use the normalization parameters determined by the training set; the normalization parameters are not recalculated for the validation, test, or online data.

[0058] The original process channel, physical derived channel, vibration feature, image feature, and channel validity mask corresponding to each anchor point are concatenated into a positional state vector in a fixed order. Every four consecutive anchor points constitute a positional term, with each positional term corresponding to a 0.4m solidification process of the cast billet. The positional state vectors within each term are flattened sequentially and mapped to 1280-dimensional term embeddings through a linear projection layer, followed by layer normalization.

[0059] Change of ladle, change of steel grade, section switching, electromagnetic stirring start / stop, electromagnetic braking start / stop, and nozzle replacement are each coded as different types of event tokens. The location anchor point corresponding to the event is determined based on the event occurrence time and the integral result of the casting speed, and the event token is inserted before the corresponding location token. When two or more events occur in the neighborhood of the same location anchor point, they are inserted sequentially according to their actual occurrence time; different events are not merged into the same event token.

[0060] Each term is simultaneously superimposed with a metallurgical length position code, a channel identity code, and a billet slice identity code. Among them, the metallurgical length position code is used to represent the physical location corresponding to the term, the channel identity code is used to distinguish the original process quantity, the physical derived quantity, the vibration feature, and the image feature, and the billet slice identity code is used to distinguish adjacent billet slices within the same casting cycle.

[0061] When the proportion of invalid channels in a positional term exceeds 30%, the positional term is not used as the target for mask reconstruction, but the positional term and its channel validity mask are still retained so that the model can identify the missing data state; when the proportion of invalid channels in 5 or more consecutive positional terms exceeds 30%, the corresponding billet slice is marked as an incomplete trajectory sample, and the confidence level of the corresponding prediction result is directly determined as level C during online inference.

[0062] S3. A Transformer with sparse hybrid expert layers is used as the backbone network, with a hidden layer dimension of 1280 and 20 layers. Each hybrid expert layer includes 8 expert networks, and the two expert networks with the highest activation probability for each term are selected. In addition to receiving term embeddings, the expert routers also receive location segment numbers and solids ratio interval numbers. The location segment number is determined based on whether the anchor point is in the crystallizer zone, secondary cooling zone, straightening zone, or air cooling zone. The solids ratio interval number is calculated by the phase diagram network based on the temperature of the corresponding anchor point and the actual composition of the current steel grade, and is numbered according to three intervals: less than 0.2, not less than 0.2 and less than 0.8, and not less than 0.8. When the temperature channel of the corresponding anchor point is invalid, the solids ratio interval number is calculated using the temperature of the nearest valid upstream anchor point. The same numbering rule is used in the training and online inference phases. The predicted solids ratio output from the current layer of the time-series basic model is not used to determine the routing number, allowing each expert to learn the state evolution laws of different location segments and solidification stages. The total number of model parameters is approximately 1.2B.

[0063] During pre-training, positional words are randomly occluded, and the occluded content is reconstructed based on the remaining positional words. The occlusion ratio of ordinary positional words varies between 15% and 40%, and in this embodiment, the initial occlusion ratio is set to 25%. The occlusion probability of the two positional words before and after the event word is set to twice the occlusion probability of ordinary positional words. At the same time, the subsequent positional state is predicted based on the positional words already experienced by the billet slice, and quantile values ​​of 0.05, 0.25, 0.5, 0.75, and 0.95 are output respectively.

[0064] The backbone network outputs the temperature field, solids ratio field, and solute concentration field in parallel. Using the physical time and metallurgical length of the billet slice passing through each anchor point as continuous coordinates, automatic differentiation is used to calculate the predicted temperature, predicted solids ratio, and predicted solute concentration relative to time and space. The normalized coordinate derivatives obtained from automatic differentiation are converted to derivatives with actual physical units based on the time and length scales used during training.

[0065] Physical residual calculation points are set at each anchor point, the midpoint of adjacent anchor points, and both sides of the event term. The predicted temperature change rate, predicted temperature spatial gradient, steel density, specific heat capacity, thermal conductivity, latent heat of solidification, and predicted solid fraction change rate are substituted into the solidification heat transfer equation to obtain the energy imbalance during heat transfer. The predicted solute concentration change rate, interdendritic feeding flow velocity, effective solute diffusion coefficient, predicted solute concentration spatial gradient, and solute redistribution source term are substituted into the solute conservation equation to obtain the solute imbalance. The flow proxy output of the molten steel in the crystallizer is substituted into the incompressible flow continuity equation to obtain the flow mass imbalance.

[0066] The surrogate model for molten steel flow in the crystallizer employs a coordinate-input type fully connected network. Its inputs include the immersion depth of the nozzle, the nozzle opening, the casting speed, the cross-sectional dimensions of the crystallizer, the electromagnetic stirring current and frequency, the density and viscosity of the molten steel, and the spatial coordinates and time of the physical residual calculation point. The output consists of three velocity components along the width, thickness, and casting direction of that calculation point. The surrogate model comprises four hidden layers, each with 256 neurons. The velocity field output from a multiphysics simulation platform is used as the supervision label, and training is performed using the mean square error between the predicted and simulated values ​​of the three velocity components. Training stops when the mean square error of the validation set fails to decrease after 10 consecutive training cycles. After training, the partial derivatives of the three velocity components output by the surrogate model with their corresponding spatial coordinates are calculated and summed. The resulting velocity divergence is used as the flow continuity residual. Therefore, the flow continuity residual does not depend on the solid fraction or temperature results predicted by the current layer of the time-series basic model.

[0067] The three types of imbalances are normalized using the root mean square value of the corresponding physical residuals in the training set, and then the squared average of the normalized residuals is calculated to form the partial differential equation residual loss. Physical residuals are not calculated for locations with invalid channels, incomplete trajectories, or missing corresponding physical quantities.

[0068] A phase diagram network with fewer than 1M parameters was trained based on CALPHAD calculations. The input to the phase diagram samples includes temperature and the elemental contents of C, Si, Mn, P, S, Cr, Ni, and Mo. The output includes liquidus temperature, solidus temperature, equilibrium solid fraction, equilibrium distribution coefficients of each element, density, specific heat capacity, thermal conductivity, and latent heat of solidification. The phase diagram network employs a 4-layer fully connected structure, with 256 neurons in each hidden layer.

[0069] After the phase diagram network is trained, the liquidus temperature error, solidus temperature error, and equilibrium solid fraction error are calculated on independent phase diagram validation samples. The phase diagram network is only used for physical consistency pre-training when the average absolute errors of the liquidus temperature and solidus temperature are both no greater than 2℃, and the average absolute error of the equilibrium solid fraction is no greater than 0.02. If these conditions are not met, the CALPHAD sampling density is increased and the network is retrained.

[0070] A phase diagram network is used to impose equilibrium solid fraction bandwidth, lever law, Scheil limit, and liquidus monotonicity constraints on the model output. The model-predicted solid fraction is compared with the equilibrium solid fraction given by the phase diagram network, and deviations exceeding the non-equilibrium correction bandwidth are included in the phase diagram network's calculations. Figure 1 Consistency loss; compare the predicted liquid phase composition with the composition range jointly defined by the lever law and the Scheil model, and include the portion exceeding the range in the phase composition. Figure 1 Consistency loss; when a combination of solute enrichment and a predicted increase in liquidus temperature occurs, the corresponding change is included in the phase... Figure 1 Causes sexual harm.

[0071] The non-equilibrium correction bandwidth is determined according to steel grade and temperature range: For the state points corresponding to the normal simulation conditions in the training set, the absolute difference between the simulated solid fraction and the equilibrium solid fraction output by the phase diagram network is calculated, and the 95th percentile of this absolute difference for each steel grade and temperature range is determined as the corresponding non-equilibrium correction bandwidth. The upper limits of allowable deviation for energy conservation and solute conservation are respectively taken as the 95th percentiles of the normalized energy deviation and normalized solute deviation corresponding to the qualified samples in the validation set; when the allowable deviation specified in the process specification for the corresponding steel grade is smaller, the value specified in the process specification is adopted. The prediction error being within the allowable range of the task means that the absolute error of the center segregation prediction does not exceed 0.05, the absolute error of the solidification endpoint position prediction does not exceed 0.30m, and after classification using the internal crack risk threshold determined when the F1 value of the validation set is the largest, the internal crack prediction result is consistent with the quality inspection result; when a validation sample simultaneously meets the above conditions of its corresponding prediction task, it is determined as a sample whose prediction error is within the allowable range of the task.

[0072] For each billet slice, the total heat removed by each cooling loop in the secondary cooling zone is calculated and compared with the enthalpy drop of the billet from the inlet to the outlet of the secondary cooling zone. The normalized difference between the two is included in the energy conservation loss. The initial total solute is calculated based on the initial composition and volume of the billet slice, and the predicted total solute is calculated based on the predicted solute concentration field. The normalized difference between the two is included in the solute conservation loss.

[0073] The total pre-training loss is:

[0074] ;

[0075] in, It is the total pre-training loss corresponding to the current training batch; The data loss consists of positional word mask reconstruction loss and subsequent positional state prediction loss; the required true values ​​come from the unmasked original words and subsequent positional states in the dual-source continuous casting time-series corpus. This is the residual loss from the partial differential equations. The temperature field, solid fraction field, and solute concentration field output by the model are substituted into the solidification heat transfer equation, the solute conservation equation, and the flow continuity equation to obtain the physical residuals. The root mean square value of the residuals corresponding to the training set is used for normalization, and then the squared average of the normalized residuals is calculated. It is phase Figure 1 Consistency loss. The model prediction results are compared with the equilibrium solid fraction bandwidth, lever law, Scheil limit, and liquidus monotonicity relationship output by the phase diagram network, and the deviations exceeding the allowable range are summarized. It is the global conservation loss, which consists of the normalized difference between the heat exchange in the secondary cooling zone and the enthalpy drop of the billet, and the normalized difference between the predicted total solute and the initial total solute. It is the weight of the residual loss of the partial differential equation; It is phase Figure 1 Weight of inconsistency loss; It is the weight of the global conservation loss.

[0076] Each loss weight is updated using a homoscedastic uncertainty weighting method. Every 500 training steps, the gradient norm of each physical loss with respect to the shared network parameters is calculated. When the gradient norm of any physical loss is lower than 1% of the gradient norm of the data loss, the upper limit of the weight of that physical loss is increased to twice the original upper limit; when its gradient norm recovers to more than 1% of the gradient norm of the data loss and remains so for 3 statistical periods, the original upper limit of the weight is restored.

[0077] When the physical violation rate on the validation set exceeds 0.5%, increase the phase... Figure 1 The consistency loss weights are applied, and the network parameters are rolled back to the training checkpoint where the most recent physical violation rate does not exceed 0.5%. The physical violation rate is the proportion of the number of state points in the validation set whose predicted states exceed the phase diagram limits, energy conservation tolerance, solute conservation tolerance, or solidity monotonicity limit, out of the total number of validation state points.

[0078] Equipment embedding, steel grade embedding, and cross-section embedding are established separately. The inputs for equipment embedding include casting machine type, arc radius, metallurgical length, crystallizer cross-sectional range, roller arrangement, secondary cooling zone, and electromagnetic stirring device parameters, which are mapped to 256-dimensional equipment embedding through a two-layer fully connected network. The inputs for steel grade embedding include the actual smelting composition and the liquidus temperature, solidus temperature, two-phase zone width, element balance distribution coefficient, and thermophysical characteristics output from the phase diagram network, which are mapped to 256-dimensional steel grade embedding through a two-layer fully connected network. The inputs for cross-section embedding include cross-section width, thickness, width-to-thickness ratio, and cross-section shape encoding, which are mapped to 256-dimensional cross-section embedding through a two-layer fully connected network.

[0079] A 768-dimensional condition vector is formed by splicing the vectors in the order of equipment embedding, steel type embedding, and cross-section embedding. Each layer of the backbone network is equipped with a scaling parameter generation layer and an offset parameter generation layer. The condition vector is mapped to the scaling coefficient and offset coefficient corresponding to the 1280-dimensional hidden state of that layer. The hidden state of that layer is multiplied by the scaling coefficient item by item and then added to the offset coefficient item by item to obtain the hidden state after modulation by the current device, steel type, and cross-section conditions.

[0080] Following the backbone network, a defect risk prediction head, a solidification endpoint prediction head, and a process parameter inversion head are set up. The defect risk prediction head outputs the risks and corresponding locations of center segregation, internal cracks, surface longitudinal cracks, corner cracks, slag entrapment, and bulging; the solidification endpoint prediction head outputs the solidification endpoint location and its quantile interval; the process parameter inversion head receives the projected temperature field, solids fraction field, and solute concentration field, as well as the defect risk, solidification endpoint location, and current control status, and outputs candidate control quantities for secondary cooling water distribution, casting speed, electromagnetic stirring, and end-of-pipe light pressure.

[0081] The task head is trained using control variables, physical states, and quality labels from simulated corpora, and process performance data from measured corpora. During training, the partial differential equation residuals and phase... Figure 1 Consistency and global conservation losses are jointly involved in optimization, ensuring that the quality prediction results and candidate control variables generated by the task head are constrained by physical laws. This embodiment was trained on a 32-GPU cluster for 11 days. After training, the physical violation rate on the validation set was 0.31%, resulting in a large-scale time-series model that can be used across casting machines, steel grades, and cross-sections.

[0082] S4. The time-series basic large model obtained in S3 is used in a steel company to build a new No. 8 arc-shaped slab continuous casting machine. The arc radius of the continuous casting machine is 10m, the metallurgical length is about 28m, and the slab cross-section is 220mm×1800mm. The first batch of products produced is high-carbon cord steel SWRH82B with a carbon content of 0.80% to 0.83%. The main control targets are center segregation, internal cracks, and the location of the solidification endpoint.

[0083] The equipment embedding network is generated by inputting the parameters of the No. 8 continuous casting machine, including its model, arc radius, metallurgical length, roll arrangement, secondary cooling zone, and electromagnetic stirring device. The actual smelting composition of SWRH82B is input into the phase diagram network and steel grade embedding network to generate the steel grade embedding. The 220mm × 1800mm cross-sectional dimensions and slab shape code are input into the cross-sectional embedding network to generate the cross-sectional embedding. The generation time for all three embeddings is less than 1 minute, after which zero-sample inference can be performed.

[0084] Unlabeled time-series data from approximately 120 heats during the first week of operation of the No. 8 continuous casting machine were used. The core parameters of the basic time-series model were frozen, and low-rank adapters with rank 16 were set in the query matrix, value matrix, and input and output projection matrices of each attention layer. The initial learning rate was set to 1 / 100,000, with 64 slab slice sequences constituting a training batch. Adaptation was stopped when the validation physical loss did not decrease after five consecutive training cycles.

[0085] The adaptation process does not use central segregation or internal crack annotation; it only calculates the partial differential equation residuals and phases corresponding to the new scene data. Figure 1Consistency loss and global conservation loss are applied, and anchoring regularization is applied to the low-rank adapter parameters. Each adaptation batch is mixed with at least 20% of the base corpus replay samples, and the physical violation rate of the model on the replay samples before and after adaptation is compared.

[0086] If the physical violation rate after adaptation is higher than that of the base model, the adaptation learning rate is reduced by 50%, and the replay sample ratio is increased to 30%. If the physical violation rate does not recover after reducing the learning rate twice consecutively, the adaptation is terminated and the parameters before adaptation are restored. The adaptation training in this embodiment uses a single industrial GPU and lasts for 2.5 hours.

[0087] The real-time process data of the No. 8 continuous casting machine is input into the material point trajectory alignment word segmenter to form the position word sequence of the current billet slice. After adaptation, the model outputs the center segregation risk, internal crack risk, solidification endpoint position, temperature field, solid fraction field, and solute concentration field, and simultaneously outputs the quantile intervals of each predicted quantity.

[0088] Based on the equilibrium solid fraction bandwidth of the current steel grade, the lever law and Scheil composition limit, the energy conservation of the secondary cooling zone, the total solute conservation, and the monotonically non-decreasing relationship of the solid fraction along the downward direction, a physically feasible set Ω is established, and the original output of the model is used to establish the physical feasibility set Ω. Project according to the following formula:

[0089] ;

[0090] in, To satisfy all physical constraints and minimize the weighted distance from the original model output, i.e. the projected physical legal output, W is a weight matrix determined based on the dimensions of different physical quantities and the confidence level of sensor measurements. Its weights are determined based on the dimensions of different physical quantities, the variance of sensor measurements, and the variance of model predictions. The first vector is the physical state vector directly output from the time-series basic model, before physical projection. It is obtained by splicing the predicted temperature, predicted solid fraction, and predicted solute concentration at each anchor point in a fixed order. The second vector is the candidate physical state vector during the projection process, whose dimension and the order of its components are the same as those of the first vector. same.

[0091] The physical feasible set must simultaneously meet the following conditions: the predicted temperature, solid fraction, and solute concentration are within the allowable range of the current steel grade phase diagram; the solid fraction is between 0 and 1; the solid fraction at the adjacent downstream position must not be lower than the solid fraction at the adjacent upstream position by more than 0.001; the normalized deviation between the predicted secondary cooling heat transfer and the enthalpy drop of the billet does not exceed the upper limit allowed in the validation set; and the normalized deviation between the predicted total solute and the initial total solute does not exceed the upper limit allowed in the validation set.

[0092] In the specific projection, firstly, the temperature, solid fraction, and solute concentration are corrected to the allowable range of the phase diagram; secondly, the overall deviation of the temperature field and solid fraction field is distributed and corrected according to the energy conservation relationship; thirdly, the solute concentration field is corrected as a whole according to the total solute conservation relationship; finally, the solid fraction is corrected by the isotonic projection method of merging adjacent disordered blocks, so that it remains monotonically non-decreasing along the metallurgical length direction.

[0093] The projection weights are determined based on the measurement variance and model prediction variance of different physical quantities. Physical quantities with higher measurement accuracy and narrower prediction intervals are assigned larger retention weights, while physical quantities with lower measurement accuracy or wider prediction intervals are assigned smaller retention weights, in order to prioritize minimizing modifications to highly reliable physical quantities.

[0094] The above projection sequence is executed cyclically. Projection stops when the weighted relative change between two adjacent projection results is no greater than 0.01%. If the stopping condition is not met after 5 iterations, the prediction is marked as a non-convergent projection. The physical projection residual is obtained by dividing the weighted difference between the projected output and the original model output by the weighted magnitude of the original model output. The additional computation time for a single projection layer is less than 5ms.

[0095] The confidence grading threshold is determined based on the offline validation set. First, samples that are physically valid and whose prediction errors are within the allowable range of the task are selected from the validation set. The width of the normalized 0.05 to 0.95 quantile interval and the 50th and 95th quantile values ​​of the physical projection residuals are then calculated.

[0096] During online prediction, a grade A is defined as follows: when both the quantile interval width and the physical projection residual do not exceed the corresponding 50th quantile value, the projection converges within 5 iterations, and the proportion of invalid input channels does not exceed 5%; when both indicators do not exceed the corresponding 95th quantile value, the projection converges within 5 iterations, and the proportion of invalid input channels does not exceed 20%; and when any indicator exceeds the corresponding 95th quantile value, the projection does not converge, the proportion of invalid input channels exceeds 20%, or the trajectories of five consecutive positions are incomplete.

[0097] During online operation, a rolling time window of 10 minutes is used to count the proportion of samples whose output exceeds the physical feasible set before projection to the total number of predicted samples within the window. When this proportion exceeds 0.5%, the correction priority of phase diagram constraints and conservation constraints in physical projection is increased, a condition drift alarm is issued to the operation and maintenance platform, and Class A predictive direct drive control is suspended. Class A predictive direct drive control is only restored when the physical violation rate does not exceed 0.5% for two consecutive rolling time windows.

[0098] S5. Input the projected temperature field, solid fraction field, solute concentration field, defect risk and solidification endpoint position obtained in S4, as well as the current secondary cooling water supply, current casting speed, current electromagnetic stirring parameters and current end light pressure parameters into the process parameter inversion head.

[0099] A candidate control variable set is constructed centered on the current control variable. For each section of the secondary cooling water flow rate, candidate values ​​are generated at ±1%, ±2%, and ±3% of the current value, respectively; for the pulling speed, candidate values ​​are generated at ±0.02 m / min and ±0.04 m / min, respectively; for the electromagnetic stirring current, candidate values ​​are generated at ±2% and ±5% of the rated current, respectively; and for the terminal light pressure reduction, candidate values ​​are generated at 0.1 mm intervals within the equipment's process allowable range. Candidate control variables that exceed the safety upper and lower limits of the L2 process automation system, exceed the equipment's allowable rate of change, or trigger PLC hard interlock conditions are directly deleted.

[0100] The remaining candidate control variables are combined one by one with the current projected physical state and input into the time-series basic large model for forward extrapolation to obtain the center segregation risk, internal crack risk, other defect risk, solidification endpoint location, shell thickness and temperature field corresponding to each candidate control variable.

[0101] For each group of candidate control quantities, candidate evaluation results are calculated. First, candidate control quantities whose projection results do not meet the physical feasibility set are deleted, and candidate control quantities whose risk for any non-target defect increases by more than 5% relative to the projection result of maintaining the current control quantity are also deleted. Target defects are determined according to the current steel grade quality control procedures. When there are two or more target defects, the predicted risk of each target defect is normalized by dividing it by the 95th percentile risk value of that defect on the validation set, and the maximum value among the normalized risks of each target defect is taken as the comprehensive target defect risk of the candidate control quantity. This avoids a situation where the risk of one target defect decreases while the risk of another target defect significantly increases. Candidate control quantities with lower comprehensive target defect risks are prioritized. When the difference in comprehensive target defect risks does not exceed 0.01, candidate control quantities whose solidification endpoint is closer to the target light reduction range and whose billet shell thickness meets safety requirements are selected. When the quality status remains the same, candidate control quantities with smaller adjustment ranges relative to the current control quantity are selected. When the adjustment ranges remain the same, candidate control quantities with smaller differences from the control quantity of the previous control cycle are selected. The finally selected candidate control quantity is determined as the target control quantity, and the difference between the target control quantity and the current control quantity is determined as the target control increment.

[0102] The effective range of the final light reduction is determined based on the projected solids fraction field. In this embodiment, the position where the solids fraction reaches 0.3 is determined as the candidate starting position for light reduction, and the position where the solids fraction reaches 0.9 is determined as the candidate ending position for light reduction. When the above positions exceed the sector range within which light reduction can be performed on the No. 8 continuous casting machine, they are restricted to the nearest boundary of the sector. The amount of light reduction is determined by the evaluation results of the candidate control quantities, while simultaneously satisfying the upper limit of roll gap adjustment, single adjustment rate, and equipment load limit.

[0103] For Level A predictions, the target control increment is compared again with the PLC hard interlock limits, the equipment's allowable rate of change, and the safety range of the L2 process automation system. Once all verifications pass, the solidification endpoint position, the start and end positions of the light reduction, the light reduction amount, the secondary cooling water distribution increment, the pull speed increment, and the electromagnetic stirring increment are sent to the PLC. If any verification fails, the prediction is not sent directly but downgraded to Level B.

[0104] For Level B predictions, the target control increment is input into the digital twin model for verification. The digital twin model receives the current temperature field, solid fraction field, solute concentration field, billet shell thickness, casting speed, secondary cooling water supply, electromagnetic stirring parameters, and light reduction parameters, respectively simulating the billet state after maintaining the current control quantity and executing the target control increment.

[0105] Simultaneously, the input temperature, boundary heat transfer intensity, and casting speed are adjusted to the upper and lower limits of their prediction ranges, respectively, to form a disturbance condition for verifying the stability of the control proposal under input uncertainty. The digital twin model outputs the following parameters after executing the control proposal: center segregation risk, internal crack risk, other defect risk, solidification endpoint location, billet shell thickness at the straightening location, billet surface temperature, energy conservation deviation, and solute conservation deviation.

[0106] The digital twin verification is deemed successful only if all of the following conditions are met: the risk of center segregation or internal cracking after implementing the target control increment is reduced by at least 10% relative to the simulation result of maintaining the current control; the solidification endpoint is located within the executable light reduction range; the billet shell thickness at the straightening position is not less than the safe thickness specified in the corresponding steel grade process specification; the billet surface temperature has not entered the brittle temperature range of the corresponding steel grade; the risk of any other defect increases by no more than 5%; the secondary cooling water supply, casting speed, roll gap, and electromagnetic stirring parameters have not exceeded the equipment safety boundaries; and the phase diagram limits, energy conservation, and solute conservation requirements are met under both normal and disturbed operating conditions.

[0107] After the digital twin verification is passed, the target control increment is sent to the PLC for execution. If any of the above conditions are not met, or if the conditions are met under normal operating conditions but not under any disturbance condition, the digital twin verification is deemed to have failed, the control recommendation is not executed, and the current safety control state is maintained.

[0108] For C-level predictions, the control variables generated by the model are not sent to the PLC. The system issues a quality risk alarm and reverts to validated expert rule control. Simultaneously, the corresponding billet slices, missing input data, quantile intervals, physical projection residuals, reasons for projection non-convergence, and digital twin verification results are marked as difficult examples and sent to the incremental pre-training queue.

[0109] After the PLC executes the control increment, it continuously collects the water flow rate, pulling speed, roller gap, electromagnetic stirring, and temperature response in subsequent control cycles. When the normalized deviation between the actual response and the model predicted response exceeds 10% for three consecutive control cycles, it stops increasing the control quantity, restores the control parameters to the most recent safe state before execution, and marks the corresponding data as an abnormal execution response sample.

[0110] The model prediction results, physical projection results, confidence levels, candidate control variables, digital twin verification results, PLC execution records, drill sample segregation results, ultrasonic flaw detection results, and manual re-inspection results are correlated according to the primary key of the billet slice. New data is imported into the dual-source continuous casting time-series corpus every 24 hours, prioritizing samples with large physical projection residuals, a confidence level of C, changes in steel grade or sudden changes in casting speed, inconsistencies between actual and model predicted responses, and inconsistencies between quality predictions and inspection results.

[0111] When performing incremental pre-training based on new samples, the weights of each physical loss are increased by 20% compared to regular pre-training. The updated model is then tested for prediction accuracy, physical violation rate, historical operating condition playback, extreme simulation operating condition playback, and inference latency. The updated model is only released to the edge inference server in a canary manner if the physical violation rate of the updated model is not higher than that of the unupdated model, the prediction accuracy of center segregation, internal cracks, and solidification endpoints is not reduced, the historical safe operating condition control results do not change in the opposite direction, and the C-level rollback can be triggered normally under extreme operating conditions.

[0112] The 1.2B parameter temporal basis model is distilled into a production-level inference model with approximately 0.2B parameters. During the distillation process, the production-level inference model simultaneously fits the positional term hidden states, quantile outputs, temperature field, solidity field, solute concentration field, and defect risk of the temporal basis model. The physically feasible manifold projection layer does not participate in the distillation, maintaining the original constraints and alternating projection computational structure.

[0113] Only when the accuracy of each quality task of the distillation model decreases by no more than 2%, the physical violation rate does not exceed 0.5%, and the end-to-end processing time of material point trajectory alignment word segmentation, model inference, physical projection, and confidence classification does not exceed 100ms, will it be deployed to the edge server of the No. 8 continuous casting machine.

[0114] After 30 consecutive days of operation, the pre-projection physical violation rate was 0.42%, and the post-projection output did not violate phase diagram limits, energy conservation, solute conservation, or solidity monotonicity constraints. Compared with the results of 86 drill samples from weeks 2 to 4, the mean absolute error of center segregation prediction was 0.04, the prediction error of solidification endpoint position was ±0.28m, and the F1 value of internal crack prediction was 0.91. The pure data-driven model obtained by removing physical losses and using the same training data had a physical violation rate of 12.7%.

[0115] It took 5 days from the commissioning of the No. 8 continuous casting machine to the model meeting quality control requirements. In the 5th week, when the production steel grade was switched to Q345B, the steel grade embedding was updated only based on the actual smelting composition of Q345B, without retraining the backbone network. The F1 value for predicting internal cracks reached 0.88 on the first day. In the initial stage of production, the amount of downgraded or scrapped due to uncertainties in quality judgment was reduced by approximately 310 tons, and an unplanned shutdown caused by a misadjustment of the secondary cooling water supply was avoided.

[0116] Example 2: Figure 2 As shown, this embodiment provides a continuous casting quality prediction and control system based on physical constraints, including:

[0117] The system consists of a dual-source corpus construction module, a trajectory lexical generation module, a basic model training module, a scene prediction and projection module, and a control decision update module, connected sequentially.

[0118] The dual-source corpus construction module is used to acquire measured time-series data, quality inspection results, and multi-physics simulation data of the continuous casting process. Based on the casting speed integral, the physical location of the quality defect is back-linked to the process state experienced by the corresponding billet slice during the continuous casting process. The fields, physical units, timestamps, and metallurgical length coordinates of the measured data and simulation data are unified to establish a dual-source continuous casting time-series corpus. The dual-source continuous casting time-series corpus is then transmitted to the trajectory word generation module.

[0119] The trajectory lexical generation module is used to set position anchor points along the metallurgical length. Based on the time when the billet slice passes through each position anchor point, it aggregates the corresponding process data, physical derived features, vibration features, image features and validity masks from the dual-source continuous casting time series corpus to generate position lexicals. It inserts event lexicals according to the occurrence position of the continuous casting event, and superimposes metallurgical length position code, channel identity code and billet slice identity code on each lexical to form a material point trajectory aligned lexical sequence. The lexical sequence is then transmitted to the basic model training module.

[0120] The basic model training module is used to input the lexical sequence into the sparse mixture expert Transformer, perform mask reconstruction and subsequent position and state prediction, and combine solidification heat transfer residuals, solute conservation residuals, flow continuity residuals, and phase... Figure 1Physical consistency pre-training is performed on the consistency loss and global conservation loss, and the hidden state is conditionally modulated by equipment embedding, steel type embedding and cross section embedding. A time series basic model with defect risk prediction head, solidification endpoint prediction head and process parameter inversion head is established, and the time series basic model is transmitted to the scene prediction projection module.

[0121] The scene prediction projection module is used to acquire unlabeled time series data of the target continuous casting scene, perform low-rank adaptation on the time series basic model, input the material point trajectory of the target continuous casting scene into the adapted time series basic model with aligned word sequence, obtain defect risk, solidification endpoint and physical state field, project the physical state field to the physical feasible set defined by phase diagram boundary, energy conservation, solute conservation and monotonic relationship of solid fraction, and determine the confidence level according to the prediction interval, projection residual, projection convergence state and input validity, and transmit the projected physical state field, defect risk, solidification endpoint and confidence level to the control decision update module;

[0122] The control decision update module is used to input the projected physical state field, defect risk, solidification endpoint location, and current control quantity into the process parameter inversion head. Within the equipment safety boundary, candidate control quantities are constructed. The candidate control quantities are then forward-engineered and physically feasible-screened using the time-series basic model to determine the target control increment. For different confidence levels, the target control increment is issued to the programmable logic controller (PLC) after safety verification, after digital twin verification, or it is rolled back to expert rule control. The control execution response and quality inspection results are then associated with the corresponding billet slices and fed back to the dual-source continuous casting time-series corpus for incremental updates to the time-series basic model.

[0123] Example 3: This example provides a computer program product, which includes computer program code. When the computer program code is run on a computer, the computer implements the continuous casting quality prediction and control method based on physical constraints as described in Example 1.

[0124] The above formulas are calculated using the actual units of their corresponding physical quantities or the normalized scale described in this embodiment. Specifically, the time in the pulling speed integral formula is expressed in seconds (s), the pulling speed in meters per second (m / s), and the calculated metallurgical length in meters (m). The simulation corpus proportion, minimum proportion, and exponential decay term in the simulation corpus proportion formula are dimensionless. The physical residuals in the pre-training total loss are weighted after normalization based on the root mean square value of the corresponding residuals in the training set. The physical projection formula uses a weight matrix W to unify the calculation scale for temperature, solid fraction, and solute concentration. All preset parameters and thresholds are determined according to the equipment safety range, steel grade process specifications, and statistical results of the training or validation set described in this embodiment. The above formulas are not uniformly considered as empirical formulas obtained through data fitting.

[0125] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0126] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0127] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 method of controlling the quality of continuous casting based on physical constraints, characterized in that, Includes the following steps: S1. Obtain the measured time series data, quality inspection results and multi-physics simulation data of continuous casting, associate the defect location with the process state of the corresponding billet slice based on the casting speed integral, unify the timestamp and metallurgical length coordinates of the two types of data, and establish a dual-source continuous casting time series corpus. S2. Set position anchor points along the metallurgical length, collect process data, physical derivation features, vibration features, image features and validity masks of the aggregated billet slices when they pass through each position anchor point, generate position words and insert event words to form a material point trajectory aligned word sequence; S3. Based on the word sequence and solidification heat transfer, solute conservation, flow continuity, phase diagram consistency and global conservation loss, the sparse mixing expert Transformer is pre-trained for physical consistency, and the equipment, steel grade and cross section are embedded with the modulation hidden state to establish a time series basic model with defect risk prediction, solidification endpoint prediction and process parameter inversion functions. S4. Adapt the time series basic model using the unlabeled time series data of the target continuous casting scenario to obtain the defect risk, solidification endpoint and physical state field. Project the physical state field to the physical feasible set defined by the phase diagram boundary, energy conservation, solute conservation and monotonic relationship of solid fraction, and determine the confidence level based on the prediction interval, projection residual, projection convergence state and input validity.

2. The continuous casting quality prediction and control method based on physical constraints according to claim 1, characterized in that, Also includes: S5. Generate candidate control quantities based on the projected physical state field and the current control quantity. Determine the target control increment through time series basic model deduction and physical feasibility screening. After performing safety verification according to the confidence level, issue the control quantity, issue the control quantity after digital twin review, or roll back the control according to expert rules. Then, feed the execution response and quality inspection results back to the dual-source continuous casting time series corpus to incrementally update the time series basic model.

3. The continuous casting quality prediction and control method based on physical constraints according to claim 1, characterized in that, S1 specifically includes: The process parameters, equipment status, vibration data, image data and quality inspection results of the continuous casting process are obtained. Correspondence is established by heat number, billet slice and timestamp. Missing data, abnormal data and data with inconsistent dimensions are processed to form the measured continuous casting time series data. Based on the billet cutting position and the integral of the casting speed, the movement trajectory of the billet slice along the metallurgical length is reconstructed. The defect position obtained by quality inspection is reverse-correlated to the process state experienced by the corresponding billet slice in the crystallizer, secondary cooling zone, straightening zone and air cooling zone, and incomplete trajectory samples are marked. The initial and boundary conditions of the multiphysics simulation are set according to the actual working conditions. Simulation data including normal working conditions, extreme working conditions and defect injection working conditions are generated. The field names, physical units, timestamp formats and metallurgical length coordinates of the simulation data and the actual continuous casting time series corpus are unified to form a dual-source continuous casting time series corpus.

4. The continuous casting quality prediction and control method based on physical constraints according to claim 1, characterized in that, S2 specifically includes: Multiple anchor points are set along the billet pulling direction. The time when the billet slice passes through each anchor point is determined according to the motion trajectory. The process data in the corresponding neighborhood is extracted, and the vibration data is segmented and frequency band features are extracted to obtain the aggregated data of each anchor point. Based on the aggregated data, the estimated values ​​of superheat, crystallizer heat flux density, secondary cooling intensity and billet shell thickness are calculated. Normalization or robust standardization is performed on different channels respectively. The processed process data, physical derived features, vibration features, image features and validity mask are concatenated into a position state vector and projected to form position words. Based on the occurrence time of the continuous casting event and the integral result of the casting speed, the corresponding event term is inserted into the position term sequence, and the metallurgical length position code, channel identity code and billet slice identity code are superimposed. Invalid position terms and incomplete trajectory samples are marked according to the proportion of invalid channels to form a material point trajectory aligned term sequence.

5. The continuous casting quality prediction and control method based on physical constraints according to claim 1, characterized in that, S3 specifically includes: The material point trajectory aligned word sequence is input into a Transformer with a sparse hybrid expert layer. The expert network is selected according to the position segment and solidity interval. Position word mask reconstruction and subsequent position state prediction are performed, and the temperature field, solidity field and solute concentration field are output. The solidification heat transfer residual and solute conservation residual are calculated based on the temperature field, solid fraction field and solute concentration field. The flow continuity residual is calculated based on the output of the molten steel flow proxy model in the crystallizer. Phase diagram consistency constraints are applied using the phase diagram network. Global conservation losses are formed based on the heat exchange in the secondary cooling zone, the enthalpy drop of the billet and the total amount of solute. The Transformer is trained together with the various losses. Based on the continuous casting equipment parameters, steel composition, and billet cross-sectional parameters, equipment embedding, steel embedding, and cross-sectional embedding are generated respectively. The condition vector formed by the three embeddings is mapped to the hidden state scaling coefficient and offset coefficient. A defect risk prediction head, a solidification endpoint prediction head, and a process parameter inversion head are set to obtain the time series basic model.

6. The continuous casting quality prediction and control method based on physical constraints according to claim 1, characterized in that, S4 specifically includes: The equipment parameters, steel composition, and cross-sectional parameters of the target continuous casting scenario are input into the time series basic model to generate corresponding embeddings; the backbone parameters of the time series basic model are frozen, and a low-rank adapter is trained using the unlabeled time series data of the target continuous casting scenario and the playback samples of the basic corpus. The adaptation parameters are then determined to be retained or restored based on the physical violation rate before and after adaptation. The material point trajectory of the target continuous casting scenario is aligned with the word sequence and input into the adapted time series basic model to obtain the defect risk, solidification endpoint and physical state field. The physical state field is then projected according to the phase diagram boundary, energy conservation, solute conservation and solid fraction monotonic relationship until the projection stopping condition is met, and the physical projection residual is calculated. Based on the validation samples, the classification thresholds for the prediction interval width and physical projection residual are determined. The prediction interval width, physical projection residual, projection convergence status, and proportion of invalid input channels are compared with the classification thresholds to determine the confidence level of A, B, or C. The decision to suspend A-level prediction direct drive control is based on the physical violation rate within the rolling time window.

7. The continuous casting quality prediction and control method based on physical constraints according to claim 2, characterized in that, S5 specifically includes: The projected physical state field, defect risk, solidification endpoint and current control quantity are input into the process parameter inversion head. Candidate control quantities are generated with the current control quantity as the center. Candidate control quantities that exceed the equipment safety range, allowable rate of change or hard interlock conditions are deleted. After forward extrapolation of the time series basic model and physical feasibility screening, the target control increment is determined. For Level A confidence, the target control increment is sent to the programmable logic controller after safety verification; for Level B confidence, the current operating condition and disturbance operating condition are reviewed using a digital twin model, and the increment is sent after the review is passed; for Level C confidence, the increment is stopped and the system reverts to expert rule control, and the corresponding data is marked as a difficult example sample; the equipment response after control execution is collected, and the system is restored to the safe state before execution if the actual response and the predicted response are inconsistent.

8. The continuous casting quality prediction and control method based on physical constraints according to claim 7, characterized in that, It also includes: associating model prediction results, control execution records, and quality inspection results according to billet slices and feeding them back to the dual-source continuous casting time series corpus; using the incremental addition of new samples to train the time series basic model; and deploying the updated model when the physical violation rate, prediction accuracy, historical working condition playback, and inference latency meet the release conditions.

9. A continuous casting quality prediction and control system based on physical constraints, employing the continuous casting quality prediction and control method based on physical constraints as described in any one of claims 1 to 8, characterized in that, include: The module consists of a dual-source corpus construction module, a trajectory lexical generation module, a basic model training module, a scene prediction and projection module, and a control decision update module, which are connected in sequence.

10. A computer program product, characterized in that, The computer program product includes computer program code, which, when run on a computer, causes the computer to implement the physical constraint-based continuous casting quality prediction and control method as described in any one of claims 1 to 8.