Inclusion unqualified influence factor analysis method and equipment

Abstract

Embodiments of the present application provide a kind of inclusion unqualified influence factor analysis method and equipment, belong to material science and technology field.The method includes: the real-time acquisition of multiple-source heterogeneous data in production environment, the multiple-source heterogeneous data include the influence factor of inclusion;The first preprocessed data is obtained by pre-processing the multiple-source heterogeneous data;Predictive features are extracted from the first preprocessed data;Prediction is carried out based on the predictive features and the prediction model deployed in advance, to obtain the prediction result, the prediction result includes inclusion level;Model interpretability technology analysis is carried out based on the prediction model, predictive features and the prediction result, to obtain the causal relationship between each prediction feature and the prediction result;And the influence factor analysis result is obtained based on the causal relationship, and the analysis result is visually displayed.The method is used to solve at least one problem existing in the prior art when dealing with inclusion.

Description

Technical Field

[0001] This invention relates to the field of materials science and technology, and specifically to a method and equipment for analyzing the influencing factors of inclusion defects. Background Technology

[0002] Inclusions, as common defects in metallic materials, significantly affect their mechanical properties, processing performance, and service performance. During the smelting and processing of iron and steel, molten steel easily interacts with slag, refractory materials, and reaction products to form non-metallic inclusions. These inclusions disrupt the continuity of the steel matrix, increase the inhomogeneity of the steel structure, and consequently affect the mechanical properties of steel materials, such as plasticity, toughness, and fatigue strength. They can also easily cause product defects during the forming process.

[0003] Traditional inclusion analysis methods rely heavily on manual inspection and experience-based judgment, which has several limitations. On one hand, qualitative detection of inclusions is generally performed using scanning electron microscopy (SEM) and energy dispersive spectroscopy (EDS), but this method is time-consuming and highly random, failing to meet the efficiency requirements of modern production. On the other hand, the limitations of researchers' expertise and the level of software automation in traditional image analysis methods are major challenges. Artifacts and noise in images can render traditional analysis methods ineffective, making analysis more difficult. Manual classification methods rely heavily on the analyst's expertise and are extremely inefficient. Furthermore, traditional methods often only allow for post-hoc analysis of existing inclusions, failing to achieve predictive quality control, leading to significantly increased quality costs and ineffective control.

[0004] Therefore, existing technologies for handling inclusions suffer from low efficiency, poor accuracy, insufficient predictive performance, and low level of intelligence. Furthermore, the various analytical methods are relatively independent and lack systematic data integration and comprehensive analysis capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a method and equipment for analyzing the influencing factors of inclusion defects, in order to solve at least one problem existing in the prior art when dealing with inclusions.

[0006] To achieve the above objectives, in a first aspect, embodiments of the present invention provide a method for analyzing the influencing factors of inclusion nonconformity. The method includes: acquiring multi-source heterogeneous data in a production environment in real time, the multi-source heterogeneous data including influencing factors of inclusions; preprocessing the multi-source heterogeneous data to obtain first preprocessed data; extracting predictive features from the first preprocessed data; making predictions based on the predictive features and a pre-deployed predictive model to obtain prediction results, the prediction results including inclusion levels; performing model interpretability analysis based on the predictive model, predictive features, and prediction results to obtain causal relationships between each predictive feature and the prediction results; and obtaining influencing factor analysis results based on the causal relationships, and visualizing the analysis results.

[0007] Optionally, real-time acquisition of multi-source heterogeneous data in the production environment includes: real-time acquisition of multi-source heterogeneous data in the production environment from the MES system and the testing and analysis system through OGG-based stream processing, wherein the multi-source heterogeneous data includes process parameter data, material composition data, equipment operation data, environmental data, and quality testing data.

[0008] Optionally, the preprocessing includes data cleaning, data integration, data transformation, and data reduction. Specifically, data cleaning includes detecting outliers in the data and processing them based on deletion, interpolation, or model prediction; data integration includes standardizing the data format, data units, and data standards; data transformation includes standardizing, normalizing, and discretizing the data; and data reduction includes dimensionality reduction of the data.

[0009] Optionally, the deployment method of the prediction model includes: acquiring multi-source heterogeneous data from historical production, each of the multi-source heterogeneous data corresponding to an inclusion-level label; preprocessing the multi-source heterogeneous data to obtain second preprocessed data; extracting features from the second preprocessed data based on feature engineering, and selecting prediction features from the extracted features; constructing a dataset based on the prediction features and their corresponding inclusion-level labels; training multiple pre-selected models based on the dataset, and optimizing the parameters of the multiple models during the training process using grid search, random search, or Bayesian optimization to obtain the optimal parameters of each model, wherein the multiple models include XGBoost, random forest, support vector machine, and neural network; combining the multiple models based on an ensemble strategy to obtain the prediction model; and deploying the prediction model to the production environment.

[0010] Optionally, extracting features from the second preprocessed data and selecting predictive features from the extracted features includes: based on feature extraction, extracting basic features from the second preprocessed data, the basic features including statistical features, time-domain features, and frequency-domain features; based on feature construction, constructing elemental ratio features, process parameter interaction features, and historical statistical features from the second preprocessed data, wherein the elemental ratio features include Ca / Al ratio and Mn / S ratio, the process parameter interaction features include the product of temperature and time and the relationship between cooling rate and temperature gradient, and the historical statistical features include the moving average and rate of change of the parameters; and based on feature selection, selecting the most relevant feature subset from the features obtained by feature extraction and feature construction to obtain the predictive features, the predictive features including the final carbon content of the electric furnace, the Al content at the beginning of LF refining, the VD vacuum time, the calcium treatment amount, and the LF refining temperature.

[0011] Optionally, the method further includes periodically acquiring multi-source heterogeneous data to evaluate the prediction accuracy of the prediction model after deploying the prediction model to the production environment.

[0012] Optionally, model interpretability analysis is performed based on the prediction model, prediction features, and prediction results to obtain the causal relationship between each prediction feature and the prediction result, including: calculating the importance score of each prediction feature to the prediction result based on feature importance, permutation importance, or SHAP value; analyzing the prediction features and prediction results based on partial dependency graphs, individual conditional expectation graphs, or cumulative local effects graphs to obtain the relationship pattern between each prediction feature and the prediction result; analyzing the prediction features and prediction results based on SHAP interaction values, two-dimensional partial dependency graphs, or statistical interaction tests to obtain the synergistic and antagonistic effects between the prediction features; and analyzing the importance score, relationship pattern, synergistic effect, and antagonistic effect based on causal discovery algorithms or causal forests to obtain the causal relationship between the prediction feature and the prediction result.

[0013] Optionally, the method further includes: receiving multi-source heterogeneous data input or adjusted by the user based on an interactive interface; making predictions based on the multi-source heterogeneous data input or adjusted by the user to obtain prediction results; and visually displaying the prediction results.

[0014] Optionally, the method further includes: receiving constraint information input by the user through an interactive interface; making predictions in the prediction model based on the constraint information to obtain multiple prediction results; selecting the optimal value of multi-source heterogeneous data from the multiple prediction results based on the target at the minimum inclusion level and the constraint information; and visually displaying the selected value and providing optimization suggestions based on a preset rule base.

[0015] On the other hand, embodiments of the present invention provide an apparatus for analyzing the influencing factors of inclusion nonconformity. The apparatus includes a memory and a processor, the processor being used to run a program, wherein the program, when run, is used to execute any of the methods described above.

[0016] The above technical solution can utilize data mining technology to analyze the main influencing factors of inclusion defects. By integrating multi-source heterogeneous data, constructing predictive models, and realizing quantitative analysis and visualization of influencing factors, it can effectively solve the technical problems of low efficiency, poor accuracy, and strong subjectivity in traditional inclusion analysis methods. It can significantly improve the intelligence level and prediction accuracy of inclusion level control, provide a brand-new technical solution for quality control in the metal material manufacturing process, and help achieve the goal of controlling inclusion levels from the source.

[0017] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the method for analyzing the influencing factors of inclusion non-compliance provided in this embodiment of the invention; Figure 2 This is a flowchart illustrating another method for analyzing factors affecting inclusion defects provided in this embodiment of the invention. Figure 3 This is a schematic diagram of the structure of the device provided in an embodiment of the present invention.

[0019] Explanation of reference numerals in the attached figures 101 Processor 102 Memory 103 bus 10 devices Detailed Implementation The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0020] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0021] Studies have shown that inclusions are an important factor affecting the surface quality of IF steel. The rolling effect of hot rolling process causes inclusions that are concentrated at 1 / 4 of the width of the continuously cast billet to migrate to the edge, ultimately resulting in the highest inclusion index at the edge of the hot-rolled plate. This indicates that the inclusion aggregation zone has a hereditary nature during the rolling process.

[0022] Existing technologies mainly employ traditional methods for inclusion detection and analysis, including metallographic rating, ASPEX method, non-destructive testing, inclusion extraction, and fatigue testing.

[0023] Metallographic grading and the ASPEX method primarily involve observing metallographic samples using an optical microscope or scanning electron microscope (SEM) and grading inclusions using a standard atlas comparison method. The ASPEX method is an automated SEM analysis method that uses a large-area automated SEM to detect inclusions, providing information on their size, quantity, distribution, and composition. The specific operation includes sample preparation, microscopic observation, image acquisition, inclusion identification, and classification grading. However, both methods have drawbacks in terms of detection accuracy and applicability. Since the observation plane is not necessarily the primary stress surface, it is impossible to guarantee the observation of the largest size area of ​​the inclusion, resulting in generally smaller detection results. The frosted glass projection method and direct observation through the microscope eyepiece are cumbersome to operate, requiring comparison with standard atlases, and are prone to significant errors due to visual limitations, leading to low measurement accuracy. The ASPEX method has high equipment costs, which cannot meet the needs of most manufacturers in China. Furthermore, these methods require highly skilled operators, making it difficult to guarantee the consistency and repeatability of the analytical results.

[0024] Non-destructive testing (NDT) methods include ultrasonic testing, magnetic testing, and X-ray diffraction. These methods identify internal defects and inclusions in steel through non-destructive testing. Ultrasonic testing utilizes the reflection and attenuation characteristics of sound waves propagating through materials to detect inclusions; magnetic testing identifies defects by measuring changes in the material's magnetism; and X-ray diffraction detects internal structural anomalies by analyzing diffraction patterns. However, these methods have limited accuracy and identification capabilities, cannot distinguish between defect categories, and are insufficient in identifying the type of inclusion. While NDT methods such as ultrasonic testing can detect larger steel volumes, reducing the probability of missing large inclusions, they are insensitive to small inclusions and cannot meet the quality control requirements for high-cleanliness materials. Furthermore, NDT methods typically only provide information about the presence of defects and cannot provide detailed information such as the specific composition and morphology of inclusions.

[0025] Inclusion extraction methods include chemical dissolution and electrolytic dissolution. Chemical dissolution mainly uses acid dissolution, which dissolves the steel matrix with an acid solution to extract inclusions for analysis. Electrolytic dissolution includes large-sample electrolysis and non-aqueous solution electrolysis, which separates inclusions through electrolysis, followed by morphological observation and compositional analysis. In particular, non-aqueous solution electrolysis, under conditions of electrolyte pH maintained between 7 and 8, voltage of 22.5V, and current density of 20mA / cm², can stably extract fine inclusions of approximately 5μm. The disadvantages of this method involve several aspects, including sample representativeness, analytical efficiency, and completeness. Inclusion concentration methods are only suitable for detecting small-volume steel samples; furthermore, the original distribution of inclusions in the sample cannot be determined after extraction. In acid dissolution, large inclusions and clustered inclusions are broken down, losing their original morphology, and unstable phases within the inclusions dissolve in the acidic environment. Non-aqueous solution electrolysis is less accurate than large-sample electrolysis for detecting randomly distributed large inclusions, and the electrolysis process is time-consuming, requiring strict control of electrolysis parameters. These limitations make this method unsuitable for the real-time requirements of modern production.

[0026] Fatigue testing is considered the most reliable method for detecting large inclusions, especially suitable for high-strength steel. This method involves conducting fatigue tests on materials to observe the initiation and propagation of fatigue cracks, thereby inferring the presence and impact of inclusions. The statistical extreme value method (SEV) can predict the maximum inclusion size and fatigue strength in steels of different volumes. The main drawbacks of this method are efficiency, cost, and destructiveness. It requires numerous tests, is time-consuming and costly. The long testing cycle cannot meet the needs of real-time quality control. It is a destructive test and cannot be used for online monitoring. Furthermore, fatigue testing typically requires specialized testing equipment and professional technicians, which is difficult for most companies to afford, limiting its widespread application.

[0027] Weld-based inclusion analysis utilizes tungsten inert gas (TIG) welding to cause inclusions to float, forming visible "floating objects." The process involves: cutting strip-shaped steel samples with a thickness of 1-10 mm; welding with argon gas at a flow rate of 8-15 L / min, a welding current of 80-120 A, and a welding speed of 200-550 mm / min; visually observing the number of "floating objects" on the weld surface; observing the morphological distribution of inclusions using a scanning electron microscope (SEM); and analyzing the composition using energy dispersive X-ray spectroscopy (EDS). The main drawbacks of this method are its applicability and technical limitations. It is primarily suitable for oxide inclusion analysis, with limited application to sulfide and nitride analysis. The applicable sample thickness range is limited; thin samples require parameter adjustments to prevent complete penetration. The welding process may introduce new inclusions or alter the morphology of existing inclusions, affecting the accuracy of the analysis results. Equipment costs are relatively high; although lower than electron beam melting, specialized welding equipment is still required, increasing the technical threshold and investment costs for enterprises.

[0028] In summary, existing analytical methods have the following drawbacks: (1) Inefficient: Traditional analysis methods take several days or even weeks to complete the analysis, which is difficult to meet the real-time requirements of modern production.

[0029] (2) Poor accuracy: Manual analysis methods have problems of strong subjectivity and poor consistency, making it difficult to guarantee the reliability of the analysis results.

[0030] (3) Insufficient predictability: Existing technologies mainly perform post-hoc analysis on the inclusions that have already been generated, and cannot achieve predictive quality control.

[0031] (4) Low integration: Various analysis methods are relatively independent and lack systematic data integration and comprehensive analysis capabilities.

[0032] (5) Low level of intelligence: There is a lack of application of advanced data mining and machine learning technologies, making it impossible to automatically discover hidden patterns and influencing factors from massive amounts of data.

[0033] Due to the aforementioned shortcomings, traditional methods for detecting and analyzing inclusions are increasingly unable to meet the efficiency and quality requirements of modern production.

[0034] With the rapid development of big data and artificial intelligence technologies, data-driven materials research and development has rapidly evolved into a new paradigm that transforms traditional trial-and-error methods—the so-called fourth paradigm of materials research and development. Machine learning technology has made significant progress in material feature construction and screening, material microstructure and performance prediction, and material optimization design, providing a new technical path for inclusion quality analysis. However, most existing data mining applications focus on general quality control, without addressing the specific problem of inclusion non-compliance through specialized data mining methods and systems. Therefore, existing data mining applications suffer from the following problems: (1) The problem of being unable to effectively identify and quantify the main influencing factors that cause inclusions to be unqualified: Traditional methods can only describe and classify the inclusions themselves, and it is difficult to analyze in depth the root causes and key influencing factors that lead to the formation of inclusions, and cannot provide accurate guidance for process optimization.

[0035] (2) The problem of difficulty in predictive control of the quality of inclusions and process optimization: Existing technologies mainly remain at the post-event detection stage, and cannot predict future quality problems based on historical data and production status, so as to realize the transformation from "post-event detection" to "pre-event prevention".

[0036] (3) Problems of insufficient integration and utilization of multi-source heterogeneous data: Existing technologies cannot make full use of the massive amount of multi-source heterogeneous data accumulated in the production process, and lack effective methods to deeply integrate materials science knowledge with data mining algorithms.

[0037] To address the above problems, embodiments of the present invention provide a method for analyzing the influencing factors of inclusion defects, such as... Figure 1 As shown, the method includes the following steps S1 to S6.

[0038] Step S1: Real-time acquisition of multi-source heterogeneous data in the production environment, including influencing factors of inclusions.

[0039] Step S2: Preprocess the multi-source heterogeneous data to obtain the first preprocessed data.

[0040] Step S3: Extract predictive features from the first preprocessed data.

[0041] Step S4: Based on the predicted features and the pre-deployed prediction model, a prediction is made to obtain the prediction result, which includes the inclusion level. The inclusion level is related to the inclusion quality, and different steel grades correspond to different standards, used to determine whether the steel grade is qualified.

[0042] Step S5: Based on the prediction model, prediction features, and prediction results, perform model interpretability analysis to obtain the causal relationship between each prediction feature and the prediction results.

[0043] Step S6: Obtain the analysis results of influencing factors based on the causal relationship, and visualize the analysis results.

[0044] The method proposed in this invention first acquires raw data from various data sources at the production site, and then preprocesses the collected raw data. During feature extraction, specific predictive features are extracted from the preprocessed data based on materials science knowledge and domain expert experience. The extracted predictive features are then input into a predictive model to obtain prediction results for inclusion levels. After obtaining the prediction results, model interpretability analysis can further identify key factors affecting inclusion levels and their mechanisms of action, obtaining the causal relationship between the predictive features and the prediction results. Finally, based on the causal relationship, further analysis results of influencing factors are obtained and a visual representation is provided to the user. The above technical solution utilizes data mining technology to analyze the main influencing factors of inclusion non-compliance. By integrating multi-source heterogeneous data, constructing a predictive model, and achieving quantitative analysis and visualization of influencing factors, it effectively solves the technical problems of low efficiency, poor accuracy, and strong subjectivity in traditional inclusion analysis methods. It significantly improves the intelligence level and prediction accuracy of inclusion level control, providing a new technical solution for quality control in the metal material manufacturing process, and helping to achieve the goal of controlling inclusion levels from the source. In practical implementation, a layered modular architecture can be adopted between each step to ensure the scalability, maintainability, and upgradeability of the system, providing a good architectural foundation for continuous technological innovation and iterative optimization.

[0045] Each step of the method proposed in this invention integrates materials science knowledge and data mining techniques. In step S1, comprehensive and accurate raw data is acquired from various data sources at the production site, including process parameter data, material composition data, equipment operation data, environmental data, and quality inspection data. Data is collected in real-time from multiple sources, such as the MES system and testing systems, using OGG technology and a streaming processing framework. A data quality monitoring mechanism is established to ensure the integrity, accuracy, and traceability of the data. This multi-source data acquisition mechanism breaks through the limitations of traditional data acquisition, laying a solid data foundation for subsequent in-depth analysis.

[0046] The data quality monitoring mechanism is used to unify data formats across different systems, optimize the performance of real-time data acquisition, and monitor and ensure data quality. In practical applications, by developing standardized data interface protocols, designing efficient data transmission mechanisms, and establishing a comprehensive data quality monitoring system, efficient acquisition and integration of multi-source heterogeneous data can be achieved.

[0047] In step S2, the collected raw data is preprocessed to provide a high-quality data foundation for subsequent analysis. Data preprocessing includes four main stages: data cleaning, data integration, data transformation, and data reduction. The data cleaning stage handles missing values, outliers, and noisy data. Statistical and machine learning-based methods are used to detect outliers, and these outliers are processed through deletion, interpolation, or model prediction. The data integration stage addresses the inconsistency problem of multi-source data by standardizing data formats, units, and standards, and resolving data conflicts. The data transformation stage standardizes, normalizes, and discretizes the data to make data with different dimensions and distributions comparable. The data reduction stage uses methods such as principal component analysis and feature selection to reduce data redundancy and improve the efficiency of subsequent analysis. This systematic data preprocessing process significantly improves data quality and provides a reliable guarantee for subsequent model prediction.

[0048] In step S4, a pre-deployed prediction model is used for prediction. The deployment method of the prediction model includes: acquiring multi-source heterogeneous data from historical production, each multi-source heterogeneous data corresponding to an inclusion level label; preprocessing the multi-source heterogeneous data to obtain second preprocessed data; extracting features from the second preprocessed data based on feature engineering, and selecting prediction features from the extracted features; constructing a dataset based on the prediction features and their corresponding inclusion level labels; training multiple pre-selected models based on the dataset, and optimizing the parameters of the multiple models during the training process using grid search, random search, or Bayesian optimization to obtain the optimal parameters of each model. The multiple models include XGBoost, random forest, support vector machine, and neural network; combining the multiple models based on an ensemble strategy to obtain the prediction model; and deploying the prediction model to the production environment.

[0049] During model training, cross-validation can be used to evaluate model performance. When combining multiple models, the performance characteristics of different algorithms in inclusion prediction are compared. Integration strategies such as voting, weighted averaging, and stacking are employed to combine the prediction results of multiple base models, improving prediction accuracy and stability. This multi-algorithm integrated prediction model construction method significantly improves prediction accuracy and stability. By integrating multiple algorithms into the prediction model, the complementary advantages of each algorithm are achieved, overcoming the limitations of a single algorithm and improving the stability and reliability of model performance.

[0050] In some embodiments of the present invention, a model performance monitoring mechanism is also established. After the prediction model is deployed to the production environment, multi-source heterogeneous data is periodically acquired to evaluate the prediction accuracy of the prediction model.

[0051] Furthermore, feature engineering, based on materials science knowledge and domain expert experience, extracts and constructs predictive features from preprocessed data. Based on feature engineering, the process of extracting features from the second preprocessed data and selecting predictive features from the extracted features includes three main steps: feature extraction, feature construction, and feature selection.

[0052] Feature extraction is used to directly extract basic features from the second preprocessed data. The basic features include statistical features, time-domain features, and frequency-domain features. Statistical features include mean, variance, skewness, kurtosis, and quantiles, etc. Time-domain features include trend, periodicity, and abrupt change points, etc. Frequency-domain features include dominant frequency and spectral energy, etc.

[0053] Feature construction, based on domain knowledge, creates new features used to construct elemental ratio features, process parameter interaction features, and historical statistical features from the second preprocessed data. The elemental ratio features include the Ca / Al ratio and Mn / S ratio; the process parameter interaction features include the product of temperature and time and the relationship between cooling rate and temperature gradient; and the historical statistical features include the moving average and rate of change of the parameters. In this step, by establishing a mapping relationship between the features and the inclusion formation mechanism, the constructed features are based on the physicochemical mechanism of inclusion formation and possess strong explanatory power.

[0054] Feature selection can employ filtering, wrapping, or embedding methods to select the most relevant subset of features from the features extracted and constructed, thereby obtaining the predicted features, which reduces data dimensionality and improves model performance. In some embodiments of the present invention, the predicted features include the final carbon content of the electric furnace, the Al content at the initial stage of LF refining, the VD vacuum time, the calcium treatment amount, and the LF refining temperature.

[0055] Through the feature engineering methods based on materials science knowledge, the metallurgical principles of inclusion formation were analyzed in depth, demonstrating the deep integration of data science and materials science. By combining materials science principles and expert experience, physically meaningful predictive features were constructed, effectively improving the model's prediction accuracy, interpretability, and physical significance.

[0056] Step S5 involves influencing factor analysis, which uses model interpretability techniques to analyze the mechanism of action and contribution of each influencing factor to the inclusion level, including four main steps: feature importance analysis, feature effect analysis, interaction effect analysis, and causal relationship inference.

[0057] Feature importance analysis employs model-based methods such as feature importance, permutation importance, or SHAP value to calculate the importance score of each predicted feature to the prediction result.

[0058] Feature effect analysis employs methods such as partial dependency graphs, individual conditional expectation graphs, or cumulative local effect graphs to analyze the relationship between the predicted features and the predicted results, thereby obtaining the relationship patterns between each predicted feature and the predicted results. For example, a partial dependency graph can be used to demonstrate how changes in the eigenvalues ​​of predicted features affect the predicted results.

[0059] Interaction effect analysis employs methods such as SHAP interaction values, two-dimensional partial dependency graphs, or statistical interaction tests to analyze the predictive features and the predictive results, revealing the synergistic and antagonistic effects between the predictive features. In other words, it analyzes the combined impact of two or more feature combinations on the predictive results, revealing the interactions between features.

[0060] Causal relationship inference employs causal discovery algorithms, causal forests, and other methods to further explore causal relationships from correlations. It analyzes the importance scores, relationship patterns, synergistic effects, and antagonistic effects to obtain the causal relationship between the predicted features and the predicted results. The causal inference results are then verified by combining domain knowledge and experimental design.

[0061] This model-interpretive analysis method combines data mining results with materials science principles to construct an interpretable framework for influencing factor analysis. It transforms the process from a "black box" to a "white box," enabling complex machine learning models to be understood and accepted by domain experts and significantly improving the interpretability, credibility, and practical value of the analysis results.

[0062] In step S6, the analysis results are presented in an intuitive and visual form, such as through charts, dashboards, etc., including a ranking chart of the importance of influencing factors, a relationship chart between features and inclusion levels, or a chart of process parameter optimization suggestions.

[0063] The method for obtaining the process parameter optimization suggestion diagram includes the following steps S601 to S604.

[0064] Step S601: Receive constraint information input by the user based on the interactive interface.

[0065] Step S602: Based on the constraint information, make a prediction in the prediction model to obtain multiple prediction results.

[0066] Step S603: Based on the target at the minimum inclusion level and the constraint information, select the optimal value of the multi-source heterogeneous data from the multiple prediction results.

[0067] Step S604: Visualize the selected values ​​and provide optimization suggestions based on a preset rule base.

[0068] In the above steps, optimization objectives and constraints were defined. The optimization objective is to minimize the predicted inclusion level. Constraint information includes multiple types, which limit the value range of the multi-source heterogeneous data. Constraint information can be classified into process constraints, composition constraints, cost constraints, and production constraints. Process constraints include the adjustable range of various parameters, such as the temperature must be between 1550-1650°C. Composition constraints include restrictions on the composition to ensure that the chemical composition of the final product meets the standards. Cost constraints include restrictions on certain operations, such as extending refining time or increasing alloy additions, which may increase costs. Production constraints mainly constrain the total processing time to avoid excessively long total processing time.

[0069] After defining the optimization objective and constraints, an optimization algorithm is invoked to solve the problem. The trained prediction model is treated as a "function," and a multi-objective optimization algorithm is used to search within this "function" and under the aforementioned constraints. The algorithm repeatedly tries different combinations of process parameters—that is, combinations of feature values ​​of different prediction features—inputting them into the prediction model to obtain prediction results, and calculating the merit score of each combination based on the optimization objective and constraints. Ultimately, the algorithm finds one or more Pareto optimal solutions as the optimal values ​​for the multi-source heterogeneous data. These values ​​satisfy the requirement that, under the current constraints, no further improvement in one objective can be achieved without harming other objectives. During the multi-objective optimization search, multiple objectives such as quality, cost, and time are simultaneously weighed. For example, if the search finds a solution where extending the refining time by 5 minutes could increase the yield by 1%, considering the capacity loss, the system ultimately recommends a solution that increases the yield by 0.8% without increasing the time.

[0070] The rule base includes various hard and soft constraints, including materials science principles such as "a Ca / Al ratio below 0.08 easily produces liquid calcium aluminate inclusions"; historical quality incident attribution analysis such as "when the final carbon content is <0.03% and the temperature is >1620°C, the probability of non-compliance is >70%"; and expert experience rules such as "when producing steel grades with high surface finish requirements, the calcium treatment amount needs to be increased by 10%". After obtaining the optimal values ​​from multi-source heterogeneous data, a radar chart can be used to compare a single set of recommended parameters with historical average parameters or standard parameters. Based on the rule base, the adjustment direction is clearly displayed, such as which parameters should be increased and which should be decreased.

[0071] Steps S601 to S604 explain the optimization suggestions based on user input. In actual production, after obtaining the causal relationship in step S5, the analysis results of influencing factors can be derived based on this relationship. Specifically, this involves comparing the parameters of the current batch / problem batch with the optimal range recommended by the model or the historical normal range. Parameters with significant deviations are automatically identified. For example, "Diagnosis found: The Al content (0.025%) in the initial stage of LF refining is significantly lower than the optimal lower limit (0.035%), which is the main risk point." Improvement suggestions are then given based on the rule base. For example, if the low Al content is diagnosed, a rule is directly triggered, such as, "It is recommended to control the Al content in the initial stage of LF refining between 0.035% and 0.045%." Afterwards, a structured analysis report is automatically generated and visualized based on the influencing factor analysis results. This visualization and decision support function enables production personnel to understand and apply complex analysis results, realizing the transformation of technological value into production value.

[0072] Furthermore, the method also includes: receiving multi-source heterogeneous data input or adjusted by the user based on an interactive interface; making predictions based on the multi-source heterogeneous data input or adjusted by the user to obtain prediction results; and visually displaying the prediction results.

[0073] This section provides interactive analysis, allowing users to perform hypothesis testing and scenario simulations. Users can adjust process parameter values ​​and view real-time changes in predicted inclusion levels. The interactive interface displays all key process parameter values ​​for the current production batch or a typical batch. Users can adjust these parameters using sliders or input boxes to obtain the model's prediction of inclusion levels under the current parameters.

[0074] Through the aforementioned decision support mechanism combining data-driven approaches and mechanistic analysis, and the human-machine collaborative optimization decision-making process, the feasibility and effectiveness of optimization suggestions are verified, multi-objective conflicts are coordinated and resolved, and the interpretability of decision results is ensured. By combining quantitative analysis from data mining with mechanistic analysis from materials science, a balance between scientific rigor and practicality is achieved, providing scientifically sound and practical decision support for process optimization. In practical applications, this decision support technology can help enterprises achieve significant improvements at the inclusion level, creating substantial economic benefits.

[0075] The method provided by this invention significantly improves the efficiency of analyzing factors affecting inclusion non-conformities through automated data acquisition and intelligent analysis algorithms. Analysis tasks that traditional manual methods require days or even weeks to complete can be finished in hours with this invention, and the accuracy of the analysis results is greatly improved, effectively avoiding analysis errors caused by human factors. In the prediction step, analysis based on user input is supported, thereby predicting potential future quality problems based on historical data and production status, realizing a shift from "post-event detection" to "pre-event prevention." By monitoring and analyzing various parameters in the production process, enterprises can identify potential quality problems in advance and take corresponding preventative measures. Unlike traditional qualitative analysis methods, this invention can quantify the contribution of each influencing factor to the inclusion level, providing precise guidance for process optimization. Through model interpretability technology, key factors affecting the inclusion level and their mechanisms of action can be accurately identified. This quantitative analysis provides a scientific basis for targeted process improvements, shifting process optimization from experience-driven to data-driven, greatly improving the accuracy and effectiveness of optimization. By mining and analyzing process parameters during production, this invention can identify the optimal combination of process parameters, thereby improving production efficiency and product quality. This data mining-based process parameter optimization method is more efficient and accurate than traditional trial-and-error methods, significantly shortening the optimization cycle, reducing optimization costs, and improving optimization results. Enhanced process parameter optimization capabilities enable enterprises to continuously improve production processes, thereby continuously improving product quality and production efficiency. Data mining transforms tacit expert knowledge and experience into explicit mathematical models and rules, realizing the digitization and systematization of production knowledge. This knowledge accumulation mechanism not only avoids knowledge loss due to personnel turnover but also provides a solid foundation for continuous improvement and innovation within enterprises. As production data continues to accumulate, the model's predictive accuracy and interpretability will continue to improve, forming a virtuous cycle of knowledge accumulation.

[0076] Furthermore, this invention can be implemented through a modular architecture design, with modules exchanging data and calling functions through standardized interfaces, ensuring the system's scalability and maintainability. The system can easily integrate new data sources, add new analytical algorithms, and expand into new application scenarios, exhibiting strong adaptability and scalability. This architecture design allows the system to be continuously upgraded and improved as enterprise needs change and technology develops, protecting the enterprise's technological investment. The enhanced system scalability also enables the technical solution to be replicated and promoted across different enterprises, production lines, and products, greatly improving the application value and promotion efficiency of this invention.

[0077] Example 1: Figure 2 This is a flowchart illustrating another method for analyzing the influencing factors of inclusion defects provided in this embodiment of the invention, such as... Figure 2 As shown, this embodiment includes the following steps.

[0078] Step S101: Multi-source heterogeneous data acquisition.

[0079] During the data acquisition phase, production scheduling and material input data were collected through the MES system, including key production information such as heat number, production time, and type and quantity of materials input. Real-time parameter data of the steelmaking process were collected through the IoT system, including key process parameters directly affecting inclusion formation such as electric arc furnace final carbon content, LF refining temperature, VD vacuum time, calcium treatment amount, cooling rate, and stirring intensity. Chemical verification data, including the content of elements such as C, Si, Mn, P, S, and Al, were collected through the testing and analysis system; these element contents directly affect the type and quantity of inclusions. Actual control parameters and equipment operation data of steelmaking production were collected through the plant database to ensure data completeness and accuracy. Simultaneously, quality inspection data, including quality indicators such as inclusion level, size distribution, and ultrasonic flaw detection results, were collected. A total of six months of production data, covering approximately 500 heats, was collected, with each heat containing over 200 characteristic variables, providing a rich data foundation for subsequent analysis.

[0080] Step S102: Data preprocessing.

[0081] The data preprocessing stage involved comprehensive cleaning and organization of the collected raw data. For the approximately 5% of missing values, different processing strategies were employed based on the missing value proportion and data characteristics: median imputation was used for smaller missing values, while random forest-based prediction imputation was used for larger missing values ​​to ensure data integrity. Box plots were used to detect outliers, and boundary value replacement was applied to data exceeding the boundaries to eliminate the impact of outliers on the analysis results. Z-score standardization was used for numerical data, and one-hot encoding was used for categorical data to unify data format and standards. Data formats and standards across different systems were standardized, and data conflicts were addressed to ensure data consistency and comparability. This systematic data preprocessing laid a high-quality data foundation for subsequent feature engineering and model building.

[0082] Step S103: Feature engineering.

[0083] The feature engineering phase, based on materials science knowledge and domain expert experience, constructed physically meaningful feature variables. More than 50 derived features were constructed, including the Ca / Al ratio, Mn / S ratio, temperature-time product, and cooling gradient. These features, based on the physicochemical mechanisms of inclusion formation, possess strong explanatory power. Using recursive feature elimination and XGBoost-based feature importance assessment, 30 key features were selected from a large pool of features, including the carbon content at the electric furnace endpoint, Al content at the initial stage of LF refining, VD vacuum time, calcium treatment amount, and LF refining temperature—factors with the most significant impact on inclusion levels. This feature engineering method based on materials science knowledge not only improved the model's predictive accuracy but also enhanced its interpretability, making the analysis results more easily understood and accepted by domain experts.

[0084] Step S104: Integration of intelligent prediction models.

[0085] The intelligent prediction model construction phase employed advanced machine learning techniques and a rigorous model validation process. The dataset was divided into training, validation, and test sets at a ratio of 80% / 10% / 10% to ensure the objectivity and reliability of model evaluation. A mixed-level prediction model was built based on the XGBoost algorithm, while random forest and support vector machine models were trained simultaneously, forming a multi-algorithm integrated prediction system. The optimal parameters for XGBoost were determined through grid search: learning rate 0.1, maximum tree depth 6, subsample ratio 0.8, and column sampling ratio 0.8. This parameter combination performed best on the validation set. A weighted average method was used to integrate the prediction results of multiple models, further improving prediction accuracy and stability. This multi-algorithm integrated modeling method fully utilizes the advantages of different algorithms, significantly improving the model's generalization ability and prediction accuracy.

[0086] Step S105: Analysis of influencing factors.

[0087] After obtaining the predictive model, it is deployed to the production environment, and the prediction results are obtained based on the real-time acquired data. The influencing factors are then analyzed in this step.

[0088] The influencing factor analysis phase employed advanced model interpretation techniques to deeply analyze the impact mechanisms of various factors on inclusion levels. SHAP value analysis revealed that the top five factors affecting inclusion levels were, in descending order: final carbon content at the electric furnace endpoint, Al content at the initial stage of LF refining, VD vacuum time, calcium treatment amount, and LF refining temperature. Partial dependency plot analysis revealed a positive correlation between the final carbon content at the electric furnace endpoint and inclusion levels, meaning that excessively low carbon content leads to higher inclusion levels. Interaction effect analysis found a significant interaction between the final carbon content at the electric furnace endpoint and the Al content at the initial stage of LF refining; inclusion control was optimal when both were within appropriate ranges. These analytical results provide precise quantitative guidance for process optimization.

[0089] Step S106: Visualization and Decision Support.

[0090] The visualization and decision support phase transforms complex analysis results into intuitive and easy-to-understand decision recommendations. It generates visual charts such as importance ranking charts of influencing factors, relationship diagrams between features and inclusion levels, and process parameter optimization suggestion diagrams, enabling production personnel to intuitively understand the analysis results.

[0091] This invention also provides an apparatus for analyzing factors affecting inclusion nonconformity, such as... Figure 3 As shown, the device includes a memory and a processor, the processor being used to run a program, wherein the program, when run, is used to execute the method for analyzing the influencing factors of inclusion defects.

[0092] The processor contains a kernel, which retrieves the corresponding program unit from memory. One or more kernels can be configured, and the analysis of factors affecting inclusion defects can be achieved by adjusting kernel parameters.

[0093] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0094] This application also provides a computer program product that, when executed on a data processing device, is suitable for executing a program that initializes an analysis method for factors affecting nonconformities due to inclusions.

[0095] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0099] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0100] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0101] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0102] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0103] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for analyzing the influencing factors of inclusion defects, characterized in that, The method includes: Real-time acquisition of multi-source heterogeneous data in the production environment, including influencing factors of inclusions; The multi-source heterogeneous data is preprocessed to obtain the first preprocessed data; Extract predictive features from the first preprocessed data; Based on the predicted features and the pre-deployed prediction model, a prediction result is obtained, which includes the inclusion level. Based on the prediction model, prediction features, and prediction results, model interpretability analysis is performed to obtain the causal relationships between each prediction feature and the prediction results; and The influencing factors analysis results are obtained based on the causal relationship, and the analysis results are visualized.

2. The method according to claim 1, characterized in that, Real-time acquisition of multi-source heterogeneous data in the production environment includes: By using OGG-based stream processing, multi-source heterogeneous data from the production environment can be collected in real time from the MES system and the testing and analysis system. The multi-source heterogeneous data includes process parameter data, material composition data, equipment operation data, environmental data, and quality testing data.

3. The method according to claim 1, characterized in that, The preprocessing includes data cleaning, data integration, data transformation, and data reduction, wherein... The data cleaning includes detecting outliers in the data and processing the outliers based on deletion, interpolation, or model prediction. The data integration includes unifying the data format, data unit, and data standard; The data transformation includes standardization, normalization, and discretization of the data; and The data reduction includes dimensionality reduction processing of the data.

4. The method according to claim 1, characterized in that, The deployment method of the prediction model includes: Acquire multi-source heterogeneous data from historical production, with each of the multi-source heterogeneous data corresponding to an inclusion-level label; The multi-source heterogeneous data is preprocessed to obtain the second preprocessed data; Features are extracted from the second preprocessed data based on feature engineering, and predictive features are selected from the extracted features. A dataset is constructed based on the predicted features and their corresponding inclusion level labels; Based on the dataset, a variety of pre-selected models are trained, and during the training process, grid search, random search, or Bayesian optimization are used to optimize the parameters of the various models to obtain the optimal parameters of each model. The various models include XGBoost, random forest, support vector machine, and neural network. The prediction model is obtained by combining the multiple models based on an ensemble strategy; and Deploy the prediction model to the production environment.

5. The method according to claim 4, characterized in that, Based on feature engineering, features are extracted from the second preprocessed data, and predictive features are selected from the extracted features, including: Based on feature extraction, basic features are extracted from the second preprocessed data, including statistical features, time-domain features, and frequency-domain features; Based on feature construction, elemental ratio features, process parameter interaction features, and historical statistical features are constructed from the second preprocessed data. The elemental ratio features include the Ca / Al ratio and the Mn / S ratio; the process parameter interaction features include the product of temperature and time and the relationship between cooling rate and temperature gradient; and the historical statistical features include the moving average and rate of change of the parameters. Based on feature selection, the most relevant feature subset is selected from the features obtained by feature extraction and feature construction to obtain the predicted features. The predicted features include the final carbon content of the electric furnace, the Al content at the beginning of LF refining, the VD vacuum time, the amount of calcium processed, and the LF refining temperature.

6. The method according to claim 4, characterized in that, The method also includes periodically acquiring multi-source heterogeneous data to evaluate the prediction accuracy of the prediction model after deploying the prediction model to the production environment.

7. The method according to claim 1, characterized in that, Based on the prediction model, prediction features, and prediction results, model interpretability analysis is performed to obtain the causal relationships between each prediction feature and the prediction results, including: Based on feature importance, permutation importance, or SHAP value, the predicted features and the predicted results are calculated to obtain the importance score of each predicted feature to the predicted results; Based on partial dependency graphs, individual condition expectation graphs, or cumulative local effect graphs, the prediction features and prediction results are analyzed to obtain the relationship patterns between each prediction feature and the prediction results. Based on SHAP interaction values, two-dimensional partial dependency graphs, or statistical interaction tests, the predicted features and the predicted results are analyzed to obtain the synergistic and antagonistic effects between the predicted features; and Based on causal discovery algorithms or causal forests, the importance scores, relationship patterns, synergistic effects, and antagonistic effects are analyzed to obtain the causal relationship between the predicted features and the predicted results.

8. The method according to claim 1, characterized in that, The method further includes: Based on the interactive interface, it receives multi-source heterogeneous data that users input or adjust; Prediction is performed based on multi-source heterogeneous data input or adjusted by the user, resulting in prediction outcomes; and The prediction results are then visualized.

9. The method according to claim 1, characterized in that, The method further includes: Based on the interactive interface, it receives constraint information input by the user; Based on the constraint information, predictions are made in the prediction model to obtain multiple prediction results; Based on the target at the minimum inclusion level and the constraint information, the optimal value of the multi-source heterogeneous data is selected from the multiple prediction results; and The selected values ​​are visualized, and optimization suggestions are given based on a preset rule base.

10. An apparatus for analyzing the influencing factors of inclusion nonconformity, the apparatus comprising a memory and a processor, characterized in that, The processor is used to run a program, wherein the program is run to perform: the method as described in any one of claims 1-9.

Images