CN113591362A

CN113591362A - Clinker proportion optimization and regulation method based on big data intelligent control algorithm

Info

Publication number: CN113591362A
Application number: CN202110451718.2A
Authority: CN
Inventors: 马天雨; 白政文; 盛语嫣; 杨茗茹; 郑之伟; 刘金平
Original assignee: Hunan Normal University
Current assignee: Hunan Normal University
Priority date: 2021-04-26
Filing date: 2021-04-26
Publication date: 2021-11-02

Abstract

The 28-day strength, namely the cement 'label', represents the final production quality of cement, production adjustment cannot be realized according to the tested 28-day strength in the production process, but the 3-day strength of the cement has a strong linear relation with the 28-day strength, and the aim of stabilizing the 28-day strength can be realized by controlling the 3-day strength. The method comprises the steps of firstly finding out key factors influencing the 3-day strength by adopting a rough set resolution matrix algorithm, forming a rough decision table, then obtaining a clean decision table by adopting preprocessing methods such as filtering and exception removal, then training a deep neural network to establish a prediction model of the 3-day strength, considering that empirical data lacks current working condition information, correcting a prediction result by adopting the latest 3-day strength, and adjusting the proficiency ratio by adopting a fuzzy PID control algorithm based on the predicted 3-day strength. The verification result shows that the method can stably control the 3-day strength of the cement and completely meet the actual production requirement.

Description

Clinker proportion optimization and regulation method based on big data intelligent control algorithm

Technical Field

The invention belongs to the technical field of cement proportioning optimization based on big data analysis and an intelligent control algorithm, and particularly relates to a decision table splicing model, a rough set attribute reduction model, a cement 3-day strength prediction model and a prediction control algorithm.

Background

The cement formula mainly comprises four types of clinker, limestone, gypsum and mineral powder, the quality of the cement is determined by different proportions of the clinker, the limestone, the gypsum and the mineral powder, and the key quality parameters of the cement are controlled by adjusting the percentages of the materials: loss on ignition, SO3 content, 3-day strength of cement. The ratio of gypsum is adjusted according to the content of SO3 SO as to adjust the setting time, the ratio of limestone is controlled according to the loss on ignition, and the ratio of clinker and mineral powder are adjusted according to the 3-day strength of cement. The method mainly considers the adjustment of clinker and mineral powder, the test result of the 3-day strength of the cement can be obtained after 3 days, and the method belongs to a large hysteresis system, so that the clinker proportion cannot be directly adjusted according to the test result in a feedback mode. However, the clinker quality, the ratio, the cement chemical components, the cement ratio and the fineness are measured, and a prediction model can be established for prediction adjustment. At present, the ingredient regulation process in China still mainly adopts manual regulation, but the influence factors of 3-day strength of cement are complex, the manual regulation fluctuation is large, and in order to stabilize the cement quality and reduce the cost, the method optimally regulates the clinker proportion and the mineral powder based on a model prediction framework.

Disclosure of Invention

In order to solve the problem of large strength fluctuation of cement in 3 days, the patent provides a big data analysis method. Establishing a 3-Day strength prediction model of cement, predicting the 3-Day strength of the cement in real time according to test data, adjusting the proportion of clinker and mineral powder through a fuzzy control algorithm based on the prediction result, considering the working condition influence of the prediction result, and feeding back the correction model prediction result based on the actual test result of 3-Day strength of the cement of Day-4 (Day-0 represents today, Day-4 is 4 days ago).

In order to achieve the above object, the scheme shown in fig. 1 is proposed, and the present invention mainly includes the following contents:

1. the process analysis finds out the relevant influence factors of the 3-day strength of the cement, the influence factors and the 3-day strength (label data)

Splicing into an original splicing table;

2. the method comprises the following steps of carrying out data preprocessing on an original splicing table, providing various combined data processing methods according to the actual situation of field data, and removing interference data:

a. for artificially recorded abnormal data, the abnormal data are cleaned before analysis, and the specific abnormal data are as follows:

(1) the time of the submission and the time of the submission are not matched,

(2) the proportion of the formula is mutated,

(3) the data is filled in with an error,

(4) the adjustment is too frequent and the adjustment is too frequent,

(5) the proportion of the formula is not 100%;

b. analyzing the correlation among different attribute data by utilizing a PCA algorithm so as to clear the redundancy attribute;

3. and (3) performing attribute reduction and importance ranking on the data after the interference is removed by adopting a rough set resolution matrix algorithm to obtain key influence factors of 3-day intensity, and combining the key factors and the 3-day intensity into a decision table, wherein the decision table has the following basic contents:

a. discretizing the preprocessed splicing table, removing redundant data and contradictory data to obtain a discrete splicing table,

b. a relative attribute reduction algorithm and a resolution matrix algorithm are adopted to find out key attributes influencing the 3-day intensity,

c. sorting the importance of the key attributes by adopting a resolution matrix importance sorting method, and mining the relationship between the important attributes and the 3-day strength rule by adopting upper and lower approximate sets to obtain an expert regulation rule table;

4. the method for establishing the 3-day cement strength prediction model based on the decision table training deep neural network comprises the following main contents:

a. the formula proportion, tricalcium silicate, tricalcium aluminate, specific surface area, fineness of 3-32um, 1-day strength, CaO, ignition loss and vertical liter weight are taken as attributes, the 3-day strength of the cement is taken as a label,

b. randomly extracting data of the decision table, dividing the data into 10 equal parts, circularly using each equal part as a verification set to train the network to obtain a statistical result of the prediction accuracy of the model,

c. extracting three training set results with the highest prediction precision, and outputting the three training set results as model prediction results after averaging;

5. considering the influence of working condition difference, based on the latest actually measured 3-day intensity (before 3 days), adopting attribute difference data to establish a multivariate linear regression model, predicting the difference value between the current 3-day intensity and the actual 3-day intensity, correcting the depth network prediction result by using the predicted difference value plus the actual 3-day intensity (before 3 days), eliminating the working condition influence, and obtaining the predicted current 3-day intensity, wherein the method mainly comprises the following contents:

a. finding out the actual 3-day intensity test value of the last 5 days, and removing outliers according to a 3sigma principle;

b. finding out a decision attribute value of a date corresponding to the actual test intensity for 3 days, and finding out a decision attribute value of the current date;

c. subtracting the decision attribute value of the actual testing date from the decision attribute value of the current date to obtain an attribute difference value;

d. training a multiple linear regression model by adopting the attribute difference value, and predicting the 3-day intensity difference value;

e. assay 3 day intensity + prediction difference = correct 3 day intensity;

f. and correcting the depth network prediction result by using the weighted corrected 3-day intensity so as to overcome the influence of working condition fluctuation.

And adjusting the clinker ratio based on the prediction result by adopting a fuzzy control algorithm, and stabilizing the strength of the cement for 3 days, wherein the method comprises the following steps:

a. performing domain segmentation on key attributes influencing the 3-day strength of the cement to obtain fuzzy variables;

b. giving a rule table according to the difference between the predicted value of the 3-day strength and the target value and the change direction of the difference, verifying the association rule between the key attribute given by the field expert and the 3-day strength of the cement by adopting an upper and lower approximation set method, and storing the expert rule supported by the data into an adjustment rule according to a certain format;

c. the fuzzy control algorithm gives out the optimal adjustment quantity of the clinker proportion according to the rule table, if the recommended adjustment suggestion conforms to the rule, the adjustment is accepted, otherwise, new data are accumulated to update the rule table.

Economic benefits are as follows: the technical scheme who adopts through this patent can reach following beneficial effect:

a. the clinker usage can be reduced by 0.5% compared with manual regulation (feasibility), and the cost per month of 1 mill is reduced by 0.5 (clinker reduction ratio) 2.3 ton/h (flow corresponding to 1% clinker ratio, unit ton) 24 hours 30 days clinker unit price (300 yuan/ton);

b. the formula proportion is changed from a manual recommendation mode to system intelligent recommendation, so that the manual workload is reduced, and the labor intensity is reduced;

c. the material cost is saved, the product percent of pass is increased by 10%, the yield is increased by 8%, the formula is adjusted in a standardized manner, the fluctuation range of the product quality is obviously reduced, and the human error is greatly reduced.

Drawings

FIG. 1 is a diagram of the compounding process of the present invention.

FIG. 2 is a block diagram of an overall recipe adjustment scheme for a cement mill process of the present invention.

FIG. 3 is a diagram showing the prediction results of the prediction model of the present invention.

FIG. 4 is a system implementation interface of the present invention.

The following detailed description of the invention refers to the accompanying drawings

FIG. 1 is a process diagram of the cement mill batching of the present invention, the cement batching mainly comprises four parts of materials according to different proportions: clinker, mineral powder, gypsum and limestone, wherein different materials affect different quality parameters of cement:

a. the quality of the clinker mainly affects the short-term and long-term strength of the cement, the representative data of the clinker quality comprises KH, AL2O3, C3A, C3S, CaO-F and cubic liter weight, the price of the clinker is higher, the clinker is a main cost source for cement production, the cement production cost can be greatly reduced by reducing the proportion of the clinker as much as possible while ensuring that the strength (cement label) of the cement in 3 days and 28 days reaches the internal control index of an enterprise, the proportion of 425 cement clinker is generally 65%, and the proportion of 525 cement clinker is generally 85%;

b. the main aim of adding the cement into the gypsum is to delay the setting time and ensure the shortest setting time of the cement in summer, and the adding proportion of the gypsum is generally 6 percent;

c. limestone mainly affects the loss on ignition of cement, wherein the proportion of 425 cement limestone is generally 10 percent, and the proportion of 525 cement limestone is generally 3 percent;

d. the main purposes of adding the mineral powder are to reduce the cost and the addition amount of clinker, wherein the proportion of 425 cement mineral powder is generally 20 percent, and the proportion of 525 cement mineral powder is generally 5 percent;

e. the mixture in the formula is ground into final cement by a roller press and a ball mill, and the specific surface area, the grinding fineness of 3-32um and less than 3um of the cement can also influence the quality of the cement.

Fig. 2 is a framework diagram of the scheme of the invention, which mainly comprises the following contents:

1. splicing the big data decision table to obtain a big data decision table,

a. historical data of 3-day strength of cement, clinker quality, clinker ratio and mineral powder ratio are collected. (including KH, AL2O3, C3A, C3S, CaO-F and a cubic liter weight. the fineness data of the cement include 3-32um content, specific surface area, <3um content and >45um content, the proportion of ingredients: clinker ratio, limestone ratio, gypsum ratio and mineral powder ratio. the quality data of the cement: SO3, LOSS and CaO).

And considering the time delay relation between each attribute data sampling point and the cement quality test data, and performing time alignment matching. Taking the clinker quality as an example: the clinker quality data KH, AL2O3, C3A, C3S were delayed by approximately 4 hours into the cement mill, and therefore, the clinker quality assay results for the 9 month No. 1 20 to 24 point blend should be considered as the first set of data for 9 month No. 2, but not as the last set of data for 9 month No. 1.

And splicing and aligning according to time by taking the 3-day strength as a decision label and the quality, the ratio, the cement fineness and the specific surface area of the clinker as attributes to obtain a decision data table (shown in table 1) of the relation of 'formula ratio + clinker quality + mill fineness- > 3-day strength'.

TABLE 1 decision table

2. Carrying out data preprocessing, denoising and outlier removing;

a. replacing non-artificially-caused abnormal constant data with the most possible value for the spliced data, and carrying out fitting or regression analysis on all data to conjecture the most possible value;

b. judging the correlation among different attribute data by using chi-square test so as to clear redundant data;

c. removing formula mutation values and deficiency values, and removing row data of which each formula is empty;

d. removing bad point data, and screening by judging whether the formula ratio accumulation is 100%;

e. outliers in the attributes and decisions that exceed or fall below normal are removed.

3. Excavating key attributes influencing the 3-day strength of the cement by using a rough set resolution matrix and a relative reduction algorithm;

1) performing discrete normalization processing on the attribute values;

2) attribute reduction, wherein conditional attributes of a decision table shown in table 1 are artificially extracted, for a decision table without artificial participation, the conditional attributes are more, and conditional core attributes for decision are required to be found out through an attribute reduction algorithm, and on the premise of determining the decision table to be a consistent decision table, the attribute reduction algorithm mainly comprises a resolution matrix and a relative reduction method, and the invention adopts a resolution matrix heuristic attribute reduction algorithm, and the specific steps are as follows:

a. forming a consistent decision table, deleting two records when the condition attributes are the same and the decision attributes are different, and deleting one of the two records if the condition attributes are the same and the decision attributes are also the same;

b. calculating a distinguishing matrix according to the discrete decision table;

c. the importance of each attribute is obtained according to the formula (1), the number of the attributes contained in each item of the resolution matrix is represented, and the formula shows that the attribute importance is high when the frequency of the attributes appearing in the resolution matrix is high, the attribute importance is high when the items of the attributes in the resolution matrix are short, the core attribute is found by integrating the importance ordering and the single-item attributes in the resolution matrix;

（1）

d. finding out all attribute combinations which do not contain the core attribute, and expressing all attribute combinations which do not contain the core attribute into a disjunctive normal form;

e. reducing the attribute with the minimum importance;

f. if the reduction is successful (redundant sample/original sample < threshold), deleting the redundant sample and inconsistent sample caused by attribute reduction, and turning to the step e, if the reduction is unsuccessful (redundant sample/original sample > threshold), recovering the reduced attribute, and ending the reduction;

g. and mining association rules of key attributes and 3-day strength by adopting a vertical approximation set method, verifying the expert experience collected on site, and forming a rule table by the expert experience supported by data.

Training a DNN network to predict the 3-day strength of the cement, wherein the training of the DNN prediction model mainly comprises the following steps:

a. dividing the decision table into 10 equal parts according to the number of records;

b. selecting the number of layers, the number of nodes of each layer, the learning rate and the number of batches trained at each time according to experience;

c. taking all data as a training set and simultaneously as a test set, and adjusting DNN model parameters according to MSE minimum: continuously training the model to adjust parameters of each layer, evaluating the accuracy once every 10 times of training, and stopping when the accuracy begins to decline (overfitting occurs);

d. setting a DNN model according to the optimized parameters, randomly selecting 8 parts of 10 parts as a training set, and selecting the other 2 parts as a test set for cross validation;

e. selecting three models with the best prediction precision, storing parameters of the three models, sending real-time data into the three models for simultaneous prediction during online prediction, and outputting the average value of the three results as the final result;

f. and (4) updating the decision table in a rolling manner, continuously adding new records, removing old records, and regularly training a DNN neural network on line to realize the gradual change of the self-adaptive working condition.

Establishing an attribute difference value multivariable regression model, correcting a depth network prediction result, and overcoming the defects that the experience error of a DNN statistical model is minimum and the real-time working condition is not considered, wherein the multivariable regression model establishing process mainly comprises the following key steps:

a. acquiring a cement strength test value of the actual 3 days of the last 4-8 days, and removing outliers according to a 3sigma principle;

c. subtracting the decision attribute value of the actual 3-day strength test before 4-8 days from the decision attribute value of the cement strength at the current date to obtain an attribute difference value, as shown in Table 2;

TABLE 2 attribute difference table corresponding to the current day attribute and the actual assay intensity

d. For historical data, all data on the same day are known, so that a multivariate linear regression model with attribute difference as an x variable and three-day intensity difference as a y variable can be established as shown in formula (2), and 3-day intensity prediction difference is calculated according to the formula (2)

。

-----------（2）

e. The parameters of a multivariable linear regression model can be optimized by adopting optimization algorithms such as particle swarm optimization, heredity and the like, but different from a pure mathematical model, the industrial model cannot simply target the MSE minimum, the model parameters need to meet process constraints, in order to obtain the process constraint range of each coefficient, the univariate step response shown in the table 3 is made on a historical data decision table, the univariate step response refers to finding out records with the same other variables, the quantitative coefficient of the attribute and the 3-day intensity is obtained by analyzing the proportional relation of the attribute and the 3-day intensity, and the interval range with the coefficient as the central point is the process constraint of the parameter.

TABLE 3 univariate step corresponding table

f. And (3) optimizing the optimal parameters of the model by using the step coefficients as constraints and adopting a particle swarm optimization algorithm to obtain an optimal parameter combination:

[cao=0.46，c3s=0.18，c3a=0.17，f0 =0.54，f1=0.04，lisheng=0.01，loss=-1.48，

=-0.662

intercept=16.42]；

g. assay 3 days intensity + predicted difference = corrected 3 days intensity, assay value for 3 days intensity 4-8 days ago + corresponding predicted difference

Obtaining corrected 3-day intensity;

h. correcting the 3-day intensity result of the depth network prediction by using the weighted corrected 3-day intensity, thereby overcoming the influence of working condition fluctuation;

FIG. 3 shows the predicted results of the model of the present invention, which shows the comparison curves of the 3-day strength measured value and the predicted value of the cement for 3#, 4#, 5# and 6# mills, respectively, and the statistical results of the model prediction are shown in Table 4: the proportion of the relative error in the range of +/-5% can basically reach 80%, the MSE of the model is relatively small, R2 is positive and basically close to 1, and the model has relatively good following performance.

TABLE 4 model prediction resultsFruit

	3#	4#	5#	6#
					MSE	1.18	0.77	1.59	1.47
R2	0.42	0.48	0.41	0.46
					±5%	0.78	0.8	0.8	0.81

FIG. 4 is a formula optimization interface diagram of the present invention, which has functions of mill selection, formula recommendation, predictive display, and historical review, and specifically includes the following main contents:

(1) the first part is mill selection, the intelligent optimization recommendation system for the cement formula currently comprises 6 mills, an equipment list column is arranged on the leftmost side of a main interface, and different mills can be selected for operation;

(2) the second part is a formula basis part, the latest average data of each attribute are displayed in real time, a change curve of the corresponding attribute in a given time range can be observed by clicking historical trend, the replacement time option of the second part is a historical review function, and the predicted value and the recommended value of each time point can be obtained by changing time;

(3) the third part is a part for displaying the ratio of clinker, gypsum, limestone and stone powder at the current time to the recommended formula at the current stage;

(4) the fourth part is a cement quality prediction part, the reference prediction is that yesterday 3-day strength is predicted according to yesterday clinker quality and proportion, and the real-time prediction is that part of data is today, and part of data is yesterday, and then today 3-day strength is predicted;

(5) the fifth part is a target setting part and can enter a target setting value interface, the interface has 3-day strength, cement loss on ignition and 3-day strength setting values of cement SO3, and the prediction and display of the recommended formula proportion and the 3-day strength can be obtained by clicking to start recommendation;

(6) and the last part is a formula recommendation display part, which can give a formula recommendation value according to the prepared data of the current ratio of clinker, limestone and gypsum to mineral powder, and display the clinker ratio, the limestone ratio and the gypsum ratio recommendation value in detail.

Compared with the prior art, the method provided by the invention has the advantages that the prediction and recommendation are carried out according to the real-time production data, the operation is simple, the environment adaptation capability is strong, the 3-day strength of the cement is accurately predicted, the optimal clinker proportion is recommended, the cement production operation cost is reduced, the energy is saved, the enterprise management is enhanced, the production efficiency is improved, and the enterprise benefit maximization is facilitated.

The foregoing shows and describes the general principles and features of the present invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims

1. Making a decision table for data preparation: and splicing and aligning the historical data of the cement strength for 3 days to remove redundancy, and obtaining a big data decision table from abnormal and contradictory data.

2. Carrying out data preprocessing: filtering, denoising and outlier removing are carried out on the decision table data by using a data preprocessing technology; processing the problems of data entry lag, data entry error and the like to obtain a new decision table; the screened problems comprise inspection delivery, unmatched inspection reporting time, formula proportion mutation, data filling errors, frequent adjustment and formula accumulation of less than 100 percent; the problem processing method comprises data cleaning, data integration and the like.

3. Rough set attribute reduction finds key attributes: and (3) finding out key attributes influencing the three-day strength by using a rough set and attribute reduction algorithm and sequencing, wherein the key attributes comprise clinker quality, a ratio table, a formula and cement fineness of a mill.

4. Training a neural network to establish a three-day intensity prediction model:

(1) training a neural network according to the rough set and the key attributes found by the reduction algorithm;

(2) predicting the strength of the clinker in 3 days in the previous day by combining the real-time strength data in one day and the quality and ratio of the clinker;

(3) and combining the 3-day intensity increment of the two adjacent days predicted by the regression model with the 3-day intensity of the previous day predicted to obtain the 3-day intensity of the current day.

5. And adjusting the formula of the cement mill in real time according to the ratio of clinker to mineral powder in the prediction model.

6. The regression model feedback adjustment rolling optimization rule base comprises the following steps: collecting actual test cement strength data, performing adjacent sampling by using a decision tree and an apriori algorithm to establish a difference value regression model, and optimizing a three-day strength model by combining a neural network prediction model to realize rolling optimization.

7. The clinker formula adjustment idea is as follows:

(1) firstly, obtaining a strength prediction model of 3 days in the previous day through a neural network, obtaining actual assay data by interpolation regression sampling, establishing a regression model to obtain 3-day strength of two adjacent days, obtaining the latest, namely corrected 3-day strength, and predicting by a rolling optimization rule to obtain a new clinker adjustment scheme;

(2) adjusting the ratio of clinker to mineral powder to obtain the actually measured cement strength;

(3) and comparing the predicted intensity value with the measured value, correcting and adjusting the clinker proportion by using a fuzzy PID algorithm when the error value of the predicted intensity value and the measured value exceeds an error allowable range, and giving a clinker and mineral powder adjusting scheme by using a feedback correction algorithm.