A method and system for on-line prediction of continuous casting quality defects and process optimization

Abstract

The application provides a method and system for online prediction of continuous casting quality defects and process optimization, the method comprising collecting production data related to the quality of the casting blank in the continuous casting production process, and summarizing the collected data for calling; pre-processing and feature extraction are performed on the production data, and one or more quality defects in the continuous casting process are predicted based on a prediction model, and the corresponding quality defect risk score is output; based on the risk score, it is determined whether there is a quality defect according to the determination condition, and when there is a quality defect, the corresponding optimization signal is output; according to the optimization signal, the preset target and the constraint condition are combined to optimize the calculation of the adjustable process parameters, and the process optimization suggestion is output. Through the above scheme, online prediction and process optimization of continuous casting quality defects can be realized, and the quality risk early warning capability and process control efficiency are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology for steel continuous casting processes, and particularly relates to a method and system for online prediction of continuous casting quality defects and process optimization. Background Technology

[0002] Continuous casting is a critical process in steel production, involving numerous process and equipment parameters with complex interrelationships and susceptibility to on-site disturbances. During casting, the billet is prone to quality defects such as cracks and segregation, which negatively impact the yield and performance of subsequent rolling. Since these defects are often difficult to observe directly in their early stages, they are frequently only detected during quality inspection or subsequent rolling, easily leading to rework or even scrapping. With the increasing demands for steel quality in high-end markets and the application of efficient production models such as hot charging and continuous casting and rolling, on-site operations not only need online identification and early warning of quality risks but also the ability to quickly provide actionable process adjustments when risks materialize, thereby reducing defect rates and improving the controllability of the production process.

[0003] Several data-driven continuous casting defect prediction schemes have emerged in the prior art. For example, patent document CN118656735A discloses a method for predicting corner transverse crack defects in continuously cast slabs based on a Stacking algorithm model. This method collects continuous casting production data, filters characteristic factors affecting corner transverse crack defects, and establishes a prediction model to predict these defects. Patent document CN119539155A discloses a multi-task prediction scheme based on deep learning, which can output multiple quality defect prediction results within the same model framework. These schemes can improve the timeliness of continuous casting quality defect prediction to a certain extent and expand the scope of prediction objects. However, these schemes are mainly used to obtain the prediction results of quality defects, and there is a lack of further connection between the prediction results and subsequent process adjustments. They fail to optimize adjustable process parameters based on the prediction results.

[0004] Furthermore, existing technologies also include solutions that combine prediction with process parameter optimization for specific defects. For example, patent document CN118296476A discloses a method for optimizing the process of longitudinal cracks on the surface of continuous casting based on support vector machines and genetic algorithms. This method establishes a prediction model for longitudinal cracks on the surface using support vector machines and optimizes relevant process parameters using genetic algorithms. While this solution can optimize process parameters based on prediction results, it primarily targets the specific object of longitudinal cracks on the surface. The optimization variables and process are all set around this object, resulting in a relatively fixed scope of application. It is difficult to uniformly configure the prediction and optimization process according to different quality risks and on-site application requirements. Summary of the Invention

[0005] To address the technical problems mentioned above, this invention proposes a system for online prediction and process optimization of continuous casting quality defects. This system integrates quality defect prediction and process parameter optimization within the same online system through the cooperation of a data acquisition module, a quality prediction module, a process optimization module, and a data storage and parameter configuration module. The system outputs quality defect prediction results and threshold judgment results based on the collected production data. When preset trigger conditions are met, the prediction results are used as the basis for process optimization calculations, and process optimization suggestions are output under process parameter constraints. Simultaneously, the data storage and parameter configuration module stores, configures, and updates relevant data and operating parameters, thereby improving the quality risk early warning capability and process control efficiency of continuous casting production.

[0006] To achieve the above objectives, this application provides the following solution:

[0007] A method for online prediction of continuous casting quality defects and process optimization includes:

[0008] S1. Collect production data related to billet quality during continuous casting and summarize the collected data for future reference;

[0009] S2. The production data is preprocessed and features are extracted, and one or more quality defects in the continuous casting process are predicted based on the prediction model, and the corresponding quality defect risk score is output.

[0010] S3. Based on the risk score, determine whether there is a quality defect according to the judgment conditions. When a quality defect exists, output the corresponding optimization signal.

[0011] S4. Based on the optimization signal, combined with the preset target and constraints, optimize the adjustable process parameters and output process optimization suggestions.

[0012] As a further explanation of the present invention, in S1, the production data includes equipment operation data, process parameter data, and billet quality inspection data.

[0013] As a further explanation of the present invention, in S2, defect categories are divided according to process influence parameters, and each defect category includes one or more quality defects; a prediction model is set for each defect category, taking the production data related to the defect category as input, and outputting the risk score corresponding to each quality defect type; the quality defect types include six categories: center segregation, center porosity, intermediate crack, triangular area crack, corner crack and surface crack, and the six quality defects are divided into three defect categories according to process influence parameters.

[0014] As a further explanation of the present invention, the prediction model adopts a machine learning model, including a one-dimensional convolutional feature extraction network and a multilayer perceptron output network; wherein, the one-dimensional convolutional feature extraction network is used to perform convolution operations and feature extraction on the input feature sequence, and the multilayer perceptron output network is used to map the extracted features and output the risk score corresponding to each quality defect.

[0015] As a further explanation of the present invention, the one-dimensional convolutional feature extraction network includes two convolution operations and two pooling operations; the multilayer perceptron output network includes multiple fully connected layers and ReLU activation layers set between adjacent fully connected layers, and a Dropout layer is set after each ReLU activation layer to suppress overfitting.

[0016] As a further explanation of the present invention, the prediction model has a rolling update mechanism; the initial prediction model is trained by constructing samples based on historical production data; during the use of the prediction model, samples are constructed using real-time collected data and risk scores, and the sample set is updated; at fixed intervals, the prediction model is retrained using the updated sample set to obtain a new prediction model, and the original online prediction model is replaced after verification that it meets the preset requirements.

[0017] As a further explanation of the present invention, the determination conditions in S3 include:

[0018] 1) The risk score p of the kth type of quality defect k ≥ Threshold T of the kth type of quality defect k If so, then the quality defect is determined to exist;

[0019] 2) If the overall risk score R ≥ the preset overall defect threshold T, then an overall quality defect is determined to exist. The overall risk score R is calculated using the following formula:

[0020]

[0021] Where K is the total number of quality defect types. Let be the weight corresponding to the k-th type of quality defect, and have . ;

[0022] The judgment process includes: first, based on the first judgment condition, comparing the risk scores of each type of quality defect with their corresponding thresholds to determine whether a quality defect exists; then calculating the comprehensive risk score R, and based on the second judgment condition, determining whether an overall quality defect exists; and finally, sending the corresponding optimization signal based on the judgment result.

[0023] As a further explanation of the present invention, S4 specifically involves: determining the process parameters to be optimized based on the optimization signal; and constructing the following objective function with minimizing the overall risk score R as the optimization objective:

[0024]

[0025] Where u is the combination of process parameters, p k (u) represents the predicted risk score of the k-th type of quality defect under the process parameter combination u, and λ is the penalty coefficient. A multi-objective optimization algorithm is used to solve the above optimization objectives, determine the final process parameter combination, and output process optimization suggestions. The multi-objective optimization algorithm is selected from genetic algorithm, particle swarm optimization, and Bayesian optimization.

[0026] As a further explanation of the present invention, the optimized process parameters are connected to the continuous casting control system, and the corresponding process parameters are adjusted.

[0027] This invention also provides a system for online prediction and process optimization of continuous casting quality defects, used to implement the above method, comprising:

[0028] The data acquisition module is used to collect production data related to billet quality during the continuous casting process and to summarize the collected data for future use.

[0029] The quality prediction module is used to preprocess and extract features from the production data, and predict one or more quality defects in the continuous casting process based on the prediction model, and output the corresponding quality defect risk score.

[0030] The quality assessment module is used to determine whether there is a quality defect based on the risk score and the assessment criteria. When a quality defect is found, the corresponding optimization signal is output.

[0031] The process optimization module is used to perform optimization calculations on adjustable process parameters based on optimization signals, combined with preset targets and constraints, and output process optimization suggestions.

[0032] As a further explanation of the present invention, the system also includes a data storage module and a parameter configuration module. The data storage module is used to store production data, quality defect risk scores, quality defect judgment results, and process optimization suggestions. The parameter configuration module is used to configure the parameters of the data acquisition module, quality prediction module, quality judgment module, and process optimization module.

[0033] The beneficial effects of this invention are:

[0034] This invention integrates continuous casting quality defect prediction and process parameter optimization into the same online system. This allows the prediction results to be further used for process optimization calculations, and outputs process optimization suggestions that meet the constraints when trigger conditions are met. This enhances the guidance role of quality risk warning results in on-site process adjustments. Simultaneously, the data storage and parameter configuration module manages relevant data and operating parameters, and the prediction and optimization processes can be configured according to on-site application needs, improving the system's applicability and operational stability. Attached Figure Description

[0035] Figure 1 This is a flowchart of the online prediction and process optimization system for continuous casting quality defects in this invention.

[0036] Figure 2 This is a schematic diagram of the prediction model in this invention. Detailed Implementation

[0037] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] This invention provides an online prediction and process optimization system for continuous casting quality defects. The system includes a data acquisition module, a quality prediction module, a process optimization module, and a data storage and parameter configuration module. These modules work together: the data acquisition module acquires process parameters and equipment data in real time during continuous casting; the quality prediction module preprocesses and extracts features from the acquired production data, and uses a machine learning model to predict and determine thresholds for multiple types of quality defects online; when preset trigger conditions are met, the process optimization module outputs process optimization suggestions with the goal of minimizing the risk score; and the data storage and module configuration module stores the production data, prediction results, and process optimization suggestions, and configures the operating parameters of each module.

[0039] In this implementation case, the No. 3 slab continuous casting machine of a steel plant is used as the implementation object. This continuous casting machine mainly produces high-quality steel grades such as pipeline steel and automotive steel, with slab cross-sectional dimensions of (230-250) mm × (900-1650) mm. Considering the production characteristics of this production line, the data related to slab quality mainly include process parameters, equipment operating status parameters, and production scheduling information. These data come from different levels of systems, including automation systems, process control systems, and manufacturing execution systems. During the quality control process of this production line, the focus is on detecting and evaluating six typical defects: center segregation, center porosity, intermediate cracks, triangular cracks, corner cracks, and surface cracks. For quality fluctuations occurring on this production line, the slab quality can be improved by adjusting process parameters such as casting speed, secondary cooling regime, and reduction regime.

[0040] This embodiment provides a method for online prediction of continuous casting quality defects and process optimization, the process of which is as follows: Figure 1 As shown, it includes:

[0041] Step 101. Collect production data related to billet quality during the continuous casting process, and summarize the collected data for future reference; the production data includes equipment operation data, process parameter data, and billet quality inspection data.

[0042] In this embodiment, taking the production data of the No. 3 slab continuous casting machine of a steel plant as an example, quality-related production data are collected. The production data includes: steel grade information, such as the steel grade and the mass percentage of the main chemical components C, Si, Mn, P, S, etc.; crystallizer operation data, such as the crystallizer cooling water flow rate, crystallizer liquid level fluctuation value, crystallizer vibration frequency, and crystallizer amplitude; secondary cooling operation data, such as the secondary cooling water main pipe temperature, secondary cooling water main pipe pressure, and water volume and pressure of each secondary cooling zone; casting temperature; billet pulling speed; electromagnetic stirring data, such as electromagnetic stirring current and frequency; equipment alignment data, such as the roll gap value and arc value of each sector roll; reduction data, such as the starting position and reduction amount of light reduction; and billet quality inspection data.

[0043] Step 102. Preprocess and extract features from the production data, and predict one or more quality defects in the continuous casting process based on the prediction model, and output the corresponding quality defect risk score.

[0044] In this embodiment, six types of quality defects can be divided into three categories based on process influence parameters: the first category includes center segregation and center porosity, which are affected by parameters such as steel grade information, billet drawing speed, and secondary cooling operation; the second category includes center cracks, triangular cracks, and corner cracks, which are affected by parameters such as crystallizer operating parameters, reduction parameters, and electromagnetic stirring; and the third category includes surface cracks, which are affected by parameters such as pouring temperature and equipment alignment. A prediction model is constructed for each type of quality defect, using the corresponding parameters as input features, and outputting the risk score corresponding to each quality defect category.

[0045] In this embodiment, for the prediction model of the first type of quality defect, steel grade information, billet pulling speed, and secondary cooling operation data can be selected as input features; for the prediction model of the second type of quality defect, crystallizer operation data, electromagnetic stirring data, and pressing data can be selected as input features; for the prediction model of the third type of quality defect, pouring temperature and equipment alignment data can be selected as input features.

[0046] Before inputting production data into the prediction model, the data is preprocessed, including one or more of the following: missing value handling, outlier handling, data transformation, and discrete feature encoding. Taking the production speed and secondary cooling zone water volume data as an example, this embodiment employs outlier handling, standardization, and missing value handling, as detailed below:

[0047] (1) Outlier handling: In this embodiment, the 3σ principle is used to identify outliers in continuous data. If a data point exceeds the mean ± 3 times the standard deviation, it is determined to be an outlier and replaced with a preset standard value.

[0048] For example, the average water flow rate of the arc center nozzle loop in the second cooling zone is 273.93, with a standard deviation of 90.30. Therefore, 3 The range is [3.03, 544.83]. If the water volume sample value of this loop is 563.72 at a certain moment, which is greater than 544.83, it is determined to be an outlier. The system will replace it with a preset standard value to reduce the impact of outlier data on subsequent feature extraction and model prediction.

[0049] (2) Standardization: Standardize the two different dimensions of data, namely throwing speed and water volume, and transform the dataset into data with a mean of 0 and a variance of 1.

[0050] Taking the actual water volume data of a narrow area in Zone 1 as an example, the mean was calculated using the formulas for mean and standard deviation. =80.96, standard deviation 35.28. Finally, by substituting the values ​​into the standardized formula, we obtain the calculated results for each data point, thus reducing the impact of these data with excessively large dimensional differences on the model.

[0051] (3) Missing value handling: For the two continuous parameters, pull speed and water volume, the missing value handling method is set as follows: if the number of missing values ​​accounts for less than 5% of the total number of samples, the standardized data will be filled with 0 to fill the missing values; if the number of missing values ​​accounts for more than 5% of the total number of samples, the sample will be deleted directly.

[0052] In this embodiment, the prediction model employs a machine learning model, including a one-dimensional convolutional feature extraction network and a multilayer perceptron output network, such as... Figure 2 As shown in the diagram, a one-dimensional convolutional feature extraction network is used to perform convolution operations and feature extraction on the input feature sequence, while a multilayer perceptron output network is used to map the extracted features and output the risk score corresponding to each quality defect. The one-dimensional convolutional feature extraction network includes two convolution operations and two pooling operations. The multilayer perceptron output network includes multiple fully connected layers and ReLU activation layers placed between adjacent fully connected layers, with a Dropout layer placed after each ReLU activation layer to suppress overfitting.

[0053] Taking the prediction model for the first type of quality defect as an example, this prediction model selects a total of 166 relevant production parameters, corresponding to the length of the input feature sequence. The value is 166. The first convolution uses a length of... Convolution kernel of 4, stride 2. Fill If the value is 0, the length of the output sequence can be obtained from the formula:

[0054]

[0055] The first pooling operation uses a window size of 4 and a stride of 2, resulting in an output sequence length of 40. The second convolution uses a kernel of length 2, a stride of 2, and padding of 0, resulting in an output sequence length of 20. The third pooling operation uses a window size of 2 and a stride of 2, resulting in an output sequence length of 10. After flattening, a 160-dimensional feature vector is obtained.

[0056] The first fully connected layer of the multilayer perceptron output network maps 160-dimensional features to 120-dimensional features, the second fully connected layer maps 120-dimensional features to 84-dimensional features, and the third fully connected layer outputs the risk score corresponding to each quality defect.

[0057] Specifically, in this embodiment, the predicted risk scores for the six quality defects are p1=0.42, p2=0.52, p3=0.33, p4=0.41, p5=0.52, and p6=0.32.

[0058] The training process of the prediction model is as follows: historical production data is collected to construct samples, and the samples are labeled with risk level tags corresponding to six types of quality defects. The samples are divided into three categories according to the defect category, and training sets and validation sets are constructed and preprocessed. The samples are divided into training sets and validation sets, and sent to the corresponding prediction models in batches to output risk scores. The model parameters are continuously updated during iterative training, and the model performance is verified by relying on the validation set. An early stopping mechanism is used to prevent overfitting. The mean squared error loss function is used to calculate the deviation between the predicted score and the labeled tag. The model parameters are optimized layer by layer through backpropagation. The iteration is repeated until the model parameters converge, and the training of the three types of prediction models is completed.

[0059] In other embodiments, the prediction model has a rolling update mechanism. The initial prediction model is trained based on historical production data and corresponding quality labels; during the use of the prediction model, the system archives the real-time collected process parameters, equipment parameters, and subsequently obtained risk scores, forming an updated sample set; the system retrains the prediction model using the updated sample set at fixed intervals (e.g., one month) to obtain a new prediction model, and replaces the original online prediction model after verifying that it meets preset requirements, thereby improving the system's adaptability to new operating conditions and prediction accuracy.

[0060] Step 103. Based on the risk score, determine whether there is a quality defect according to the judgment conditions. When a quality defect exists, output the corresponding optimization signal.

[0061] The determination criteria include:

[0062] 1) The risk score p of the kth type of quality defect k ≥ Threshold T of the kth type of quality defect k If the threshold T is found to be present, then the quality defect is determined to exist. k Set according to steel type and quality control requirements;

[0063] 2) If the overall risk score R ≥ the preset overall defect threshold T, then an overall quality defect is determined to exist. The overall risk score R is calculated using the following formula:

[0064]

[0065] Where K is the total number of quality defect types. Let be the weight corresponding to the k-th type of quality defect, and have . .

[0066] The judgment process includes: first, based on the first judgment condition, comparing the risk scores of each type of quality defect with their corresponding thresholds to determine whether a quality defect exists; then, calculating the comprehensive risk score R, and based on the second judgment condition, determining whether an overall quality defect exists; and based on the judgment result, sending the corresponding optimization signal to optimize and adjust the corresponding process parameters.

[0067] Specifically, in this embodiment, the steel grade is St12. The steel grade category and quality control requirements set the thresholds for center segregation and center porosity to 1.0, and the thresholds for intermediate cracks, triangular cracks, corner cracks, and surface cracks to 0.5. According to the first judgment condition, the risk score p5 = 0.52 for corner crack defects is higher than its corresponding threshold T5 = 0.5, indicating the presence of a corner crack quality defect. The weights are set as follows: center segregation weight w1 = 0.1855, center porosity weight w2 = 0.1835, intermediate crack weight w3 = 0.1722, triangular crack weight w4 = 0.1611, corner crack weight w5 = 0.1444, and surface crack weight w6 = 0.1533. The preset comprehensive defect threshold T = 0.5, resulting in a comprehensive risk score R = 0.420. According to the second judgment condition, the comprehensive risk score does not exceed the threshold. Based on the combined judgment results of the two conditions, an optimization signal for the corner crack defect is issued.

[0068] Step 104. Based on the optimization signal, combined with the preset target and constraints, perform optimization calculations on the adjustable process parameters and output process optimization suggestions.

[0069] Based on the optimization signals, the process parameters to be optimized are determined; with minimizing the overall risk score R as the optimization objective, the following objective function is constructed:

[0070]

[0071] Where u is the combination of process parameters, p k (u) represents the predicted risk score of the k-th type of quality defect under the process parameter combination u, and λ is the penalty coefficient.

[0072] The above optimization objectives are solved using a multi-objective optimization algorithm, wherein the multi-objective optimization algorithm is selected from genetic algorithm, particle swarm optimization, and Bayesian optimization.

[0073] Specifically, in this embodiment, based on the optimization signal for corner crack defects issued in the previous step, the process parameters to be optimized are determined to include the average drawing speed u1 (m / min), the starting position of light reduction u2 (represented by sector number), the ending position of light reduction u3 (represented by sector number), and the amount of light reduction u4 (mm); referring to the process specifications, the value ranges for each process parameter to be optimized are set as follows:

[0074]

[0075]

[0076]

[0077]

[0078] A genetic algorithm is used to solve the above optimization objective: First, within the preset constraints of each process parameter, multiple sets of process parameters are randomly generated as candidate process parameter combinations, wherein the preset constraint range of each process parameter is set with reference to the process specification; the candidate process parameter combinations are input as input parameters to each prediction module to obtain the risk scores of various quality defects and calculate the comprehensive risk score; the candidate process parameter combinations are screened according to the objective function calculation results, and a new generation of parameter combinations is generated through selection, crossover, and mutation operations; the prediction evaluation and parameter update process is repeated until the preset termination condition is reached, the termination condition including reaching the maximum number of iterations or the objective function satisfying the convergence condition; according to the finally determined process parameter combination, process optimization suggestions are output. Under the premise of satisfying the preset constraint range of each process parameter, the optimization results of the average drawing speed, the starting position of light pressing, the ending position of light pressing, and the amount of light pressing are output as process optimization suggestion values.

[0079] In other embodiments, after obtaining the optimization results, the optimized process parameters are connected to the continuous casting control system, and the corresponding process parameters are adjusted to achieve closed-loop control of quality prediction, process optimization and process parameter adjustment.

[0080] This embodiment also provides a system for online prediction and process optimization of continuous casting quality defects, used to implement the above method, including:

[0081] The data acquisition module is used to collect production data related to billet quality during the continuous casting process and to summarize the collected data for future use.

[0082] The quality prediction module is used to preprocess and extract features from the production data, and predict one or more quality defects in the continuous casting process based on the prediction model, and output the corresponding quality defect risk score.

[0083] The quality assessment module is used to determine whether there is a quality defect based on the risk score and the assessment criteria. When a quality defect is found, the corresponding optimization signal is output.

[0084] The process optimization module is used to perform optimization calculations on adjustable process parameters based on optimization signals, combined with preset targets and constraints, and output process optimization suggestions.

[0085] In other embodiments, the system further includes a data storage module for storing production data, quality defect risk scores, quality defect judgment results, and process optimization suggestions.

[0086] In the data storage module, production data is managed using a combination of recent and historical tables to accommodate both real-time querying and historical traceability needs. The recent table stores production data from the most recent period, facilitating daily data retrieval and quality inspection data additions; the historical table is used for long-term archiving and querying. In this embodiment, for example, the recent table can store 200,000 production data entries, roughly equivalent to two months' worth of data. The system automatically checks the recent table's data volume weekly; when the recent table's data volume exceeds a preset threshold, the excess data is moved to the historical table to maintain the efficiency of accessing the recent table.

[0087] In other embodiments, the system further includes a parameter configuration module for configuring parameters of the data acquisition module, quality prediction module, quality judgment module, and process optimization module.

[0088] The data acquisition module can be configured with the types of data to be collected, the preset standard values ​​corresponding to the data, etc.; the quality prediction module can be configured with the selection of prediction model hyperparameters and feature parameters, and the update cycle can be configured for the rolling update mechanism; the quality judgment module can be configured with the threshold parameters for various quality defects, etc.; and the process optimization module can be configured with the process parameters to be adjusted for each optimization signal, the process parameter constraint range, and the optimization algorithm parameters, etc.

[0089] The above description represents a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for online prediction and process optimization of continuous casting quality defects, characterized in that, include: S1. Collect production data related to billet quality during continuous casting and summarize the collected data for future reference; S2. The production data is preprocessed and features are extracted, and one or more quality defects in the continuous casting process are predicted based on the prediction model, and the corresponding quality defect risk score is output. S3. Based on the risk score, determine whether there is a quality defect according to the judgment conditions. When a quality defect exists, output the corresponding optimization signal. S4. Based on the optimization signal, combined with the preset target and constraints, optimize the adjustable process parameters and output process optimization suggestions.

2. The method according to claim 1, characterized in that, In step S2, defect categories are divided according to process influence parameters, and each defect category includes one or more quality defects. A prediction model is set for each defect category, which takes the production data related to the defect category as input and outputs the risk score corresponding to each type of quality defect.

3. The method according to claim 2, characterized in that, The prediction model employs a machine learning model, comprising a one-dimensional convolutional feature extraction network and a multilayer perceptron output network. The one-dimensional convolutional feature extraction network is used to perform convolution operations and feature extraction on the input feature sequence, while the multilayer perceptron output network is used to map the extracted features and output the risk score corresponding to each quality defect.

4. The method according to claim 3, characterized in that, The one-dimensional convolutional feature extraction network includes two convolution operations and two pooling operations; the multilayer perceptron output network includes multiple fully connected layers and ReLU activation layers set between adjacent fully connected layers, and a Dropout layer is set after each ReLU activation layer to suppress overfitting.

5. The method according to claim 2, characterized in that, The prediction model has a rolling update mechanism; the initial prediction model is trained by constructing samples based on historical production data; during the use of the prediction model, samples are constructed using real-time collected data and risk scores, and the sample set is updated; at fixed intervals, the prediction model is retrained using the updated sample set to obtain a new prediction model, and the original online prediction model is replaced after verification that it meets the preset requirements.

6. The method according to claim 1, characterized in that, In S3, the determination conditions include: 1) The risk score p of the kth type of quality defect k ≥ Threshold T of the kth type of quality defect k If so, then the quality defect is determined to exist; 2) If the overall risk score R ≥ the preset overall defect threshold T, then an overall quality defect is determined to exist. The overall risk score R is calculated using the following formula: Where K is the total number of quality defect types. Let be the weight corresponding to the k-th type of quality defect, and have . ; The judgment process includes: first, based on the first judgment condition, comparing the risk scores of each type of quality defect with their corresponding thresholds to determine whether a quality defect exists; then calculating the comprehensive risk score R, and based on the second judgment condition, determining whether an overall quality defect exists; and finally, sending the corresponding optimization signal based on the judgment result.

7. The method according to claim 1, characterized in that, Specifically, S4 involves: determining the process parameters to be optimized based on the optimization signal; and constructing the following objective function with minimizing the overall risk score R as the optimization objective: Where u is the combination of process parameters, p k (u) represents the predicted risk score of the k-th type of quality defect under the process parameter combination u, and λ is the penalty coefficient; A multi-objective optimization algorithm is used to solve the above optimization objectives, determine the final combination of process parameters, and output process optimization suggestions.

8. The method according to claim 7, characterized in that, The optimized process parameters are then integrated into the continuous casting control system, and the corresponding process parameters are adjusted.

9. A system for online prediction and process optimization of continuous casting quality defects, used to implement the method as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to collect production data related to billet quality during the continuous casting process and to summarize the collected data for future use. The quality prediction module is used to preprocess and extract features from the production data, and predict one or more quality defects in the continuous casting process based on the prediction model, and output the corresponding quality defect risk score. The quality assessment module is used to determine whether there is a quality defect based on the risk score and the assessment criteria. When a quality defect is found, the corresponding optimization signal is output. The process optimization module is used to perform optimization calculations on adjustable process parameters based on optimization signals, combined with preset targets and constraints, and output process optimization suggestions.

10. The system according to claim 9, characterized in that, The system also includes a data storage module and a parameter configuration module. The data storage module is used to store production data, quality defect risk scores, quality defect judgment results, and process optimization suggestions. The parameter configuration module is used to configure the parameters of the data acquisition module, quality prediction module, quality judgment module, and process optimization module.

Images