A flow field real-time prediction method combining Latin hypercube sampling

Abstract

The application discloses a kind of flow field real-time prediction methods of combining Latin hypercube sampling, it is related to the field of foundry production.Provide a kind of fast prediction model of continuous casting three-dimensional crystallizer flow field combining Latin hypercube sampling and machine learning technology.The model aims to fully cover the parameter space by Latin hypercube sampling method, excavate the potential mapping relationship between key process parameters and the distribution of molten steel flow field in crystallizer, realize the fast and accurate prediction of flow field form.The model has the characteristics of high precision, low computing resource demand and fast response, can effectively overcome the shortcomings of long time-consuming and large resource consumption of traditional numerical simulation technology.Meanwhile, it makes up for the defects of traditional sampling methods such as poor representativeness of randomly sampled samples and exponential growth of sample size in high-dimensional space under uniform sampling.

Description

Technical Field

[0001] This invention relates to the field of casting production, and in particular to a method for real-time prediction of flow fields using Latin hypercube sampling. Background Technology

[0002] In continuous casting production, the violent fluctuations of molten steel within the crystallizer often lead to instability of the free liquid surface and related defects such as inclusions. This is the core cause of billet surface quality problems and has long been a key challenge for major steel companies in controlling steel surface quality. The occurrence of these defects is closely related to the flow behavior and surface fluctuations of the molten steel within the crystallizer, with flow velocity and surface fluctuation amplitude being key parameters reflecting flow field stability. In actual production, some key process parameters, such as casting speed and argon blowing rate, significantly affect the morphology of the molten steel flow field within the crystallizer. However, the molten steel flow field inside the crystallizer is invisible and highly nonlinear, making online velocity measurement extremely difficult.

[0003] Currently, numerical simulation technology based on software such as Fluent has been widely used in the flow field analysis of continuous casting molds. It can identify high-speed flow regions by simulating the flow field distribution of the free surface under different process parameters, thereby predicting slag entrapment risks and optimizing process parameters to improve product quality. However, numerical simulation has a long calculation cycle, making it difficult to meet the needs of real-time online control. Existing methods can achieve a certain degree of online monitoring, but their response speed is slow, computation time is long, and resource consumption is high, making it difficult to truly meet real-time requirements.

[0004] In recent years, machine learning technology has achieved remarkable results in various fields such as facial recognition and autonomous driving, demonstrating its ability to automatically learn patterns and rules from large-scale data. Leveraging its superior performance in object detection, recognition, and prediction, machine learning has been gradually applied to industrial production processes, helping to improve production efficiency and product quality. Its powerful nonlinear mapping capabilities and efficient data processing advantages provide a new technical path for real-time prediction in continuous casting processes. The training effect of machine learning models depends on the quality and representativeness of the training samples. Compared to traditional uniform or random sampling, Latin hypercube sampling can achieve uniform coverage of the entire parameter space with a smaller number of samples, effectively avoiding sample point clustering or missing data, thus obtaining a statistically representative training dataset at a lower computational cost.

[0005] Therefore, those skilled in the art are dedicated to developing a real-time flow field prediction method combining Latin hypercube sampling. This paper proposes a rapid flow field prediction model for continuous casting three-dimensional crystallizers that combines Latin hypercube sampling with machine learning techniques. This model aims to fully cover the parameter space using the Latin hypercube sampling method, uncovering the potential mapping relationship between key process parameters and the distribution of molten steel flow field within the crystallizer, thereby achieving rapid and accurate prediction of the flow field morphology. The model features high accuracy, low computational resource requirements, and rapid response, effectively overcoming the shortcomings of traditional numerical simulation techniques, such as long processing time and high resource consumption. It also compensates for the deficiencies of traditional sampling methods, such as poor representativeness of random samples and the exponential growth of sample size in high-dimensional spaces due to uniform sampling. Summary of the Invention

[0006] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is that the real-time flow field prediction technology is time-consuming and resource-intensive.

[0007] To achieve the above objectives, this invention provides a real-time flow field prediction method that combines Latin hypercube sampling, including using machine learning technology to mine the nonlinear mapping relationship between process parameters and the flow field behavior of molten steel, and constructing a prediction model.

[0008] Furthermore, a dataset covering various combinations of process parameters is generated in advance using numerical simulation data; the machine learning model is trained on the dataset to learn the nonlinear mapping relationship between parameters and results; after training, in actual production, the model no longer performs any physical iterative calculations, but directly maps the predicted results through computation.

[0009] Furthermore, the Latin hypercube sampling method is adopted to stratify each process parameter dimension and randomly select sample points within each layer to ensure that the sampling points are uniformly distributed in the multidimensional parameter space.

[0010] Furthermore, the Latin hypercube sampling method divides the value range of each process parameter into several equally probable intervals, extracts a sample point in each interval, and then randomly combines the sample points of different parameters to form a complete sampling scheme.

[0011] Furthermore, it includes the following steps: Step 1: Construct the process parameter space using the Latin hypercube sampling method; Step 2: Numerical simulation batch calculation and flow field data export; Step 3: Data preprocessing and dataset partitioning; Step 4: Train and optimize the machine learning model to build a prediction model; Step 5: Predict the flow field conditions under different process parameters.

[0012] Furthermore, the process parameters include casting speed, sprue immersion depth, and argon blowing rate.

[0013] Furthermore, in step 2, a numerical simulation model of the crystallizer is used to perform batch simulation calculations on the combination of process parameters corresponding to each sampling point; after the calculation is completed, the velocity distribution data of the flow field at the free liquid surface of the molten steel in the crystallizer under the conditions of each sampling point are exported to form the original high-dimensional dataset.

[0014] Furthermore, in step 3, the flow field velocity data is preprocessed, including data cleaning and outlier removal; a normalization method is used to map the data to a uniform scale to eliminate the influence between different units; the normalized data is randomly shuffled to eliminate the interference of sample order on model training.

[0015] Furthermore, in step 4, the process parameters are used as input, the flow field velocity is used as output, the fully connected neural network is used as the backbone network, the absolute average error is used as the loss function, and the training set is used to train the machine learning model. After training, the model is optimized within the Optuna framework.

[0016] Furthermore, the optimization includes the activation function of the neural network, the number of neurons, the number of hidden layers in the neural network, and the type of optimizer.

[0017] Traditional numerical simulations in continuous casting production are time-consuming, typically taking hours or even days, making them unsuitable for the real-time demands of changing process conditions in actual production. This invention utilizes machine learning technology to uncover the nonlinear mapping between process parameters and molten steel flow field behavior, constructing a rapid prediction model to replace time-consuming numerical simulations and achieve millisecond-level output, meeting the timeliness requirements of production sites. This invention pre-generates a dataset covering various combinations of process parameters using numerical simulation data. The machine learning model is trained on this dataset, learning the complex nonlinear mapping between parameters and results. After training, in actual production, the model no longer performs any physical iteration calculations but directly maps the prediction results to milliseconds through simple computation. This invention utilizes machine learning technology to uncover the nonlinear mapping between process parameters and molten steel flow field, constructing a rapid prediction model to replace time-consuming traditional numerical simulations, reducing prediction time from hours or even days to milliseconds. It can output flow field cloud map results instantly as process parameters change, enabling online monitoring and rapid feedback of the continuous casting process, effectively improving production efficiency, stabilizing molten steel flow rate, and reducing billet scrap rate.

[0018] Traditional sampling methods, such as random sampling and uniform sampling, result in exponential growth of sample size in high-dimensional spaces, leading to high computational resource demands for numerical simulations. This invention employs the Latin hypercube sampling method for experimental design of key process parameters. By stratifying each parameter dimension and randomly selecting sample points within each layer, it ensures a uniform distribution of sampling points across the multidimensional parameter space. This Latin hypercube sampling method divides the value range of each process parameter into several equally probable intervals, extracting one sample point from each interval. Then, the sampling points from different parameters are randomly combined to form a complete sampling scheme. This method achieves uniform coverage of the high-dimensional parameter space with a fixed sample size, and the sample size grows linearly with the parameter dimension rather than exponentially, significantly reducing the number of samples required to achieve equivalent representativeness. Compared to traditional sampling methods, the Latin hypercube sampling method used in this invention provides better uniformity of parameter space coverage with the same sample size, or significantly reduces the required sample size while achieving the same coverage quality. This effectively reduces the computational resource consumption and time cost of subsequent numerical simulations, providing an efficient and economical sampling scheme for constructing high-quality training datasets.

[0019] Compared with the prior art, the present invention has the following obvious substantive features and significant advantages: 1. This invention employs Latin hypercube sampling to efficiently cover the space of key process parameters, obtaining a statistically representative training dataset with a small sample size. This effectively avoids the problem of the sample size expanding rapidly with increasing dimension in high-dimensional parameter spaces, which is faced by traditional sampling methods.

[0020] 2. Latin hypercube sampling significantly reduces the number of numerical simulations required to achieve equivalent sample representativeness, greatly reducing computational resource consumption and time costs during the data generation phase. The trained machine learning model is lightweight, and a single prediction can be completed within milliseconds. It does not rely on high-performance computing servers and can be directly deployed on ordinary industrial computers in the production site, avoiding data transmission latency and remote dependency issues, making it particularly suitable for production line environments with limited computing resources.

[0021] 3. By collecting key process parameters in real time, the model can instantly output the distribution of the molten steel flow field and the stability assessment results of the free liquid surface within the crystallizer, providing quantitative basis for operators to adjust process parameters and prevent flow field-related defects. This solution has advantages such as fast response speed, low resource requirements, and ease of deployment, and can be widely applied to various continuous casting production lines, possessing good promotional value and industrialization prospects.

[0022] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0023] Figure 1 This is a flowchart of a machine learning model according to a preferred embodiment of the present invention; Figure 2 This is a scatter plot comparison of machine learning model results and numerical simulation calculation results from a preferred embodiment of the present invention. Figure 3 This is a cloud map comparison of machine learning model results and numerical simulation calculation results in a preferred embodiment of the present invention. Detailed Implementation

[0024] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0025] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0026] like Figure 1 The diagram shows a flowchart of the machine learning method constructed in this invention. The specific steps of this invention are as follows: 1. Construct the critical process parameter space using the Latin hypercube sampling method: The Latin hypercube sampling method was used to design experiments for key process parameters such as casting speed, sprue immersion depth, and argon blowing rate. The value range of each parameter was divided into several equally probable intervals. A sample point was randomly selected within each interval, and then the sample points for different parameters were randomly combined to generate a sampling scheme covering the entire parameter space. In this embodiment, the sampling range for the key process parameters was: casting speed 0.5 - 1.7 m / min, sprue immersion depth 130 - 270 mm, and argon blowing rate 3 - 17 L / min. A total of 1000 samples were collected.

[0027] 2. Numerical simulation batch calculation and flow field data export: Using an experimentally validated numerical simulation model of the crystallizer, batch simulation calculations were performed on the process parameter combinations corresponding to each sampling point based on the previously generated sampling scheme. After the calculations were completed, the velocity distribution data of the flow field at the free surface of the molten steel in the crystallizer under each set of sampling point conditions were exported to form the original high-dimensional dataset.

[0028] 3. Data preprocessing and dataset partitioning: The exported flow velocity data is preprocessed, including data cleaning and outlier removal. Normalization is used to map the data to a uniform scale, eliminating the influence of different units of measurement. The normalized data is then randomly shuffled to eliminate potential interference from sample order during model training. The dataset is divided into training and test sets in an 8:2 ratio, with the training set used for training and optimizing the machine learning model, and the test set used for performance evaluation after model training.

[0029] 4. Train and optimize the machine learning model to build a fast prediction model: Using process parameters as input, flow velocity as output, a fully connected neural network as the backbone, and absolute average error as the loss function, the machine learning model is trained using a training set. After training, optimization is performed within the Optuna framework. Optimization includes adjusting the neural network's activation function, number of neurons, number of hidden layers, and the type of optimizer. The optimization iterations are 50 to 100 times, with the loss function value as the objective.

[0030] 5. Predicting the flow field under different process parameters: The trained machine learning model is applied to actual prediction tasks. By inputting any new combination of process parameters (casting speed, immersion depth of the nozzle, argon blowing rate, etc.), the model can quickly output the corresponding velocity distribution of the molten steel flow field within the crystallizer. By changing the input parameters, the flow field conditions under different process conditions can be continuously predicted, enabling rapid evaluation of the variation of flow field morphology with process parameters and providing a quantitative basis for process parameter optimization.

[0031] Four random combinations of process parameters were input, and numerical simulation results and machine learning model prediction results were obtained and reconstructed into a contour map. For example... Figure 2 As shown in the figure, the scatter plots of the data from the machine learning model and the numerical simulation results are compared. It can be observed that all data points are closely distributed around the diagonal of y=x, indicating a high degree of consistency between the model predictions and the numerical simulation results. Figure 3As shown, the left side presents the numerical simulation results, and the right side presents the machine learning model prediction results. Specifically, (a) the process parameter combination is: casting speed 1.046 m / min, sprue immersion depth 133.8 mm, and argon blowing rate 13.974 L / min. (b) the process parameter combination is: casting speed 1.412 m / min, sprue immersion depth 154.7 mm, and argon blowing rate 14.187 L / min. (c) the process parameter combination is: casting speed 0.989 m / min, sprue immersion depth 193.6 mm, and argon blowing rate 13.960 L / min. (d) the process parameter combination is: casting speed 1.098 m / min, sprue immersion depth 182.3 mm, and argon blowing rate 6.548 L / min. It can be seen that the machine learning model prediction results are highly consistent with the numerical simulation results, indicating that the machine learning model can effectively predict flow field characteristics.

[0032] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A real-time flow field prediction method combining Latin hypercube sampling, characterized in that, This includes using machine learning techniques to uncover the nonlinear mapping relationship between process parameters and the flow field behavior of molten steel, and constructing a predictive model.

2. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 1, characterized in that, A dataset covering various combinations of process parameters is generated in advance using numerical simulation data; the machine learning model is trained on the dataset to learn the nonlinear mapping relationship between parameters and results. After training, in actual production, the model no longer performs any physical iteration calculations, but directly maps the prediction results through computation.

3. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 1, characterized in that, The Latin hypercube sampling method is adopted to stratify each process parameter dimension and randomly select sample points within each layer to ensure that the sample points are uniformly distributed in the multidimensional parameter space.

4. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 3, characterized in that, The Latin hypercube sampling method divides the value range of each process parameter into several equally probable intervals, extracts a sample point in each interval, and then randomly combines the sample points of different parameters to form a complete sampling scheme.

5. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 1, characterized in that, Includes the following steps: Step 1: Construct the process parameter space using the Latin hypercube sampling method; Step 2: Numerical simulation batch calculation and flow field data export; Step 3: Data preprocessing and dataset partitioning; Step 4: Train and optimize the machine learning model to build a prediction model; Step 5: Predict the flow field conditions under different process parameters.

6. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 5, characterized in that, The process parameters include casting speed, sprue immersion depth, and argon blowing rate.

7. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 5, characterized in that, In step 2, a crystallizer numerical simulation model is used to perform batch simulation calculations on the combination of process parameters corresponding to each sampling point. After the calculation is completed, the velocity distribution data of the flow field at the free liquid surface of the molten steel in the crystallizer under each sampling point condition are exported to form the original high-dimensional dataset.

8. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 5, characterized in that, Step 3 involves preprocessing the flow field velocity data, including data cleaning and outlier removal. A normalization method is used to map the data to a uniform scale to eliminate the influence between different units; the normalized data is then randomly shuffled to eliminate the interference of sample order on model training.

9. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 5, characterized in that, In step 4, the process parameters are used as input, the flow field velocity is used as output, the fully connected neural network is used as the backbone network, the absolute average error is used as the loss function, and the training set is used to train the machine learning model. After training, the model is optimized in the Optuna framework.

10. The real-time flow field prediction method combining Latin hypercube sampling as described in claim 9, characterized in that, The optimization includes the activation function of the neural network, the number of neurons, the number of hidden layers in the neural network, and the type of optimizer.

Images