Cement clinker f-caco adaptive prediction method

Abstract

The application discloses a cement clinker f-Cao self-adaptive prediction method, first, the obtained auxiliary variable data and free calcium oxide data are normalized, then the window length and the update step are set and the sample set is divided, the new data is processed and the similarity index of the new and old data is calculated, then the training set is determined according to the similarity index to train the neural network, and by adopting the sliding window strategy, the window length is adjusted, the model with the best effect is selected as the final prediction model, and the self-adaptive update is realized. The application fully considers the problems of different working conditions and time elapse in the new dry process cement burning process, so that the prediction accuracy of the model is improved, the self-adaptive prediction of the model is realized, and the multi-working condition prediction accuracy of the cement clinker free calcium oxide content is improved.

Description

Technical Field

[0001] This invention relates to the field of methods for detecting f-Cao content in cement clinker, specifically an adaptive prediction method for f-Cao content in cement clinker. Background Technology

[0002] f-CaO in cement clinker is one of the key indicators, significantly impacting cement quality and performance. Traditionally, the measurement of f-CaO content in cement clinker has been achieved using chemical analysis methods, such as acid hydrolysis-titration and calcification methods. These methods offer advantages like high accuracy and reliability, but they are time-consuming and require specialized chemical analysis techniques and equipment, making them unsuitable for rapid online analysis. Furthermore, while existing technologies favor neural network prediction methods, soft measurement models built using neural networks cannot adapt to the complex and variable operating conditions of cement clinker. Summary of the Invention

[0003] This invention provides an adaptive prediction method for f-CaO content in cement clinker to solve the problem that existing technologies using neural networks to detect f-CaO content in cement clinker cannot adapt to changes in new working conditions.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] An adaptive prediction method for cement clinker f-Cao, characterized by the following steps:

[0006] An adaptive prediction method for cement clinker f-Cao, characterized by the following steps:

[0007] Step 1: Obtain auxiliary variable data for m sampling periods within a continuous time period and free calcium oxide data for h sampling periods within the continuous time period from the cement production line, wherein the sampling period of the free calcium oxide data is different from the sampling period of the auxiliary variable data.

[0008] Based on the acquired free calcium oxide data, a normalized matrix B for free calcium oxide is constructed, and a normalized matrix A is constructed by averaging the auxiliary variable data according to the sampling period of the free calcium oxide data. * The normalized matrix B of free calcium oxide and the normalized matrix A of auxiliary variable data are then compared. * Combined into matrix C, C = (A * ,B);

[0009] Step 2: Using the matrix C obtained in Step 1 as the sample dataset, set the window length L, and divide the matrix C into multiple sample subsets C of length L based on the set window length L. * ;

[0010] Step 3: Obtain the newly added auxiliary variable data and free calcium oxide data from the cement production line, and set the update step size b. When the newly added free calcium oxide data reaches the set update step size b, construct the normalized matrix B of the newly added free calcium oxide data. o The newly added auxiliary variable data were averaged according to the sampling period of the newly added free calcium oxide data before a normalized matrix A was constructed. o Then, the normalized matrix B of the newly added free calcium oxide data is combined. o The normalized matrix A of the newly added auxiliary variable data o Construct matrix E, and the i-th row of matrix E is E i From matrix A o The i-th row A o i Matrix B o The i-th row B o i Composition, namely E i =[A o i B o i ];

[0011] Compare matrix E with each sample subset C obtained in step 2. * By comparing the data, a subset C of samples with the required similarity is obtained. * As a training set;

[0012] Step 4: Use the training set obtained in Step 2 to train the neural network model, and use the trained neural network model as a preliminary prediction model for free calcium oxide.

[0013] Step 5: Set different window lengths and repeat steps 2, 3, and 4 to obtain multiple sets of preliminary prediction models for free calcium oxide. Select the set of preliminary prediction models for free calcium oxide with the best prediction performance as the final prediction model for free calcium oxide, and use the final prediction model for free calcium oxide to make predictions.

[0014] Step 6: After completing a prediction, discard the final prediction model for free calcium oxide used in the current prediction and wait for the next batch of newly added free calcium oxide data to arrive at b. Then repeat steps 3-5 to obtain a new final prediction model for free calcium oxide, thereby achieving adaptive updating of the model.

[0015] In the further step 3, matrix E and each sample subset C are calculated using the similarity index calculation formula and the similarity index formula. * The similarity index is used to select a subset C of samples that meet the requirements. * As a training set.

[0016] Furthermore, in step 3, the subset C of samples with the highest similarity index is selected. * As a training set.

[0017] In a further step 4, the neural network model employs a Long Short-Term Memory (LSTM) network.

[0018] Furthermore, during training in step 4, the neural network model is trained using the training set to obtain prediction data, and the newly added normalized auxiliary variable data is used as the test set data. The prediction data and the test set data are compared to obtain the goodness of fit and root mean square error. The neural network model that meets the requirements for goodness of fit and root mean square error is used as the preliminary prediction model for free calcium oxide that has been trained.

[0019] In the further step 5, the goodness of fit and root mean square error of the preliminary prediction models of free calcium oxide trained under different window lengths are compared, and the preliminary prediction model of free calcium oxide with the best goodness of fit and root mean square error is selected as the final prediction model of free calcium oxide.

[0020] This invention is an adaptive prediction method for f-CaO content in cement clinker based on an improved JITL strategy. It can fully consider the different characteristics of auxiliary variables under different working conditions, classify and model them according to the different characteristics of auxiliary variables, and construct an adaptive prediction model. This can improve the stability and robustness of the prediction model in long-term learning and provide an accurate basis for predicting the f-CaO content in cement clinker.

[0021] Compared with the prior art, the advantages of the present invention are:

[0022] 1. Compared with conventional direct detection methods for f-CaO content in cement clinker, this invention uses a data-driven modeling method to predict f-CaO content in real time. This solves the problems of traditional chemical analysis methods being time-consuming, requiring specialized chemical analysis techniques and equipment, and not being suitable for rapid online analysis. It meets the actual needs for accuracy and speed, and provides real-time guidance for production.

[0023] 2. This invention addresses the problem of inaccurate prediction of f-CaO content in cement clinker due to varying operating conditions during the new dry-process cement firing process. An adaptive strategy model is proposed. A sliding window is used to improve the Just-In-Time Learning (JITL) strategy, thereby enhancing the model's adaptability and resolving the issues of the model's inability to adapt to changes in operating conditions and the JITL strategy's inability to accurately partition the dataset.

[0024] 3. This invention takes into account the time delay between f-CaO and auxiliary variables. It utilizes a Long Short Memory Gate Neural Network (LSTM) to better capture dynamic changes in time series data. By treating the auxiliary variables and the f-CaO content of cement clinker as time series, the model's ability to handle time series problems is improved. Attached Figure Description

[0025] Figure 1 This is a curve showing the predicted f-CaO content when L=100 in an embodiment of the present invention.

[0026] Figure 2 This is a prediction curve of f-CaO content when L=150 in an embodiment of the present invention.

[0027] Figure 3 This is a curve showing the predicted f-CaO content when L=200 in an embodiment of the present invention.

[0028] Figure 4 This is a curve showing the predicted f-CaO content when L=250 in an embodiment of the present invention.

[0029] Figure 5 This is a curve showing the predicted f-CaO content when L=300 in an embodiment of the present invention. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0031] This embodiment discloses an adaptive prediction method for cement clinker f-Cao, characterized by the following steps:

[0032] Step 1: According to the cement production process mechanism, obtain auxiliary variable data for m consecutive sampling periods t1 from the DCS distributed control system of the cement production line and the artificial chemical laboratory detection, as well as free calcium oxide data for h consecutive sampling periods t2. The auxiliary variable data for each sampling period is n-dimensional data, and the n-dimensional auxiliary variable data for the i-th sampling period is denoted as row vector a. i The free calcium oxide data in the j-th sampling period is denoted as column vector b. j , i∈[1,m], j∈[1,h], and have

[0033] In this embodiment, taking the online prediction of free calcium oxide content in cement clinker produced by the new dry process cement plant as an example, based on the mechanism model of free calcium oxide in cement clinker and auxiliary variables, kiln current, kiln tail flue temperature, kiln speed, kiln head coal feed rate, kiln head hood temperature, raw meal feed rate, kiln head negative pressure, decomposer outlet temperature, decomposer coal feed rate, tertiary air temperature, and primary grate pressure are selected as auxiliary variables. A sampling period of 5 seconds is used to obtain 691,200 data points of auxiliary variables for 40 consecutive days in the DCS, and 480 data points of free calcium oxide content for the corresponding 40 consecutive days of auxiliary variable data are obtained from manual testing of cement clinker at 2-hour intervals. Due to the high sampling frequency of auxiliary variable data (5 seconds), the data volume is enormous. In addition, due to the influence of equipment wear and aging, abnormal operating conditions, and changes in the on-site environment, there is interference information and outliers in the detected auxiliary variable and f-CaO content data. Therefore, filtering and outlier removal are required. After outlier removal and filtering, the number of auxiliary variables is 480,000 and the corresponding 400 free calcium oxide content data.

[0034] Then, calculate the row vector a. i The xth feature a in i×x The normalized value is a′ i×x Thus, a normalized matrix A = (a′) is constructed for the auxiliary variable data. i×x ) m×n Calculate column vector b j The normalized value of the j-th value is b′ j×1 Thus, a normalized matrix B = (b′) for free calcium oxide data is constructed. j×1 ) h×1 x∈[1,n];

[0035] In this embodiment, the normalized value is calculated using equation (1):

[0036]

[0037] In formula (1): a i×p These are the raw values ​​(i.e., the auxiliary variables and free calcium oxide data that have not been normalized), a′ i×p It is the normalized value, max(a) i×p ), min(a i×p ) is a i×p The maximum and minimum values ​​of all data corresponding to the feature.

[0038] Since the sampling periods for free calcium oxide data and auxiliary variable data are different, this embodiment performs mean processing on the auxiliary variable data within the sampling period t2 of the free calcium oxide data. Therefore, the matrix A′ is obtained by averaging the auxiliary variables within one sampling period of the free calcium oxide data. i =[a″i×1 ,a″ i×2 ,…,a″ i×n Thus, h results are obtained. And the normalization matrix B = (b′) corresponding to h free calcium oxide data j×1 ) h×1 ,where a i×x Let A′ be the mean of the x-th auxiliary variable in the i-th sampling period. i This represents the result of taking the average value in the i-th sampling period.

[0039] Next, the matrix A obtained after processing the auxiliary variables is... * The normalized matrix B of free calcium oxide is (b′) j×1 ) h×1 Combined into matrix C = (A * B), which serves as the sample dataset for subsequent use, wherein:

[0040]

[0041] Taking the online prediction of free calcium oxide content in cement clinker produced by the new dry process cement production method in cement plants as an example, the 480,000 auxiliary variable data are averaged according to the sampling period of free calcium oxide in 2-hour units. Then, the average of 11 auxiliary variables within each 2-hour period is calculated and matched with the free calcium oxide content at that time, thus obtaining 400 sets of data on auxiliary variables and free calcium oxide, which are used as subsequent datasets.

[0042] Step 2: Set the window length L, and divide the matrix C obtained in Step 1 into multiple sample subsets of length L. Where D i For the i-th sample subset, D i =(A′) i ,b 1×i ).

[0043] Taking the online prediction of free calcium oxide content in cement clinker from the new dry process cement production method in cement plants as an example, based on the window length of 50 set above, the 400 sets of sample data are divided into 8 sample subsets, D1 to D8, according to L=50, and arranged in chronological order.

[0044] Step 3: Set the update step size 'b' for the newly added auxiliary variable sample dataset. When the number of newly added free calcium oxide data reaches 'b', normalize the newly added auxiliary variable data and the free calcium oxide data. Specifically:

[0045] The newly added free calcium oxide data were normalized to obtain the normalized matrix B. oThe newly added auxiliary variable data were averaged according to the sampling period of the newly added free calcium oxide data, and then normalized to obtain the normalized matrix A. o Then, the normalized matrix B° of the newly added free calcium oxide data and the normalized matrix A of the newly added auxiliary variable data are combined. o Construct matrix E, and the i-th row of matrix E is E i From matrix A o The i-th row A o i Matrix B o The i-th row B o i Composition, namely E i =[A o i B o i ].

[0046] Taking the online prediction of free calcium oxide content in cement clinker from the new dry process cement production method in cement plants as an example, the update step size b is set to 1 day, which means 17,280 auxiliary variables and 12 corresponding free calcium oxide content data. The window length of the dataset is set to 50, which means 72,000 auxiliary variables and 50 free calcium oxide data. After mean normalization, there are 50 sets of auxiliary variables and 50 sets of free calcium oxide data.

[0047] Then, matrix E is compared with each sample subset D. i Similarity calculations are performed, and the relationship between the newly added data matrix E and each sample subset D is calculated using the similarity index calculation formula. i Similarity index S qi .

[0048] Further research on distance-based similarity index calculation methods revealed that measuring similarity solely by the distance between two points is somewhat one-sided. Researchers have shifted their focus to spatial orientation, proposing a similarity index calculation method that combines distance and angle. This approach provides a more comprehensive understanding of the correlation between sample data and the database. The formula for calculating the angle and the similarity index S based on distance and angle are detailed below. i The formula is expressed as:

[0049] cos(θ xi )= <x q ,x i > / (‖x i ||2||x q ||2) (2),

[0050]

[0051] In the formula, cos(θ)xi ) represents the newly added sample data x q With data points x in the sample subset i Angular similarity, cos(θ) xi )≥0, i=1,...,L, w is the weight.

[0052] Obtain the similarity index S qi Then, according to the similarity index S qi The samples are sorted, and the subset with the highest similarity is selected as the training set for the immediate learning strategy.

[0053] Step 4: Feed the training set obtained in Step 3 into the Long Short-Term Memory (LSTM) network for training. During training, the LSTM network is trained with the training set to obtain prediction data, and the data of the newly added auxiliary variables after normalization is used as the test set data. The prediction data and the test set data are compared to obtain the goodness of fit R. 2 And the root mean square error (RMSE). The goodness of fit R obtained from training. 2 The Long Short-Term Memory (LSTM) network, which is closest to 1 and has the smallest root mean square error (RMSE), was used as a preliminary prediction model for free calcium oxide.

[0054] Wherein, the fit R 2 The calculation formula is:

[0055]

[0056] In equation (4), For prediction data; y i For the actual values ​​in the test set. In formula (4), the numerator represents the total amount of information that the model did not capture, and the denominator represents the amount of information carried by the real labels, so R 2 It measures the proportion of information that the model fails to capture relative to the information contained in the true labels. If model R... 2 A result of 0 indicates a poor model fit; the closer the result is to 1, the better the regression fit, and a result of around 0.8 is generally considered to indicate a high model fit.

[0057] The root mean square error (RMSE) is the square root of the ratio of the square of the deviation between the predicted and actual values ​​to the number of observations, n. It measures the deviation between the predicted and actual values ​​and is particularly sensitive to outliers in the data. The RMSE is shown in equation (5).

[0058]

[0059] Taking the online prediction of free calcium oxide content in cement clinker produced by the new dry process cement plant as an example, the newly added data is used as a test set and fed into the prediction model for prediction. The final result is R. 2=0.202, RMSE=0.067.

[0060] Step 5: Adjust the window length L to the changed window length L by adopting a sliding window strategy. i And repeat steps 2, 3, and 4 to obtain multiple sets of preliminary prediction models for free calcium oxide. Among them, L i Let be the window length that is changed for the i-th time.

[0061] In this example, the sliding window strategy is combined with the real-time learning strategy to address the shortcoming of the real-time learning strategy in accurately dividing the dataset. By obtaining the goodness-of-fit performance index under different L lengths, the optimal window length L is finally determined. Therefore, the cases of L = 100, 150, 200, 250, and 300 are selected to verify the optimal window length.

[0062] By comparing different window lengths L i The obtained goodness of fit R 2 The root mean square error (RMSE) is used to determine the optimal window length L. u and with window length L u The corresponding preliminary prediction model for free calcium oxide is used as the final prediction model for free calcium oxide. The optimal window length L is utilized. u The final prediction model for free calcium oxide under the given conditions completed the prediction and obtained the optimal prediction result.

[0063] Taking the online prediction of free calcium oxide content in cement clinker from the new dry process cement production method of the cement plant as an example, the prediction model of free calcium oxide content in cement clinker trained above is combined with the sliding window strategy. Different window prediction accuracy is tested in 400 sets of sample data. Window lengths L=100, L=150, L=200, L=250 and L=300 are selected to verify the appropriate window length. The model parameters are kept unchanged, and the step size is 50 sets of data for evidence. Figures 1-5 The prediction curves of f-CaO content in cement clinker using the LSTM model under different window lengths are presented. Table 1 shows the root mean square error and goodness of fit of the prediction results under different window lengths. Table 1 is as follows:

[0064] Table 1. Mean squared error and goodness of fit of prediction results under different window lengths.

[0065]

[0066] The calculations and analysis in Table 1 show that when the window length is L=100, L=200, and L=300, the prediction error of f-CaO content in cement clinker is relatively large. This is mainly because the window is too small and the amount of data within the window is too small to reflect the changes in the current cement production conditions. Especially when L=100, the insufficient training sample data cannot establish an effective model. As the window length L increases, the prediction accuracy of f-CaO content in cement clinker improves. This is mainly because the window contains more data that can reflect the characteristics of the prediction samples for training the model. When the window length is 300, the mean square error is the smallest and the goodness of fit is large, resulting in the highest prediction accuracy of the model. Further increases in window length lead to a decrease in prediction accuracy, mainly because too much similar interference data is used to train the model. Based on the above analysis, the experimental results indicate that the optimal window length is 250.

[0067] Step 6: Implement adaptive model updates:

[0068] After completing a prediction, the final prediction model for free calcium oxide used in the current prediction is abandoned, and the system waits for the next batch of newly added free calcium oxide data to arrive at b. Then, steps 3-5 are repeated to obtain a new final prediction model for free calcium oxide, thereby achieving adaptive updating of the model.

[0069] After the prediction ends, wait for the next batch of new data to reach the specified update step size of 12, that is, accumulate 12 free calcium oxide data points and their corresponding auxiliary variables. Then, discard the already constructed model, put the previously added data into the historical database in chronological order, and update the database. Based on the updated database and the new data, perform the above operations again to make a new prediction, and the model adaptively updates.

[0070] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. These embodiments are merely descriptions of preferred embodiments and are not intended to limit the scope or concept of the invention. The specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. Such combinations, as long as they do not violate the spirit of the present invention, should also be considered as part of this disclosure. To avoid unnecessary repetition, the present invention will not further describe the various possible combinations.

[0071] This invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this invention and without departing from the design idea of ​​this invention, all modifications and improvements made by those skilled in the art to the technical solutions of this invention should fall within the protection scope of this invention. The technical content for which protection is sought in this invention has been fully described in the claims.

Claims

1. An adaptive prediction method for cement clinker f-Cao, characterized in that, Includes the following steps: Step 1: Obtain auxiliary variable data for m sampling periods within a continuous time period and free calcium oxide data for h sampling periods within the continuous time period from the cement production line, wherein the sampling period of the free calcium oxide data is different from the sampling period of the auxiliary variable data. Based on the acquired free calcium oxide data, a normalized matrix B for free calcium oxide is constructed, and a normalized matrix A is constructed by averaging the auxiliary variable data according to the sampling period of the free calcium oxide data. * The normalized matrix B of free calcium oxide and the normalized matrix A of auxiliary variable data are then compared. * Combined to form a matrix , ; Step 2: Using the matrix C obtained in Step 1 as the sample dataset, set the window length L, and based on the set window length L, divide the matrix... Divide into multiple sample subsets of length L. ; Step 3: Obtain the newly added auxiliary variable data and free calcium oxide data from the cement production line, and set the update step size b. When the newly added free calcium oxide data reaches the set update step size b, construct the normalized matrix of the newly added free calcium oxide data. The newly added auxiliary variable data were averaged according to the sampling period of the newly added free calcium oxide data before a normalized matrix was constructed. Then, the normalization matrix of the newly added free calcium oxide data was combined. Normalized matrix of newly added auxiliary variable data Construct matrix E, and the i-th row of matrix E... From the matrix The i-th row composition, Compare matrix E with each sample subset obtained in step 2. The data was compared to obtain a subset of samples that met the similarity requirements. As a training set; Step 4: Train the neural network model using the training set obtained in Step 3, and use the trained neural network model as the preliminary prediction model for free calcium oxide; the neural network model adopts a Long Short-Term Memory (LSTM) network. Step 5: Set different window lengths and repeat steps 2, 3, and 4 to obtain multiple sets of preliminary prediction models for free calcium oxide. Select the set of preliminary prediction models for free calcium oxide with the best prediction performance as the final prediction model for free calcium oxide, and use the final prediction model for free calcium oxide for prediction. Step 6: After completing a prediction, discard the final prediction model for free calcium oxide used in the current prediction and wait for the next batch of newly added free calcium oxide data to arrive at b. Then repeat steps 3-5 to obtain a new final prediction model for free calcium oxide, thereby achieving adaptive updating of the model.

2. The adaptive prediction method for cement clinker f-Cao according to claim 1, characterized in that, In step 3, matrix E and the distance calculation formula for each sample subset are obtained by using the similarity index angle calculation formula and the distance calculation formula. The similarity index is used to select a subset of samples that meet the requirements. As a training set.

3. The adaptive prediction method for cement clinker f-Cao according to claim 2, characterized in that, In step 3, select the subset of samples with the highest similarity index. As a training set.

4. The adaptive prediction method for cement clinker f-Cao according to claim 1, characterized in that, In step 4, during training, the neural network model is trained using the training set to obtain prediction data, and the newly added normalized auxiliary variable data is used as the test set data. The prediction data and the test set data are compared to obtain the goodness of fit and root mean square error. The neural network model that meets the requirements for goodness of fit and root mean square error is used as the preliminary prediction model for free calcium oxide that has been trained.

5. The adaptive prediction method for cement clinker f-Cao according to claim 4, characterized in that, In step 5, the goodness of fit and root mean square error of the preliminary prediction models of free calcium oxide trained under different window lengths are compared, and the preliminary prediction model of free calcium oxide with the best goodness of fit and root mean square error is selected as the final prediction model of free calcium oxide.

Images