CN111461171A - A data optimization method and system for building a prediction model of silicon content in blast furnace hot metal - Google Patents

Abstract

The invention discloses a data selection method and system for constructing a prediction model of silicon content in blast furnace hot metal. The unsupervised learning algorithm k-means++ is used to perform clustering, and input variable samples are clustered according to the degree of similarity. Abnormal data is eliminated in the 2019-2010 to strengthen the mapping relationship between input variable samples and silicon content data; at the same time, based on continuous time period indicators, high-confidence samples are screened to effectively reduce the interference of abnormal samples on the mapping relationship; based on the frequency histogram Determine the high-frequency range of silicon content, overcome the conservativeness of the traditional averaging method, and provide the most matching silicon content data for the input variable samples. After the evaluation model is verified, the data optimization method has better performance in model training than the traditional mean method. The invention solves the problem that the mapping relationship between the original input data and the silicon content is weak, and can effectively improve the training effect of the prediction model.

Description

A data optimization method for constructing a prediction model of silicon content in blast furnace hot metal and system

technical field

The application belongs to the field of prediction of silicon content in blast furnace molten iron, and particularly relates to a data optimization method and system for building a prediction model for silicon content in blast furnace molten iron.

Background technique

Prediction of silicon content in hot metal is one of the keys to optimal control of blast furnaces. At present, most researchers use data-driven ideas to build prediction models for silicon content in blast furnaces. These models have high requirements on the quality of training data sets. However, due to the multi-scale characteristics of blast furnace smelting and the harsh data collection environment, the collected raw data has serious abnormalities and missing problems, especially the raw silicon content data is unevenly distributed and has serious noise, resulting in the difference between the input variables and the silicon content. The mapping relationship is weak. The specific performance is as follows: the input variables are collected by sensors and recorded at hourly intervals, while the silicon content data is manually collected and tested due to process limitations. During the recording interval of some input variables, the silicon content has many values, large fluctuations and time lags. At this time, it is difficult to reasonably determine the silicon content value that best matches the input variable, which greatly hinders the acquisition of high-quality training sets and the construction of stable prediction models. In order to enhance the mapping relationship between input variables and silicon content, some scholars have proposed a method to calculate the arithmetic mean value within a certain time interval, such as Song Jinghua. Research on multi-scale characteristics of blast furnace smelting process and prediction method of silicon content [D]. Zhejiang University , 2016 and Chu Y, Gao C.Data-based multiscale modeling for blastfurnace system[J].AIChE Journal,2014,60(6):2197-2210. discloses a bag sample analysis method, that is, in the process of iron tapping Collect two silicon content values, and take their arithmetic mean; Liu Min, Research and prediction of blast furnace silicon content based on fuzzy model [D]. Inner Mongolia University of Science and Technology, 2012, the data were analyzed with 30min as the sampling interval for each input. Fusion, that is, calculating the arithmetic mean of the data within 30 minutes.

However, we found that the method of calculating the arithmetic mean is effective only for uniformly sampled time series data, but not for non-uniformly spaced data due to the small amount of time series data; when the input variable record interval memory When there are multiple silicon content values and the fluctuation is large, the mean value method is more conservative, and it is easy to be disturbed by noise and cause the silicon content to deviate from the correct range.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a data optimization method for prediction model of silicon content in blast furnace hot metal. Solve the problems of non-uniform recording time interval, large data fluctuation and weak mapping relationship between input variable samples and silicon content data in the original data of blast furnace molten iron, and establish a reasonable correlation value between input variables and matching reasonable silicon content, so as to improve Silicon content provides a high-quality training set for the predictive performance of data-driven predictive models.

The technical scheme provided by the present invention is as follows:

In one aspect, a data optimization method for constructing a prediction model for silicon content in blast furnace molten iron, comprising:

Obtain blast furnace molten iron production sample data, including production process input variable samples and silicon content samples;

Use the clustering model to cluster the input variable samples in the production process, and eliminate the abnormal input variable samples;

Based on the set continuous time length index, the representative time period of each cluster is extracted from the input variable samples of each cluster;

Draw the frequency histogram of the silicon content values contained in the representative time period of each cluster, and determine the high frequency interval of the silicon content value of each cluster;

Based on the high-frequency interval of the silicon content value of each cluster, for each input variable sample of each cluster, the best silicon content value is selected from all the corresponding silicon content samples.

By using the clustering method, this scheme divides the input variable samples into clusters according to the degree of similarity, and removes abnormal data from the complex blast furnace molten iron data; at the same time, based on the indicators of continuous time periods and the frequency histogram, the corresponding reasonableness of the input variables of each cluster is shown. The silicon content range, based on which the optimal silicon content value is associated with the input sample. It solves the problem of weak mapping relationship between the original input data and silicon content, which can effectively improve the effect of the prediction model.

Further, the specific process of selecting the optimal silicon content value for each cluster input variable sample is as follows:

The set Si is formed by the silicon content test time from the time of recording the corresponding input variable to the unit recording interval, and it is judged whether the silicon content value in Si is in the high-frequency range of the silicon content value of the cluster where the input variable is located, If it is, the silicon content value in Si whose recording time is arranged at the front is selected as the best silicon content value corresponding to the input variable; The silicon content value is used as the best silicon content value corresponding to the input variable, and at the same time, the selected silicon content value is no longer used as the best silicon content candidate value of other input variables.

Further, the high-frequency interval of silicon content includes two intervals, namely a first high-frequency interval and a second high-frequency interval, wherein the first high-frequency interval is the interval with the highest frequency in the frequency histogram of the silicon content value, The second high frequency interval is the interval with the second highest frequency in the frequency histogram of the silicon content value; and when the frequency histogram of the silicon content value is drawn, the silicon content value is uniformly discretized to obtain 10 discrete intervals;

When judging whether the silicon content value in Si is in the high frequency range of the silicon content value of the cluster where the input variable is located, the first high frequency range and the second high frequency range are judged in sequence.

Further, the specific process of extracting the representative time period of each cluster from the input variable samples of each cluster is as follows:

Step 3.1), sort all the samples in the cluster according to the order of recording time, and obtain the sample time series T of the cluster;

Step 3.2), according to the continuity of the recording time, divide the sequence T into different continuous time subsequences; T={T ₁ ,T ₂ ,...,T _l }, and the continuous time length of T _i is denoted as L(T _i ), 1≤i≤l, l represents the total number of continuous time subsequences;

L(T)为T中所有输入样本记录的时间长度之和，β为设定的持续时间长度，单位为小时，初始值设置为5；Step 3.3), calculate the proportion of continuous time subsequences with time length greater than β in the cluster

L(T) is the sum of the time lengths of all input sample records in T, β is the set duration, the unit is hours, and the initial value is set to 5;

Step 3.4), if ρ<0.6, reduce the set duration, let β=β-1, and return to step 3.3), otherwise, delete all continuous time subsequences whose duration is less than β from T; this , the recording time in T is taken as the representative time period of the cluster.

Further, in the process of using the clustering model to cluster the input variable samples in the production process and eliminating the abnormal input variable samples, the k-means++ clustering model is used for clustering, and the details are as follows:

Step 2.1), use the k-means++ algorithm to cluster the input variable samples in the input variable sample set U into k clusters. After the clustering is completed, if the total number of samples in a certain cluster accounts for less than 2% of the total number of samples, delete the all input variable samples of the cluster;

Step 2.2), repeat step 2.1), until the total number of samples in all clusters accounts for no less than 2% of the total number of samples.

Based on the k-means++ clustering algorithm, the samples are classified to distinguish different furnace condition information, and the mapping relationship between the input variable samples and the silicon content data is strengthened;

The specific content of the k-means++ algorithm can be found in the literature: Arthur D, Vassilvitskii S. K-Means++: The Advantages of Careful Seeding[C]//Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2007, New Orleans, Louisiana, USA , January 7-9, 2007. ACM, 2007.

The specific value of k is to keep the silhouette coefficient in the k-means++ algorithm between [-1, 1], and preferably as close to 1 as possible.

Further, the input variable samples include recording time c ₁ , oxygen enrichment rate c ₂ , carbon monoxide c ₃ , hydrogen c ₄ , carbon dioxide c ₅ , standard wind speed c ₆ , oxygen enrichment flow c ₇ , cold air flow c ₈ , blast air Kinetic energy c ₉ , bolly gas volume c ₁₀ , bosh gas index c ₁₁ , theoretical combustion temperature c ₁₂ , top pressure c ₁₃ , oxygen-enriched pressure c ₁₄ , cold air pressure c ₁₅ , total pressure difference c ₁₆ , hot air pressure c ₁₇ , actual wind speed c ₁₈ , hot air temperature c ₁₉ , top temperature northeast c ₂₀ , top temperature southwest c ₂₁ , top temperature northwest c ₂₂ , top temperature southeast c ₂₃ , resistance coefficient c ₂₄ , blast humidity c ₂₅ , set coal injection The amount c ₂₆ and the coal injection amount c ₂₇ in the previous hour, the silicon content sample includes the assay time d ₁ and the silicon content data d ₂ , wherein the recording period of the input variable sample is 1 hour.

On the other hand, a data optimization system for constructing a prediction model of silicon content in blast furnace molten iron, comprising:

The sample data acquisition unit is used to obtain sample data of blast furnace molten iron production, including samples of input variables in the production process and samples of silicon content;

The sample clustering and elimination unit is used to cluster the input variable samples of the production process by using the clustering model, and eliminate abnormal input variable samples;

The representative time period extraction unit, based on the set continuous time length index, extracts the representative time period of each cluster from the input variable samples of each cluster;

The high-frequency interval drawing and determining unit of silicon content is used to draw the frequency histogram of silicon content values contained in the representative time period of each cluster, and determine the high-frequency interval of silicon content values of each cluster;

The best silicon content selection unit, based on the high frequency interval of the silicon content value of each cluster, selects the best silicon content value from all the corresponding silicon content samples for each input variable sample of each cluster.

Further, the optimal silicon content selection unit includes:

The silicon content value building module within the unit sampling interval is used to form the set Si from the silicon content assay time from the recording time of the corresponding input variable to the unit recording interval time;

The judgment module is used to judge whether the silicon content value in Si is in the high frequency range of the silicon content value of the cluster where the input variable is located. The optimal silicon content value corresponding to the input variable; if not, select the silicon content value closest to the midpoint value of the high-frequency interval from Si as the optimal silicon content value corresponding to the input variable, and at the same time, the selected silicon content value is no longer a candidate for optimal silicon content for other input variables.

Further, the representative time period extraction unit includes:

The sample time series acquisition module is used to sort all the samples in the cluster according to the order of recording time, and obtain the sample time series T of the cluster;

_The continuous time subsequence division unit is used to _divide the sequence T _into different continuous time subsequences according to the continuity of the recording time _; is L(T _i ), 1≤i≤l, where l represents the total number of continuous time subsequences;

The proportion calculation module is used to calculate the proportion ρ of continuous time subsequences whose time length is greater than β in the cluster;

Among them, L(T) is the sum of the time lengths of all input sample records in T, β is the set duration length, the unit is hours, and the initial value is set to 5;

The representative time period selection module, according to the proportion value obtained by the proportion calculation module, makes a judgment. If ρ<0.6, reduce the set duration, let β=β-1, and re-call the proportion calculation module to calculate the proportion, Call the representative time period selection module again, otherwise, delete all continuous time subsequences whose duration is less than β from T; at this time, the recorded time in T is taken as the representative time period of the cluster.

Further, use k-means++ clustering model in the sample clustering and culling unit for clustering, by using the k-means++ algorithm, the input variable samples in the input variable sample set U are clustered into k clusters, after the clustering is completed. , if the total number of samples in a certain cluster accounts for less than 2% of the total number of samples, delete all input variable samples of the cluster; repeatedly call the k-means++ clustering model for clustering until the total number of samples in all clusters is in the total number of samples The proportion is not less than 2%.

beneficial effect

The invention provides a data selection method and system for constructing a prediction model of silicon content in blast furnace hot metal. The unsupervised learning algorithm k-means++ is used for clustering, and the input variable samples are clustered according to the degree of similarity. Abnormal data is eliminated in the 2019-2010 to strengthen the mapping relationship between input variable samples and silicon content data; at the same time, based on continuous time period indicators, high-confidence samples are screened to effectively reduce the interference of abnormal samples on the mapping relationship; based on the frequency histogram Determine the high-frequency range of silicon content, overcome the conservativeness of the traditional averaging method, and provide the most matching silicon content data for the input variable samples. After verification of the evaluation model, the data optimization method can improve the accuracy of silicon content prediction compared with the traditional mean method. The invention solves the problem that the mapping relationship between the original input data and the silicon content is weak, and can effectively improve the effect of the prediction model.

Description of drawings

1 is a schematic flowchart of Embodiment 1 of the present invention;

Figure 2 is a schematic diagram of the k-means++ clustering result after removing abnormal clusters;

FIG. 3 is a frequency histogram of silicon content in each cluster, wherein (a) is ClusterA, (b) is ClusterB, (c) is ClusterC, (d) is ClusterD, and (e) is ClusterE.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

Example 1

In this example, the actual blast furnace generation data collected from 0:00 on October 1, 2017 to 23:00 on October 31st in a ^2650m3 blast furnace in a domestic steel plant is used as an example for illustration.

As shown in Figure 1, a data optimization method for prediction model of silicon content in blast furnace molten iron, the process is as follows:

Step 1: Obtain blast furnace molten iron production sample data, including production process input variable samples and silicon content samples.

The input variable sample set U includes 27 input variables, as shown in Table 1, including: recording time, oxygen enrichment rate, carbon monoxide, hydrogen, carbon dioxide, standard wind speed, oxygen-enriched flow, cold air flow, blast kinetic energy, belly gas volume, bolly gas index, theoretical combustion temperature, top pressure, oxygen-rich pressure, cold air pressure, total pressure difference, hot air pressure, actual wind speed, hot air temperature, top temperature northeast, top temperature southwest, top temperature northwest, top temperature southeast, The resistance coefficient, blast humidity, set coal injection volume and coal injection volume in the last hour are collected by the sensor, and the recording interval is 1 hour; the silicon content sample set V, as shown in Table 2, includes the test time and the test value of silicon content, The test was collected manually, and the recording interval was irregular. The sample set contains a total of 744 input variable samples and 1478 silicon content data.

Table 1 Sample set of input variables in production process

Table 2 Silicon content sample set

Step 2: Use the clustering model to cluster the input variable samples in the production process, and eliminate abnormal input variable samples.

2.1) Cluster the samples into k clusters, perform multiple tests, calculate the silhouette coefficient of the clustering results, and finally select k=5 with a larger silhouette coefficient for clustering. After the clustering is completed, if the total number of samples in a certain cluster is within the total If the proportion of the number is less than 2%, delete all input variable samples of the cluster

2.2) Repeat step 2.1) until the total number of samples in all clusters accounts for no less than 2% of the total number of samples, and the final sample clustering result is obtained, as shown in Figure 2.

Step 3: Based on the set continuous time period length index, extract each cluster representative time period from each cluster input variable sample.

3.1) Mark the clustering results as ClusterA, ClusterB, ClusterC, ClusterD and ClusterE respectively, and the total number of samples in each cluster is 26, 129, 71, 393 and 116 respectively. Sort all the samples in the cluster according to the order of recording time to obtain The sample time series T={t ₁ , t ₂ ,..., _th } of the cluster;

3.2) According to the continuity of the recording time, the sequence T is divided into different continuous time subsequences; T={T ₁ , T ₂ ,...,T _l }, where T _i (1≤i≤1) continuous time The length is recorded as L(T _i );

其中ClusterA、ClusterB、ClusterD和ClusterE，连续时间子序列总和分别为17、109、366和103，在对应簇中占比65％、84％、93％和89％。3.3) Take β=5, and calculate the proportion of continuous time subsequences whose time length is greater than β in the cluster

Among them, ClusterA, ClusterB, ClusterD and ClusterE, the sums of continuous time subsequences are 17, 109, 366 and 103 respectively, accounting for 65%, 84%, 93% and 89% of the corresponding clusters.

3.4) When β=5, the proportion of continuous time subsequences of ClusterC in the cluster is ρ<0.6, so the set duration is reduced until the duration is β=3, ρ=0.63. Time subsequences with duration less than β in each cluster T are eliminated. At this time, the time recorded in T is taken as the representative time period of the cluster. The above statistics are recorded in Table 3.

Table 3 Statistics of each cluster

Step 4: Draw a frequency histogram of the silicon content values included in the representative time period of each cluster, and determine the "high frequency interval" of the silicon content value of each cluster.

则将y_m2加入数据集D_T，y_m1和y_m2分别表示记硅含量的记录时间和对应的硅含量值。4.1) From the silicon content sample set V, screen out the silicon content data set D _T whose recording time is within the representative time period T of each cluster, that is,

Then y _m2 is added to the data set D _T , and y _m1 and y _m2 respectively represent the recording time of recording the silicon content and the corresponding silicon content value.

4.2) Draw the frequency distribution histogram of each cluster _DT . The abscissa is the silicon content value, and the ordinate is the frequency, as shown in Figure 3. Considering that the actual silicon content is an analog quantity, the silicon content value is uniformly discretized to obtain 10 discrete intervals, and the “first high-frequency interval” and the “second high-frequency interval” are counted. The "first high frequency interval" and "second high frequency interval" of each cluster are ([0.301, 0.342], [0.342, 0.383]), ([0.516, 0.58], [0.452, 0.516]), ([ 0.458,0.534],[0.382,0.458]),([0.536,0.605],[0.467,0.536]),([0.49,0.528],[0.414,0.452]);

Step 5: Based on the high frequency interval of the silicon content of each cluster, select the best silicon content value from all the corresponding silicon content samples for each input variable sample of each cluster.

5.1) Table 4 shows the proportion of the number of different silicon content values in the total number of samples in each cluster in the recording interval of each cluster input variable. According to the order of the recording time of the samples, the corresponding silicon content value is selected for it.

Table 4 The proportion of silicon content in the total number of samples in the cluster in the recording interval of different input variables in each cluster

After the selection is completed, a new data sample is generated as a data set for subsequent prediction work. According to statistics, among the 1478 silicon content values, a total of 731 values are in the high frequency range, which matches the silicon content of 735 samples after anomaly removal. The matching results are listed in Table 5.

Table 5 Statistics of the number of samples that can be matched with silicon content in the high-frequency interval of each cluster

The preferred data set selected in this example enhances the mapping relationship between the input variable samples and the silicon content data through the clustering algorithm; reduces noise interference based on the set continuous time period indicators, and has higher reliability than the traditional mean method. Overcome the conservatism of traditional averaging methods.

Embodiment 2

Based on the above method, the embodiment of the present invention also provides a data optimization system for constructing a prediction model of silicon content in blast furnace hot metal, including:

The sample data acquisition unit is used to obtain sample data of blast furnace molten iron production, including samples of input variables in the production process and samples of silicon content;

The sample clustering and elimination unit is used to cluster the input variable samples of the production process by using the clustering model, and eliminate abnormal input variable samples;

The representative time period extraction unit, based on the set continuous time length index, extracts the representative time period of each cluster from the input variable samples of each cluster;

The high-frequency interval drawing and determining unit of silicon content is used to draw the frequency histogram of silicon content values contained in the representative time period of each cluster, and determine the high-frequency interval of silicon content values of each cluster;

The best silicon content selection unit, based on the high frequency interval of the silicon content value of each cluster, selects the best silicon content value from all the corresponding silicon content samples for each input variable sample of each cluster.

Among them, the optimal silicon content selection unit includes:

The silicon content value building module within the unit sampling interval is used to form the set Si from the silicon content assay time from the recording time of the corresponding input variable to the unit recording interval time;

The judgment module is used to judge whether the silicon content value in Si is in the high frequency range of the silicon content value of the cluster where the input variable is located. The optimal silicon content value corresponding to the input variable; if not, select the silicon content value closest to the midpoint value of the high-frequency interval from Si as the optimal silicon content value corresponding to the input variable, and at the same time, the selected silicon content value is no longer a candidate for optimal silicon content for other input variables.

Wherein, the representative time period extraction unit includes:

The sample time series acquisition module is used to sort all the samples in the cluster according to the order of recording time, and obtain the sample time series T of the cluster;

_The continuous time subsequence division unit is used to _divide the sequence T _into different continuous time subsequences according to the continuity of the recording time _; is L(T _i ), 1≤i≤l, where l represents the total number of continuous time subsequences;

The proportion calculation module is used to calculate the proportion ρ of continuous time subsequences whose time length is greater than β in the cluster;

Among them, L(T) is the sum of the time lengths of all input sample records in T, β is the set duration length, the unit is hours, and the initial value is set to 5;

The representative time period selection module, according to the proportion value obtained by the proportion calculation module, makes a judgment. If ρ<0.6, reduce the set duration, let β=β-1, and re-call the proportion calculation module to calculate the proportion, Call the representative time period selection module again, otherwise, delete all continuous time subsequences whose duration is less than β from T; at this time, the recorded time in T is taken as the representative time period of the cluster.

Wherein, the k-means++ clustering model is used in the sample clustering and elimination unit for clustering, and the input variable samples in the input variable sample set U are clustered into k clusters by using the k-means++ algorithm. After the clustering is completed, If the total number of samples in a certain cluster accounts for less than 2% of the total number of samples, delete all input variable samples of the cluster; repeatedly call the k-means++ clustering model for clustering, until the total number of samples in all clusters is less than the total number of samples in the total number of samples The proportion is not less than 2%.

It should be understood that the functional unit modules in various embodiments of the present invention may be centralized in one processing unit, or each unit module may exist physically alone, or two or more unit modules may be integrated into one unit module, It can be implemented in the form of hardware or software.

The above is only a description of the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments, and all technical solutions that belong to the idea of the present invention belong to the protection scope of the present invention. Those skilled in the art to which the present invention pertains can make various modifications or supplements to the described specific embodiments without departing from the spirit and principle of the present invention, or substitute in similar ways, which are regarded as the protection scope of the present invention.

Claims (10)

1. A data optimization method for constructing a prediction model of the silicon content of blast furnace molten iron is characterized by comprising the following steps:

obtaining blast furnace molten iron production sample data, including a production process input variable sample and a silicon content sample;

clustering the input variable samples in the production process by using a clustering model, and removing abnormal input variable samples;

extracting each cluster representative time period from each cluster input variable sample based on the set continuous time length index;

drawing a frequency histogram of the silicon content values contained in each cluster representative time period, and determining a high-frequency interval of the silicon content values of each cluster;

and selecting the optimal silicon content value from all corresponding silicon content samples for each input variable sample of each cluster based on the high-frequency interval of the silicon content values of each cluster.

2. The method of claim 1, wherein the selection of the optimal silicon content value for each cluster of input variable samples is performed by:

forming a set Si by the silicon content values from the time when the silicon content test time is in the recording time of the corresponding input variable to the time of unit recording interval, judging whether the silicon content value in the Si is in the high-frequency interval of the silicon content value of the clustering cluster in which the input variable is positioned, if so, arranging the recording time of the silicon content value in the Si at the forefront, and selecting the recording time as the optimal silicon content value corresponding to the input variable; if not, the silicon content value closest to the point value in the high-frequency region is selected from Si as the optimal silicon content value corresponding to the input variable, and the selected silicon content value is no longer used as the optimal silicon content candidate value for the other input variable.

3. The method of claim 2, wherein the silicon content bin comprises two bins, a first bin and a second bin, wherein the first bin is the next highest bin in the frequency histogram of silicon content values and the second bin is the next highest bin in the frequency histogram of silicon content values; when the frequency histogram of the silicon content value is drawn, carrying out uniform discretization treatment on the silicon content value to obtain 10 discrete intervals;

and when judging whether the silicon content value in the Si is in the high-frequency interval of the silicon content value of the clustering cluster where the input variable is positioned, sequentially judging the first high-frequency interval and the second high-frequency interval.

4. The method of claim 1, wherein the specific process of extracting the representative time period of each cluster from the input variable samples of each cluster is as follows:

step 3.1), sequencing all samples in the cluster according to the recording time sequence to obtain a sample time sequence T of the cluster;

step 3.2), dividing the sequence T into different continuous time subsequences according to the continuity of the recording time; t ═ T₁,T₂,…,T_l}，T_iIs recorded as L (T)_i) I is 1. ltoreq. l, l representing the total number of consecutive time subsequences;

step 3.3), calculating the occupation ratio of continuous time subsequences with the time length longer than β in the cluster

L (T) is the sum of the time lengths of all input sample records in T, β is the set duration length in hours, the initial value is set to 5;

step 3.4), if rho <0.6, reducing the set duration, making β equal to β -1, and returning to step 3.3), otherwise, deleting all continuous-time subsequences with duration less than β from T, and at this time, taking the recording time in T as the representative time period of the cluster.

5. The method according to claim 1, wherein the clustering for the process input variable samples and the abnormal input variable samples is performed by using a k-means + + clustering model, and the method comprises the following steps:

step 2.1), clustering the input variable samples in the input variable sample set U into k clusters by using a k-means + + algorithm, and deleting all input variable samples of a cluster if the total number of the samples of the cluster is less than 2% of the total number of the samples after clustering is finished;

step 2.2), repeating step 2.1) until the total number of samples in all clusters is not less than 2% of the total number of samples.

6. The method of any of claims 1-5, wherein the input variable sample comprises a recording time c₁Oxygen enrichment rate c₂Carbon monoxide c₃Hydrogen gas c₄Carbon dioxide c₅Standard wind speed c₆Oxygen-enriched flow rate c₇Flow rate of cold air c₈And blast kinetic energy c₉Gas flow of furnace bosh c₁₀Gas index of furnace bosh c₁₁Theoretical combustion temperature c₁₂C, top pressure₁₃Oxygen-enriched pressure c₁₄Pressure of cold air c₁₅Total pressure difference c₁₆Pressure of hot air c₁₇Actual wind speed c₁₈Temperature c of hot air₁₉Top temperature northeast c₂₀Southwest c top temperature₂₁Top temperature northwest c₂₂Southeast C, top temperature₂₃Coefficient of resistance c₂₄Blast air humidity c₂₅Setting the coal injection quantity c₂₆And the amount of coal injected in last hour c₂₇The silicon content sample includes an assay time d₁And silicon content data d₂Wherein the recording period of the input variable sample is 1 hour.

7.A data optimization system for constructing a prediction model of the silicon content of blast furnace molten iron is characterized by comprising the following steps:

the sample data acquisition unit is used for acquiring sample data of the molten iron production of the blast furnace, and comprises a production process input variable sample and a silicon content sample;

the sample clustering and rejecting unit is used for clustering the input variable samples in the production process by using a clustering model and rejecting abnormal input variable samples;

a representative time period extraction unit which extracts each cluster representative time period from each cluster input variable sample based on the set continuous time length index;

the silicon content high-frequency interval drawing and determining unit is used for drawing a frequency histogram of the silicon content values contained in each cluster of representative time periods and determining the high-frequency interval of the silicon content values of each cluster;

and the optimal silicon content selecting unit selects the optimal silicon content value from all the corresponding silicon content samples for each input variable sample of each cluster based on the high-frequency interval of the silicon content value of each cluster.

8. The system of claim 7, wherein the optimum silicon content selecting unit comprises:

the silicon content value construction module in unit sampling interval time is used for constructing a set Si from the time when the silicon content assay time is in the corresponding input variable recording time to the time when the silicon content value in the unit recording interval time;

the judgment module is used for judging whether the silicon content value in the Si is in the high-frequency interval of the silicon content value of the clustering cluster where the input variable is located, if so, the recording time of the silicon content value in the Si is arranged at the forefront silicon content value, and the best silicon content value corresponding to the input variable is selected; if not, the silicon content value closest to the point value in the high-frequency region is selected from Si as the optimal silicon content value corresponding to the input variable, and the selected silicon content value is no longer used as the optimal silicon content candidate value for the other input variable.

9. The system according to claim 7, wherein the representative period extracting unit includes:

the sample time sequence acquisition module is used for sequencing all samples in the cluster according to the recording time sequence to obtain a sample time sequence T of the cluster;

a continuous time subsequence dividing unit for dividing the sequence T into different continuous time subsequences according to the continuity of the recording time; t ═ T₁,T₂,…,T_l}，T_iIs recorded as L (T)_i) I is 1. ltoreq. l, l representing the total number of consecutive time subsequences;

the occupation ratio calculation module is used for calculating the occupation ratio rho of the continuous time subsequences with the time length larger than β in the cluster;

wherein L (T) is the sum of the time lengths of all input sample records in T, β is a set duration length in hours, and the initial value is set to 5;

and the representative time period selection module is used for judging according to the ratio value obtained by the ratio calculation module, if rho is less than 0.6, reducing the set duration length, making β equal to β -1, calling the ratio calculation module again to calculate the ratio, calling the representative time period selection module again, and otherwise, deleting all continuous time subsequences with the duration length less than β from T, wherein the recording time in T is used as the representative time period of the cluster.

10. The system according to claim 7, wherein the sample clustering and removing unit performs clustering using a k-means + + clustering model, clusters the input variable samples in the input variable sample set U into k clusters by using a k-means + + algorithm, and deletes all input variable samples in a cluster if the total number of samples in the cluster is less than 2% of the total number of samples after the clustering is completed; and repeatedly calling the k-means + + clustering model for clustering until the total number of samples in all clusters is not less than 2% of the total number of samples.

Images