Iron-making data sampling method fusing time decay and multi-parameter stability

Abstract

The invention belongs to the technical field of blast furnace ironmaking data processing, and discloses an ironmaking data sampling method fusing time decay and multi-parameter stability, and the method comprises the following steps: firstly, carrying out dual quality filtering on ironmaking historical data; partitioning according to time, and retaining the latest block data; and calculating a furnace condition stability index based on the average value of variable coefficients of multiple process parameters, obtaining a comprehensive weight in combination with a time attenuation coefficient, and periodically and dynamically updating a sampling strategy. The method overcomes the defect that the existing method is not combined with the characteristics of the blast furnace ironmaking process, improves the representativeness of the sampled sample, adapts to data mining and modeling requirements, is simple and convenient to implement, and can be used for data preprocessing of the full life cycle of the blast furnace.

Description

Technical Field

[0001] This invention belongs to the field of blast furnace ironmaking data processing technology, and relates to sampling blast furnace data during the ironmaking production process. It is applicable to the data preprocessing stage of blast furnace data mining, regression analysis and machine learning model training. Specifically, it is an ironmaking data sampling method that integrates time decay and multi-parameter stability. Background Technology

[0002] Sampling of blast furnace data is required during the ironmaking process. However, blast furnace ironmaking data exhibits significant time effects and non-stationarity, as specifically shown below:

[0003] (1) Impact of furnace age cycle: As the furnace age increases, furnace lining erosion and changes in thermal state lead to continuous drift in the distribution of process parameters, and the adaptability of long-term data to the current furnace condition decreases;

[0004] (2) Timeliness difference: Recent data directly reflects the current furnace condition characteristics and has higher reference value, while the reference value of long-term data naturally decreases over time;

[0005] (3) Stability differences: Historical data during the stable period of furnace conditions (such as the normal smelting stage) still have certain reference value, while the data during the fluctuating period of furnace conditions (such as after raw material replacement or equipment maintenance) have a rapidly declining reference value due to abnormal distribution.

[0006] Current data sampling methods for blast furnace ironmaking have the following shortcomings:

[0007] (1) The drawbacks of simple random sampling: it does not distinguish the time sequence and timeliness of the data, and is prone to including a large amount of low-value long-term data, resulting in insufficient representativeness of the sample;

[0008] (2) The drawback of fixed-proportion sampling: Data from each period are sampled at a uniform proportion, without taking into account the differences in furnace condition stability. The high proportion of data from the fluctuation period will interfere with the modeling results.

[0009] (3) The defects of general time-weighted sampling: it only relies on time decay to adjust the weights, without combining the characteristics of blast furnace process to correct the data stability, and cannot adapt to the parameter distribution changes caused by furnace condition fluctuations.

[0010] The aforementioned shortcomings make it difficult for existing sampling methods to remove redundant / interfering data while ensuring data representativeness, ultimately affecting the accuracy of blast furnace data mining (such as furnace temperature prediction and energy consumption analysis) and modeling. Summary of the Invention

[0011] This invention aims to provide a sampling method for ironmaking data that integrates time decay and multi-parameter stability, taking into account both time effects and furnace condition stability. It reflects the timeliness of data through time segmentation, ensuring high weight for recent data; calculates the furnace condition stability index based on the coefficient of variation (CV) of multiple process parameters, and corrects the weight of data from different periods; and outputs a high-quality sample set to adapt to the training needs of models such as linear regression, locally weighted regression, and machine learning, thereby improving modeling accuracy.

[0012] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0013] A method for sampling ironmaking data that integrates time decay and multi-parameter stability includes the following steps:

[0014] S1. Perform quality filtering on historical ironmaking data to remove outliers and low-quality data;

[0015] S2. Divide the filtered data into N blocks (N≥5) in ascending order of time, and retain the entire Nth block;

[0016] S3. Calculate the furnace condition stability index for each block based on the average coefficient of variation (CV) of multiple process parameters;

[0017] S4. Combine the time decay coefficient and stability index to calculate the comprehensive weight of each block;

[0018] S5. Randomly sample the first N-1 blocks without replacement according to the overall weight;

[0019] S6. Periodically re-execute S1~S5 to dynamically update the sampling dataset.

[0020] Furthermore, the quality filtering described in S1 employs a dual mechanism: first, significant outliers are removed through IQR, and then extreme values ​​are compressed by filtering through quantile ranges.

[0021] Furthermore, the method for removing significant outliers using IQR is as follows:

[0022] For each process parameter x, calculate the 25th quantile Q1 and the 75th quantile Q3, and obtain the interquartile range IQR = Q3 - Q1. Keep the data that satisfy Q1 - λ × IQR ≤ x ≤ Q3 + λ × IQR, and filter out significant outliers, where λ ranges from 1.5 to 4.5.

[0023] The method for filtering the quantile range is as follows:

[0024] Configure a uniform quantile threshold (lower quantile q_l and upper quantile q_h), and retain data whose parameter values ​​are in the range [q_l, q_h].

[0025] Furthermore, the calculation process for the furnace condition stability index in S3 is as follows:

[0026] S31. Calculate the CV of the j-th parameter in the i-th block: CVij = σij / μij (μj > 0);

[0027] S32. Calculate the global CV of the j-th parameter:

[0028] S33. Calculate the stability index of the j-th parameter in the i-th block: Sij = max(0, 1 - CVij / );

[0029] S34. Obtain the multi-parameter stability index of the i-th block by weighting according to preset weights: Ii= , where wj≥0 and .

[0030] Furthermore, the comprehensive weight formula in S4 is: W i =r Δt_i ×exp(-β·I i ), where Δt_i=Ni is the time interval, r∈(0,1) is the time decay coefficient, and β>0 is the stability influence coefficient.

[0031] Furthermore, r is preferably 0.7~0.95, and β is preferably 0.2~1.0.

[0032] Furthermore, the sampling size C in S5 i =max(N min ,⌊len(block i )×W i ⌋), where N min Typically, this is set to 5% to 10% of the data volume per block (to ensure that each block retains at least some data to avoid the loss of critical information).

[0033] Furthermore, the dynamic update cycle in S6 is adjusted according to the frequency of changes in blast furnace conditions.

[0034] The beneficial effects of this invention are:

[0035] Compared with existing technologies, the beneficial effects of the present invention are as follows:

[0036] 1. More representative sample: By combining the "multi-parameter CV stability index" with the characteristics of blast furnace process, the interference of data during the fluctuation period is avoided; at the same time, the value of recent data is highlighted by the time decay weight, which solves the defect of existing methods that "only focus on time and not on process".

[0037] 2. Better adaptability: The dynamic update mechanism can optimize the sampling strategy as the furnace age increases and the furnace condition changes, avoiding long-term bias caused by fixed sampling, and adapting to the data processing needs of the entire blast furnace life cycle;

[0038] 3. Lower implementation cost: Based on time-blocking rather than complex event detection, no additional event monitoring module needs to be deployed, and it can be integrated into existing data systems, making it easy to implement;

[0039] 4. Better modeling support: The high-quality output sample set can be directly used for training models such as linear regression and machine learning, reducing the impact of redundant data on model training efficiency and improving the model's prediction accuracy for the current furnace condition.

[0040] 5. Cost reduction and efficiency improvement with significant benefits: This technology is the foundation of blast furnace modeling and analysis. After establishing a model using this technology, it can be used to explore smelting patterns, improve blast furnace stability, and help reduce costs and increase efficiency. Detailed Implementation

[0041] The present invention will be further described below with reference to embodiments:

[0042] The ironmaking data sampling method integrating time decay and multi-parameter stability described in this invention includes the following steps: S1. Perform quality filtering on historical ironmaking data to remove outliers and low-quality data; For the input historical ironmaking data (including time column + multiple process parameter column), anomalies and low-quality data are removed through a dual mechanism to ensure the reliability of the basic data; the dual mechanism includes first removing significant outliers through IQR, and then filtering and compressing extreme values ​​through quantile range.

[0043] The IQR method for removing significant outliers is as follows: For each process parameter x, calculate the 25th quantile Q1 and the 75th quantile Q3, and obtain the interquartile range IQR = Q3 - Q1. Keep the data that satisfy Q1 - λ × IQR ≤ x ≤ Q3 + λ × IQR, and filter out significant outliers, where λ ranges from 1.5 to 4.5. The method for filtering quantile ranges is as follows: Configure a uniform quantile threshold (e.g., lower quantile q_l=0.005, upper quantile q_h=0.995) to retain data with parameter values ​​in the range [q_l, q_h], and further compress the impact of instantaneous fluctuations (such as sensor failure) and extreme abnormal furnace conditions; S2. Divide the filtered data into N blocks (N≥5) in ascending order of time, and retain the entire Nth block; Among them: (1) The Nth block (latest data block) is completely retained to ensure that the current furnace condition information is not lost; (2) The first N-1 blocks are historical data blocks, which are used for subsequent weighted sampling; S3. Calculate the furnace condition stability index for each block based on the average coefficient of variation (CV) of multiple process parameters to quantify the degree of fluctuation of the data for that block; The calculation process for the furnace condition stability index is as follows: S31. CV Calculation within a Single-Parameter Block: For the j-th process parameter within the block, calculate its mean μij and sample standard deviation σij to obtain the CV of that parameter: CV ij =σ ij / μ ij (μj>0, avoid dividing by 0); S32. Single-parameter global CV calculation: For the j-th process parameter, calculate its average CV across all blocks, which serves as the global fluctuation benchmark for that parameter.

[0044] S33. Calculation of the stability index within a single-parameter block: For the j-th parameter in the i-th block, define S... ij =max(0,1-CV ij / The larger the value, the smaller the fluctuation of the parameter in this block compared to the global value (the more stable it is). S34. Calculation of Multi-Parameter Stability Index: For the i-th block, the stability indices of each parameter are weighted and averaged according to preset weights to obtain the comprehensive stability index of the block: I i = , where w j ≥0 represents the weight of the j-th parameter, and The weights are set according to the importance of the parameters to the furnace condition.

[0045] S4. Combine the time decay coefficient and stability index to calculate the comprehensive weight of each block; The formula for calculating the overall weight is: W i =r Δt_i ×exp(-β·I i ), where Δt_i=Ni is the time interval, r∈(0,1) is the time decay coefficient, and β>0 is the stability influence coefficient.

[0046] The overall weight of each data point is composed of the product of "time decay weight" and "stability correction factor", achieving a two-dimensional weighting: Time decay weight W time,i : Reflects the timeliness of the data; the more recent the time, the higher the weight: W time,i =r Δt_i , where Δt_i=Ni (the time interval between the i-th block and the latest block, where Δt_N=0 for the latest block), and r∈(0,1) is the time decay coefficient (preferably 0.7~0.95); Stability correction factor W stab,i : Reflects furnace condition stability; the smaller the fluctuation, the closer the correction factor is to 1 (higher weight): W stab,i =exp(-β·I i), where β>0 is the stability influence coefficient (preferably 0.2~1.0), which controls the penalty for fluctuations on the weights; Overall weight W i =W time,i ×W stab,i .

[0047] S5. Randomly sample the first N-1 blocks without replacement according to the overall weight; (1) Calculation of basic sample size: For the i-th block, the sample size C i =max(N min ,⌊len(block i )×W i ⌋); where N min Typically, this is set to 5% to 10% of the data volume of each block, ensuring that at least 10 to 20 data entries are retained in each block to avoid information loss; (2) Result merging: The first N-1 blocks of sampled data are merged with the Nth block of full data, sorted in ascending order of time, to obtain the final sampled dataset.

[0048] S6. Periodically re-execute S1-S5 to dynamically update the sampling dataset.

[0049] The dynamic update cycle is adjusted according to the frequency of changes in blast furnace conditions. S1-S5 are re-executed at fixed intervals (such as daily or weekly) to update the sampling dataset, ensuring that the sampling strategy always adapts to the current furnace conditions and avoiding modeling bias caused by long-term use of fixed samples.

[0050] Example

[0051] Taking the company's 4350m³ blast furnace as an example, the feasibility of this method was verified. Specific parameters: blast furnace production data from September 8, 2022 to May 28, 2025, with a sampling frequency of hourly, and a total of 23,365 records after removing maintenance data; key process parameters: output (t / d), coke ratio (kg / tFe), blast volume (Nm³ / min), and total heat load (10MJ / h).

[0052] The data sampling for fusing time decay and multi-parameter stability of the above data includes the following steps:

[0053] S1. Import the above historical data of blast furnace ironmaking, perform quality filtering on the historical data, and remove outliers and low-quality data.

[0054] For the input historical data, including the time column and four key process parameter columns (including output, coke ratio, air volume, and total heat load), a dual mechanism is used to remove abnormal and low-quality data to ensure the reliability of the basic data:

[0055] The data format is a structured table; the parameter weights are configured as follows: output: 0.3, coke ratio: 0.25, air volume: 0.25, total heat load: 0.2; IQR elimination + quantile filtering are used for the four parameters, specifically as follows:

[0056] S11. IQR outlier removal: For each process parameter x, calculate the 25th percentile (Q1) and 75th percentile (Q3) to obtain the interquartile range IQR = Q3 - Q1. Retain the data that satisfy Q1 - λ × IQR ≤ x ≤ Q3 + λ × IQR, and filter out significant outliers. The value of λ ranges from 1.5 to 4.5.

[0057] S12. Quantile Range Filtering: Configure a uniform quantile threshold (e.g., lower quantile q_l=0.005, upper quantile q_h=0.995) to retain data with parameter values ​​in the range [q_l, q_h], further compressing the impact of instantaneous fluctuations (such as sensor failure) and extreme abnormal furnace conditions;

[0058] Therefore, 1162 data entries were deleted, resulting in 22204 valid data entries.

[0059] S2. Divide the filtered 22204 data entries into 10 blocks in ascending order of time, with each block containing 2220-2221 data entries. Keep all entries in the 10th block.

[0060] S3. Calculate the furnace condition stability index for each block based on the average coefficient of variation (CV) of multiple process parameters;

[0061] The calculation process for the stability index is as follows:

[0062] S31. Calculate the CV of the j-th parameter in the i-th block: CVij = σij / μij (μj > 0);

[0063] S32. Calculate the global CV of the j-th parameter:

[0064] S33. Calculate the stability index of the j-th parameter in the i-th block: Sij = max(0, 1 - CVij / );

[0065] S34. Obtain the multi-parameter stability index of the i-th block by weighting according to preset weights: Ii= , where wj≥0 and .

[0066] S4. Combine the time decay coefficient and stability index to calculate the comprehensive weight of each block;

[0067] The comprehensive weighting formula is: W i =r Δt_i ×exp(-β·I i), where Δt_i=Ni is the time interval, r∈(0,1) is the time decay coefficient, and β>0 is the stability influence coefficient.

[0068] With r set to 0.95 and β set to 0.9, the time decay weight, stability parameter, and overall weight for each block were calculated sequentially, yielding the following results:

[0069] Block number Time interval Δt <![CDATA[Time decay weight W time > <![CDATA[Stability Index I i > <![CDATA[Stability correction factor W stab > <![CDATA[Comprehensive weight W i > Original number of rows in the block Actual sampling size 1 9 0.6302 0.0681 0.9406 0.5928 2221 1316 2 8 0.6634 0.1577 0.8677 0.5756 2221 1278 3 7 0.6983 0 1 0.6983 2221 1550 4 6 0.7351 0.0226 0.9799 0.7203 2221 1599 5 5 0.7738 0.1443 0.8782 0.6796 2220 1508 6 4 0.8145 0.176 0.8535 0.6952 2220 1543 7 3 0.8574 0.1786 0.8515 0.7301 2220 1620 8 2 0.9025 0.0072 0.9935 0.8966 2220 1990 9 1 0.95 0 1 0.9500 2220 2109 10 0 1 0.0076 0.9932 0.9932 2220 2220

[0070] S5. Randomly sample the first 9 blocks without replacement according to the overall weight, with a sample size C. i =max(N min ,⌊len(block i )×W i ⌋), and merge all the sampling results to obtain 16,733 final samples;

[0071] S6. Set the update cycle to weekly. Every Monday, reread the data from the previous week and re-execute S1~S5 to dynamically update the sampled dataset for training the prediction model for the current week. Dynamic updates ensure that the sampling strategy always adapts to the current furnace conditions, avoiding modeling bias caused by using fixed samples for a long time.

Claims

1. A method for sampling ironmaking data that integrates time decay and multi-parameter stability, characterized in that: Includes the following steps: S1. Perform quality filtering on historical ironmaking data to remove outliers and low-quality data; S2. Divide the filtered data into N blocks in ascending order of time, where N≥5, and retain the entire Nth block. S3. Calculate the furnace condition stability index for each block based on the average coefficient of variation of multiple process parameters; S4. Combine the time decay coefficient and stability index to calculate the comprehensive weight of each block; S5. Randomly sample the first N-1 blocks without replacement according to the overall weight; S6. Periodically re-execute S1~S5 to dynamically update the sampling dataset.

2. The ironmaking data sampling method integrating time decay and multi-parameter stability according to claim 1, characterized in that: The quality filtering described in S1 employs a dual mechanism: first, significant outliers are removed through IQR, and then extreme values ​​are compressed through quantile range.

3. The ironmaking data sampling method integrating time decay and multi-parameter stability according to claim 2, characterized in that: The method for removing significant outliers using IQR is as follows: For each process parameter x, calculate the 25th quantile Q1 and the 75th quantile Q3, and obtain the interquartile range IQR = Q3 - Q1. Keep the data that satisfy Q1 - λ × IQR ≤ x ≤ Q3 + λ × IQR, and filter out significant outliers, where λ ranges from 1.5 to 4.

5. The method for filtering the quantile range is as follows: Configure a uniform quantile threshold (lower quantile q_l and upper quantile q_h), and retain data whose parameter values ​​are within the range of [q_l, q_h].

4. The ironmaking data sampling method integrating time decay and multi-parameter stability according to claim 1, characterized in that: The calculation process for the furnace condition stability index described in S3 is as follows: S31. Calculate the CV of the j-th parameter in the i-th block: CVij = σij / μij (μj > 0); S32. Calculate the global CV of the j-th parameter: S33. Calculate the stability index of the j-th parameter in the i-th block: Sij=max(0.1-CVij / ); S34. Obtain the multi-parameter stability index of the i-th block by weighting according to preset weights: Ii= , where wj≥0 and .

5. The ironmaking data sampling method integrating time decay and multi-parameter stability according to claim 1, characterized in that: The formula for calculating the overall weight described in S4 is as follows: W i =r Δt_i ×exp(-β·I i ), where Δt_i=Ni is the time interval, r∈(0,1) is the time decay coefficient, and β>0 is the stability influence coefficient.

6. The ironmaking data sampling method integrating time decay and multi-parameter stability according to claim 5, characterized in that: The value of r is preferably 0.7~0.95, and the value of β is preferably 0.2~1.

0.

7. The ironmaking data sampling method integrating time decay and multi-parameter stability according to claim 1, characterized in that: Sampling size C in S5 i =max(N min ,⌊len(block i )×W i ⌋), where N min Set it to 5% to 10% of the data volume per block.

8. The ironmaking data sampling method integrating time decay and multi-parameter stability according to claim 1, characterized in that: The dynamic update cycle in S6 is adjusted according to the frequency of changes in blast furnace conditions.

Images