How to build a score calculation model

Abstract

To construct a score calculation model using a machine learning algorithm to determine an instability score for diagnosing stable or unstable furnace conditions.SOLUTION: By performing data processing on index data (operation control index), an instability score indicating the degree of instability of a furnace condition is determined. Of a plurality of instability scores determined, the instability score indicating the stability of furnace conditions is extracted. For each of a plurality of thresholds, the index data from the extracted instability score is used as teacher data to construct a score calculation model. Deviations between the instability score determined by an operator from the index data and the instability score determined using each of the plurality of score calculation models constructed is obtained. A score calculation model when the deviations are minimal is determined as a score calculation model to determine the instability score when diagnosing stability or instability of the furnace condition. The threshold at which the determined score calculation model is constructed is determined as a threshold for diagnosing stable or unstable furnace conditions.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a method for constructing a score calculation model for determining an instability score to be compared with a threshold value when diagnosing whether the furnace state of a blast furnace is stable or unstable. [Background technology]

[0002] In the method for determining the furnace condition of a blast furnace described in Patent Document 1, individual instability scores for determining the instability of the furnace condition are calculated based on index data of operation management indexes in the blast furnace, and when a total instability score calculated from the individual instability scores falls within a predetermined range defined by a predetermined score, the furnace condition is determined to be unstable. Here, the individual instability score is calculated from a calculation formula including a reference value and index data. Furthermore, the total instability score is the sum or average value of the individual instability scores calculated for multiple types of operation management indexes. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-080556 Summary of the Invention [Problem to be solved by the invention]

[0004] The method for determining reactor conditions described in Patent Document 1 does not utilize a machine learning model. The inventors of the present application focused on utilizing a machine learning model in diagnosing whether the reactor conditions are stable or unstable, and as a result of extensive research, have completed the present invention. [Means for solving the problem]

[0005] In the method for constructing a score calculation model of the present invention, index data related to blast furnace operation management indexes is acquired, and data processing is performed on the index data to determine an instability score indicating the degree of instability of the blast furnace's furnace conditions. Then, from the multiple instability scores determined by the data processing, an instability score indicating the stability of the furnace conditions determined based on a magnitude relationship with a threshold is extracted. For each of the multiple thresholds, the index data used to determine the extracted instability score is used as training data, and a score calculation model for determining the instability score from the index data is constructed using a machine learning algorithm.

[0006] Next, the deviation between the instability score determined by the operator from the index data and the instability score determined using each of the constructed score calculation models is determined. The score calculation model that minimizes the deviation is determined as the score calculation model for determining the instability score when diagnosing the stability or instability of the reactor state. In addition, the threshold value used when the determined score calculation model was constructed is determined as the threshold value when diagnosing the stability or instability of the reactor state.

[0007] The instability score indicating the stability of the reactor condition can be set to an instability score below a threshold value. The instability score can be set within a range of 0.0 to 1.0. Here, the reactor condition when the instability score is 0.0 is more stable than the reactor condition when the instability score is 1.0.

[0008] The multiple thresholds can be in the range of 0.0 to 1.0. When multiple types of machine learning algorithms are prepared, a threshold for diagnosing whether the reactor condition is stable or unstable can be determined for each type of machine learning algorithm. The total number of index data used to construct the score calculation model is preferably greater than the total number of index data when the operator determines the instability score.

[0009] The above-mentioned data processing can be performed using a score calculation algorithm. The deviation can be calculated using the root mean square error. The stability or instability of the reactor state can be diagnosed by comparing the instability score obtained from the determined score calculation model with the determined threshold value. [Effects of the Invention]

[0010] According to the present invention, it is possible to construct a score calculation model (machine learning model) that determines an instability score to be compared with a threshold value when diagnosing whether the reactor condition is stable or unstable. [Brief explanation of the drawings]

[0011] [Figure 1] 10 is a flowchart illustrating a method for diagnosing the furnace condition. [Figure 2] 10 is a flowchart illustrating a method for constructing a score calculation model. [Figure 3] 1 is a block diagram illustrating a system for performing a method for building a score calculation model. [Figure 4] FIG. 10 is a diagram showing the frequency distribution of instability scores. [Figure 5] FIG. 10 is a diagram showing the relationship between the threshold Sth and RMSE in two types of machine learning algorithms. [Figure 6] 10 is a flowchart illustrating a method (modification) for constructing a score calculation model. DETAILED DESCRIPTION OF THE INVENTION

[0012] This embodiment is a method for constructing a score calculation model that determines an instability score to be compared with a threshold value when diagnosing whether the furnace condition of a blast furnace is stable or unstable. By determining the instability score using the constructed score calculation model, it is possible to perform a diagnosis of whether the furnace condition is stable or unstable, similar to a diagnosis by an operator. Hereinafter, the method for constructing the score calculation model of this embodiment will be specifically described.

[0013] (Score calculation model) The score calculation model in this embodiment calculates an instability score St, which is a determination factor for instability of the furnace state of a blast furnace, by inputting data of operation management indicators (hereinafter referred to as index data) described later. As will be described later, by comparing the instability score St with a threshold value Sth, it is possible to diagnose whether the furnace state is stable or unstable. By diagnosing the stability or instability of the furnace state using the instability score St calculated from the score calculation model, it is possible to diagnose the stability or instability of the furnace state without depending on the ability, experience, etc. of the operator.

[0014] The score calculation model is constructed based on a known machine learning algorithm. For example, One Class SVM (Support Vector Machine) or autoencoder (autoencoder) can be used as this machine learning algorithm. One Class SVM is a method that applies SVM, a machine learning classification algorithm, to unsupervised one-class classification, in which one class's worth of normal data is trained, a classification boundary is determined, and outliers are detected based on that boundary. An autoencoder is an algorithm that uses a neural network in machine learning. One Class SVM and autoencoder are known technologies, so detailed explanations will be omitted.

[0015] (Operational management indicators) Operational control indicators are indicators that affect the diagnosis of the stability or instability of furnace conditions, and are generally used to manage the operation of a blast furnace. Examples of operational control indicators include those shown in Table 1 below. As shown in Table 1 below, operational control indicators include indicators obtained at the furnace top, indicators obtained inside the furnace or on the furnace wall, and indicators obtained at the tuyere or tap hole. Note that the operational control indicators are not limited to the indicators shown in Table 1 below. Furthermore, the operational control indicators do not need to be all the indicators listed in Table 1 below, and may be only a portion (multiple) of the indicators listed in Table 1 below that are highly important in diagnosing the stability or instability of furnace conditions.

[0016] [Table 1]

[0017] The index data can be measured by a sensor installed in the blast furnace, or can be obtained by performing arithmetic processing on the measurement data of the sensor. The index data can also be acquired at a predetermined interval. The predetermined interval can be determined as appropriate, but can be, for example, any time between 10 minutes and 60 minutes, for example, 30 minutes.

[0018] The following describes the details of each operation control index shown in Table 1. Note that the measurement method or calculation method for each operation control index described below is publicly known, so detailed explanation will be omitted.

[0019] "Daily charging count" refers to the total number of charges (coke or ore) charged per day into a blast furnace. "Cantilever probe temperature" refers to the furnace gas temperature [°C] measured by a cantilever probe installed at the furnace throat. "Average furnace top temperature" refers to the average furnace top gas temperature [°C] measured by the riser pipe at the top of the furnace. "Number of abnormal index finger descents" refers to the number of times abnormal burden descents (such as slips or drops) occur in the blast furnace. "Charging depth at landing" refers to the depth [m] when the sounding lands on the burden. "Charging depth at hoisting" refers to the depth [m] of the burden when the sounding is hoisted up. "Sounding descent rate" refers to the descent rate [mm / min] of the burden measured by sounding. The "index finger layer thickness ratio" is the ratio of the ore layer thickness to the coke layer thickness [%] measured by sounding.

[0020] "Top gas CO analysis value," "Top gas CO2 analysis value," "Top gas H2 analysis value," "Top gas N2 analysis value," and "Top gas CH4 analysis value" are the composition [mol%] of each gas component (CO, CO2, H2, N2, CH4) contained in the top gas measured at the top of the furnace. "Top gas flow rate" is the flow rate [Nm 3 / min]. "Furnace throat gas flow rate" is the furnace top gas flow rate [m / sec] calculated by dividing the furnace top gas flow rate by the furnace throat cross-sectional area. "Furnace top spray flow rate" is the flow rate [t / h] of cooling water sprayed inside the furnace to adjust the furnace top temperature.

[0021] "Average gas flow velocity in the furnace" is the average gas flow velocity [m / sec] obtained by dividing the average of the gas volume and bosh gas volume at the furnace top by the average cross-sectional area in the furnace, and then correcting it by the average temperature and pressure at the tuyere tip and furnace top. "Shaft pressure" is the pressure [kPa] measured by a pressure sensor installed in the shaft. "Shaft pressure fluctuation time" is the time [sec] during which the pressure measured by each of the multiple pressure sensors installed around the furnace circumference in the shaft is equal to or greater than a predetermined value set at each level. "Stave temperature" is the temperature [°C] measured at the stave in the furnace wall.

[0022] The "K value (total)" is the permeability resistance index [-] for the entire blast furnace calculated by the following formula (1).

[0023]

number

[0024] In the above formula (1), P blast is the air pressure [kg / cm 2 ], P tоp is the furnace top pressure [kg / cm 2 ], V is the Bosch gas volume [Nm 3 The K values ​​(top, middle, and bottom) shown in Table 1 above were calculated by applying the applied formula (top pressure P tоp is the airflow resistance index calculated from the formula (where the pressure at the top, middle, and bottom are replaced with the pressure at the top, middle, and bottom, respectively).

[0025] "SLC amount" is the amount of solution loss carbon [kg / tp] in the blast furnace calculated from the blast conditions at the tuyere end and the furnace top gas conditions. "Total heat load" is the total value [MW] of the heat load calculated from the inlet and outlet temperatures of the stave cooling water, the specific heat of the water, and the flow rate of the cooling water for the entire blast furnace. "Blast pressure" is the pressure [kPa] of the hot blast measured in front of the annular pipe. "Tuyere end pulverized coal ratio" is the amount of pulverized coal injected per ton of pig iron [kg / tp]. "PCI injection amount" is the amount of pulverized coal injected per blast amount (actual value) [g / Nm 3 "InputH2" is the amount of hydrogen input per ton of pig iron [kg / tp].

[0026] "Amount of pig iron produced at the tuyere" is the amount of pig iron produced per day [t / D] calculated from the raw materials that have descended to the tuyere. "Amount of slag produced at the tuyere" is the amount of slag produced per day [t / D] calculated from the raw materials that have descended to the tuyere. "PC replacement rate" is the replacement rate [-] of pulverized coal and coke calculated from the carbon content in the pulverized coal and coke. "Hot metal temperature" is the temperature [°C] of the molten iron measured by a temperature sensor in the hot metal runner.

[0027] Among the operation control indices shown in Table 1 above, there are operation control indices that include a plurality of index data acquired at a plurality of different positions of the blast furnace (a plurality of different positions in the furnace circumferential direction and the furnace height direction). In this case, the average value of the plurality of index data can be used as the index data of the operation control indices. For example, the average value of the charging depth at time of landing at a plurality of positions in the furnace circumferential direction can be used as the index data of the charging depth at time of landing, and the average value of the charging depth at time of hoisting at a plurality of positions in the furnace circumferential direction can be used as the index data of the charging depth at time of hoisting.

[0028] Index data for operation control indicators such as shaft pressure, shaft pressure fluctuation time, and stave temperature may be measured at multiple positions in the furnace circumferential direction and furnace height direction, and in this case, the average value of the index data at multiple positions can be used as the index data for each operation control indicator.Since the abnormal descent phenomenon in the number of abnormal index finger descents may occur at multiple positions in the furnace circumferential direction, the total value of the number of abnormal index finger descents at multiple positions (e.g., four directions) in the furnace circumferential direction can be used as the index data for the number of abnormal index finger descents.

[0029] (Unstable score St) The instability score St is a score indicating the degree of instability of the reactor state, and is a score compared with a threshold value to diagnose the stability or instability of the reactor state. In this embodiment, the larger the instability score St, the more likely the reactor state is to become unstable; in other words, the smaller the instability score St, the more likely the reactor state is to become stable. In this embodiment, the instability score St is set within a range of 0.0 or more and 1.0 or less, but is not limited to this. Note that the instability score St in this embodiment corresponds to the total instability score St described in Patent Document 1.

[0030] (Furnace condition diagnosis method) The method for diagnosing the furnace condition will be described with reference to the flowchart shown in Fig. 1. In the method for diagnosing the furnace condition, as will be described below, it is determined whether the furnace condition is stable or unstable.

[0031] In step S101, index data of operation management indexes is acquired. In step S102, the index data acquired in step S101 is input into the score calculation model to calculate the instability score St_model. As described above, the score calculation model is configured to output the instability score St_model when index data is input.

[0032] Here, a method (one example) for calculating the instability score St_model using the score calculation model will be described below.

[0033] When One Class SVM is used as the machine learning algorithm that forms the basis of the score calculation model, a discrimination boundary for a region where the reactor condition is considered to be stable (hereinafter referred to as the "stable region") for the index data is determined in advance. Then, the distance between the discrimination boundary and the index data for which the instability score St_model is to be determined can be defined as the instability score St_model. Here, for the discrimination boundary of the stable region, the instability score St_model can be defined as 0.0, and the instability score St_model can be defined as 1.0 for the Euclidean distance when the distance from the discrimination boundary is maximum. Then, depending on the distance from the discrimination boundary, the instability score St_model can be defined within a range of 0.0 or more and 1.0 or less.

[0034] When an autoencoder is used as the machine learning algorithm that forms the basis of the score calculation model, index data when the reactor condition is stable is used as input data, and a model is created in which the output data reproduces the input data as accurately as possible. Then, the degree of reproduction of the output data when the index data for which the instability score St_model is to be calculated is input to the autoencoder can be defined as the instability score St_model. Here, when the degree of reproduction is highest, the instability score St_model can be set to 0.0, and when the degree of reproduction is lowest, the instability score St_model can be set to 1.0.

[0035] In step S103, it is determined whether the instability score St_model calculated in step S102 is equal to or less than a threshold value Sth. As will be described in the score calculation model construction method described later, the threshold value Sth is determined depending on the type of machine learning algorithm that forms the basis of the score calculation model, and therefore the threshold value Sth corresponding to the machine learning algorithm used in step S102 is used as the threshold value Sth in step S103. If the instability score St_model is equal to or less than the threshold value Sth, the process proceeds to step S104, and if the instability score St_model is greater than the threshold value Sth, the process proceeds to step S105.

[0036] In step S104, the reactor conditions are diagnosed as stable. As described above, in this embodiment, the smaller the instability score St, the more likely the reactor conditions are to become stable. Therefore, when the instability score St_model is equal to or less than the threshold value Sth, the reactor conditions are diagnosed as stable.

[0037] In step S105, the reactor condition is diagnosed as unstable. As described above, in this embodiment, the larger the instability score St, the more likely the reactor condition becomes unstable. Therefore, when the instability score St_model is larger than the threshold value Sth, the reactor condition is diagnosed as unstable.

[0038] Depending on the content of the definition of the instability score St, the magnitude relationship between the instability score St and the threshold value St may be reversed when diagnosing the stability or instability of the reactor condition. That is, if the smaller the instability score St, the more likely the reactor condition is to become unstable, then it can be diagnosed that the reactor condition is unstable when the instability score St_model is equal to or less than the threshold value Sth. On the other hand, if the larger the instability score St, the more likely the reactor condition is to become stable, then it can be diagnosed that the reactor condition is stable when the instability score St_model is greater than the threshold value Sth.

[0039] The diagnostic method shown in FIG. 1 can be performed using a predetermined diagnostic device (not shown). This diagnostic device may include an acquisition unit (S101 in FIG. 1) that acquires index data, a calculation unit (S102 in FIG. 1) that calculates an instability score St_model using a score calculation model, and a diagnostic unit (S103 to S105 in FIG. 1) that diagnoses the stability / instability of the reactor condition based on a comparison between the instability score St_model and a threshold value Sth. Here, the machine learning algorithm and the threshold value Sth that form the basis of the score calculation model can be stored in a memory. The memory may be provided inside the diagnostic device or outside the diagnostic device. When the memory is provided outside the diagnostic device, the memory and the diagnostic device may be capable of communicating with each other wirelessly or via a wired connection.

[0040] When the furnace condition is diagnosed as unstable, the unstable furnace condition can be resolved by taking operational action to appropriately adjust the operational factors of the blast furnace operation. Here, examples of the operational factors of the blast furnace operation include the air blowing conditions from the tuyere, the conditions for charging raw materials (coke and ore) from the furnace top, and the properties of the raw materials. Whether the unstable furnace condition has been resolved can be determined by performing the process shown in Figure 1 after taking the above-mentioned operational action.

[0041] (How to build a score calculation model) The method for constructing a score calculation model will be described with reference to the flowchart shown in Fig. 2. Fig. 3 is a block diagram showing the configuration of a system that executes the method for constructing a score calculation model. The system 10 shown in Fig. 3 can execute the processing shown in Fig. 2.

[0042] In step S201, an operator determines an instability score St_оpe based on a plurality of index data acquired in period A. Here, in system 10 shown in Fig. 3, the instability score St_оpe determined by the operator is acquired.

[0043] The period A can be determined as appropriate, for example, one month. Furthermore, the index data can be acquired at a predetermined cycle Δt1 within the period A, and this cycle Δt1 can be, for example, one hour. The operator can be an operator with sufficient ability and experience. Each time index data is acquired, the operator determines an instability score St_оpe from the index data based on his or her own experience. The instability score St_оpe is generated for each index data acquired.

[0044] In step S202, the instability score calculation unit 11 uses a rule-based model to calculate the instability score St_r from the index data acquired at a predetermined cycle Δt2 within the period B. The rule-based model is a model that calculates the instability score St_r from the index data based on rules created from the empirical rules of blast furnace operation experts. In the rule-based model, rules indicating the correlation between the index data and the instability score St_r are predetermined, and therefore the instability score St_r can be calculated from the index data according to these rules.

[0045] The above-mentioned period B may be the same period as the period A described in step S201, or may be a period different from period A. Here, period B is preferably longer than period A. Furthermore, the above-mentioned cycle Δt2 is a cycle shorter than the cycle Δt1 described in step S201, and can be, for example, one minute. The instability score St_r is generated by the number of index data acquired within period B, but by making period B longer than period A or making the cycle Δt2 shorter than the cycle Δt1, the total number of instability scores St_r can be increased beyond the total number of instability scores St_оpe determined in step S201.

[0046] The unstable score St_r obtained in step S202 is used to construct a score calculation model in step S205, which will be described later. Specifically, as will be described in detail below, index data used when the unstable score St_r, which is equal to or less than the threshold value Sth, is obtained is used as training data for constructing the score calculation model. In constructing a score calculation model (machine learning algorithm), the use of more training data can improve the calculation accuracy of the score calculation model. Therefore, in order to increase the training data (index data), the total number of unstable scores St_r is made larger than the total number of unstable scores St_оpe, as described above.

[0047] By performing the processing of step S202, it is possible to obtain the frequency distribution (one example) shown in Fig. 4. Fig. 4 shows the frequency distribution (one example) for all the instability scores St_r obtained in step S202. In Fig. 4, the horizontal axis represents the instability score St_r (0.0 or more and 1.0 or less), and the vertical axis represents the relative frequency [-] of each instability score St_r.

[0048] In step S203, the threshold setting unit 12 sets an arbitrary threshold Sth as a candidate value. The threshold Sth set as a candidate value is a plurality of thresholds St included in a predetermined range, and the upper and lower limit values ​​defining the predetermined range can be determined appropriately. In this embodiment, since the instability score St can take a value of 0.0 or more and 1.0 or less, the predetermined range is a range included in 0.0 or more and 1.0 or less. As will be described later, the threshold Sth is used to identify index data that can be considered to indicate stable reactor conditions, so it is preferable to set the predetermined range to a range in which the instability score St is relatively small. For example, the predetermined range can be set to 0.0 or more and 0.5 or less.

[0049] In step S204, the score calculation model construction unit 13 extracts index data that indicates that the reactor condition is stable, based on the multiple instability scores St_r obtained in step S202 and the threshold value Sth (candidate value) set in step S203. Specifically, first, from all the instability scores St_r obtained in step S202, the instability scores St_r that are equal to or less than the threshold value Sth (candidate value) are extracted. The extracted instability scores St_r are at least a part of the frequency distribution shown in Fig. 4. Here, if the threshold value Sth (candidate value) set in step S203 changes, the extracted instability scores St_r also change.

[0050] Next, the index data used when the extracted instability score St_r was calculated is identified. In step S202, when the instability score St_r is calculated from the index data using the rule-based model, if the index data and the instability score St_r are stored in a linked state in memory, the index data used when the extracted instability score St_r was calculated can be identified.

[0051] In step S205, the score calculation model construction unit 13 constructs a score calculation model using the index data extracted in step S204. Specifically, the index data is used as training data for the score calculation model, and the score calculation model is constructed from a machine learning algorithm. A known method can be used to construct the score calculation model (machine learning algorithm) using the training data, so a detailed description will be omitted. The score calculation model constructed in step S205 is any one of the multiple types of machine learning algorithms described above (One Class SVM or autoencoder), and can be determined in advance.

[0052] In step S206, the instability score calculation unit 14 uses the score calculation model constructed in step S205 to calculate the instability score St_model from the index data acquired within the period A described in step S201. As described in step S201, multiple index data are acquired within the period A, and therefore the instability score St_model is calculated by inputting each index data into the score calculation model. Here, the total number of instability scores St_model is the same as the total number of instability scores St_ope determined in step S201.

[0053] In step S207, the RMSE calculation unit 15 calculates the RMSE (Root Mean Squared Error) based on the instability score St_оpe determined in step S201 and the instability score St_model calculated in step S206. Specifically, the RMSE is calculated based on the following formula (2). In the following formula (2), n is the total number of instability scores St_model and St_оpe.

[0054]

number

[0055] In this embodiment, the RMSE is calculated, but the present invention is not limited to this. Any method may be used as long as it can grasp the difference between the instability score St_ope determined in step S201 and the instability score St_model calculated in step S206.

[0056] In step S208, the threshold setting unit 12 determines whether or not all thresholds Sth included in the predetermined range described in step S203 have been set as candidate values. If all thresholds Sth have been set as candidate values, the process proceeds to step S209. On the other hand, if all thresholds Sth have not been set as candidate values, the process returns to step S203, and the thresholds Sth that have not been set are set as candidate values. Then, the processes from step S204 to step S207 are performed. In this way, the processes from step S203 to step S208 are repeated until all thresholds Sth have been set as candidate values. As a result, RMSE is calculated for each of the thresholds Sth (candidate values) set in step S203.

[0057] Figure 5 shows an example of the relationship between RMSE and threshold Sth (candidate value) for two types of machine learning algorithms. One Class SVM and autoencoder are used as the machine learning algorithms. The candidate values ​​for threshold Sth range from 0.1 to 0.5. As shown in Figure 5, RMSE changes depending on threshold Sth (candidate value), and a minimum value of RMSE occurs for each machine learning algorithm.

[0058] In step S209, the threshold determination unit 16 determines the threshold Sth that indicates the smallest RMSE calculated in step S207 from among all thresholds Sth set as candidate values. Here, when the RMSE is calculated in step S207, if this RMSE is associated with the threshold St (candidate value) set in step S203 and stored in memory, the threshold Sth that indicates the smallest RMSE can be identified.

[0059] In the example shown in Figure 5, when the machine learning algorithm is One Class SVM, the threshold Sth at which the RMSE is minimized is 0.30, whereas when the machine learning algorithm is an autoencoder, the threshold Sth at which the RMSE is minimized is 2.25.

[0060] The threshold value Sth determined in step S209 is used as the threshold value Sth to be compared with the instability score St_model obtained from the score calculation model constructed in step S205 when diagnosing whether the reactor condition is stable or unstable. Here, the threshold value Sth determined in step S209 can be stored in memory in a state linked to the score calculation model constructed in step S205.

[0061] As described above, once a score calculation model is constructed (S205 in FIG. 2) and the threshold value Sth used in this score calculation model is determined (S209 in FIG. 2), the furnace condition can be diagnosed as described in the flowchart shown in FIG. 1. In this furnace condition diagnosis, the score calculation model (step S205 in FIG. 2) constructed when the threshold value Sth determined in step S209 in FIG. 2 is set is used. Furthermore, the threshold value Sth determined in step S209 in FIG. 2 is used as the threshold value Sth for diagnosing whether the furnace condition is stable or unstable.

[0062] The RMSE calculated in step S207 of Fig. 2 indicates the difference between the instability score St_оpe determined by the operator (S201 of Fig. 2) and the instability score St_model calculated from the score calculation model (S206 of Fig. 2). Therefore, when the RMSE is minimum, it means that the difference between the instability score St_оpe and the instability score St_model is the smallest.

[0063] As described above, by diagnosing the reactor condition using the combination of the threshold value Sth and the score calculation model that minimizes the RMSE, the instability score St_model obtained from the index data can be made closest to the instability score St_ope (S201 in FIG. 2) determined by the operator. This makes it possible to derive the same instability score St as that determined by the operator using the score calculation model.

[0064] When the instability score St_оpe is determined by an operator, there is a limit to the determination work, and therefore it is not possible to determine a large number of instability scores St_оpe. On the other hand, if a score calculation model is constructed as in this embodiment, the instability score St_оdel can be calculated simply by inputting index data into the score calculation model, and there is no limit that arises from the determination work of an operator.

[0065] Furthermore, the threshold value Sth at which the RMSE is minimized is a value suitable for diagnosing the stability or instability of the reactor state using the instability score St_model obtained from the score calculation model. That is, the result of comparing the instability score St_model obtained from the score calculation model with the threshold value Sth (stable / unstable reactor state) tends to be equivalent to the result (stable / unstable reactor state) diagnosed by the operator.

[0066] (Variation) A modified example of the method for constructing a score calculation model will be described with reference to the flowchart shown in Fig. 6. In Fig. 6, the same processes as those described in Fig. 2 are assigned the same reference numerals as in Fig. 2, and detailed descriptions thereof will be omitted. Below, differences from the processes described in Fig. 2 will be mainly described.

[0067] In Fig. 2, the instability score St_r is calculated using a rule-based model in step S202, but in Fig. 6, the instability score St_model is calculated using a score calculation model in step S210. In Fig. 2, index data is extracted when the instability score St_r is equal to or less than the threshold value Sth in step S204, but in Fig. 6, index data is extracted when the instability score St_model calculated in step S210 is equal to or less than the threshold value Sth in step S211.

[0068] As the score calculation model used in step S210, it is possible to use the score calculation model constructed when the threshold value Sth determined in step S209 is set in the process shown in Fig. 2. That is, when constructing a score calculation model for the first time, the process shown in Fig. 2 is performed, and after constructing a score calculation model used for diagnosing the furnace condition by the process shown in Fig. 2, the process shown in Fig. 6 can be performed. This makes it possible to improve the calculation accuracy of the score calculation model.

[0069] Furthermore, when a new score calculation model for calculating the instability score St_model used to diagnose the reactor condition is constructed by the processing shown in FIG. 6, and when the processing shown in FIG. 6 is subsequently performed, the newly constructed score calculation model can be used as the score calculation model used in step S210.

[0070] Note that the score calculation model used in step S210 is not limited to the score calculation model described above, and a score calculation model previously constructed from any index data can also be used. Furthermore, in step S210, instead of a score calculation model (machine learning algorithm), a predetermined arithmetic expression can be used to calculate the unstable score St from the index data. For example, the unstable score St (total unstable score St) can be calculated using the method (calculation expression) described in Patent Document 1. Similar to step S202 shown in FIG. 2, step S210 is a process that only increases the total number of unstable scores St. Therefore, in step S210, it is sufficient to be able to increase the total number of unstable scores St, and the accuracy of the unstable score St_model is not required. In other words, in steps S202 and S210, it is sufficient to be able to calculate a large number of unstable scores from the index data by predetermined data processing.

[0071] The processes (so-called functions) described in FIGS. 1, 2, and 6 can be realized by a program. Here, in step S201 of FIGS. 2 and 6, a process for acquiring the instability score St_оpe determined by the operator can be performed. A computer program prepared in advance to realize each function is stored in an auxiliary storage device, and a control unit such as a CPU reads the program stored in the auxiliary storage device into a main storage device, and the control unit executes the program read into the main storage device, thereby operating each function. Each function can be operated by one control device, or by multiple control devices connected to each other.

[0072] The above program can also be provided to a computer in a state in which it is recorded on a computer-readable recording medium. Examples of recording media include optical disks such as CD-ROMs, phase-change optical disks such as DVD-ROMs, magneto-optical disks such as MOs (Magnet Optical) and MDs (Mini Disks), magnetic disks such as floppy disks and removable hard disks, and memory cards such as CompactFlash (registered trademark), SmartMedia, SD memory cards, and memory sticks. Also included as recording media are hardware devices such as integrated circuits (e.g., IC chips) specially designed and configured for the purposes of the present invention. [Explanation of symbols]

[0073] 10: System, 11: Instability score calculation part (rule-based model), 12: threshold setting unit, 13: score calculation model construction unit, 14: Instability score calculation unit (score calculation model), 15: RMSE calculation unit, 16: Threshold determination unit

Claims

1. Obtaining indicator data related to blast furnace operation management indicators, Data processing is performed on the index data to obtain an instability score indicating the degree of instability of the reactor condition; extracting the instability score that indicates the stability of the reactor condition determined based on a magnitude relationship with a threshold value from the plurality of instability scores obtained by the data processing; constructing a score calculation model for determining the instability score from the index data using a machine learning algorithm by using the index data used when determining the extracted instability score for each of the plurality of thresholds as training data; A deviation between the instability score determined by an operator from the index data and the instability score calculated using each of the plurality of score calculation models constructed is calculated; The score calculation model when the deviation is minimum is determined as the score calculation model for obtaining the instability score when diagnosing the stability or instability of the reactor state, and the threshold value when the determined score calculation model is constructed is determined as the threshold value when diagnosing the stability or instability of the reactor state, In diagnosing the stability or instability of the furnace condition, In a case where the instability score is decreased as the reactor state becomes more stable, the reactor state is diagnosed as stable when the instability score is equal to or less than the threshold value, and the reactor state is diagnosed as unstable when the instability score is greater than the threshold value; A method for constructing a score calculation model, characterized in that, when the instability score is increased as the reactor condition becomes more stable, the reactor condition is diagnosed as unstable when the instability score is below the threshold value, and the reactor condition is diagnosed as stable when the instability score is greater than the threshold value.

2. the instability score is equal to or greater than 0.0 and equal to or less than 1.0; 2. The method for constructing a score calculation model according to claim 1, wherein a reactor state when the instability score is 0.0 is more stable than a reactor state when the instability score is 1.

0.

3. 3. The score calculation model construction method according to claim 2, wherein the plurality of threshold values ​​are in the range of 0.0 to 1.

0.

4. 4. The score calculation model construction method according to claim 1, further comprising determining a threshold value for diagnosing whether the reactor condition is stable or unstable for each type of machine learning algorithm.

5. A method for constructing a score calculation model described in any one of claims 1 to 4, characterized in that the total number of index data used to construct the score calculation model is greater than the total number of index data when the operator determines the instability score.

6. 6. The score calculation model construction method according to claim 1, wherein the data processing is performed by the score calculation model.

7. 7. The score calculation model construction method according to claim 1, wherein the deviation is a root mean square error.

Images