Intelligent control system for casting machine based on digital twinning

Abstract

The application discloses a kind of based on digital twinning's casting machine intelligent control system, comprising: acquisition module, for obtaining the multidimensional original data of casting machine whole-process equipment;Pretreatment module, for the multidimensional original data of casting machine whole-process equipment is pretreated, and obtains pretreatment data;Construction module, for constructing digital twin model;Prediction module, for the pretreatment data is input into digital twin model and is predicted, and output target decision scheme;Control module, for based on the target decision scheme to casting machine whole-process equipment is intelligently controlled;Based on digital twinning model, precision prediction is realized to casting machine whole-process, and closed-loop intelligent control is formed, improves casting machine control efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of casting machines, and in particular to an intelligent control system for casting machines based on digital twins. Background Technology

[0002] As a core piece of equipment in the metallurgical industry connecting steelmaking and rolling, the continuous casting machine (hereinafter referred to as the casting machine) directly determines the quality of the cast billet, production efficiency, energy consumption, and service life of the equipment. With the transformation of the steel industry towards intelligent and green production, traditional casting machine control technology can no longer meet the production requirements of high quality, high efficiency, and low energy consumption, and has gradually exposed many technical bottlenecks.

[0003] In existing technologies, traditional casting machine control systems often focus on monitoring local parameters of a single process (such as crystallizer temperature and straightening machine speed) or core equipment, lacking multi-dimensional data coverage of the entire casting process (steel casting, crystallization, secondary cooling, billet cutting, etc.), resulting in low casting machine control accuracy and relatively lagging overall control.

[0004] Therefore, there is an urgent need for an intelligent control system for casting machines based on digital twins to solve the above problems. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in the aforementioned technologies. Therefore, the objective of this invention is to propose an intelligent control system for casting machines based on digital twins, which achieves accurate prediction and forms closed-loop intelligent control of the entire casting machine process based on a digital twin model, thereby improving the control efficiency of the casting machine.

[0006] To achieve the above objectives, embodiments of the present invention propose an intelligent control system for casting machines based on digital twins, comprising: The acquisition module is used to acquire multi-dimensional raw data of the entire casting machine process equipment; The preprocessing module is used to preprocess the multi-dimensional raw data of the entire casting machine process equipment to obtain preprocessed data; Modules for building digital twin models; The prediction module is used to input the preprocessed data into the digital twin model for prediction and output the target decision scheme; The control module is used to intelligently control the entire process equipment of the casting machine based on the target decision scheme.

[0007] Preferably, the acquisition module includes: The first acquisition submodule is used to acquire the ladle car position / speed, molten steel temperature and molten steel flow rate data during the molten steel pouring process, as the first data; The second acquisition submodule is used to acquire data on crystallizer liquid level, copper plate temperature, and cooling water flow / pressure during the crystallizer solidification process, as second data. The third acquisition submodule is used to acquire the vibration table amplitude / frequency and vibration table cylinder pressure data in the vibration table coordination process, as the third data. The fourth acquisition submodule is used to acquire data on the fan-shaped section pulling speed, roller pressure / speed and billet temperature in the fan-shaped section conveying process, as the fourth data. The fifth acquisition submodule is used to acquire data on slab surface defects, internal density / defects, and slab dimensions during the slab inspection process, as the fifth data. The sixth acquisition submodule is used to acquire the roller conveyor speed and slab position / attitude data in the roller conveyor transfer process, as the sixth data; The timing annotation submodule is used to uniformly annotate the first to sixth data with a 10Hz sampling frequency timestamp. The determination submodule is used to use the first data, second data, third data, fourth data, fifth data, and sixth data as multi-dimensional raw data for the entire casting machine process.

[0008] Preferably, the preprocessing module includes: The data cleaning submodule is used to process the missing and outlier values ​​in the multi-dimensional raw data of the entire casting machine process equipment to obtain the multi-dimensional raw data of the entire casting machine process equipment after data cleaning. The noise reduction submodule is used to reduce the noise of multi-dimensional raw data from the entire casting machine process after data cleaning, and obtain noise-reduced data. The timing synchronization submodule is used to perform linear interpolation alignment of the noise reduction data according to the timestamp, generating a 10Hz timing data stream as preprocessed data.

[0009] Preferred methods for constructing digital twin models include: Obtain a multi-dimensional training dataset for the entire casting machine process; the training dataset includes equipment parameters and operating parameters for the entire casting machine process. A virtual mapping model is constructed based on the equipment parameters of the entire casting process described in the training dataset and the mapping between the model and the casting equipment. The VAE-GAN-Transformer hybrid model is trained based on the training dataset, and anomaly detection is performed based on the set loss function and training parameters to obtain a full-process anomaly perception model. Based on the improved NSGA-II algorithm, a full-process multi-objective decision-making model is obtained by setting the objective function and constraints, designing the algorithm structure, and iteratively optimizing the output of the Pareto optimal solution set. An initial digital twin model is obtained based on the virtual mapping model, the full-process anomaly perception model, and the full-process multi-objective decision-making model. Obtain a multi-dimensional test dataset covering the entire casting machine process; The initial digital twin model is tested based on the test dataset. When the test results meet the requirements, a trained digital twin model is obtained.

[0010] Preferred methods for constructing virtual mapping models include: The virtual mapping model includes device-level refined modeling, system-level coupled modeling, and operating condition-level adaptation modeling. Digital atomic models of individual devices are constructed, and geometric modeling, physical behavior modeling and virtual sensor mapping modeling are completed for each device. The device structure, physical characteristics and data acquisition functions of each device are digitally mapped to complete the device-level fine modeling. The physical coupling relationship between devices is constructed based on the device-level atomic model, and the virtual twin is integrated and assembled according to a unified coordinate system to complete the system-level coupling modeling. We acquire production scenarios for various steel grades, establish parametric working condition templates, and set up a dynamic correction mechanism to eliminate the discrepancy between virtual and physical conditions, thus completing working condition-level adaptation modeling.

[0011] Preferably, a digital atomic model of each individual device is constructed, and geometric modeling, physical behavior modeling, and virtual sensor mapping modeling are completed for each device. The device structure, physical characteristics, and data acquisition functions of each device are digitally mapped to achieve refined device-level modeling, including: The geometric modeling uses SolidWorks for mechanical structure modeling and Unity3D for visualization rendering optimization; the casting machine is divided into 6 core pieces of equipment: tundish car, crystallizer, vibrating table, sector section, slab inspection device, and roller conveyor. Each type of equipment is constructed in layers according to core components, auxiliary components, and connecting structures. The physical behavior modeling includes a molten steel flow model, a solidification temperature field model, and a shaking table dynamics model. The molten steel flow model was constructed using the SPH smoothed particle fluid dynamics algorithm; The solidification temperature field model was constructed using the FEM finite element algorithm. The governing equations were established based on thermophysical parameters, and transient thermal analysis was employed. The solver was the conjugate gradient method. The dynamic model of the shaking table adopts the multibody dynamics algorithm. The crystallizer and the shaking table are rigidly connected. The motion equation is established based on the hydraulic cylinder driving force model and solved using the Newmark-β method. The virtual sensor mapping modeling involves placing virtual sensors at the locations of physical sensors corresponding to the virtual devices, with each sensor number corresponding to a physical sensor. Key parameters such as sensor sampling frequency are set, and the output format is sensor number-timestamp-value.

[0012] Preferred methods for constructing a full-process anomaly detection model include: A hybrid VAE-GAN-Transformer structure was built based on TensorFlow 2.15. The VAE encoder was used for feature extraction and dimensionality reduction, the Transformer decoder improved the ability to capture temporal features, and the discriminator was used to distinguish anomalies. Construct the total loss function; Obtain the training dataset, use the AdamW optimizer and cosine annealing learning rate to train the model, set an early stopping mechanism to avoid overfitting, and obtain a full-process anomaly perception model.

[0013] Preferred methods for constructing a full-process multi-objective decision-making model include: Define a multi-objective function; Set constraints based on equipment parameters and process parameters respectively: The model structure is constructed using an improved NSGA-II algorithm; The population is initialized using Latin hypercube sampling. Based on twin simulation evaluation, non-dominated sorting and crowding calculation, selection-crossover-mutation operation, and Pareto optimal solution set is iteratively output to obtain a full-process multi-objective decision model.

[0014] Preferably, the control module includes: The parsing submodule is used to parse the target decision scheme to obtain the target control command; The execution submodule is used to perform intelligent control of the entire casting machine process based on the target control command.

[0015] Preferably, the total loss function is constructed as follows: ; in, Represents the total loss function; This represents the VAE reconstruction loss; This represents the VAEKL divergence loss; Indicates GAN adversarial loss; Indicates cross-device collaboration loss; , , , These are the weighting coefficients, all of which are greater than 0 and have a sum of 1.

[0016] This invention provides an intelligent control system for casting machines based on digital twins. By acquiring multi-dimensional raw data from the entire casting machine process in real time through an acquisition module, the system can comprehensively and promptly understand the equipment's operating status. For example, real-time monitoring of key parameters such as temperature, pressure, and speed allows the prediction module to quickly analyze and predict based on the digital twin model once abnormalities are detected. The control module can then immediately adjust the equipment according to the target decision-making plan, reducing downtime caused by equipment failures or parameter deviations, thereby improving production continuity and efficiency. The constructed digital twin model can simulate the entire casting machine production process. The system can identify bottlenecks and potential optimization points in the production process through simulation and analysis of different production scenarios; it can perform real-time evaluation and prediction of equipment operating status using the digital twin model; and it can simulate the production effects under different equipment configurations and operating strategies.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a block diagram of an intelligent control system for a casting machine based on digital twin according to an embodiment of the present invention; Figure 2 This is a block diagram of a preprocessing module according to an embodiment of the present invention; Figure 3 This is a block diagram of a control module according to an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] Example 1, such as Figure 1 As shown, an intelligent control system for a casting machine based on digital twins includes: The acquisition module is used to acquire multi-dimensional raw data of the entire casting machine process equipment; The preprocessing module is used to preprocess the multi-dimensional raw data of the entire casting machine process equipment to obtain preprocessed data; Modules for building digital twin models; The prediction module is used to input the preprocessed data into the digital twin model for prediction and output the target decision scheme; The control module is used to intelligently control the entire process equipment of the casting machine based on the target decision scheme.

[0022] In this embodiment, the multi-dimensional raw data includes multi-dimensional data from the steel pouring process, the vibration table coordination process, the sector section conveying process, the slab inspection process, and the roller conveyor transfer process.

[0023] In this embodiment, preprocessing includes data cleaning, data noise reduction, and time synchronization processing.

[0024] In this embodiment, the target decision scheme refers to the Pareto optimal solution that achieves a dynamic balance between the three major objectives of minimizing slab defect rate, minimizing energy consumption, and maximizing casting speed throughout the entire casting process, while satisfying all production constraints.

[0025] The working principle and beneficial effects of the above technical solution are as follows: By acquiring multi-dimensional raw data of the entire casting machine process in real time through the acquisition module, the system can comprehensively and timely understand the operating status of the equipment. For example, real-time monitoring of key parameters such as temperature, pressure, and speed allows the prediction module to quickly analyze and predict based on a digital twin model once abnormal parameters are detected. The control module can then immediately adjust the equipment according to the target decision plan, reducing downtime caused by equipment failure or parameter deviation, thereby improving production continuity and efficiency. The constructed digital twin model can simulate the entire casting machine production process. The system can identify bottlenecks and potential optimization points in the production process through simulation and analysis of different production scenarios; it can use the digital twin model to evaluate and predict the operating status of the equipment in real time; and it can simulate the production effects under different equipment configurations and operating strategies.

[0026] Example 2: Acquisition module, including: The first acquisition submodule is used to acquire the ladle car position / speed, molten steel temperature and molten steel flow rate data during the molten steel pouring process, as the first data; The second acquisition submodule is used to acquire data on crystallizer liquid level, copper plate temperature, and cooling water flow / pressure during the crystallizer solidification process, as second data. The third acquisition submodule is used to acquire the vibration table amplitude / frequency and vibration table cylinder pressure data in the vibration table coordination process, as the third data. The fourth acquisition submodule is used to acquire data on the fan-shaped section pulling speed, roller pressure / speed and billet temperature in the fan-shaped section conveying process, as the fourth data. The fifth acquisition submodule is used to acquire data on slab surface defects, internal density / defects, and slab dimensions during the slab inspection process, as the fifth data. The sixth acquisition submodule is used to acquire the roller conveyor speed and slab position / attitude data in the roller conveyor transfer process, as the sixth data; The timing annotation submodule is used to uniformly annotate the first to sixth data with a 10Hz sampling frequency timestamp. The determination submodule is used to use the first data, second data, third data, fourth data, fifth data, and sixth data as multi-dimensional raw data for the entire casting machine process.

[0027] In this embodiment, the frequency of the vibration table is measured in cpm; the internal quality assessment index of the slab is obtained by converting ultrasonic flaw detection data.

[0028] The working principle and beneficial effects of the above technical solution are as follows: By comprehensively acquiring data from each stage, the system can perform quality control over the entire casting machine production process; and it uniformly marks the first to sixth data points with a 10Hz sampling frequency timestamp. This ensures that the data collected from different stages have a unified time reference, facilitating data synchronization and analysis; the unified timestamps enable the system to monitor the operating status of the entire casting machine process in real time.

[0029] Example 3: As Figure 2 As shown, the preprocessing module includes: The data cleaning submodule is used to process the missing and outlier values ​​in the multi-dimensional raw data of the entire casting machine process equipment to obtain the multi-dimensional raw data of the entire casting machine process equipment after data cleaning. The noise reduction submodule is used to reduce the noise of multi-dimensional raw data from the entire casting machine process after data cleaning, and obtain noise-reduced data. The timing synchronization submodule is used to perform linear interpolation alignment of the noise reduction data according to the timestamp, generating a 10Hz timing data stream as preprocessed data.

[0030] The working principle and beneficial effects of the above technical solution are as follows: by introducing wavelet denoising and time synchronization mechanism, the multi-source data are accurately aligned in the time dimension, solving the time sequence breakage problem of the original solution and providing highly reliable input for the anomaly perception model.

[0031] Example 4: A method for constructing a digital twin model, including: Obtain a multi-dimensional training dataset for the entire casting machine process; the training dataset includes equipment parameters and operating parameters for the entire casting machine process. A virtual mapping model is constructed based on the equipment parameters of the entire casting process described in the training dataset and the mapping between the model and the casting equipment. The VAE-GAN-Transformer hybrid model is trained based on the training dataset, and anomaly detection is performed based on the set loss function and training parameters to obtain a full-process anomaly perception model. Based on the improved NSGA-II algorithm, a full-process multi-objective decision-making model is obtained by setting the objective function and constraints, designing the algorithm structure, and iteratively optimizing the output of the Pareto optimal solution set. An initial digital twin model is obtained based on the virtual mapping model, the full-process anomaly perception model, and the full-process multi-objective decision-making model. Obtain a multi-dimensional test dataset covering the entire casting machine process; The initial digital twin model is tested based on the test dataset. When the test results meet the requirements, a trained digital twin model is obtained.

[0032] In this embodiment, the initial digital twin model also includes a model interaction module, which is used to establish a closed-loop interaction between the virtual mapping model, the full-process anomaly perception model, and the full-process multi-objective decision model: the virtual mapping model outputs simulation data to the anomaly perception model every 50ms; the anomaly perception model outputs anomaly probabilities in real time, and when the probability > 0.3, it injects dynamic constraints into the multi-objective decision model; the multi-objective decision model inputs the Pareto solution set into the virtual mapping model for physical feasibility verification every 10 iterations.

[0033] In this embodiment, the test results meet the requirements as follows: when the anomaly detection F1 value is ≥0.92, the physical feasibility of the solution reaches 100%, and the decision delay is ≤500ms.

[0034] The working principle and beneficial effects of the above technical solution are as follows: A virtual mapping model is constructed based on the equipment parameters of the entire casting machine process in the training dataset, mapping them to the casting machine equipment. This model can accurately reproduce the physical structure and characteristics of the casting machine in virtual space. The mapping relationship between the virtual mapping model and the actual casting machine equipment allows the system to reflect the operating status of the actual equipment in the virtual model in real time. A VAE-GAN-Transformer hybrid model is trained based on the training dataset, and anomaly detection is performed based on a set loss function and training parameters, resulting in a full-process anomaly perception model. This hybrid model combines the advantages of VAE (Variational Autoencoder), GAN (Generative Adversarial Network), and Transformer, enabling it to learn complex patterns and rules in the entire casting machine operation data. By setting appropriate loss functions and training parameters, the model can accurately identify the differences between normal and abnormal operating states, thereby accurately detecting anomalies in the entire casting machine process, such as equipment failures and parameter anomalies. The full-process anomaly perception model can issue early warnings at the early stages of anomalies, providing operators with sufficient time to take appropriate measures. For example, when the model detects abnormal fluctuations in the temperature or pressure of the crystallizer, it promptly issues an alarm to remind operators to check the equipment operation, prevent the abnormality from worsening, and reduce production losses. Based on the improved NSGA-II algorithm, by setting the objective function and constraints, designing the algorithm structure, and iteratively optimizing the output Pareto optimal solution set, a full-process multi-objective decision model is obtained.

[0035] Example 5: A method for constructing a virtual mapping model, including: The virtual mapping model includes device-level refined modeling, system-level coupled modeling, and operating condition-level adaptation modeling. Digital atomic models of individual devices are constructed, and geometric modeling, physical behavior modeling and virtual sensor mapping modeling are completed for each device. The device structure, physical characteristics and data acquisition functions of each device are digitally mapped to complete the device-level fine modeling. The physical coupling relationship between devices is constructed based on the device-level atomic model, and the virtual twin is integrated and assembled according to a unified coordinate system to complete the system-level coupling modeling. We acquire production scenarios for various steel grades, establish parametric working condition templates, and set up a dynamic correction mechanism to eliminate the discrepancy between virtual and physical conditions, thus completing working condition-level adaptation modeling.

[0036] In this embodiment, system-level coupling modeling involves constructing physical coupling relationships between devices. Crystallizer-vibration table coupling model: Derivation of the correlation formula: Based on historical data, the coupling relationship between liquid level (h) and amplitude (A) and frequency (f) is fitted, and the formula is as follows: Goodness of fit R² ≥ 0.95; Coupling parameter settings: Implemented in Unity3D through script binding—when the amplitude of the virtual vibration table is adjusted from 2mm to 2.5mm, the liquid level of the virtual crystallizer is automatically updated according to the formula 780 + 0.6 × 2.5 - 0.02 × 150 = 780 + 1.5 - 3 = 778.5mm; Sector-roller coupling model: Synchronization speed formula: Roller speed ( = 1.05 × sector segment pulling speed ( (Reserve buffer space for slab), for example: When =1.2m / min, =1.26 m / min; Synchronous control logic: Establish signal linkage between the virtual sector speed encoder and the roller drive motor. When the speed changes by ±0.1 m / min, Synchronous change ±0.105m / min, response delay ≤50ms; Virtual twin integration and assembly: Coordinate system one: Establish a global coordinate system with the crystallizer center as the origin (X-axis: casting machine length direction, Y-axis: width direction, Z-axis: height direction). All equipment models are imported according to the actual installation coordinates, with alignment accuracy ≤ ±2mm; Physics engine integration: Export the physical behavior model simulated by ANSYS as a dynamic parameter file, import it into the NVIDIA PhysX physics engine, and bind the simulation parameters to the virtual model components through the parameter mapping interface.

[0037] In this embodiment, working condition-level adaptation modeling is used, and working conditions are classified into three core working conditions based on steel type (low carbon steel LC, high carbon steel HC, and special steel SS). Each working condition has a dedicated physical parameter template table (Table 1). Table 1 Dynamic correction mechanism: The deviation between the measured value of the physical sensor and the output of the virtual model is calculated every 5 seconds. When the deviation of liquid level or temperature is greater than 5%, the thermophysical parameters of the virtual model are automatically adjusted.

[0038] Template calling logic: Set the working condition switching button in the virtual twin control interface. When low carbon steel LC is selected, the template parameters are automatically called - the steel temperature is set to 1550℃, the pulling speed is set to 1.3m / min, the vibration table amplitude is set to 2.2mm, and the thermophysical parameters of the physical model are updated synchronously (such as the latent heat of solidification 270kJ / kg).

[0039] The working principle and beneficial effects of the above technical solution are as follows: Digital atomic models of individual devices are constructed, and geometric modeling, physical behavior modeling, and virtual sensor mapping modeling are completed for each device. Geometric modeling accurately presents the shape, size, and structure of the device, allowing operators to intuitively understand the spatial layout. Physical behavior modeling simulates the operating patterns of the device under different working conditions, such as heat conduction and mechanical performance, which helps to deepen the understanding of the device's working principle. Virtual sensor mapping modeling maps sensor data from the actual device to the virtual model, realizing real-time monitoring of the device's operating status. Through this comprehensive digital mapping, the structure, physical characteristics, and data acquisition functions of the device can be accurately restored, providing an accurate foundation for subsequent analysis and optimization. Physical coupling relationships between devices are constructed based on the device-level atomic model and modeled according to a unified coordinate system. The system completes the integration and assembly of virtual twins; this clearly presents the interactions and influences between various devices in the casting machine system, such as the flow of molten steel and energy transfer between different devices; by understanding the physical coupling relationships between devices, operators can better grasp the operating rules of the entire casting machine system, predict the synergistic effects between devices, and avoid production problems caused by incoordination between devices; it acquires production scenarios for various steel grades and establishes parametric chemical condition templates; different steel grades have different physical and chemical properties, and their production processes and equipment operation requirements for casting machines are also different; the parametric chemical condition templates can adjust the parameters and operating modes of the virtual model according to different steel grades and production scenarios, so that the virtual model can accurately simulate the operation of the casting machine under various working conditions; this makes the virtual mapping model more versatile and adaptable, and can meet the production needs of different steel grades.

[0040] Example 6: Construct digital atomic models for individual devices, completing geometric modeling, physical behavior modeling, and virtual sensor mapping modeling for each device. Digitally map the device structure, physical characteristics, and data acquisition functions of each device to achieve refined device-level modeling, including: The geometric modeling uses SolidWorks for mechanical structure modeling and Unity3D for visualization rendering optimization; the casting machine is divided into 6 core pieces of equipment: tundish car, crystallizer, vibrating table, sector section, slab inspection device, and roller conveyor. Each type of equipment is constructed in layers according to core components, auxiliary components, and connecting structures. The physical behavior modeling includes a molten steel flow model, a solidification temperature field model, and a shaking table dynamics model. The molten steel flow model was constructed using the SPH smoothed particle fluid dynamics algorithm; The solidification temperature field model was constructed using the FEM finite element algorithm. The governing equations were established based on thermophysical parameters, and transient thermal analysis was employed. The solver was the conjugate gradient method. The dynamic model of the shaking table adopts the multibody dynamics algorithm. The crystallizer and the shaking table are rigidly connected. The motion equation is established based on the hydraulic cylinder driving force model and solved using the Newmark-β method. The virtual sensor mapping modeling involves placing virtual sensors at the locations of physical sensors corresponding to the virtual devices, with each sensor number corresponding to a physical sensor. Key parameters such as sensor sampling frequency are set, and the output format is sensor number-timestamp-value.

[0041] In this embodiment, the equipment-level refined geometric modeling (tools: SolidWorks + Unity3D) involves the breakdown of key equipment: the casting machine is divided into 6 core pieces of equipment (mid-ladle car, crystallizer, vibrating table, sector section, slab inspection device, and roller conveyor). Each type of equipment is modeled in layers: core component - auxiliary component - connecting structure. Crystallizer: core component → auxiliary component → connecting structure, modeling error ≤ 0.1mm; Vibrating table: core component → auxiliary component → connecting structure, modeling error ≤ 0.2mm; Model export and format conversion: after SolidWorks modeling is completed, it is exported as .STEP format, imported into Unity3D, and the MeshSimplify plugin is used to optimize the number of faces to ensure a rendering frame rate ≥ 30fps.

[0042] Physical behavior modeling (tools: ANSYS Mechanical + NVIDIA PhysX) Steel flow model (SPH algorithm): Particle parameter settings: Particle number = 50000, Particle radius = 2mm, Density = 7850kg / m³, Viscosity coefficient = 6.5×10 -3 Pa·s (molten steel at 1550℃), surface tension coefficient = 1.8 N / m; boundary conditions: the inner wall of the crystallizer is set as a no-slip boundary, the inlet velocity of molten steel = 0.5 m / s (matching the actual pouring speed), and the outlet is connected to the inlet of the fan-shaped section; simulation output: the velocity field and liquid level change curve of molten steel in the crystallizer. Solidification temperature field model (FEM algorithm): mesh generation: the copper plate of the crystallizer adopts a hexahedral structured mesh, and the molten steel region adopts a tetrahedral unstructured mesh; thermophysical parameter assignment: thermal conductivity of copper plate λ = 380 W / (m·K), specific heat capacity c = 385 J / (kg·K); latent heat of solidification of molten steel Q = 270 kJ / kg, liquid thermal conductivity =45W / (m·K), thermal conductivity of solid state =50W / (m·K); Governing equations and solutions: Transient thermal analysis is used, and the governing equations are as follows: (ρ is density, T is temperature, t is time), time step = 0.5s, the solver uses the conjugate gradient method, and outputs the curve of the thickness of the solidified layer of molten steel changing with time (1 data point is output every 1s).

[0043] Vibration table dynamics model (multibody dynamics algorithm): Degrees of freedom setting: The vibration table is set to 6 degrees of freedom of translation + rotation, and the crystallizer is rigidly connected to the vibration table (constraining 6 degrees of freedom); Driving force model: Cylinder driving force F(t) = P × S (P is the cylinder pressure (Pa), S is the piston area = π × (0.12m / 2)² = 0.0113m²), the pressure P is based on the actual production fluctuation law (±5 × 10 5 Pa) settings; solving the equations of motion: the equations are (m=5000kg, x is displacement, c=1×10) 4 N·s / m, k=5×10 5 N / m), the solver uses the Newmark-β method (β=0.25, γ=0.5), and outputs amplitude and frequency variation curves over time (sampling frequency 10Hz). Virtual sensor mapping modeling and sensor position matching: Virtual sensors are placed at the corresponding physical sensor positions on the virtual equipment (such as thermocouples on the crystallizer copper plate, pressure interfaces of the vibration table cylinder), with their numbers corresponding one-to-one with the physical sensors (e.g., virtual sensor V-T1 corresponds to physical sensor P-T1); sampling parameter synchronization: the sampling frequency (1Hz for key parameters, 5Hz for general parameters) and accuracy (temperature ±1℃, pressure ±0.1MPa) of the virtual sensors are completely consistent with those of the physical sensors, and the data output format is sensor number-timestamp-value (e.g., V-T1-20241001100000-1550℃).

[0044] The working principle and beneficial effects of the above technical solution are as follows: Through digital modeling, equipment can be designed and optimized in a virtual environment, avoiding frequent physical prototyping and modifications in the traditional design process, greatly shortening the design cycle and reducing design costs; physical behavior modeling can accurately simulate the operation process of the equipment, helping engineers to deeply understand the physical characteristics and operating laws of the equipment; through the analysis and optimization of the model, the optimal operating parameters can be determined, improving the production efficiency and product quality of the equipment; virtual sensor mapping modeling can acquire the operating data of the equipment in real time and compare and analyze it with the data of the actual equipment; once abnormal data is detected, fault diagnosis and prediction can be carried out in a timely manner, and measures can be taken in advance to reduce equipment downtime and maintenance costs.

[0045] Example 7: A method for constructing a full-process anomaly detection model, including: A hybrid VAE-GAN-Transformer structure was built based on TensorFlow 2.15. The VAE encoder was used for feature extraction and dimensionality reduction, the Transformer decoder improved the ability to capture temporal features, and the discriminator was used to distinguish anomalies. Construct the total loss function; Obtain the training dataset, use the AdamW optimizer and cosine annealing learning rate to train the model, set an early stopping mechanism to avoid overfitting, and obtain a full-process anomaly perception model.

[0046] In this embodiment, a VAE-GAN-Transformer hybrid model structure is adopted, built based on TensorFlow 2.15: The generator consists of a VAE encoder and a Transformer decoder. The VAE encoder comprises an input layer (fully connected, 10 nodes), a feature extraction layer 1 (fully connected, 256 nodes, ReLU activation, He normal initialization, L2 regularization λ=0.001), a feature extraction layer 2 (fully connected, 128 nodes, ReLU activation, He normal initialization, L2 regularization λ=0.001), a mean layer (fully connected, 64 nodes, Xavier normal initialization), and a standard deviation layer (fully connected, 64 nodes, Softplus activation, Xavier normal initialization). It utilizes reparameterization techniques. , The latent vector is sampled; the Transformer decoder consists of an input layer (fully connected, 64 nodes), a feature mapping layer (fully connected, 128 nodes, GELU activation, He normal initialization), Transformer encoder 1 (self-attention layer, 128 nodes, GELU activation, Xavier normal initialization, multi-head attention), Transformer encoder 2 (self-attention layer, 128 nodes, GELU activation, Xavier normal initialization, multi-head attention), and an output layer (fully connected, 10 nodes, Sigmoid activation, Xavier normal initialization). Attention is calculated as follows: ; Each is a 128-dimensional matrix; Discriminator: Includes input layer (fully connected, 10 nodes), feature layer 1 (fully connected, 256 nodes, GELU activation, He normal initialization, dropout ratio 0.2), feature layer 2 (fully connected, 128 nodes, GELU activation, He normal initialization, dropout ratio 0.2), temporal fusion layer (fully connected, 64 nodes, GELU activation, Xavier normal initialization), and output layer (fully connected, 1 node, Sigmoid activation, Xavier normal initialization). Temporal fusion uses average pooling to compress the 10-dimensional features (50 data points) from the first 5 seconds into 10 dimensions. Training model: Each sample contains 10-dimensional features + 1-dimensional label (0=normal, 1=abnormal); the optimizer is set to AdamW, the learning rate is scheduled using cosine annealing, the batch size is 64, the iteration is 200 rounds, and an early stopping mechanism is set. Anomaly detection: Input real-time 10-dimensional features. If MSE≤0.06 and discriminator score≤0.25, it is judged as normal; otherwise, it is judged as abnormal. The 10-dimensional features are defined as: crystallizer liquid level (mm), copper plate temperature (°C), vibration table amplitude (mm), vibration table frequency (cpm), fan-shaped section pulling speed (m / min), roller pressure (MPa), billet surface temperature (°C), billet width (mm), roller speed (rpm), and cooling water flow rate (m³ / h).

[0047] The working principle and beneficial effects of the above technical solution are as follows: The VAE-GAN-Transformer hybrid structure combines the feature extraction capability of VAE, the temporal feature capture capability of Transformer, and the adversarial training mechanism of GAN, enabling more accurate identification of abnormal patterns in the entire process data. Whether it's a small abnormal fluctuation or a complex combination of abnormalities, the model exhibits high detection accuracy. The use of the Transformer decoder allows the model to handle data with complex temporal relationships effectively. In many real-world scenarios, the occurrence of anomalies is often closely related to the historical information and temporal order of the data; this model can effectively capture this information, improving the reliability of anomaly detection. By constructing a total loss function, employing regularization methods, and applying an early stopping mechanism, the model has strong generalization ability. It maintains good performance across different datasets and scenarios, reducing the risk of overfitting and improving the model's practicality.

[0048] Example 8: A method for constructing a full-process multi-objective decision-making model, including: Define a multi-objective function; Set constraints based on equipment parameters and process parameters respectively: The model structure is constructed using an improved NSGA-II algorithm; The population is initialized using Latin hypercube sampling. Based on twin simulation evaluation, non-dominated sorting and crowding calculation, selection-crossover-mutation operation, and Pareto optimal solution set is iteratively output to obtain a full-process multi-objective decision model.

[0049] In this embodiment, the decision objective and constraints are constructed. Multi-objective function: Minimize slab defect rate (Q), minimize energy consumption (E), maximize casting speed (Q / E). The function is: Parameter definition: Q is the slab defect rate (%), and E is the actual energy consumption (kWh / t). =80 is the baseline energy consumption. This represents the actual pulling speed (m / min). =1.2 is the reference pulling speed; Weight determination: Based on the analytic hierarchy process (AHP), five experts were invited to score the quality, energy consumption, and efficiency, and the final weights were determined to be 0.5:0.3:0.2.

[0050] Constraints: Equipment parameter constraints: Vibration table amplitude A∈[1,4]mm, frequency f∈[100,200]Hz; sector segment tension speed ∈[0.6,1.5]m / min; Process constraints: Cooling water flow rate F∈[50,100]m³ / h.

[0051] Improved NSGA-II algorithm structure design Population initialization (generation 1): Population size: 50 decision variables (decision variable = [A, f, ...) ,F】); Initialization method: Latin hypercube sampling is used to ensure that the sample uniformly covers the constraint space. For example, amplitude A is uniformly generated with 50 values ​​within [1,4] mm, and the same applies to other variables.

[0052] Twin simulation evaluation: Simulation input: Input each set of decision variables into the virtual mapping model, and simulate for 10 seconds (covering one pulling cycle). Simulation output: Extract Q (defect rate, output by the virtual slab inspection device) and E (energy consumption, output by the virtual electricity meter) after simulation. (Actual pulling speed, output by the virtual encoder), generate the target vector: for each set of decision variables, generate the vector F = [Q, E, - ] Non-dominated sorting and crowding calculation: Non-dominated ranking: Based directly on the objective vector F, Pareto dominance relationships are determined for 50 sets of decision variables, and non-dominated levels are divided (Level 1: the optimal solution that is not dominated by other solutions, Level 2: the solution that is dominated only by Level 1, and so on). Crowding level calculation: The formula is: (m=3 targets, Let m be the objective value of the i-th solution. / (where the maximum / minimum value is the m-th objective), the greater the crowding, the better the diversity of solutions.

[0053] Selection-crossover-mutation operation: Selection: A tournament selection method is used, randomly selecting 2 sets of solutions, and selecting the solution with higher non-dominated level and greater crowding to enter the next generation; Crossover: Simulated binary crossover (SBX) is used, with a crossover probability of 0.8 and a distribution index of 20. Mutation: Multinomial mutation is used, with a mutation probability of 0.1 and a distribution index of 10.

[0054] Algorithm iteration and optimal solution output: Iteration process: Number of iterations: 50 rounds, with 50 new solutions generated in each round. After merging with the parent solution, the 50 optimal solutions are selected for the next round through non-dominated sorting and crowding calculation. Simulation feedback: After every 10 iterations, the current Pareto optimal solution (the first layer of non-dominated solution) is input into the physical-virtual mapping model for verification to ensure that all solutions meet the physical feasibility constraints.

[0055] Optimal solution output: Pareto optimal solution set: After 50 iterations, output the first-level non-dominated solutions (approximately 10-15 sets); Optimal solution selection: Prioritize the smallest Q-value + the second largest. The rule automatically selects one solution (e.g., Q=0.3%, E=82kWh / t). =1.3m / min), which was used as the final decision-making option.

[0056] The working principle and beneficial effects of the above technical solution are as follows: By setting a multi-objective function, the model can simultaneously consider multiple conflicting objectives. For example, in a production process, it may simultaneously pursue cost minimization, output maximization, and quality optimization. This makes decision-making no longer limited to the optimization of a single objective, but more in line with the needs of actual complex systems, and can provide decision-makers with more comprehensive and integrated solutions. Constraints are set based on equipment parameters and process parameters respectively to ensure the feasibility and practicality of the model. For example, the maximum production capacity of the equipment, the temperature and pressure limits of the process, etc., can all be incorporated into the model as constraints. The decision-making scheme obtained by the model constructed in this way is within the range allowed by actual production conditions, avoiding theoretically feasible but practically unfeasible schemes. This model construction method, by comprehensively considering multiple objectives, combining actual constraints, and employing efficient algorithms and reasonable iteration mechanisms, can provide decision-makers with a set of high-quality Pareto optimal solutions. These solutions achieve a good balance among multiple objectives, helping decision-makers make more scientific and reasonable decisions.

[0057] Example 9: As Figure 3 As shown, the control module includes: The parsing submodule is used to parse the target decision scheme to obtain the target control command; The execution submodule is used to perform intelligent control of the entire casting machine process based on the target control command.

[0058] The working principle and beneficial effects of the above technical solution are as follows: A detailed analysis of the target decision-making scheme transforms what might otherwise be a macroscopic and abstract decision into specific target control instructions. This avoids potential deviations in information transmission and understanding, ensuring that all equipment in the casting machine process receives precise and clear operational instructions, thereby improving the accuracy of equipment task execution. Intelligent control of the entire casting machine process based on target control instructions enables real-time monitoring and response to the equipment. When abnormal situations or parameter changes occur during casting machine production, the system can automatically adjust control instructions according to preset rules and algorithms, ensuring that the equipment is always in optimal operating condition.

[0059] Example 10: Constructing the total loss function, including: ; in, Represents the total loss function; This represents the VAE reconstruction loss; This represents the VAEKL divergence loss; Indicates GAN adversarial loss; Indicates cross-device collaboration loss; , , , These are the weighting coefficients, all of which are greater than 0 and have a sum of 1.

[0060] In this embodiment, ;in, Indicates the number of samples; Represents the weight coefficient of the j-th feature; This represents the original feature value of the j-th dimension of the i-th sample; Let represent the reconstructed feature value of the j-th dimension of the i-th sample.

[0061] In this embodiment, ; This represents the variance value of the i-th sample output by the encoder; This represents the mean of the i-th sample output by the encoder.

[0062] In this embodiment, ; This represents the probability that the discriminator classifies the i-th real sample; This represents the probability that the discriminator will classify the i-th generated sample.

[0063] In this embodiment, ; This represents the measured liquid level in the physical crystallizer; This represents the simulation value of the virtual mapping model.

[0064] The working principle and beneficial effects of the above technical solutions are as follows: different loss functions constrain and optimize the model from different perspectives; VAE reconstruction loss enables the model to learn the specific features and details of the data, VAEKL divergence loss allows the model to learn the potential distribution of the data, GAN adversarial loss can enhance the realism and diversity of the data generated by the model, and cross-device collaborative loss considers the consistency and collaboration of data in a multi-device environment; by combining these loss functions, the model can learn the features of the data from multiple dimensions, avoid model overfitting, and thus improve the model's generalization ability on different datasets and tasks.

[0065] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A digital twin-based intelligent control system for casting machines, characterized in that, include: The acquisition module is used to acquire multi-dimensional raw data of the entire casting machine process equipment; The preprocessing module is used to preprocess the multi-dimensional raw data of the entire casting machine process equipment to obtain preprocessed data; Modules for building digital twin models; The prediction module is used to input the preprocessed data into the digital twin model for prediction and output the target decision scheme; The control module is used to intelligently control the entire process equipment of the casting machine based on the target decision scheme.

2. The intelligent control system for casting machines based on digital twins as described in claim 1, characterized in that, The acquisition module includes: The first acquisition submodule is used to acquire the ladle car position / speed, molten steel temperature and molten steel flow rate data during the molten steel pouring process, as the first data; The second acquisition submodule is used to acquire data on crystallizer liquid level, copper plate temperature, and cooling water flow / pressure during the crystallizer solidification process, as second data. The third acquisition submodule is used to acquire the vibration table amplitude / frequency and vibration table cylinder pressure data in the vibration table coordination process, as the third data. The fourth acquisition submodule is used to acquire data on the fan-shaped section pulling speed, roller pressure / speed and billet temperature in the fan-shaped section conveying process, as the fourth data. The fifth acquisition submodule is used to acquire data on slab surface defects, internal density / defects, and slab dimensions during the slab inspection process, as the fifth data. The sixth acquisition submodule is used to acquire the roller conveyor speed and slab position / attitude data in the roller conveyor transfer process, as the sixth data; The timing annotation submodule is used to uniformly annotate the first to sixth data with a 10Hz sampling frequency timestamp. The determination submodule is used to use the first data, second data, third data, fourth data, fifth data, and sixth data as multi-dimensional raw data for the entire casting machine process.

3. The intelligent control system for casting machines based on digital twins as described in claim 1, characterized in that, The preprocessing module includes: The data cleaning submodule is used to process the missing and outlier values ​​in the multi-dimensional raw data of the entire casting machine process equipment to obtain the multi-dimensional raw data of the entire casting machine process equipment after data cleaning. The noise reduction submodule is used to reduce the noise of multi-dimensional raw data from the entire casting machine process after data cleaning, and obtain noise-reduced data. The timing synchronization submodule is used to perform linear interpolation alignment of the noise reduction data according to the timestamp, generating a 10Hz timing data stream as preprocessed data.

4. The intelligent control system for casting machines based on digital twins as described in claim 1, characterized in that, Methods for constructing digital twin models include: Obtain a multi-dimensional training dataset for the entire casting machine process; the training dataset includes equipment parameters and operating parameters for the entire casting machine process. A virtual mapping model is constructed based on the equipment parameters of the entire casting process described in the training dataset and the mapping between the model and the casting equipment. The VAE-GAN-Transformer hybrid model is trained based on the training dataset, and anomaly detection is performed based on the set loss function and training parameters to obtain a full-process anomaly perception model. Based on the improved NSGA-II algorithm, a full-process multi-objective decision-making model is obtained by setting the objective function and constraints, designing the algorithm structure, and iteratively optimizing the output of the Pareto optimal solution set. An initial digital twin model is obtained based on the virtual mapping model, the full-process anomaly perception model, and the full-process multi-objective decision-making model. Obtain a multi-dimensional test dataset covering the entire casting machine process; The initial digital twin model is tested based on the test dataset. When the test results meet the requirements, a trained digital twin model is obtained.

5. The intelligent control system for casting machines based on digital twins as described in claim 4, characterized in that, Methods for constructing virtual mapping models include: The virtual mapping model includes device-level refined modeling, system-level coupled modeling, and operating condition-level adaptation modeling. Digital atomic models of individual devices are constructed, and geometric modeling, physical behavior modeling and virtual sensor mapping modeling are completed for each device. The device structure, physical characteristics and data acquisition functions of each device are digitally mapped to complete the device-level fine modeling. The physical coupling relationship between devices is constructed based on the device-level atomic model, and the virtual twin is integrated and assembled according to a unified coordinate system to complete the system-level coupling modeling. We acquire production scenarios for various steel grades, establish parametric working condition templates, and set up a dynamic correction mechanism to eliminate the discrepancy between virtual and physical conditions, thus completing working condition-level adaptation modeling.

6. The intelligent control system for casting machines based on digital twins as described in claim 5, characterized in that, Digital atomic models of individual devices are constructed, and geometric modeling, physical behavior modeling, and virtual sensor mapping modeling are completed for each device. The device structure, physical characteristics, and data acquisition functions of each device are digitally mapped, completing refined device-level modeling, including: The geometric modeling uses SolidWorks for mechanical structure modeling and Unity3D for visualization rendering optimization; the casting machine is divided into 6 core pieces of equipment: tundish car, crystallizer, vibrating table, sector section, slab inspection device, and roller conveyor. Each type of equipment is constructed in layers according to core components, auxiliary components, and connecting structures. The physical behavior modeling includes a molten steel flow model, a solidification temperature field model, and a shaking table dynamics model. The molten steel flow model was constructed using the SPH smoothed particle fluid dynamics algorithm; The solidification temperature field model was constructed using the FEM finite element algorithm. The governing equations were established based on thermophysical parameters, and transient thermal analysis was employed. The solver was the conjugate gradient method. The dynamic model of the shaking table adopts the multibody dynamics algorithm. The crystallizer and the shaking table are rigidly connected. The motion equation is established based on the hydraulic cylinder driving force model and solved using the Newmark-β method. The virtual sensor mapping modeling involves placing virtual sensors at the locations of physical sensors corresponding to the virtual devices, with each sensor number corresponding to a physical sensor. Key parameters such as sensor sampling frequency are set, and the output format is sensor number-timestamp-value.

7. The intelligent control system for casting machines based on digital twins as described in claim 4, characterized in that, The construction method of the full-process anomaly detection model includes: A hybrid VAE-GAN-Transformer structure was built based on TensorFlow 2.

15. The VAE encoder was used for feature extraction and dimensionality reduction, the Transformer decoder improved the ability to capture temporal features, and the discriminator was used to distinguish anomalies. Construct the total loss function; Obtain the training dataset, use the AdamW optimizer and cosine annealing learning rate to train the model, set an early stopping mechanism to avoid overfitting, and obtain a full-process anomaly perception model.

8. The intelligent control system for casting machines based on digital twins as described in claim 4, characterized in that, The method for constructing a full-process multi-objective decision-making model includes: Define a multi-objective function; Set constraints based on equipment parameters and process parameters respectively: The model structure is constructed using an improved NSGA-II algorithm; The population is initialized using Latin hypercube sampling. Based on twin simulation evaluation, non-dominated sorting and crowding calculation, selection-crossover-mutation operation, and Pareto optimal solution set is iteratively output to obtain a full-process multi-objective decision model.

9. The intelligent control system for casting machines based on digital twins as described in claim 1, characterized in that, The control module includes: The parsing submodule is used to parse the target decision scheme to obtain the target control command; The execution submodule is used to perform intelligent control of the entire casting machine process based on the target control command.

10. The intelligent control system for casting machines based on digital twins as described in claim 7, characterized in that, Construct the total loss function, including: ； in, Represents the total loss function; This represents the VAE reconstruction loss; This represents the VAEKL divergence loss; Indicates GAN adversarial loss; Indicates cross-device collaboration loss; , , , These are the weighting coefficients, all of which are greater than 0 and have a sum of 1.

Images