Continuous casting powder defect analysis method based on multi-model stacking integration algorithm

Abstract

The invention provides a continuous casting powder defect analysis method and system based on a multi-model stacking integration algorithm, and the method comprises the steps: segmenting extracted high-frequency feature data into subsequences through a sliding window, and selecting important features; and selecting a base learner and a meta learner of the stacked integrated model to process the high-frequency feature data to obtain a final prediction result. Constructing a stack integration model according to the base learner and the meta learner; and according to the stack integration model, selecting hyper-parameter configuration optimization and outputting a final prediction result. According to the method, the defect of correlation analysis between high-frequency data fluctuation and casting powder defects in the continuous casting crystallizer is overcome by adopting a method based on a multi-model stacking integration algorithm; the high-frequency feature data is segmented through the sliding window, and the prediction limitation of a single model is converted into the advantage of an integrated model in combination with multi-model cross validation, so that the prediction accuracy and generalization ability of the casting powder defects are improved, and the analysis efficiency of the casting powder defects is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent manufacturing technology, and specifically relates to a continuous casting mold slag defect analysis method and system based on a multi-model stacking integrated algorithm. Background Art

[0002] During the continuous casting process, mold slag performs core functions such as thermal insulation, lubricating the solidified shell, and absorbing inclusions. As the steel industry moves toward shorter processes, higher casting speeds, and lower carbonization, mold slag must possess greater thermal stability and adaptability. Substandard mold slag performance, such as poor fluidity, abnormal crystallization rates, or unstable bulk density, can easily lead to surface defects in the ingot, such as longitudinal cracks, subsurface bubbles, and slag inclusions. These defects can even cause steel breakouts, severely impacting production safety and yield rates.

[0003] Existing testing methods rely heavily on static laboratory tests (such as bulk density, viscosity, and melting point), lacking the ability to monitor dynamic process conditions in real time. Furthermore, existing testing systems struggle to support mold slag optimization under new process parameters, such as the unique crystallization requirements of ultra-low carbon steel grades.

[0004] Continuous casting mold slag defect detection is a core technical requirement for improving steel production quality and efficiency. The multi-model stacking ensemble algorithm, with its high accuracy and strong adaptability, provides an innovative path for dynamic detection and process optimization. By fusing the prediction results of multiple base models and leveraging meta-models for secondary learning, the multi-model stacking ensemble algorithm significantly improves prediction accuracy and robustness in complex industrial scenarios.

[0005] Continuous casting mold slag defect detection involves multi-source heterogeneous data (such as temperature, viscosity, bulk density, flow images, etc.), and a single model cannot fully capture the feature correlation. The stacking integration algorithm can effectively solve this problem by extracting features in layers and complementing the model. However, existing research has mostly focused on single parameter optimization. In the case of multi-source heterogeneous data, the data volume is huge but the degree of structure is low, which requires higher requirements for data cleaning and feature alignment, and the processing cost is high. Breakthroughs are still needed in the real-time and generalization capabilities of multi-model fusion, especially the algorithm stability problem in high-noise industrial environments.

[0006] Industrial data suffers from high noise and a lack of labeled samples, necessitating the integration of transfer learning or semi-supervised methods to reduce reliance on labeled data. Furthermore, the black-box nature of complex integrated models can undermine the trust of process personnel, necessitating the development of visual interpretation tools and integration with process knowledge bases.

[0007] The patent document "A Multi-Round Multi-Model Fusion Detection Method for Improving Defect Detection Capabilities" (CN116740004A) discloses a method for fine-tuning defect location and classification through multi-model fusion screening. By employing a multi-round detection strategy, this method effectively ensures detection accuracy while ensuring real-time performance and significantly reduces missed defects. However, it lacks comprehensiveness, is insufficient in analyzing the correlation between high-frequency data fluctuations and mold slag defects, and has poor generalization capabilities.

[0008] In summary, there is an urgent need for a mold slag defect analysis method that can reflect the correlation between high-frequency data fluctuations in the continuous casting mold and mold slag defects, and has better generalization ability and overall prediction performance. Summary of the Invention

[0009] In view of the defects in the prior art, the purpose of the present invention is to provide a continuous casting mold slag defect analysis method based on a multi-model stacking integrated algorithm.

[0010] The continuous casting mold slag defect analysis method based on a multi-model stacking integrated algorithm provided by the present invention includes:

[0011] Step S1: Segment the extracted high-frequency feature data into subsequences through a sliding window and select important features;

[0012] Step S2: Select the base learner and meta-learner of the stacked ensemble model, process the important features, and obtain the important features after dimensionality reduction and the final prediction results;

[0013] Step S3: construct a stacked ensemble model based on the base learner and the meta learner;

[0014] Step S4: According to the stacked ensemble model, select the hyperparameter configuration optimization and output the final prediction result.

[0015] Preferably, the step S1 includes:

[0016] Step S1.1, mapping the hot coil defect location to the steelmaking high-frequency location by calculating the ratio of the slab length to the hot coil length to obtain high-frequency feature data;

[0017] Step S1.2: Select sliding window parameters that can cover the most high-frequency feature data, and use the sliding window to perform subsequence segmentation on the high-frequency feature data. The sliding window parameters include step size and window length.

[0018] Step S1.3: Further extract the subsequences within the window based on their time domain characteristics and frequency domain characteristics, and select important features as model input parameters based on importance ranking and expert experience.

[0019] Preferably, step S2 includes:

[0020] Step S2.1: Use XGBoost and decision tree as base learners and random forest as meta-learner to perform data dimensionality reduction on important features;

[0021] Step S2.2: Divide the important features after dimensionality reduction into a specified number of sub-datasets through the self-service sampling method, resample to extract a certain number of new samples, construct the same number of decision trees for the sub-datasets, count the result categories of each decision tree according to the voting method, and take the result category with the most votes as the final result category of the decision tree.

[0022] Preferably, step S3 includes:

[0023] Step S3.1: Randomly divide the important features after dimensionality reduction into 5 parts, 4 of which are used as training sets and 1 as test set, construct XGBoost and decision tree, and perform 5-fold cross-validation on the training sets to obtain the trained XGBoost and decision tree;

[0024] Step S3.2: After training, XGBoost and decision tree are each subjected to five-fold cross-validation on the test set, and the predicted labels are merged as the prediction results.

[0025] Step S3.3: The prediction results are used as new features of the training data and put into the random forest model for secondary training to obtain a trained random forest model.

[0026] In step S4, the prediction result is input into the trained random forest model with hyperparameter configuration, and the final prediction result is output.

[0027] According to the present invention, a continuous casting mold slag defect analysis system based on a multi-model stacking integrated algorithm includes:

[0028] Module M1, splits the extracted high-frequency feature data into subsequences through a sliding window and selects important features;

[0029] Module M2 selects the base learner and meta-learner of the stacked ensemble model, processes important features, and obtains important features after dimensionality reduction;

[0030] Module M3, constructs a stacked ensemble model based on base learners and meta-learners;

[0031] Module M4, based on the stacked ensemble model, selects hyperparameter configuration optimization and outputs the final prediction result.

[0032] Preferably, the module M1 includes:

[0033] Module M1.1: Calculate the ratio of slab length to hot coil length to map the hot coil defect location to the high-frequency position of steelmaking to obtain high-frequency feature data;

[0034] Module M1.2: Select sliding window parameters that can cover the most high-frequency feature data, and use the sliding window to perform subsequence segmentation on the high-frequency feature data. The sliding window parameters include step size and window length.

[0035] Module M1.3 further extracts the subsequences within the window based on time domain characteristics and frequency domain characteristics, and selects important features as model input parameters based on importance ranking and expert experience.

[0036] Preferably, the module M2 includes:

[0037] Module M2.1: Using XGBoost and decision trees as base learners and random forest as meta-learner, we perform data dimensionality reduction on important features.

[0038] Module M2.2: Divide the important features after dimensionality reduction into a specified number of sub-datasets through the self-service sampling method, resample to extract a certain number of new samples, construct the same number of decision trees for the sub-datasets, count the result categories of each decision tree according to the voting method, and use the result category with the most votes as the final result category of the decision tree.

[0039] Preferably, the module M3 includes:

[0040] Module M3.1: Randomly divide the important features after dimensionality reduction into 5 parts, 4 of which are used as training sets and 1 as test set. Build XGBoost and decision tree models, perform 5-fold cross-validation on the training sets, and obtain the trained XGBoost and decision tree models.

[0041] In module M3.2, the trained XGBoost and decision tree perform five-fold cross-validation on the test set, and merge the predicted labels as the prediction results;

[0042] Module M3.3: The prediction results are used as new features of the training data to be put into the random forest model for secondary training to obtain the trained random forest model.

[0043] The module M4 inputs the prediction result into the trained random forest model with hyperparameter configuration and outputs the final prediction result.

[0044] According to the present invention, a computer-readable storage medium storing a computer program is provided, and when the computer program is executed by a processor, the steps of the continuous casting mold slag defect analysis method based on the multi-model stacking integrated algorithm are implemented.

[0045] According to the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the continuous casting protective slag defect analysis method based on a multi-model stacking integrated algorithm are implemented.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. The present invention has better generalization ability and overall prediction performance, which not only effectively improves the efficiency of mold slag defect analysis, but also provides new ideas and methods for in-depth research on the causes of mold slag defects.

[0048] 2. The present invention solves the problem of correlation analysis between high-frequency data fluctuations in continuous casting molds and mold slag defects by adopting a multi-model stacking integration algorithm that is rarely used in steel industry research.

[0049] 3. The present invention uses a sliding window to segment high-frequency feature data and combines it with multi-model cross-validation to transform the prediction limitations of a single model into the advantages of an integrated model, thereby improving the prediction accuracy and generalization ability of mold slag defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0051] Figure 1 Schematic diagram of the stacking integration method flow.

[0052] Figure 2 Schematic diagram of high-frequency features and defect location distribution.

[0053] Figure 3 Schematic diagram of sliding window.

[0054] Figure 4 Schematic diagram of important features of data in the window.

[0055] Figure 5 Schematic diagram of hyperparameter configuration for stacked ensemble models. DETAILED DESCRIPTION

[0056] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0057] Aiming at the problem of continuous casting mold slag defects, in order to improve on-site production and reduce the occurrence of mold slag defects caused by high-frequency data fluctuations in the continuous casting crystallizer, the present invention proposes a continuous casting mold slag defect analysis method based on a multi-model stacking integration algorithm. The high-frequency feature data is segmented by a sliding window, important features are extracted, and the stacking integration algorithm is combined to realize the prediction of mold slag defects and the range of occurrence locations. The integrated algorithm is superior to a single learning model and a voting classification model, has better generalization ability and overall prediction performance, and effectively improves the efficiency of mold slag defect analysis. Figure 1 For example, the specific contents include:

[0058] Step S1: Segment the high-frequency feature data extracted from the continuous casting mold into subsequences through a sliding window.

[0059] Specifically, the steps include:

[0060] Step S1.1: To analyze whether the main cause of mold slag defects is related to the fluctuation of each variable, the ratio of slab length to hot coil length is calculated to map the hot coil defect position to the high-frequency position of steelmaking, and obtain high-frequency feature data, such as Figure 2 shown.

[0061] Step S1.2: Based on the comparison of different sliding window lengths and step size splitting effects, select the sliding window parameters to find the parameters that can cover all high-frequency fluctuation features to the greatest extent, and avoid the features being affected by the window splitting position. Finally, 10 points are selected as a window length and 2 points as the step size of the sliding window. The sliding window is used to split the data into subsequences, as shown in the following example: Figure 3 shown.

[0062] Step S1.3: Further extract the data in the window based on the time domain characteristics and frequency domain characteristics. Combined with the statistical analysis and the ranking of the importance of the features of the correlation test and the experience of experts, select Figure 4 The important features of the example are used as model input parameters for in-depth analysis.

[0063] Step S2: Selection of stacked ensemble model base learners and meta-learners.

[0064] Specifically, the steps include:

[0065] Step S2.1: Selecting a meta-learner for the stacked ensemble model: Random Forest was used as the meta-learner. Selecting base learners for the stacked ensemble model: XGBoost and decision trees were used as base learners. Combining XGBoost and decision trees can effectively reduce the risk of a single model, increasing model robustness and diversity. A simple decision tree model can reduce the risk of XGBoost overfitting and balance model complexity.

[0066] Step S2.2: Divide the important feature data after dimensionality reduction into multiple sub-datasets through the self-service sampling method. Extract a certain number of new samples from each sub-dataset through resampling, and construct multiple decision tree models for the sub-datasets. The result categories of each decision tree model are statistically analyzed through the form of average or majority voting principles. The category with the most votes is used as the overall prediction output of the model to obtain the final prediction result, which can effectively reduce the probability of model overfitting and improve the generalization ability.

[0067] By adopting a method based on a multi-model stacking integration algorithm, the shortcomings of the correlation analysis between high-frequency data fluctuations in the continuous casting mold and the protection slag defects, as well as the problem of the limited application of the stacking integration algorithm in steel industry research, are solved.

[0068] Step S3: Build a stacked ensemble model based on the model selected in step S2.

[0069] Specifically, the method includes the following steps:

[0070] Step S3.1: Divide the training data into 5 parts, 4 for training the model and 1 as a test set for prediction. Use each part as a test set for prediction and combine the predicted labels of each part as a new feature.

[0071] Specifically, 80% of the important feature data after dimensionality reduction were randomly sampled as training data and 20% as test data. The XGBoost and decision tree models in the base learner were constructed for the training data, and the two machine learning models were subjected to five-fold cross-validation on the training data.

[0072] Step S3.2: Similarly, perform two five-fold cross-validations on the test data. Merge the predicted labels of each set as new features.

[0073] Step S3.3: The prediction results of the two independent machine learning models, that is, the new features, are put into the random forest model as new features of the training data for secondary training, combining the advantages of each model to improve the overall model effect and prediction performance.

[0074] Step S4: Based on the constructed stacked ensemble model, the most appropriate hyperparameter configuration of the current model is selected through cross-validation, grid search and other methods. The hyperparameter configuration includes the number of base estimators (n_estimators), the maximum tree depth (max_depth), etc. The specific parameter configuration is as follows: Figure 5 As shown in Figure 2, it can effectively improve the model learning effect.

[0075] By integrating machine learning models such as XGBoost, decision tree, and random forest, and segmenting high-frequency feature data through sliding windows and combining multi-model cross-validation, the prediction limitations of a single model are transformed into the advantages of an integrated model, thereby improving the prediction accuracy and generalization ability of mold slag defects, and providing new methods and ideas for in-depth research on the causes of mold slag defects.

[0076] The present invention also provides a continuous casting protection slag defect analysis system based on a multi-model stacking integrated algorithm. The continuous casting protection slag defect analysis system based on a multi-model stacking integrated algorithm can be implemented by executing the process steps of the continuous casting protection slag defect analysis method based on a multi-model stacking integrated algorithm. That is, those skilled in the art can understand the continuous casting protection slag defect analysis method based on a multi-model stacking integrated algorithm as a preferred implementation of the continuous casting protection slag defect analysis system based on a multi-model stacking integrated algorithm.

[0077] According to the present invention, a continuous casting mold slag defect analysis system based on a multi-model stacking integrated algorithm includes:

[0078] Module M1, splits the extracted high-frequency feature data into subsequences through a sliding window and selects important features;

[0079] Module M2 selects the base learner and meta-learner of the stacked ensemble model, processes important features, and obtains important features after dimensionality reduction;

[0080] Module M3, constructs a stacked ensemble model based on base learners and meta-learners;

[0081] Module M4, based on the stacked ensemble model, selects hyperparameter configuration optimization and outputs the final prediction result.

[0082] In more preferred embodiments, the module M1 includes:

[0083] Module M1.1: Calculate the ratio of slab length to hot coil length to map the hot coil defect location to the high-frequency position of steelmaking to obtain high-frequency feature data;

[0084] Module M1.2: Select sliding window parameters that can cover the most high-frequency feature data, and use the sliding window to perform subsequence segmentation on the high-frequency feature data. The sliding window parameters include step size and window length.

[0085] Module M1.3 further extracts the subsequences within the window based on time domain characteristics and frequency domain characteristics, and selects important features as model input parameters based on importance ranking and expert experience.

[0086] In more preferred embodiments, the module M2 includes:

[0087] Module M2.1: Using XGBoost and decision trees as base learners and random forest as meta-learner, we perform data dimensionality reduction on important features.

[0088] Module M2.2: Divide the important features after dimensionality reduction into a specified number of sub-datasets through the self-service sampling method, resample to extract a certain number of new samples, construct the same number of decision trees for the sub-datasets, count the result categories of each decision tree according to the voting method, and use the result category with the most votes as the final result category of the decision tree.

[0089] In more preferred embodiments, the module M3 includes:

[0090] Module M3.1: Randomly divide the important features after dimensionality reduction into 5 parts, 4 of which are used as training sets and 1 as test set. Build XGBoost and decision tree models, perform 5-fold cross-validation on the training sets, and obtain the trained XGBoost and decision tree models.

[0091] In module M3.2, the trained XGBoost and decision tree perform five-fold cross-validation on the test set, and merge the predicted labels as the prediction results;

[0092] Module M3.3: The prediction results are used as new features of the training data to be put into the random forest model for secondary training to obtain the trained random forest model.

[0093] The module M4 inputs the prediction result into the trained random forest model with hyperparameter configuration and outputs the final prediction result.

[0094] According to the present invention, a computer-readable storage medium storing a computer program is provided, and when the computer program is executed by a processor, the steps of the continuous casting mold slag defect analysis method based on the multi-model stacking integrated algorithm are implemented.

[0095] According to the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the continuous casting protective slag defect analysis method based on a multi-model stacking integrated algorithm are implemented.

[0096] Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices, modules, and units provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0097] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A continuous casting mold slag defect analysis method based on a multi-model stacking integrated algorithm, characterized in that: include: Step S1: Segment the extracted high-frequency feature data into subsequences through a sliding window and select important features; Step S2: Select the base learner and meta-learner of the stacked ensemble model, process the important features, and obtain the important features after dimensionality reduction and the final prediction results; Step S3: construct a stacked ensemble model based on the base learner and the meta learner; Step S4: According to the stacked ensemble model, select the hyperparameter configuration optimization and output the final prediction result.

2. The continuous casting mold slag defect analysis method based on multi-model stacking integrated algorithm according to claim 1 is characterized in that: The step S1 comprises: Step S1.1, mapping the hot coil defect location to the steelmaking high-frequency location by calculating the ratio of the slab length to the hot coil length to obtain high-frequency feature data; Step S1.2: Select sliding window parameters that can cover the most high-frequency feature data, and use the sliding window to perform subsequence segmentation on the high-frequency feature data. The sliding window parameters include step size and window length. Step S1.3: Further extract the subsequences within the window based on their time domain characteristics and frequency domain characteristics, and select important features as model input parameters based on importance ranking and expert experience.

3. The continuous casting mold slag defect analysis method based on multi-model stacking integrated algorithm according to claim 1 is characterized in that: The step S2 comprises: Step S2.1: Use XGBoost and decision tree as base learners and random forest as meta-learner to perform data dimensionality reduction on important features; Step S2.2: Divide the important features after dimensionality reduction into a specified number of sub-datasets through the self-service sampling method, resample to extract a certain number of new samples, construct the same number of decision trees for the sub-datasets, count the result categories of each decision tree according to the voting method, and take the result category with the most votes as the final result category of the decision tree.

4. The continuous casting mold slag defect analysis method based on multi-model stacking integrated algorithm according to claim 1 is characterized in that: The step S3 comprises: Step S3.1: Randomly divide the important features after dimensionality reduction into 5 parts, 4 of which are used as training sets and 1 as test set, construct XGBoost and decision tree, and perform 5-fold cross-validation on the training sets to obtain the trained XGBoost and decision tree; Step S3.2: After training, XGBoost and decision tree are each subjected to five-fold cross-validation on the test set, and the predicted labels are merged as the prediction results. Step S3.3: The prediction results are used as new features of the training data and then put into the random forest model for secondary training to obtain a trained random forest model; In step S4, the prediction result is input into the trained random forest model with hyperparameter configuration, and the final prediction result is output.

5. A continuous casting mold slag defect analysis system based on a multi-model stacking integrated algorithm, characterized in that: include: Module M1, splits the extracted high-frequency feature data into subsequences through a sliding window and selects important features; Module M2 selects the base learner and meta-learner of the stacked ensemble model, processes important features, and obtains important features after dimensionality reduction; Module M3, constructs a stacked ensemble model based on base learners and meta-learners; Module M4, based on the stacked ensemble model, selects hyperparameter configuration optimization and outputs the final prediction result.

6. The continuous casting mold slag defect analysis system based on multi-model stacking integrated algorithm according to claim 5 is characterized in that: The module M1 includes: Module M1.1: Calculate the ratio of slab length to hot coil length to map the hot coil defect location to the high-frequency position of steelmaking to obtain high-frequency feature data; Module M1.2: Select sliding window parameters that can cover the most high-frequency feature data, and use the sliding window to perform subsequence segmentation on the high-frequency feature data. The sliding window parameters include step size and window length. Module M1.3 further extracts the subsequences within the window based on time domain characteristics and frequency domain characteristics, and selects important features as model input parameters based on importance ranking and expert experience.

7. The continuous casting mold slag defect analysis system based on multi-model stacking integrated algorithm according to claim 5 is characterized in that: The module M2 includes: Module M2.1: Using XGBoost and decision trees as base learners and random forest as meta-learner, we perform data dimensionality reduction on important features. Module M2.2: Divide the important features after dimensionality reduction into a specified number of sub-datasets through the self-service sampling method, resample to extract a certain number of new samples, construct the same number of decision trees for the sub-datasets, count the result categories of each decision tree according to the voting method, and use the result category with the most votes as the final result category of the decision tree.

8. The continuous casting mold slag defect analysis system based on multi-model stacking integrated algorithm according to claim 5 is characterized in that: The module M3 includes: Module M3.1: Randomly divide the important features after dimensionality reduction into 5 parts, 4 of which are used as training sets and 1 as test set. Build XGBoost and decision tree models, perform 5-fold cross-validation on the training sets, and obtain the trained XGBoost and decision tree models. In module M3.2, the trained XGBoost and decision tree perform five-fold cross-validation on the test set, and merge the predicted labels as the prediction results; Module M3.3: Use the prediction results as new features of the training data to enter the random forest model for secondary training to obtain the trained random forest model; The module M4 inputs the prediction result into the trained random forest model with hyperparameter configuration and outputs the final prediction result.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the continuous casting mold slag defect analysis method based on a multi-model stacking integrated algorithm according to any one of claims 1 to 4 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the continuous casting mold slag defect analysis method based on a multi-model stacking integrated algorithm according to any one of claims 1 to 4 are implemented.

Images