Abnormal feature determination method, model training method and tapping index prediction method
By automatically identifying and eliminating abnormal features in blast furnace ironmaking production and retraining the prediction model, the prediction error problem of traditional methods in the face of abnormal situations is solved, and efficient and real-time prediction of iron tapping indicators is achieved.
Patent Information
- Application Number
- CN202510909207.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-11-18
AI Technical Summary
Traditional forecasting methods are ill-suited to dealing with abnormal blast furnace production conditions. They cannot automatically identify and resolve the increased forecasting errors caused by erroneous feature data, requiring manual intervention for adjustment.
By determining the absolute residuals and SHAP values of the sample data, abnormal features are automatically identified, and these features are removed when anomalies are detected, and the prediction model is retrained.
It enables automatic location and removal of abnormal features, improves the real-time performance and robustness of predictions, reduces prediction errors, and enhances production efficiency and product quality.
Smart Images
Figure CN120974172A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of steel production and processing technology, specifically to a method for determining abnormal characteristics, a method for training a model, a method for predicting iron indicators, and processor and computer program products. Background Technology
[0002] Blast furnace ironmaking is the core hub of the "iron-steel-finished product" process in the steel industry. Accurate prediction of tapping time, molten iron volume, and molten iron temperature is crucial for ensuring efficient collaboration between the iron and steel interfaces. By reducing the prediction deviation of tapping data, the waiting time for molten iron in the steelmaking process can be reduced, directly improving converter capacity utilization. Furthermore, it can reduce molten iron temperature drop, improving product quality and having significant implications for production planning, energy consumption optimization, and equipment maintenance.
[0003] However, traditional forecasting methods rely on manual empirical formulas or static statistical models (such as multiple linear regression), which are difficult to cope with abnormal situations in blast furnace production. For example, when the instrument for collecting a certain feature data is damaged, all the collected feature data will be wrong. However, traditional methods can only identify the larger error, but cannot identify which feature data is causing the increase in prediction error, let alone solve this problem automatically. It requires human intervention (stopping the model from running) and tedious adjustments to the model. Summary of the Invention
[0004] The purpose of this invention is to provide a method for determining abnormal features, a method for training a model, a method for predicting iron indicators, a processor, and a computer program product, so as to solve or at least partially solve the above-mentioned defects of the prior art.
[0005] To achieve the above objectives, a first aspect of the present invention provides a method for determining abnormal features, the method comprising:
[0006] Based on the actual value and model prediction value of the iron output index corresponding to each sample data point in multiple sample data sets, determine the absolute residual corresponding to each sample data point; and
[0007] When the absolute residuals of multiple consecutive sample data in the plurality of sample data are all greater than the corresponding preset thresholds, the following abnormal feature screening steps are triggered:
[0008] For each modeling feature in the aforementioned consecutive sets of sample data, perform the following operations:
[0009] Determine the real-time value corresponding to the degree of influence of this modeling feature on the prediction result; and
[0010] When the difference between the real-time value and the historical benchmark value corresponding to the influence of the modeling feature on the prediction result is greater than the corresponding preset threshold, the modeling feature is judged as an abnormal feature.
[0011] Preferably, the preset threshold corresponding to the absolute residuals of the consecutive multiple sample data is determined based on the average of the prediction errors of the multiple sample data within a preset statistical period.
[0012] Preferably, the degree of influence of the modeling feature on the prediction result is determined based on the SHAP value corresponding to the modeling feature.
[0013] Preferably, the historical baseline value corresponding to the degree of influence of the modeling features on the prediction results is determined based on sample data corresponding to the absence of the abnormal feature screening step.
[0014] A second aspect of this invention provides a method for training a model, the method comprising:
[0015] Data sampling was conducted during the blast furnace ironmaking production process to obtain multiple sample data.
[0016] Each of the multiple sample data includes: multiple modeling features and actual values of multiple iron output indicators;
[0017] Input the multiple modeling features included in each sample data into the corresponding prediction model to determine the model prediction value of each iron output index corresponding to each sample data.
[0018] When abnormal features exist in the multiple sample data, the abnormal features are removed from the multiple sample data, and training data is regenerated using the multiple sample data after removing the abnormal features. The training data is then trained using a machine learning algorithm to regenerate the prediction model.
[0019] Specifically, the abnormal features in the plurality of sample data are determined according to the method for determining abnormal features.
[0020] Preferably, the modeling features include: iron content of sinter, coke-to-butadiene ratio, and air pressure; the iron tapping indicators include iron tapping time, iron tapping amount, and iron tapping temperature.
[0021] Preferably, the method further includes:
[0022] Before inputting the multiple modeling features included in each sample data into the corresponding prediction model to determine the model prediction value for each iron output index corresponding to each sample data,
[0023] For each iron output index in each sample data, perform the following operations:
[0024] Based on the coefficient of variation and skewness corresponding to the actual value of the iron tapping index, and the preset adjustment coefficients corresponding to different coefficients of variation and skewness, the adjustment coefficient corresponding to the iron tapping index is determined; and based on the adjustment coefficient, the numerical range corresponding to the actual value of the iron tapping index is determined; and
[0025] For each sample data, perform the following operations:
[0026] If the actual values of multiple iron output indicators in the sample data are not within the corresponding numerical range, the sample data is determined to be an invalid sample and discarded.
[0027] Preferably, the method further includes determining the numerical range corresponding to the actual value of the iron tapping index through the following method:
[0028] Based on the actual values of the iron tapping index, corresponding to the 75th quantile Q3 and the 25th quantile Q1, the first parameter IQR is determined as follows:
[0029] IQR = Q3 - Q1
[0030] Based on the 75th percentile Q3 corresponding to the actual value of the iron tapping index, the first parameter IQR, and the adjustment coefficient λ, the upper limit U of the numerical range corresponding to the iron tapping index is determined:
[0031] U=Q3+λ*IQR
[0032] And based on the 25th percentile Q1 corresponding to the actual value of the iron tapping index, the first parameter IQR, and the adjustment coefficient λ, the lower limit L of the numerical range corresponding to the iron tapping index is determined:
[0033] L=Q1-λ*IQR.
[0034] Preferably, the method further includes: when the amount of data reduction corresponding to the training data exceeds the corresponding preset threshold, the adjustment coefficient corresponding to each iron output index is amplified accordingly.
[0035] Preferably, the prediction model is a support vector machine, and a three-layer grid search is used to determine the kernel function corresponding to the prediction model and the parameter combination corresponding to the kernel function.
[0036] A third aspect of this invention provides a method for predicting iron production indicators, the method comprising:
[0037] Collect multiple modeling features corresponding to the steel processing equipment, input the modeling features into the prediction model, and obtain the iron output index corresponding to the steel processing equipment.
[0038] The prediction model is determined based on the method used to train the model.
[0039] A fourth aspect of the present invention provides a processor for running a program, wherein the program is executed to perform: the method for training the model or the method for predicting iron indicators.
[0040] A fifth aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method for training the model or the method for predicting iron indicators.
[0041] The training model method provided in this invention, when detecting abnormal feature data, triggers dynamic diagnosis of SHAP values based on preset threshold residual monitoring. Through precise comparison of real-time feature contribution vectors with historical benchmarks, it achieves automatic location and removal of abnormal features. The model is then retrained using the remaining sample data after removing abnormal features, realizing full automation from data anomaly detection and dynamic feature selection to rapid model reconstruction. This overcomes the bottlenecks of traditional methods in terms of real-time performance, robustness, and interpretability, providing an industrial-grade reliable solution for blast furnace iron tapping prediction.
[0042] 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
[0043] 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:
[0044] Figure 1 This is a flowchart of the method for determining abnormal features provided in the embodiments of the present invention;
[0045] Figure 2 This is a flowchart of the training model method provided in the embodiments of the present invention. Detailed Implementation
[0046] 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.
[0047] 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.
[0048] Example 1
[0049] Figure 1 This is a flowchart of the method for determining abnormal features provided in an embodiment of the present invention, such as... Figure 1 As shown, this embodiment of the invention provides a method for determining abnormal features, the method comprising S101 to S102:
[0050] S101. Determine the absolute residual corresponding to each sample data based on the actual value and model prediction value of the iron output index corresponding to each sample data in multiple sample data.
[0051] Furthermore, the tapping parameters include tapping time, tapping quantity, or tapping temperature; these parameters can be measured in real time by high-frequency sensors during the blast furnace ironmaking operation.
[0052] Specifically, absolute residual is an important concept in statistics and regression analysis, used to measure the degree of difference between observed and predicted values. Absolute residual is the absolute value of the difference between the observed value (actual value) and the predicted value obtained through a regression model or other prediction methods. Absolute residual quantifies the average deviation between the model's predicted value and the actual value. The mathematical expression for absolute residual is:
[0053]
[0054] Among them, E t Let y represent the absolute residual corresponding to the iron output index in the t-th sample data. t This indicates the actual value of the iron index. This represents the model's predicted value for the iron index.
[0055] S102. When the absolute residuals corresponding to multiple consecutive sample data in the plurality of sample data are all greater than the corresponding preset thresholds, the following abnormal feature screening steps are triggered:
[0056] For each modeling feature in the aforementioned consecutive multiple sample data, perform the following operations S201 to S202:
[0057] S201. Determine the real-time value corresponding to the degree of influence of the modeling feature on the prediction result;
[0058] S202. When the difference between the real-time value and the historical benchmark value corresponding to the degree of influence of the modeling feature on the prediction result is greater than the corresponding preset threshold, the modeling feature is determined to be an abnormal feature.
[0059] First, the modeling features in S102 will be explained in detail. For blast furnace ironmaking operations, multiple modeling features can be determined based on various production and processing parameters to predict iron production indicators. For example, in some embodiments, the modeling features include: iron content of sinter, coke-to-butane ratio, and blast pressure.
[0060] Furthermore, the preset threshold corresponding to the absolute residuals of the consecutive multiple sample data in S102 is determined based on the average of the prediction errors of the multiple sample data within a preset statistical period.
[0061] Furthermore, to illustrate S102, when the absolute residuals corresponding to the sample data in three consecutive tests are all greater than 1.5 times the historical mean absolute error (MAE), that is, when the following condition is met:
[0062] Three times in a row:
[0063] A red alert is triggered, automatically freezing the current feature set and initiating an abnormal feature screening step. This mechanism focuses on persistent prediction bias. The Mean Absolute Error (MAE) is one of the metrics for measuring model prediction performance; it assesses the accuracy of model predictions by calculating the average absolute error between predicted and true values.
[0064] In some embodiments, a sliding window (the window size is determined by the amount of iron tapping data in the past three days) is used to update the historical average MAE in real time, adapting to the periodic production fluctuations of the blast furnace, ensuring that the threshold discrimination criteria are dynamically adjusted with the operating conditions, and avoiding misjudgment or omission caused by a fixed threshold.
[0065] This invention can accurately locate abnormal features by analyzing and comparing the real-time SHAP value of the modeling features corresponding to the problem data that triggers the abnormal feature screening step with the historical baseline value.
[0066] In other embodiments, the condition is met when the absolute residuals corresponding to a single sample are all greater than twice the historical mean absolute error (MAE):
[0067]
[0068] A yellow alert is triggered, and the current single sample data is marked as an anomaly candidate.
[0069] In some embodiments, the degree of influence of the modeling features in S201 on the prediction results is determined based on the SHAP value corresponding to the modeling features.
[0070] Specifically, the SHAP value represents the marginal contribution of each feature to a single prediction result in a machine learning model. Based on the SHAP value, the prediction process of the machine learning model can be understood more intuitively, improving the model's interpretability and credibility. The SHAP value reveals the "fair" contribution of each feature to the model's prediction result by quantifying the average marginal contribution of a feature under different feature combinations. The calculation of the SHAP value is based on the concept of Shapley values, the core idea of which is to consider the marginal contribution of a feature across all possible feature combinations.
[0071] This invention visualizes feature contributions through SHAP values, such as the negative correlation between "wind pressure lag value" and iron output indicators, and filters out the top 90% of features with the highest cumulative contribution, which can provide data-driven basis for subsequent steel manufacturing process optimization.
[0072] In other embodiments, LIME, marginal effects, gradient methods, or integrated gradients can also be used to quantify the marginal contribution of a single feature to a single prediction result.
[0073] In some embodiments, the historical baseline value corresponding to the degree of influence of the modeling features on the prediction results described in S202 is determined based on sample data corresponding to the absence of the abnormal feature screening step.
[0074] Specifically, during the model training phase, the SHAP value of each feature is calculated for the normal operating condition data. Features with a cumulative contribution of ≥90% are selected and stored. Normal operating conditions refer to sample data for which the abnormal feature screening step has not been triggered.
[0075] After triggering the abnormal feature screening step, the real-time SHAP value is calculated for the currently input data and compared with the historical baseline SHAP value calculated using normal operating condition data. Features with large deviations in SHAP value are identified and judged as "data acquisition abnormal features".
[0076] In some embodiments, a ranking table of SHAP values (Top-90% features) corresponding to the modeling features is dynamically displayed.
[0077] This invention also provides an anomaly feature management system for implementing a dual-threshold residual early warning strategy. Further, the dual-threshold residual early warning strategy includes:
[0078] When the absolute residual corresponding to a single sample data is greater than twice the historical mean absolute error (MAE), a yellow flashing warning will be triggered on the interface.
[0079] If the absolute residuals of the sample data are greater than 1.5 times the historical mean absolute error (MAE) in three consecutive tests, the interface will trigger a red flashing warning and simultaneously send an SMS notification containing the abnormal feature number and residual details to the process supervisor. The corresponding abnormal handling log will also be automatically recorded. The abnormal handling log can be retrieved by time and abnormal type to facilitate subsequent tracing of abnormal data collection events.
[0080] The method for determining abnormal features provided in this invention triggers an abnormal feature screening step when multiple consecutive samples exceed a preset threshold; and determines abnormal features by comparing the real-time value corresponding to the influence of modeling features on the prediction result with historical benchmark values. This method can automatically complete abnormal feature identification and accurately locate abnormal features, thereby solving the problem of long-term error deviation caused by continuously using erroneous data for prediction.
[0081] Example 2
[0082] Example 2 is implemented based on Example 1. Figure 2 This is a flowchart of the training model method provided in an embodiment of the present invention, such as... Figure 2 As shown in the figure, this embodiment of the invention also provides a method for training a model, the method comprising:
[0083] S301. Data sampling is performed during the blast furnace ironmaking production process to obtain multiple sample data.
[0084] Each of the multiple sample data includes: multiple modeling features and actual values of multiple iron output indicators;
[0085] In some embodiments, modeling features include: iron content in sinter, coke-to-butadiene ratio, and blast pressure. Iron tapping parameters include tapping time, tapping rate, or tapping temperature.
[0086] In some embodiments, before executing S302, for each iron output index or each modeling feature in each sample data, a single-dimensional outlier filtering is performed, including the following processing steps:
[0087] For each iron output index or each modeling feature in each sample data: execute S401 to S402;
[0088] S401. Based on the coefficient of variation and skewness corresponding to the actual value of the single-dimensional indicator (each iron output indicator or each modeling feature), and the preset adjustment coefficients corresponding to different coefficients of variation and skewness, determine the adjustment coefficient corresponding to the single-dimensional indicator.
[0089] S402. Determine the numerical range corresponding to the actual value of the single-dimensional indicator based on the adjustment coefficient corresponding to the single-dimensional indicator.
[0090] For each single-dimensional indicator (each iron output indicator or each modeling feature) in each sample data, execute S403;
[0091] S403. When the actual value of the single-dimensional indicator corresponding to the sample data is not within the corresponding numerical range, the value of the single-dimensional indicator is set to null.
[0092] In some embodiments, before inputting the multiple modeling features included in each sample data into the corresponding prediction model to determine the model prediction value of each iron output index corresponding to each sample data, the following processing steps are performed:
[0093] For each iron output index in each sample data, execute S501 to S502:
[0094] S501. Based on the coefficient of variation and skewness corresponding to the actual value of the iron tapping index, and the preset adjustment coefficients corresponding to different coefficients of variation and skewness, determine the adjustment coefficient corresponding to the iron tapping index.
[0095] Specifically, the coefficient of variation (CV) is a statistical measure of the relative dispersion of data. It is defined as the ratio of the standard deviation to the mean (usually expressed as a percentage). The coefficient of variation can eliminate the influence of dimensions and compare the dispersion of different datasets or variables of different magnitudes. The smaller the coefficient of variation (CV), the more concentrated the data is relative to the mean, and the higher the stability; the larger the coefficient of variation (CV), the greater the data fluctuation and the lower the stability.
[0096] In some embodiments, the coefficient of variation (CV) is determined based on the standard deviation (σ) and the mean (μ):
[0097]
[0098] Furthermore, skewness is a statistical indicator used to describe the degree of asymmetry in data distribution. It measures the direction and extent of skewness in the data distribution and can be calculated using the Pearson skewness coefficient and the sample skewness formula.
[0099] Specifically, in the Pearson skewness coefficient calculation formula, the skewness is determined based on the mean μ, mode, and standard deviation σ:
[0100]
[0101] In the sample skewness formula, based on the sample size n and the i-th sample value x... i Sample mean And the sample standard deviation s, to determine the skewness:
[0102]
[0103] In other embodiments, the coefficient of variation and skewness can also be calculated using the std, mean, and skew functions in the Pandas Python data analysis library.
[0104] In some embodiments, the adjustment coefficient λ corresponding to any iron tapping index is determined based on the following conditions:
[0105] When the CV corresponding to the iron output index is <0.1 and |Skewness| is <0.5, λ = 1.5, which is applied to normal and stable operating conditions, using the traditional threshold.
[0106] When the iron output index corresponds to 0.1 <= CV < 0.3 or 0.5 <= |Skewness| < 1.0, λ = 1.2, which is applied to data fluctuations or slight skewness to tighten the outlier range;
[0107] When the CV >= 0.3 or |Skewness| >= 1.0 corresponding to the iron output index, λ = 1.0. This is applied to areas with severe fluctuations or strong skewness, and significant outliers are strictly eliminated.
[0108] S502. Based on the adjustment coefficient, determine the numerical range corresponding to the actual value of the iron tapping index;
[0109] Specifically, the numerical range corresponding to the actual value of the iron tapping index is determined in the following way:
[0110] Based on the actual values of the iron tapping index, corresponding to the 75th quantile Q3 and the 25th quantile Q1, the first parameter IQR is determined as follows:
[0111] IQR = Q3 - Q1
[0112] Based on the 75th percentile Q3 corresponding to the actual value of the iron tapping index, the first parameter IQR, and the adjustment coefficient λ, the upper limit U of the numerical range corresponding to the iron tapping index is determined:
[0113] U=Q3+λ*IQR
[0114] And based on the 25th percentile Q1 corresponding to the actual value of the iron tapping index, the first parameter IQR, and the adjustment coefficient λ, the lower limit L of the numerical range corresponding to the iron tapping index is determined:
[0115] L=Q1-λ*IQR.
[0116] For each sample data, execute S503:
[0117] S503. When the actual values of multiple iron output indicators in the sample data are not within the corresponding numerical range, the sample data is determined to be an invalid sample and discarded.
[0118] Specifically, a three-dimensional cross-validation process is implemented for tapping time, molten iron quantity, and molten iron temperature: a sample is only considered valid if all three parameters are normal (i.e., simultaneously satisfying the corresponding numerical ranges for each of the three parameters). By jointly filtering sample data from multiple dimensions, misjudgments caused by a single dimension can be avoided, maximizing the retention of normal sample data.
[0119] In some embodiments, the method further includes: when the amount of data reduction corresponding to the training data exceeds a corresponding preset threshold, increasing the adjustment coefficient corresponding to each iron output index accordingly.
[0120] Specifically, if the amount of training data corresponding to the filtered data decreases by more than 20%, the coefficient level is automatically relaxed by one level. For example, the adjustment coefficient λ = 1.0 is adjusted to λ = 1.2. At this time, the modified adjustment coefficient λ needs to be substituted into S502, and S502 and S503 are executed again to prevent over-filtering.
[0121] By modifying the adjustment coefficient, the numerical range for outlier determination can be narrowed or widened accordingly, thereby improving the retention rate of normal data.
[0122] In some embodiments, the present invention employs cubic spline interpolation and bidirectional boundary filling techniques in the spatiotemporal continuity filling method to complete the training data and fill in the gaps.
[0123] Specifically, cubic spline interpolation, compared to linear interpolation, can preserve the continuity of the second derivative of the data, reduce the RMSE after filling, and truly restore the parameter fluctuation trend. Bidirectional boundary filling addresses missing values at the beginning and end of time series data by first initializing with historical valid values through forward filling, and then correcting the end-point deviation through backward filling, ensuring the integrity of the first and last data points. This solves the problem of high omission rates for missing values at the beginning and end in traditional single interpolation methods.
[0124] It should be noted that the embodiments of the present invention apply cubic spline interpolation and bidirectional boundary filling to fill missing values in sample data, which can reduce the missing value rate to 0%. After the missing values are repaired, the root mean square error (RMSE) of the sample data is reduced and the trend restoration is improved, thereby providing more reliable input data for subsequent model training.
[0125] S302. Input the multiple modeling features included in each sample data into the corresponding prediction model to determine the model prediction value of each iron output index corresponding to each sample data.
[0126] S303. Determine whether the multiple sample data have abnormal features, and if the multiple sample data have abnormal features, execute S304 to S305;
[0127] Specifically, for S303, the abnormal features in multiple sample data can be determined according to the method for determining abnormal features described in Embodiment 1.
[0128] S304. When abnormal features exist in the multiple sample data, the abnormal features are removed from the multiple sample data.
[0129] By removing the abnormal features, we can avoid the problem of long-term error deviation caused by continuously using erroneous data for prediction, and we can also reduce the feature dimension and shorten the model retraining time.
[0130] S305. Then, using the multiple sample data after removing the abnormal features, training data is regenerated, and the training data is trained using a machine learning algorithm to regenerate the prediction model.
[0131] The present invention pre-processes the prediction model with lightweighting. After removing abnormal features, it retrains using the remaining data. The training speed is significantly faster than that of traditional models, and the time consumption is greatly reduced. It can be adapted to the blast furnace tapping speed and realize real-time automated problem solving and real-time prediction.
[0132] In some embodiments, the prediction model is a support vector machine, which uses a three-layer grid search to determine the kernel function and the parameter combination corresponding to the kernel function, thereby realizing a lightweight prediction model.
[0133] Specifically, for support vector machine models, the selection of kernel function and its parameter optimization are key factors affecting model performance. This invention proposes to use a three-layer grid search space to determine the optimal combination of kernel function and its parameters, optimize model parameters for dynamic feature sets, use 3-fold cross-validation to calculate the MAE corresponding to different parameter combinations, and select the parameter combination with the smallest MAE.
[0134] The three-layer grid search space refers to the further optimization of key parameters for each kernel function based on the kernel function selection, forming a three-layer nested search structure. The following is a detailed description of the three-layer grid search space proposed in the embodiments of the present invention:
[0135] First layer: Kernel function type selection:
[0136] Specifically, the first layer is used to determine the degree of nonlinearity of the problem and select the most suitable kernel function type.
[0137] In some embodiments, the first layer evaluates the performance of different kernel functions by fixing the regularization parameter C and the kernel function coefficient γ to default values (e.g., C = 1, γ = 'scale') to determine the most suitable kernel function type. For example, if the first layer outputs the 'rbf' kernel function, it performs best and proceeds to the second layer.
[0138] It should be noted that the kernel function of SVM determines how data is mapped from the original space to the high-dimensional feature space. Common kernel functions include radial basis function (RBF) kernel, linear kernel, polynomial kernel, Gaussian kernel, and Sigmoid kernel.
[0139] Second layer: Optimization of regularization parameter C:
[0140] First, it should be noted that regardless of the kernel function chosen, SVM has a common regularization parameter C, which controls the model's tolerance for misclassification.
[0141] The second layer is used to optimize the regularization parameter C in order to balance model complexity and error.
[0142] In some embodiments, the second layer fixes the kernel function to the optimal kernel function determined by the first layer (e.g., 'rbf'), fixes the kernel function coefficient γ (e.g., γ = 'scale'), and searches for the optimal regularization parameter C value. For example, the second layer achieves the best performance when C = 100, and then proceeds to the third layer.
[0143] Third layer: Kernel function coefficient fine-tuning:
[0144] In some embodiments, the third layer fixes the kernel function to the optimal kernel function determined by the first layer (e.g., 'rbf') and the regularization parameter C to the optimal value determined by the second layer (e.g., C = 100), fine-tuning the kernel function coefficient γ to control the model's generalization ability. For example, the third layer outputs γ = 0.1 with optimal performance.
[0145] According to the three-layer grid search space proposed in this embodiment of the invention, the final kernel function and corresponding parameter combination are {kernel':'rbf','C':100.0,'gamma':'0.1'}
[0146] It should be noted that the training model method provided in this embodiment of the invention can be integrated into a software system. Specifically, a corresponding software system can be developed using Python 3.9 and then packaged into an independent Docker image.
[0147] The training model method provided in this invention can be embedded in the blast furnace control system. In specific implementation, when new sampling data corresponding to the blast furnace ironmaking production process is received, the "data preprocessing → abnormal feature removal → retraining prediction model → model prediction" pipeline is automatically triggered to ensure that the prediction results are accurately synchronized with the blast furnace iron tapping cycle.
[0148] The embodiments of the present invention reduce feature dimensionality by eliminating abnormal features and determine the optimal combination of kernel function and parameters corresponding to the prediction model by using a three-layer grid search space. This can lighten the prediction model, thereby shortening the model training time, adapting to the real-time requirements of data prediction in blast furnace ironmaking scenarios, and reducing memory usage while improving CPU / GPU utilization.
[0149] This invention provides a method for predicting iron output indicators. The method includes: collecting multiple modeling features corresponding to steel processing equipment, inputting the modeling features into a prediction model, and obtaining the iron output indicators corresponding to the steel processing equipment.
[0150] The prediction model is determined based on the method used to train the model.
[0151] In some embodiments, the predicted curves of iron tapping time, iron tapping amount, and iron tapping temperature, as well as the fluctuation band of actual iron tapping data, are displayed in real time.
[0152] This invention provides a processor for running a program, wherein the program, when run, is used to execute: the method for training the model or the method for predicting iron indicators.
[0153] This invention provides a computer program product, including a computer program that, when executed by a processor, implements the method for training the model or the method for predicting iron indicators.
[0154] This invention addresses the core bottlenecks of traditional prediction methods in data anomaly handling, real-time performance, and interpretability by constructing a dynamic comparison and analysis mechanism for SHAP values. When anomalies in feature data are detected, dynamic diagnosis of SHAP values is triggered based on preset threshold residual monitoring. Through precise comparison of real-time feature contribution vectors with historical benchmarks, automatic location and removal of abnormal features are achieved. The lightweight model reconstruction process based on remaining effective features can be completed within minutes, significantly improving efficiency compared to manual intervention. This ensures that prediction errors quickly converge to normal operating conditions, fundamentally meeting the real-time prediction requirements of blast furnace iron tapping data.
[0155] At the production application level, this technology supports optimized molten iron ladle scheduling in the steelmaking process, reducing the average waiting time for molten iron ladles, decreasing molten iron temperature drop losses, and improving the qualified rate of molten steel composition, significantly reducing energy consumption losses and quality fluctuations in the converter smelting process. Its innovative technical architecture provides a replicable and universally applicable solution for blast furnace ironmaking process control. The core technical concepts can be transferred to key processes such as converter steelmaking in the steel industry, promoting the transformation of the steel industry from a "post-event manual traceability" control mode to a "real-time intelligent processing" control mode. This has significant technical demonstration value and industry promotion value for improving the automation and intelligence level of my country's steel industry.
[0156] 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.
[0157] 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 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0158] 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.
[0159] 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.
[0160] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0161] 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.
[0162] 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.
[0163] 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.
[0164] 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 of determining an abnormal feature, characterized by, The method comprises: According to the actual value and the model prediction value of the tapping index corresponding to each sample data in the plurality of sample data, the absolute residual corresponding to each sample data is determined; and When the absolute residuals corresponding to the plurality of consecutive sample data in the plurality of sample data are all greater than the corresponding preset threshold value, the following abnormal feature screening step is triggered: For each modeling feature in the plurality of consecutive sample data, the following operations are performed: Determine the real-time value corresponding to the influence degree of the modeling feature on the prediction result; and When the difference between the real-time value and the historical reference value corresponding to the influence degree of the modeling feature on the prediction result is greater than the corresponding preset threshold value, the modeling feature is determined as an abnormal feature.
2. The method of claim 1, wherein, The preset threshold value corresponding to the absolute residual of the plurality of consecutive sample data is determined according to the average value of the prediction errors corresponding to the plurality of sample data in the preset statistical period.
3. The method of claim 1, wherein, The influence degree of the modeling feature on the prediction result is determined according to the SHAP value corresponding to the modeling feature.
4. The method of claim 1, wherein, The historical reference value corresponding to the influence degree of the modeling feature on the prediction result is determined according to the sample data corresponding to the abnormal feature screening step which is not triggered.
5. A method of training a model, the method comprising: The method comprises: Data sampling is performed in the blast furnace ironmaking production operation process to obtain a plurality of sample data; Each sample data in the plurality of sample data comprises a plurality of modeling features and actual values of a plurality of tapping indexes. The plurality of modeling features included in each sample data are input into the corresponding prediction model to determine the model prediction value of each tapping index corresponding to the sample data. When the plurality of sample data has an abnormal feature, the abnormal feature is removed from the plurality of sample data, and the training data is regenerated using the plurality of sample data after removing the abnormal feature, the prediction model is regenerated by training the training data using a machine learning algorithm. The determination method of the abnormal feature according to any one of claims 1 to 4 is used to determine the abnormal feature in the plurality of sample data.
6. The method of claim 5, wherein, The modeling features comprise sinter iron content, coke butter ratio, and air pressure; and the tapping indexes comprise tapping time, tapping amount, and tapping temperature.
7. The method of claim 5, wherein, The method further comprises: Before the plurality of modeling features included in each sample data are input into the corresponding prediction model to determine the model prediction value of each tapping index corresponding to the sample data, For each tapping index in the sample data, the following operations are performed: According to the coefficient of variation and skewness corresponding to the actual value of the tapping index, and the preset adjustment coefficient corresponding to different coefficients of variation and skewness, the adjustment coefficient corresponding to the tapping index is determined; and According to the adjustment coefficient, the numerical interval corresponding to the actual value of the tapping index is determined; and For the sample data, the following operations are performed: When the actual values of the plurality of tapping indexes of the sample data are not all within the corresponding numerical interval, the sample data is determined as an invalid sample, and the sample data is discarded.
8. The method of claim 7, wherein, The method further comprises that the numerical interval corresponding to the actual value of the tapping index is determined by the following method: Based on the actual values of the iron tapping index, corresponding to the 75th quantile Q3 and the 25th quantile Q1, the first parameter IQR is determined as follows: IQR = Q3 - Q1 Based on the 75th percentile Q3 corresponding to the actual value of the iron tapping index, the first parameter IQR, and the adjustment coefficient λ, the upper limit U of the numerical range corresponding to the iron tapping index is determined: U=Q3+λ*IQR And based on the 25th percentile Q1 corresponding to the actual value of the iron tapping index, the first parameter IQR, and the adjustment coefficient λ, the lower limit L of the numerical range corresponding to the iron tapping index is determined: L=Q1-λ*IQR.
9. The method of claim 7, wherein, The method further includes: when the amount of data reduction corresponding to the training data exceeds the corresponding preset threshold, the adjustment coefficient corresponding to each iron output index is amplified accordingly.
10. The method of claim 5, wherein, The prediction model is a support vector machine, which uses a three-layer grid search to determine the kernel function and the parameter combination corresponding to the kernel function.
11. A method of predicting a tapping index, characterized by, The method includes: Collect multiple modeling features corresponding to the steel processing equipment, input the modeling features into the prediction model, and obtain the iron output index corresponding to the steel processing equipment. The prediction model is determined by the method of training a model according to any one of claims 5-10.
12. A processor, comprising: For running a program, wherein the program is run to perform: the method of training a model as described in any one of claims 5 to 10 or the method of predicting iron ore indices as described in claim 11.
13. A computer program product, characterised in that, Includes a computer program that, when executed by a processor, implements the method of training a model as described in any one of claims 5 to 10 or the method of predicting iron indices as described in claim 11.