Machine learning system for blast furnace operation
The machine learning device improves low-frequency blast furnace operation prediction accuracy by preprocessing and normalizing data, addressing the limitations of existing systems in handling infrequent data through smoothing and outlier removal, resulting in enhanced predictive models.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- NIPPON STEEL CORPORATION
- Filing Date
- 2024-11-25
- Publication Date
- 2026-06-04
AI Technical Summary
Existing machine learning systems for blast furnace operation struggle to achieve high prediction accuracy for low-frequency blast furnace operation data, as they are primarily designed for high-frequency data and fail to handle changes in operating conditions effectively.
A machine learning device that preprocesses blast furnace operation data by smoothing, normalizing, and removing outliers, followed by selecting appropriate explanatory variables, to construct a prediction model using normal distribution data for improved accuracy in low-frequency data prediction.
The proposed method enhances the prediction accuracy of low-frequency blast furnace operation data while maintaining accuracy for high-frequency data, utilizing a Hampel Identifier method for outlier removal and moving averages for smoothing.
Smart Images
Figure 2026091644000001_ABST
Abstract
Description
[Technical Field]
[0001] This invention relates to a machine learning device for blast furnace operation. [Background technology]
[0002] Systems that utilize physical models (for example, systems that predict molten iron temperature using transient models) are known for predicting and controlling blast furnace operations. While these prediction systems using physical models can achieve high prediction accuracy when boundary conditions are clear, their accuracy decreases when boundary conditions are unknown (for example, when the composition of the raw fuel is unknown, or when the packing status of the packed bed in the furnace is unknown).
[0003] Another known prediction system is the expert model, which makes predictions using a knowledge base expressed by rules. In the case of the expert model, if operating conditions change over time since the rules were created, the rules need to be reviewed.
[0004] One known method to address these problems is to search for similar cases from recent historical data and construct a local statistical model for those similar cases to make future predictions. However, this model has problems such as difficulty in predicting operations under different operating conditions than before, and a decrease in prediction accuracy if there is not enough historical similar cases accumulated to guarantee the accuracy of the local statistical model.
[0005] Patent Document 1 discloses a machine learning device for blast furnace operation, which includes: a data preprocessing unit that converts non-normally distributed data from blast furnace operation data, where the skewness exceeds a skewness threshold, into normally distributed data to unify the blast furnace operation data into normally distributed data; a data type selection unit that selects explanatory variables corresponding to a predetermined objective variable based on the data type of the individual operation data included in the blast furnace operation data; and a model generation unit that generates a blast furnace operation prediction model by performing machine learning using the blast furnace operation data corresponding to the selected explanatory variables and the blast furnace operation data corresponding to the objective variable as training data. [Prior art documents] [Patent Documents]
[0006] [Patent Document 1] Japanese Patent Publication No. 2022-153923 [Overview of the Initiative] [Problems that the invention aims to solve]
[0007] In recent years, there has been a need to improve the prediction accuracy of blast furnace operation data that occurs infrequently (low-frequency blast furnace operation data). The machine learning device for blast furnace operation described in Patent Document 1 primarily aims to improve the prediction accuracy of high-frequency blast furnace operation data, and has not adequately considered how to improve the prediction accuracy of low-frequency blast furnace operation data. [Means for solving the problem]
[0008] To solve the above problems, the machine learning device for blast furnace operation according to the present invention is characterized by having (1) a data preprocessing unit that performs a smoothing step of generating smoothed blast furnace operation data by smoothing each of a plurality of blast furnace operation data, and a normal distribution data unification step of unifying the smoothed blast furnace operation data into normal distribution data by converting non-normal distribution data whose skewness exceeds a skewness threshold among the smoothed blast furnace operation data into normal distribution data, and a data type selection unit that selects explanatory variables corresponding to a predetermined target variable based on the data type of the normal distribution data, and a prediction model construction unit that constructs a prediction model for blast furnace operation by performing machine learning using the normal distribution data corresponding to the selected explanatory variables and the normal distribution data corresponding to the target variable as training data.
[0009] (2) The machine learning apparatus for blast furnace operation according to (1) above, characterized in that the data preprocessing unit performs a removal process to remove abnormal values contained in the blast furnace operation data before performing the smoothing process on the blast furnace operation data.
[0010] (3) The machine learning apparatus for blast furnace operation according to (2) above, characterized in that the data preprocessing unit performs the removal process according to the Hampel Identifier method based on the following formula (A). median(xi)±cMAD(xi)······························································································ Formula (A) However, MAD(xi) = median(|xi - median(xi)|), where median is the median value and c is a parameter selected from values greater than or equal to 5 (but not limited to integers).
[0011] (4) The machine learning apparatus for blast furnace operation according to any one of (1) to (3) above, characterized in that when the blast furnace operation data includes operation data of the same type acquired at different timings, the data type of these operation data is treated as different data types and an explanatory variable is selected accordingly.
[0012] (5) The learning device for blast furnace operation according to any one of (1) to (3) above, characterized in that the smoothing process is a process for calculating a moving average of blast furnace operation data.
[0013] (6) The machine learning apparatus for blast furnace operation according to (1) above, characterized in that the data preprocessing unit performs a removal process to remove outliers included in the blast furnace operation data before performing the smoothing process on the blast furnace operation data.
[0014] (7) The machine learning apparatus for blast furnace operation according to (6) above, characterized in that the data preprocessing unit performs the removal process according to the Hampel Identifier method based on the following formula (A). median(xi)±cMAD(xi)······························································································ Formula (A) However, MAD(xi) = median(|xi - median(xi)|), where median is the median value and c is a parameter selected from values between 3 and 5 (but not limited to integers). [Effects of the Invention]
[0015] The machine learning device for blast furnace operation of the present invention makes it possible to construct a predictive model with excellent prediction accuracy for low-frequency blast furnace operation data. [Brief explanation of the drawing]
[0016] [Figure 1] This is a functional block diagram of a machine learning system for blast furnace operation. [Figure 2] This is a flowchart explaining how to create a blast furnace operation prediction model. [Figure 3] This bar graph shows the parameter c that minimizes the RMSE of solution loss carbon within ±2σ and the parameter c that minimizes the RMSE of solution loss carbon outside ±2σ (operational data: measured values at 1-hour intervals). [Figure 4]It is a bar graph showing parameter c that minimizes the RMSE of the hot metal temperature within ±2σ and parameter c that minimizes the RMSE of the hot metal temperature outside ±2σ (operation data: measured values at hourly intervals). [Figure 5] It is a bar graph showing the improvement effect of RMSE when parameter c is set to the optimal parameter c and the amount of solution loss carbon and hot metal temperature outside ±2σ are predicted. [Figure 6] It is a bar graph showing the improvement effect of RMSE of the first example corresponding to the first embodiment. [Figure 7] It is a bar graph showing the improvement effect of RMSE of the second example corresponding to the second embodiment. [Embodiment for Carrying Out the Invention]
[0017] (Definition of Terms) Among the blast furnace operation data, the blast furnace operation data with low occurrence frequency is defined as "low-frequency blast furnace operation data". "Blast furnace operation data with low occurrence frequency (low-frequency blast furnace operation data)" is defined as blast furnace operation data that deviates from the average value beyond the range of ±predetermined σ. "Predetermined σ" can be, for example, any value of 1.5σ or more. Therefore, blast furnace operation data beyond the range of ±1.5σ may be regarded as "low-frequency blast furnace operation data", or blast furnace operation data beyond the range of ±2.0σ may be regarded as "low-frequency blast furnace operation data". The relatively frequently occurring blast furnace operation data that does not belong to the low-frequency blast furnace operation data is defined as "high-frequency blast furnace operation data".
[0018] The purpose of the present invention is to improve the prediction accuracy of low-frequency blast furnace operation data while sacrificing the prediction accuracy of high-frequency blast furnace operation data. The configuration of the learning device for blast furnace operation for achieving such a purpose will be described while referring to FIG. 1. FIG. 1 is a functional block diagram of the machine learning device for blast furnace operation of the present embodiment. Referring to the same figure, the machine learning device 1 for blast furnace operation includes a data preprocessing unit 10, a data type selection unit 20, a prediction model construction unit 30, and a storage medium 40. These elements work together to construct a blast furnace operation prediction model with excellent prediction accuracy for low-frequency blast furnace operation data.
[0019] (First Embodiment) The method for constructing the blast furnace operation prediction model in this embodiment will be explained in detail with reference to the flowchart in Figure 2. The construction of the blast furnace operation prediction model is carried out in the following order: Step S1: Data preprocessing, Step S2: Data type selection, and Step S3: Prediction model construction. Each step will be explained in detail below.
[0020] (Step S1: Data preprocessing) In step S1, the data preprocessing unit 10 preprocesses the operational data acquired from the blast furnace 2. This preprocessing includes removing outliers (step S1-1), smoothing (step S1-2), and normalizing (step S1-3). However, if there are no outliers, step S1-1 may be omitted. In this case, the data preprocessing in step S1 starts from step S1-2.
[0021] (Step S1-1: Removal of outliers) Blast furnace operation data includes information obtained from the blast furnace during operation (including during shutdown) and information calculated from such obtained information. Blast furnace operation data may contain outliers for various reasons (e.g., system malfunctions). If outliers are included in the operation data, it will affect the selection of training data and the prediction accuracy of the prediction model. Therefore, it is desirable to remove outliers included in the operation data (equivalent to "outlier removal") according to predetermined conditions. The predetermined conditions will be described later.
[0022] Here, operational data includes, for example, blast furnace operating conditions, charging raw material conditions, slag data, sensor data, solution loss carbon amount, and tuyeres heat balance. The operating conditions of a blast furnace include the airflow rate from the tuyeres, the oxygen enrichment rate of the oxygen-enriched air blown in from the tuyeres (oxygen enrichment rate: (amount of oxygen in the airflow + amount of enriched oxygen) / (airflow rate) × 100 - oxygen concentration in the airflow (%)), the airflow temperature of the gas blown in from the tuyeres, the airflow pressure of the gas blown in from the tuyeres, the amount of pulverized coal blown in, and the furnace top pressure. The charging conditions for blast furnace raw materials include the amount of blast furnace raw materials charged and the composition of the blast furnace raw materials. Blast furnace raw materials include ore layer charging materials used to form the ore layer and coke layer charging materials used to form the coke layer. Ore layer charging materials include sintered ore, pellets, lump ore, and uncalcined carbon-containing lump ore. In addition, ore layer charging materials may include materials other than ore (for example, reducing aids such as small lump coke). Coke layer charging materials may include ferrocoke. Slag refers to the high-temperature molten material removed from the tap of a blast furnace. Slag data includes molten iron discharge volume, molten iron temperature, molten iron composition, slag discharge volume, slag composition, etc. Sensor data refers to data directly detected by sensors installed in the blast furnace, and includes data such as top gas temperature data, top gas component analysis data (CO, CO2, N2, H2, etc.), stave temperature data, bottom wall temperature data, bottom temperature data, and shaft pressure data. The heat balance above the tuyere is the difference between a predetermined input heat quantity (sensible heat of gas blown in from the tuyere and heat of combustion of carbon burning in the furnace) and a predetermined output heat quantity (heat of decomposition of moisture in the gas blown in from the tuyere, and heat of reaction from direct reduction reactions and solution loss reactions). It is an indicator used to sequentially evaluate the heat balance inside the furnace.
[0023] Outliers can be removed, for example, by the following formula (1) which follows the "Hampel Identifier method".
number
[0024] In this embodiment, parameter c is set to 5 or higher. This is because setting parameter c to 5 or higher improves the prediction accuracy of low-frequency blast reactor operation data. The inventors obtained this finding through the following analysis process using the model of Patent Document 1.
[0025] Furnace capacity 5000m 3 In a blast furnace of a specific grade with a reducing agent ratio of 500 kg / t-pig, the RMSE (Relative Mean Squared Error) was determined when predicting solution loss carbon within ±2σ range, and when predicting solution loss carbon outside ±2σ range, while varying the value of parameter c. Since this is the model described in Patent Document 1, no smoothing of the blast furnace operation data was performed; instead, measurement data at one-hour intervals (i.e., raw data) was used. "RMSE" is an index of prediction error and is also called the mean squared error.
[0026] Note that "solution loss carbon amount within ±2σ" is an example of high-frequency blast furnace operation data, and "solution loss carbon amount outside ±2σ" is an example of low-frequency blast furnace operation data. The reason for using "an example" is that in this specification, blast furnace operation data that deviates from the average value beyond the range of ± predetermined σ is defined as low-frequency blast furnace operation data, and predetermined σ can be any value greater than or equal to 1.5σ.
[0027] Furthermore, the inventors performed the same analysis process for molten iron temperature as they did for solution loss carbon, and obtained the RMSE when predicting molten iron temperatures within ±2σ range and when predicting molten iron temperatures outside the ±2σ range, while changing the parameter c.
[0028] The bar graph in Figure 3 shows the parameter c that minimizes the RMSE of solution loss carbon within the ±2σ range, and the parameter c that minimizes the RMSE of solution loss carbon outside the ±2σ range (hereinafter also referred to as the optimal parameter c). In other words, Figure 3 shows the value of parameter c that yields the highest prediction accuracy. The bar graph in Figure 4 corresponds to Figure 3 and shows the parameter c that minimizes the RMSE for molten iron temperatures within the ±2σ range, and the parameter c that minimizes the RMSE for molten iron temperatures outside the ±2σ range (hereinafter also referred to as the optimal parameter c). From Figures 3 and 4, we can conclude that "the optimal parameter c for predicting blast furnace operation data outside the ±2σ range is 5 or greater." The bar graph in Figure 5 shows the improvement in RMSE when parameter c is set to the optimal parameter c and solution loss carbon amount and molten iron temperature outside the ±2σ range are predicted. The improvement effect was evaluated by using the RMSE when parameter c is set to 4 and solution loss carbon amount and molten iron temperature outside the ±2σ range are predicted as the baseline RMSE, and calculating the difference (Δrelative RMSE) from this baseline RMSE.
[0029] From Figures 3 to 5, it can be considered that setting parameter c to 5 or higher improves the prediction accuracy of low-frequency blast reactor operation data.
[0030] Figures 3 to 5 show the prediction accuracy of a prediction model constructed by machine learning on blast furnace operation data (raw data) that has not undergone smoothing. However, the inventors hypothesized that if the prediction model was constructed by machine learning on smoothed blast furnace operation data, the effect of improving prediction accuracy by setting parameter c to 5 or higher would be even greater. The inventors confirmed that this hypothesis was correct in the first embodiment described later.
[0031] Furthermore, as mentioned above, the Hampel Identifier method determines whether a value is an outlier based on its "degree of deviation" from the median rather than the mean. Therefore, it can identify outliers even in data that does not necessarily follow a normal distribution. Consequently, the Hampel Identifier method is an excellent tool for processing blast furnace operation data.
[0032] If there are missing data in the blast furnace operation data, interpolation processing may be performed. Such interpolation processing can also be performed if data gaps occur due to the removal of outliers. The interpolation method is not particularly limited, but for example, methods such as interpolation using the previous value or the average of preceding and succeeding values, or linear interpolation based on preceding and succeeding values can be used.
[0033] (Step S1-2: Smoothing process) When aiming to predict low-frequency blast furnace operation data, small fluctuations can overlook larger fluctuations due to their influence. Therefore, a smoothing process is implemented to smooth the fluctuations in the time-series data of blast furnace operation data. By using the smoothed blast furnace operation data (hereinafter also referred to as smoothed blast furnace operation data) as training data, as described later, it is possible to construct a blast furnace operation prediction model with excellent prediction accuracy for low-frequency blast furnace operation data.
[0034] Smoothed blast furnace operation data may also be a moving average of blast furnace operation data. A moving average is calculated by shifting the average value for each interval in time series data. The "interval" is not particularly limited, but for example, in the case of blast furnace operation data measured every hour, a 2-hour "interval" can be used. In this case, the moving average at time T can be the average of the measurement taken one hour before time T and the measurement taken one hour after time T. However, data obtained by smoothing time-series blast furnace operation data using a low-pass filter may also be referred to as "smoothed blast furnace operation data."
[0035] (Step S1-3: Normalization process) As mentioned above, blast furnace operation data often exhibits a skewed distribution rather than a normal one. In other words, smoothed blast furnace operation data includes data with a non-normal distribution. Many blast furnace operation prediction models assume that the operation data used as training data for machine learning follows a normal distribution. When non-normally distributed data is used, it becomes impossible to construct an appropriate model (a model that satisfies the desired prediction accuracy). Therefore, in step S1-3, a process is performed to convert the non-normal distribution to a normal distribution. Prediction accuracy is improved by constructing a prediction model using a normal distribution. However, this is common technical knowledge, so a detailed explanation is omitted. Needless to say, if the blast furnace operation data is originally smoothed and follows a normal distribution, then normalization processing is unnecessary.
[0036] Specifically, the data preprocessing unit 10 determines whether the absolute value of the skewness in the smoothed blast furnace operation data is greater than a threshold (hereinafter also referred to as the skewness threshold). If the absolute value of the skewness exceeds the skewness threshold, the smoothed blast furnace operation data is processed to approximate a normal distribution (hereinafter also referred to as the normalization process). For the normalization process, for example, the "Box-Cox transform" and the "Yeo-Johson transform" can be used. The general formulas for the "Box-Cox transform" and the "Yeo-Johson transform" are shown in equations (2) and (3), respectively. Note that "Equation 2" corresponds to equation (2), and "Equation 3" corresponds to equation (3).
number
number
[0037] The skewness threshold is not limited to any particular type of blast furnace operation data, as described in Patent Document 1. For example, an appropriate skewness threshold can be found by determining the relationship between the skewness threshold and the RMSE while varying the skewness threshold for the predicted molten iron temperature.
[0038] (Step S2: Selection of explanatory variables) By performing step S1, all acquired blast furnace operation data are standardized to a normal distribution, resulting in a group of normally distributed data consisting of a large number of normally distributed data. This group of normally distributed data may include smoothed blast furnace operation data that has been determined not to require normalization processing.
[0039] In step S2, the data type selection unit 20 selects explanatory variables corresponding to a predetermined dependent variable based on the data type of the normally distributed data. "Data type" refers to the category of each operational data and is a different concept from the operational data (numerical value). Therefore, both explanatory variables and dependent variables in this specification refer to categories of operational data. For example, if the user sets the data type (category) as "pulverized coal ratio," the data type selection unit 20 can select the data type (category) as the explanatory variable corresponding to this "pulverized coal ratio."
[0040] When selecting explanatory variables, temporal factors can also be considered. That is, operational data acquired at different times can be treated as operational data of different data types. For example, "pulverized coal injection amount from XX o'clock to YY o'clock" and "pulverized coal injection amount from YY o'clock to YY o'clock" are different data types and can be treated as independent explanatory variables.
[0041] Significant and independent explanatory variables that explain the dependent variable may be automatically selected (automatic selection method). Examples of known automatic selection methods include Random Forest (RF), LASSO, and EN. Random Forest (RF): This method selects variables based on their importance in a Random Forest (RF) analysis. It selects the variables with the highest importance after running RF. LASSO method: LASSO is an abbreviation for "Least Absolute Shrinkage and Selection Operator." In this method, the standard regression coefficients of variables that do not contribute to the model tend to be zero, so only variables with non-zero standard regression coefficients are selected. EN method: EN is an abbreviation for "Elastic Net," and like the LASSO method, the standard regression coefficients tend to be zero. Therefore, variables that are not zero should be selected.
[0042] (Step S3: Building a predictive model for blast furnace operation) The prediction model construction unit 30 creates a blast furnace operation prediction model (in other words, a neural network) using supervised machine learning, with normally distributed data corresponding to the explanatory variables selected in step S2 and normally distributed data corresponding to a predetermined target variable as training data. It goes without saying that these normally distributed data are selected from the aforementioned group of normally distributed data.
[0043] Machine learning is an artificial intelligence (AI) technology that uses predetermined data as input to make predictions based on predetermined data analysis methods. Data analysis methods refer to analytical algorithms used to analyze data. Therefore, blast furnace operation prediction models are constructed for each target variable predetermined by the user, and the blast furnace operation prediction model is built using machine learning technology according to the analytical algorithm appropriate for each target variable. The constructed blast furnace operation prediction model can be tuned as needed. The tuning method is described below. Similar operation data with different time series from the operation data selected as training data are defined as input values and actual values. Then, predicted values based on the blast furnace operation prediction model are obtained from these input values, and if the difference between these predicted values and actual values exceeds a predetermined error, the blast furnace prediction model is tuned. Tuning can be done, for example, by adjusting hyperparameters. Hyperparameters can be adjusted using various methods that are publicly available as open source.
[0044] Two methods can be considered to account for this time-series nature. The first is to treat data from the past few hours (e.g., 8 hours) as a separate variable and select the optimal training data. The second is to incorporate past time-series information (e.g., 1 to 8 hours ago) as training data when constructing a neural network, and construct the neural network as an RNN (Recurrent Neural Network) structure. Time-series information refers to the time when the operational data was acquired. Time-series nature can be considered using either of these two methods.
[0045] When constructing a neural network, overfitting can be avoided by the following method (a known method): Since overfitting occurs when the number of weight coefficients in the neural network exceeds the number of explanatory variables, weight coefficients can be appropriately removed. This avoids overfitting and reduces the computation of the neural network's weight coefficients. For example, overfitting can be avoided by applying dropout and structured learning with forgetting.
[0046] For selecting the coefficient networks to eliminate, one can use a method such as the bootstrap-LASSO method, which selects variables that frequently have a coefficient of 0 and sets their connection coefficients to 0. This avoids overfitting and allows for more efficient construction of the learning model. Alternatively, a method of randomly selecting the coefficient networks to eliminate may also be employed.
[0047] By performing step S3, it is possible to construct a blast furnace operation prediction model with excellent prediction accuracy for low-frequency blast furnace operation data.
[0048] The processes described in steps S1 to S3 can be realized by a program. That is, a program prepared in advance to realize various processes is stored in the storage medium 40, and a CPU or the like included in the computer executes the program stored in the storage medium 40, whereby it can be realized. The data preprocessing unit 10, the data type selection unit 20, and the prediction model construction unit 30 can be realized by a CPU or the like included in the computer.
[0049] (Second Embodiment) In the first embodiment, the parameter c set in the outlier removal (step S1-1) was limited to 5 or more. However, in this embodiment, the parameter c is set to be 3 or more and less than 5. When the parameter c is 3 or more and less than 5, in addition to outliers, data deviated from the average value that cannot be said to be outliers is also removed. Therefore, in this embodiment, the process in step S1-1 is also referred to as "deviation value removal". As described in the first embodiment, when predicting low-frequency blast furnace operation data, it is desirable to set the parameter c to 5 or more. However, even if the parameter c is less than 5, by using the smoothed blast furnace operation data, the prediction accuracy of the low-frequency blast furnace operation data can be improved. The present inventors have confirmed such an effect in the second embodiment described later.
[0050] The present invention will be specifically described by showing the first embodiment and the second embodiment. The first embodiment corresponds to the first embodiment, and the second embodiment corresponds to the second embodiment. (First Embodiment) Furnace volume 5000m 3A predictive model was constructed to predict the amount of solution loss carbon and molten iron temperature in a blast furnace of a specific grade with a reducing agent ratio of 500 kg / t-pig. Parameter c was set to the "optimal parameter of 5 or greater," and the blast furnace operation data was set to a "2-hour moving average (normally distributed data)." Based on the predictive model constructed under these conditions, the amount of solution loss carbon and molten iron temperature outside the ±2σ range predicted were compared with the amount of solution loss carbon and molten iron temperature of the comparative example, respectively, to investigate the improvement effect on RMSE. The predictive model for the comparative example was constructed with parameter c set to "4" and the blast furnace operation data as "1-hour interval data (normally distributed data)." The comparative example corresponds to the predicted values when predicted according to the predictive model in Patent Document 1.
[0051] Figure 6 is a bar graph showing the improvement in RMSE compared to the comparative example. From Figure 6, it can be seen that if the prediction model is set to parameter c as the "optimal parameter of 5 or more" and the blast furnace operation data as a "moving average (normal distribution data)", the prediction accuracy of solution loss carbon amount and molten iron temperature outside the ±2σ range is significantly improved.
[0052] (Second example) Furnace capacity 5000m 3 A predictive model was constructed to predict the amount of solution loss carbon and molten iron temperature in a blast furnace of a specific grade with a reducing agent ratio of 500 kg / t-pig. Parameter c was set to "4," and the blast furnace operation data was set to a "2-hour moving average (normally distributed data)." Based on the predictive model constructed under these conditions, the amount of solution loss carbon and molten iron temperature outside the predicted ±2σ range were compared with the amount of solution loss carbon and molten iron temperature of the comparative example, respectively, to investigate the improvement effect on RMSE. The predictive model for the comparative example was the same as that used in the first example.
[0053] Figure 7 is a bar graph showing the improvement in RMSE compared to the comparative example. From Figure 7, it can be seen that with a prediction model set to parameter c as "4" and blast furnace operation data as a "moving average (normally distributed data)", the prediction accuracy of solution loss carbon amount and molten iron temperature outside the ±2σ range improves. The improvement in prediction accuracy is not as high as in the first example, but it is better than the comparative example. [Explanation of symbols]
[0054] 1. Machine learning equipment for blast furnace operation 2 blast furnace 10 Data Preprocessing 20. Data Type Selection Section 30 Predictive Model Construction Department 40 Storage medium
Claims
1. A data preprocessing unit that performs a smoothing step to generate smoothed blast furnace operation data by smoothing multiple sets of blast furnace operation data, and a normal distribution data unification step that unifies the smoothed blast furnace operation data into normal distribution data by converting non-normal distribution data whose skewness exceeds a skewness threshold among these smoothed blast furnace operation data into normal distribution data, A data type selection unit that selects explanatory variables corresponding to a predetermined dependent variable based on the data type of normally distributed data, A predictive model construction unit constructs a predictive model for blast furnace operation by performing machine learning using normally distributed data corresponding to the selected explanatory variables and normally distributed data corresponding to the objective variable as training data, and A machine learning device for blast furnace operation, characterized by having the following features.
2. The machine learning apparatus for blast furnace operation according to claim 1, characterized in that the data preprocessing unit performs a removal process to remove abnormal values contained in the blast furnace operation data before performing the smoothing process on the blast furnace operation data.
3. The machine learning apparatus for blast furnace operation according to claim 2, characterized in that the data preprocessing unit performs the removal process according to the Hampel Identifier method based on the following formula (1). median(x i )±cMAD(x i )・・・・・・・Style (1) However, MAD(x i ) = median(|x i -median(x i )|) where median is the median value and c is a parameter selected from values greater than or equal to 5 (but not limited to integers).
4. The machine learning apparatus for blast furnace operation according to any one of claims 1 to 3, characterized in that when the blast furnace operation data includes operation data of the same type acquired at different timings, the data type of these operation data is treated as different data types for explanatory variables.
5. The smoothing process described above is a process for calculating the moving average of blast furnace operation data. A learning device for blast furnace operation according to any one of features 1 to 3.
6. The machine learning apparatus for blast furnace operation according to claim 1, characterized in that the data preprocessing unit performs a removal process to remove outliers included in the blast furnace operation data before performing the smoothing process on the blast furnace operation data.
7. The machine learning apparatus for blast furnace operation according to claim 6, characterized in that the data preprocessing unit performs the removal process according to the Hampel Identifier method based on the following formula (1). median(x i )±cMAD(x i )・・・・・・・Style (1) However, MAD(x i ) = median(|x i − median(x i )|), where median is the median value, and c is a parameter selected from values greater than or equal to 3 and less than 5 (not limited to integers).