Heat exchange coefficient adaptive model construction method based on machine learning

Abstract

The invention relates to the technical field of cooling control after rolling of medium and heavy plates, in particular to a heat exchange coefficient self-adaptive model construction method based on machine learning, and specifically relates to a heat exchange coefficient self-adaptive model construction method based on machine learning, which comprises the following steps: step 1, data preparation; step 2, feature engineering: identifying key features influencing a heat exchange coefficient through correlation analysis, constructing derivative features, and realizing dynamic feature weighting based on the importance of the derivative features; 3, performing model training and optimization, comparing prediction performances of linear regression, random forest and neural network algorithms, and selecting an optimal algorithm to construct a basic model; and 4, model deployment and online self-learning are carried out. According to the method, through data driving and mechanism fusion, dynamic feature weighting and a multi-model fusion mechanism are combined, and accurate prediction of the heat exchange coefficient and self-adaptive iterative optimization of the model are achieved.

Description

Technical Field

[0001] This invention relates to the field of cooling control technology after rolling of medium and heavy plates, and specifically to a method for constructing an adaptive model of heat transfer coefficient based on machine learning. Background Technology

[0002] Post-rolling cooling of medium and heavy plates is a crucial step in steel production. It refers to the process of regulating the temperature change of steel plates after rolling through a cooling system to control the target microstructure, residual stress distribution, and dimensional accuracy. This directly affects the mechanical properties, microstructure uniformity, and product quality of the steel plates, and is a core process for enhancing the competitiveness of medium and heavy plate products.

[0003] In existing technologies, the heat transfer coefficient models of post-rolling cooling systems for medium and heavy plates mostly rely on traditional mechanism modeling methods, such as the literature "Research and Application of Heat Transfer Mechanism of Advanced Post-Rolling Cooling System for Medium and Heavy Plates". By establishing a stable communication system, a precise tracking system, and a complete data storage and query system, automatic cooling function can be achieved, thereby meeting the requirements for precise online cooling control of medium and heavy plates. However, such modeling methods have weak feedforward adaptive capabilities, poor response to sudden changes in parameters such as cooling start-up temperature and water temperature, low model generalization ability, and long learning cycle. Furthermore, the historical data has poor universality, and data needs to be re-accumulated after the parameters of different sites or equipment are modified, resulting in low deployment efficiency. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a machine learning-based adaptive model construction method for heat transfer coefficient. By integrating data-driven and mechanism-based approaches, and combining dynamic feature weighting and multi-model fusion mechanisms, it achieves accurate prediction of heat transfer coefficient and adaptive iterative optimization of the model.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: a method for constructing an adaptive heat transfer coefficient model based on machine learning, comprising the following steps:

[0006] Step 1: Data preparation. Collect key parameter data and corresponding heat transfer coefficients during the cooling process of medium and heavy plates after rolling. Clean the key process parameter data, including outlier identification and removal, and fill missing values ​​with interpolation of previous and subsequent data or process experience values. For outliers of core process parameters such as starting cooling temperature and reheating temperature, manually review and correct them based on knowledge of controlled rolling and controlled cooling processes.

[0007] Step 2: Feature engineering, which identifies key features affecting the heat transfer coefficient through correlation analysis and constructs derived features, and implements dynamic feature weighting based on the importance of the derived features;

[0008] Step 3: Model training and optimization. The predictive performance of linear regression, random forest and neural network algorithms is compared. The performance evaluation is based on five-fold cross-validation, and the coefficient of determination R², mean squared error MSE and residual standard deviation s are used as the core evaluation indicators. The optimal algorithm is selected to build the basic model. In order to further improve the generalization ability and stability of the model under complex conditions, a dynamic weighted multi-model fusion mechanism is introduced to adaptively adjust its weight in the final prediction result.

[0009] Step 4: Model Deployment and Online Self-Learning. Deploy the basic model optimized in Step 3 to the production control system in the form of microservice API or dynamic link library. At the same time, start a batch self-learning task: connect to the production database regularly every week to obtain a large number of recent data blocks, perform data preprocessing, and then perform a complete round of model retraining and validation to generate a model version with better performance and roll it up to the production environment to achieve continuous iterative evolution of the model.

[0010] Furthermore, in step 1, outlier removal adopts the 3σ principle based on the Laida criterion or box plot method, that is, when the data value deviates from the mean by more than 2 times the standard deviation, it is judged as an outlier and removed; for outliers of key process parameters, manual correction is carried out in combination with rolling process knowledge.

[0011] Furthermore, in step 2, the key features identified through correlation analysis include: using correlation heatmaps to quantify the correlation between each parameter and the heat transfer coefficient, determining that the strong positive correlation features are steel plate thickness, cooling temperature, target temperature and water volume, the strong negative correlation feature is the set speed, and the weak correlation features are water temperature and water pressure.

[0012] Furthermore, in step 2, the calculation formula for dynamic feature weighting based on the importance of derived features is as follows:

[0013]

[0014] in, Let i be the weight of the i-th feature. Let be the importance index of the i-th feature, and n be the total number of features; the feature importance index is calculated by the Gini impurity reduction of the random forest algorithm.

[0015] Furthermore, in step 3, the adaptive weight adjustment method is as follows: based on the real-time prediction error of each sub-model through a dynamic weighted multi-model fusion mechanism, the smaller the error, the higher the weight.

[0016] Furthermore, in step 3, the dynamic correction formula for the multi-model fusion mechanism is:

[0017]

[0018] in, This is the predicted value of the heat transfer coefficient after fusion. The dynamic weights of the m-th model are... The predicted value of the m-th model is given, where M is the total number of models; dynamic weights. The weight of each model is adjusted adaptively based on its real-time prediction error; the smaller the error, the greater the weight.

[0019] Furthermore, in step 4, dynamically updating the model parameters includes: when the amount of newly collected data reaches a preset threshold, triggering an incremental learning mechanism to update the local parameters of the model based on the new data, without having to retrain the entire model.

[0020] Furthermore, in step 4, the implementation process of the batch self-learning function is as follows: initialize the batch learning parameters, including the maximum number of learning data blocks N and the size of each data block. It connects to the production database to obtain data blocks, performs incremental learning after preprocessing the data blocks, records learning logs and model evaluation results, and outputs optimized model parameters when the number of learned data blocks reaches N.

[0021] Furthermore, in step 3, the evaluation metrics for model prediction performance include: coefficient of determination. , residual standard deviation s and mean square error MSE, where , and The model performance is judged to meet the standard at that time.

[0022] Furthermore, in step 4, the production control system includes TRACK, PCPM, SUPICSU, and PREMODE modules. In the PREMODE module, an online automatic model update function is constructed: it receives real-time field data streams from the production line, and when the amount of new data accumulates to a preset threshold, it automatically triggers an incremental learning mechanism to use the new data to locally update the model. The update process does not require retraining all historical data, which significantly reduces the consumption of computing resources and meets the real-time requirements.

[0023] The above approach has the following beneficial effects:

[0024] 1. This solution effectively improves the accuracy and stability of heat transfer coefficient prediction. By preprocessing data to remove outliers and correcting deviations in key parameters, combined with feature engineering to accurately identify the core factors affecting the heat transfer coefficient, and then comparing and selecting the optimal machine learning algorithm to build a basic model, the prediction error is significantly reduced. This allows the model to more accurately reflect the heat transfer patterns in the actual cooling process, providing a reliable basis for temperature control.

[0025] 2. This solution enhances the model's adaptability to complex operating conditions. It introduces a multi-model fusion mechanism, dynamically adjusting weights based on the real-time prediction errors of each model, enabling flexible responses to sudden changes in parameters such as cooling temperature and water temperature. Simultaneously, the incremental learning mechanism for dynamically updating model parameters allows for rapid model optimization as new data accumulates to a certain scale, eliminating the need for retraining and ensuring the model remains consistently adapted to actual production conditions.

[0026] 3. This solution improves the model's generalization ability and cross-scenario applicability. By using batch self-learning functionality to quantitatively analyze differences between different sites, it enables the sharing and migration of historical data across production lines, significantly reducing the model debugging cycle after new projects or equipment parameter adjustments. This allows the model to quickly adapt to different manifold layouts, product specifications, and process conditions, reducing reliance on manual parameter tuning.

[0027] 4. This solution combines machine learning with mechanistic modeling to achieve dynamic prediction and real-time control of the heat transfer coefficient, reducing manual intervention, improving production efficiency, providing core technical support for the construction of next-generation intelligent cooling systems, and enhancing technological competitiveness in the field of post-rolling cooling of medium and heavy plates.

[0028] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the method steps in an embodiment of the machine learning-based adaptive heat transfer coefficient model construction method of the present invention;

[0030] Figure 2 This is a schematic diagram of regression analysis and correlation heatmap of an embodiment of the adaptive heat transfer coefficient model construction method based on machine learning of the present invention;

[0031] Figure 3 This is a schematic diagram of the batch self-learning function process in an embodiment of the machine learning-based adaptive heat transfer coefficient model construction method of the present invention. Detailed Implementation

[0032] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0033] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0034] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0035] The following detailed description illustrates the specific implementation method:

[0036] Example:

[0037] As attached Figure 1 , Figure 2 and Figure 3 As shown: A method for constructing an adaptive heat transfer coefficient model based on machine learning, including the following steps:

[0038] Step 1: Data Preparation. Collect key parameter data and corresponding heat transfer coefficients during the post-rolling cooling process of medium and heavy plates. Perform data cleaning on the key process parameter data, including outlier identification and removal. Outlier removal adopts the 3σ principle based on the Laida criterion or box plot method, that is, when the data value deviates from the mean by more than 3 times the standard deviation, it is judged as an outlier and removed. For outliers of key process parameters, manual correction is performed in combination with rolling process knowledge. Missing values ​​are filled by interpolation of previous and subsequent data or process experience values. For outliers of core process parameters such as starting cooling temperature and reheating temperature, manual review and correction are performed in combination with controlled rolling and controlled cooling process knowledge.

[0039] Step 2: Feature engineering. Identify key features affecting the heat transfer coefficient through correlation analysis and construct derived features. Implement dynamic feature weighting based on the importance of the derived features. Key features identified through correlation analysis include: using correlation heatmaps to quantify the correlation between each parameter and the heat transfer coefficient. Strongly positively correlated features are steel plate thickness, cooling temperature, target temperature, and water volume. Strongly negatively correlated features are set speed. Weakly correlated features are water temperature and water pressure.

[0040] The formula for calculating dynamic feature weighting based on the importance of derived features is as follows:

[0041]

[0042] in, Let i be the weight of the i-th feature. Let be the importance index of the i-th feature, and n be the total number of features; the feature importance index is calculated by the Gini impurity reduction of the random forest algorithm.

[0043] Step 3: Model training and optimization. Compare the predictive performance of linear regression, random forest, and neural network algorithms. The evaluation metrics for model predictive performance include: coefficient of determination. , residual standard deviation s and mean square error MSE, where , and The model performance is judged to meet the standard in real time; and the coefficient of determination R², mean squared error MSE and residual standard deviation s are used as the core evaluation indicators. The optimal algorithm is selected to build the basic model. In order to further improve the generalization ability and stability of the model under complex working conditions, a dynamic weighted multi-model fusion mechanism is introduced to adaptively adjust its weight in the final prediction result. The adaptive adjustment of weight is based on the real-time prediction error of each sub-model through the dynamic weighted multi-model fusion mechanism. The smaller the error, the higher the weight.

[0044] The dynamic correction formula for the multi-model fusion mechanism is:

[0045]

[0046] in, This is the predicted value of the heat transfer coefficient after fusion. The dynamic weights of the m-th model are... The predicted value of the m-th model is given, where M is the total number of models; dynamic weights. The weight of each model is adjusted adaptively based on its real-time prediction error; the smaller the error, the greater the weight.

[0047] Step 4: Model Deployment and Online Self-Learning. The basic model optimized in Step 3 is deployed to the production control system in the form of a microservice API or dynamic link library. The production control system includes TRACK, PCPM, SUPICSU, and PREMODE modules. In the PREMODE module, an online automatic model update function is built: it receives real-time data streams from the production line. When the amount of new data accumulates to a preset threshold, an incremental learning mechanism is automatically triggered to locally update the model using the new data. The update process does not require retraining all historical data, significantly reducing computing resource consumption and meeting real-time requirements. At the same time, a batch self-learning task is started: it connects to the production database weekly to obtain recent large-scale data blocks, performs data preprocessing, and then performs a complete round of model retraining and validation to generate a higher-performance model version, which is then rolled out to the production environment to achieve continuous iterative evolution of the model.

[0048] By receiving on-site data in real time and dynamically updating model parameters, the dynamic updating of model parameters includes: when the amount of newly collected data reaches a preset threshold, such as 50 to 100 sets, the incremental learning mechanism is triggered, and the model is locally updated based on the new data without retraining the entire model. Combined with the batch self-learning function, the model can be continuously iterated and optimized.

[0049] The implementation process of the batch self-learning function is as follows: Initialize the batch learning parameters, including the maximum number of learning data blocks N and the size of each data block. It connects to the production database to obtain data blocks, performs incremental learning after preprocessing the data blocks, records learning logs and model evaluation results, and outputs optimized model parameters when the number of learned data blocks reaches N.

[0050] In actual production scenarios, taking a medium-thick plate manufacturer as an example, it has long faced the problem of insufficient temperature control precision in the post-rolling cooling process of medium-thick plates. This manufacturer produces a variety of medium-thick plate specifications, including steel plates of different thicknesses. During post-rolling cooling, the lack of an accurate heat transfer coefficient model leads to unstable microstructure and properties of the cooled steel plates, resulting in a high defect rate. Therefore, the above-mentioned solution was adopted, and the specific implementation process is as follows: First, a comprehensive and detailed analysis was conducted on numerous factors affecting the heat transfer coefficient, such as steel plate thickness, water temperature, initial cooling temperature, and final cooling temperature. Through parameter nodalization and interpolation, the characteristic values ​​of each heat transfer coefficient influencing factor across the entire range were obtained. When a new steel plate to be cooled enters the cooling process, the characteristic values ​​corresponding to its various influencing factors are quickly captured. Next, based on the k-NN principle, samples with high similarity to the target steel plate are accurately found in the database of cooled sample steel plates. The similarity calculation here comprehensively considers the degree of difference in the characteristic values ​​of multiple influencing factors. Finally, the IDW plate averaging algorithm is used to perform a weighted average of the heat transfer coefficients of the similar sample steel plates found, thereby predicting the heat transfer coefficient required for the target cooling steel plate.

[0051] Traditional heat transfer coefficient calculation models often consider only a few key factors or lack in-depth analysis of the complex interactions between these factors, leading to significant deviations between calculated results and actual conditions. This proposed solution, however, fully considers the interactions of all factors influencing heat transfer coefficients across the entire range, greatly improving the stability of temperature control. In practical applications, the factory has seen a significant improvement in temperature control accuracy after adopting this model. For example, when cooling 20mm thick steel plates, the previous model resulted in large fluctuations in the final cooling temperature, frequently exceeding the allowable range for the target final cooling temperature, leading to uneven steel plate performance. With this model, through precise heat transfer coefficient prediction, heat exchange during the cooling process can be controlled more accurately, ensuring the final cooling temperature is stably controlled within the target range. Overall, the system is more stable than before, effectively improving product quality and reducing the defect rate. Furthermore, this model also demonstrates good predictive ability for cooling new specifications and varieties of steel plates. When encountering a new type of steel plate with a thickness of 30mm, previous models could not quickly and accurately provide suitable cooling parameters. However, this model can quickly find similar cases in the sample database based on the characteristic values ​​of its influencing factors and predict the accurate heat transfer coefficient, providing a reliable basis for the cooling process. This greatly enhances the flexibility and adaptability of production. According to actual field verification, using this model has improved the hit rate of final cooling temperature by about 5% compared with the past, effectively ensuring the stable operation of production.

[0052] In actual operation, the temperature data for each steel plate is obtained from the actual temperature reading of an on-site pyrometer. To ensure the accuracy and representativeness of the data, the average temperature of the middle 500mm area (excluding the first and last 500mm sections) is selected, while outliers caused by residual water are filtered out. Before the steel plate enters the cooling zone, the system acquires the actual temperature of the steel plate in real time, i.e., the starting cooling temperature. Then, combined with the heat transfer coefficient estimated by the model and other relevant parameters, the cooling process is dynamically controlled. For example, based on the calculated heat transfer coefficient, the system automatically adjusts the cooling water volume, roller speed, and manifold opening method to ensure that the steel plate cools down according to the preset cooling procedure and ultimately reaches the target final cooling temperature. The entire process achieves automation and precision, reducing errors caused by human intervention and further improving production efficiency and product quality stability.

[0053] The comparison results of the machine learning models are shown in Table 1:

[0054] Table 1 Comparison of Machine Learning Modes

[0055]

[0056] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for constructing an adaptive model of heat transfer coefficient based on machine learning, characterized in that, Includes the following steps: Step 1: Data preparation. Collect key parameter data and corresponding heat transfer coefficients during the cooling process of medium and heavy plates after rolling. Clean the key process parameter data, including outlier identification and removal, and fill missing values ​​with interpolation of previous and subsequent data or process experience values. For outliers of core process parameters such as starting cooling temperature and reheating temperature, manually review and correct them based on knowledge of controlled rolling and controlled cooling processes. Step 2: Feature engineering, which identifies key features affecting the heat transfer coefficient through correlation analysis and constructs derived features, and implements dynamic feature weighting based on the importance of the derived features; Step 3: Model training and optimization. The predictive performance of linear regression, random forest and neural network algorithms is compared. The performance evaluation is based on five-fold cross-validation, and the coefficient of determination R², mean squared error MSE and residual standard deviation s are used as the core evaluation indicators. The optimal algorithm is selected to build the basic model. In order to further improve the generalization ability and stability of the model under complex conditions, a dynamic weighted multi-model fusion mechanism is introduced to adaptively adjust its weight in the final prediction result. Step 4: Model Deployment and Online Self-Learning. Deploy the basic model optimized in Step 3 to the production control system in the form of microservice API or dynamic link library. At the same time, start a batch self-learning task: connect to the production database regularly every week to obtain a large number of recent data blocks, perform data preprocessing, and then perform a complete round of model retraining and validation to generate a model version with better performance and roll it up to the production environment to achieve continuous iterative evolution of the model.

2. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 1, characterized in that, In step 1, outlier removal adopts the 3σ principle based on the Laida criterion or box plot method, that is, when the data value deviates from the mean by more than 3 times the standard deviation, it is judged as an outlier and removed; for outliers of key process parameters, manual correction is carried out in combination with rolling process knowledge.

3. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 2, characterized in that, In step 2, key features are identified through correlation analysis, including: using correlation heatmaps to quantify the correlation between each parameter and the heat transfer coefficient, identifying the strong positive correlation features as steel plate thickness, cooling temperature, target temperature and water volume, the strong negative correlation feature as set speed, and the weak correlation features as water temperature and water pressure.

4. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 3, characterized in that, In step 2, the formula for calculating dynamic feature weighting based on the importance of derived features is as follows: in, Let i be the weight of the i-th feature. Let be the importance index of the i-th feature, and n be the total number of features; the feature importance index is calculated by the Gini impurity reduction of the random forest algorithm.

5. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 4, characterized in that, In step 3, the adaptive weight adjustment method is as follows: based on the real-time prediction error of each sub-model, the smaller the error, the higher the weight.

6. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 5, characterized in that, In step 3, the dynamic correction formula for the multi-model fusion mechanism is: in, This is the predicted value of the heat transfer coefficient after fusion. The dynamic weights of the m-th model are... The predicted value of the m-th model is given, where M is the total number of models; dynamic weights. The weight of each model is adjusted adaptively based on its real-time prediction error; the smaller the error, the greater the weight.

7. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 6, characterized in that, In step 4, dynamically updating the model parameters includes: when the amount of newly collected data reaches a preset threshold, triggering the incremental learning mechanism, updating the local parameters of the model based on the new data, without retraining the entire model.

8. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 7, characterized in that, In step 4, the implementation process of the batch self-learning function is as follows: Initialize the batch learning parameters, including the maximum number of learning data blocks N and the size of each data block. It connects to the production database to obtain data blocks, performs incremental learning after preprocessing the data blocks, records learning logs and model evaluation results, and outputs optimized model parameters when the number of learned data blocks reaches N.

9. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 8, characterized in that, In step 3, the evaluation metrics for model prediction performance include: coefficient of determination. , residual standard deviation s and mean square error MSE, where , and The model performance is judged to meet the standard at that time.

10. The method for constructing an adaptive heat transfer coefficient model based on machine learning according to claim 9, characterized in that, In step 4, the production control system includes TRACK, PCPM, SUPICSU, and PREMODE modules. In the PREMODE module, an online automatic model update function is built: it receives real-time data streams from the production line, and when the amount of new data accumulates to a preset threshold, it automatically triggers an incremental learning mechanism to use the new data to locally update the model. The update process does not require retraining all historical data, which significantly reduces the consumption of computing resources and meets the real-time requirements.

Images