Cement clinker f-cao prediction method considering multiple time series

Abstract

The application discloses a cement clinker f-CaO prediction method considering multiple time sequences, and comprises the following steps: 1, auxiliary variable data acquisition and preprocessing; 2, calculating the residence time of materials in each device, according to the length of the residence time, intercepting time sequence variables of different lengths in different devices, and then compressing the intercepted time sequences into the same length; 3, establishing a grouping convolution fusion attention mechanism auxiliary variable feature extraction network, feeding the extracted features into an LSTM, and finally obtaining an optimal model auxiliary variable feature extraction network and an optimal duration. The application fully considers that the influence degrees of different variables on the final content of f-CaO are different in different time periods, and the variables influence each other and interfere with the model feature extraction, through extracting variable time point features and device grouping features, the effective features of the auxiliary variables can be quickly extracted, and the online prediction accuracy of the cement clinker f-CaO is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cement clinker f-CaO content soft measurement, in particular to a cement clinker f-CaO prediction method considering multiple time series. BACKGROUND

[0002] The free lime (f-CaO) in cement clinker is the calcium oxide in free state existing in the cement clinker which fails to react with acidic oxides such as silicon dioxide in the cement production process. The f-CaO content in cement clinker is a main evaluation parameter of the quality of cement clinker, and the level thereof is directly related to the quality of the final cement product. When the content is too high, the strength of the cement is reduced and even cracking occurs, and reducing the f-CaO content in the clinker requires more persistent calcination of the clinker, which leads to an increase in energy consumption and affects the service life of the rotary kiln refractory bricks.

[0003] At present, the commonly used f-CaO content detection method is to periodically manually sample after the clinker is formed and cooled, and then perform laboratory offline testing. Since the sampling interval is long (usually 2h), the f-CaO content detection value has a lag, which is difficult to guide the subsequent cement clinker production process, causing fluctuations in the site working conditions, and the quality and yield of the clinker product are worrisome. Another commonly used method is to establish a neural network prediction model between auxiliary variables and the f-CaO in cement clinker, which can timely predict the f-CaO content in cement clinker. However, due to the reasons that there are many auxiliary variable parameters in cement production and the influence of different variables on f-CaO is different in different time periods, after the time series data of all variables are directly sent into the neural network model, a large amount of disturbance information will interfere with the training of the neural network, resulting in low prediction accuracy of the model. SUMMARY

[0004] The present application is to solve the above-mentioned problems existing in the prior art, and proposes a cement clinker f-CaO prediction method considering multiple time series, so as to exclude a large amount of disturbance information, extract the effective features of the auxiliary variable time series data, and thus improve the prediction accuracy of the f-CaO content in cement clinker.

[0005] In order to achieve the above-mentioned application purposes, the present application adopts the following technical solutions:

[0006] The cement clinker f-CaO prediction method considering multiple time series is applied to a production device composed of a preheating and decomposition system, a rotary kiln and a grate cooler, and the characteristic thereof is that the cement clinker f-CaO prediction method is performed according to the following steps:

[0007] Step 1: Based on the process mechanism of cement clinker f-Cao generation, select n auxiliary variables with large correlation coefficients with cement clinker f-Cao. Let the sampling period of cement clinker f-Cao data be T. Obtain p cement clinker f-Cao data and their corresponding auxiliary variables from the DCS system according to the sampling period T, and perform preprocessing and column-normalization respectively to obtain the cement clinker f-Cao vector. and auxiliary variable matrix in, Indicates transpose, b i T represents the i-th sampling period. i Data on cement clinker f-Cao within the range; A i T represents the i-th sampling period. i Auxiliary variable data within, and T represents the i-th sampling period. i Auxiliary variable data collected from the preheating decomposition system, rotary kiln, and grate cooler;

[0008] Step 2, Processing the model input:

[0009] Step 2.1: Based on empirical formulas, the residence times T of cement materials in the preheating and decomposition system, rotary kiln, and grate cooler are respectively... yf T hzy and T blj And T yf < <T hzy <T blj T>T yf +T hzy +T blj ;

[0010] Step 2.2, let the i-th sampling period T i The start time is The i-th sampling period T i The end time is make In the i-th sampling period T i Auxiliary variable data of internal grate cooler The time period matched is T. blj And in the i-th sampling period T i Auxiliary variable data of internal grate cooler Extract from Start to stay time T blj The data segment up to this point;

[0011] Let A hzy In the i-th sampling period T i Auxiliary variable data of internal rotary kiln The time period matched in is bl1·T hzyand the auxiliary variable data of the preheating and decomposing system in the i th sampling period T i Auxiliary variable data of the rotary kiln The data segment from to the residence time ;

[0012] Let A yf The auxiliary variable data of the preheating and decomposing system in the i th sampling period T i The time period matched in is bl2·T yf , and the auxiliary variable data of the preheating and decomposing system in the i th sampling period T i The data from to to , wherein bl1 and bl2 represent duration coefficients, and bl1 and bl2>1;

[0013] Step 2.3, after the auxiliary variable data of the grate cooler in p sampling periods are respectively intercepted according to the process of step 2.2, the p grate cooler data are compressed into a grate cooler matrix with a length of q by the piecewise mean value method; similarly, a rotary kiln matrix and a preheating and decomposing system matrix with a length of q are obtained; thus, the auxiliary variable single-channel matrix A' with a dimension of p×q×n×1 is formed by the grate cooler matrix, the rotary kiln matrix and the preheating and decomposing system matrix;

[0014] Step 3, constructing an auxiliary variable feature extraction network based on device fusion attention, which sequentially includes a time point data weighting module, a channel conversion layer, a device weighting module and an LSTM network;

[0015] Step 3.1, the time point data weighting module sequentially includes a two-dimensional convolution layer and a time point data weighting layer;

[0016] The auxiliary variable single-channel matrix A' is input into the auxiliary variable feature extraction network, and after being processed by the two-dimensional convolution layer with a convolution kernel size of , a feature map Q 1A with a dimension of p×q'×n×1 is obtained, wherein q' represents the length of the intercepted and compressed time period;

[0017] The time point data weighting layer processes the feature map Q 1A using three convolution layers with a kernel size of 1×1, respectively, to obtain three feature maps Q 1B , Q 1C and Q 1D , respectively, and then processes Q 1B , Q 1C and Q 1DReconstruction is performed in the dimension of auxiliary variable number and time period, to obtain three feature maps Q with the dimension of p×N×1 1B,reshape , 1C,reshape ,Q 1D,reshape ;

[0018] Q is obtained by using formula (1) 1B,reshape The correlation degree of the data of the kth time point in Q k 1B,reshape and the data of the jth time point in Q 1C,reshape Thus, the correlation degree matrix of the auxiliary variable is obtained

[0019]

[0020] In formula (1), N represents the number of time point data of all auxiliary variables, and N=q′×n;

[0021] The transpose of Q 1D,reshape and Q 1S is multiplied to obtain the feature map matrix Q 1F,reshape , and then Q 1F,reshape is reconstructed to obtain the time point weight feature map Q 1A with the same dimension as Q 1F , and finally Q 1A and Q 1F are added to obtain the feature map Q 1 of the weighted auxiliary variable;

[0022] Step 3.2, the channel conversion layer exchanges the channel dimension and the dimension of the number of auxiliary variables of Q 1 by using the transpose function, to obtain the multivariable channel feature map Q 2 with the shape of p×q′×1×n;

[0023] Step 3.3, the device weighting module sequentially includes a grouping fusion module and a device attention mechanism module;

[0024] The grouping fusion module includes a grouping convolution layer and a depth separable convolution layer;

[0025] The grouping convolution layer divides the multivariable channel feature map Q 2 to obtain the preheating decomposition system feature map rotary kiln feature map and grate cooler feature map

[0026] The depth separable convolution layer processes by using point-by-point convolution to obtain the fused preheating decomposition system feature map​ Fused rotary kiln feature map Fused grate cooler feature map

[0027] Will The feature map of any u-th device and v-th device is respectively denoted as Q u 3C And Q v 3C The device attention mechanism module respectively processes Q u 3C And Q v 3C After transposition and reconstruction, the feature map matrix Q u 3D of the u-th device and the feature map matrix Q v 3D of the v-th device are obtained; thus, the influence Q uv 3X of the u-th device on the v-th device is obtained by using formula (2), and further the device influence feature map Q

[0028] (2)

[0029] The weighted feature map Q u 3E of the u-th device is obtained by using formula (3), and thus the device weighted time sequence feature feature map Q 3E is obtained:

[0030]

[0031] In formula (3), Q 3β represents a scale coefficient;

[0032] Step 3.4, the time sequence feature feature map Q 3E is sent into the LSTM network for prediction, and a dimension p×1 auxiliary variable prediction result is obtained;

[0033] Step 4, determination of the optimal duration and prediction of cement clinker f-CaO;

[0034] The auxiliary variable matrix A and the cement clinker f-Cao vector b are input into the model auxiliary variable feature extraction network, different bl1 and bl2 are selected, the network is trained by the Adam optimizer, and the mean square error loss function is calculated to update the network parameters until the maximum training times are reached, thus obtaining the optimal model auxiliary variable feature extraction network and the optimal duration coefficient And are used for predicting the cement clinker f-Cao of the auxiliary variable matrix.

[0035] The electronic device comprises a memory and a processor, and the memory is used for storing a program supporting the processor to execute the cement clinker f-CaO prediction method, and the processor is configured to execute the program stored in the memory.

[0036] The computer readable storage medium stores a computer program, and when the computer program is run by a processor, the steps of the cement clinker f-CaO prediction method are executed.

[0037] Compared with the prior art, the technical progress of the present application is:

[0038] 1、The present application calculates the residence time of the material in each device through the cement process characteristics, then intercepts the data of the corresponding time period according to the device where the variable is located, and excludes a large amount of interference information; the time series length of the auxiliary variable data is unified through mean compression, the original time series characteristics of the auxiliary variable data are retained, and the prediction accuracy of the cement clinker f-CaO is improved.

[0039] 2、The present application proposes a time point data weighting module for the problem that variables have different effects on cement clinker at different time periods, integrates each variable time period data block by using two-dimensional convolution, weights the data block by using the time point data weighting module, highlights the important data block features, and proposes a device grouping weighting module, groups the variables according to the devices, and gives greater weight to the devices with greater influence, thereby improving the ability of the grouped convolution network to extract effective features of the auxiliary variables.

[0040] 3、The present application can fully consider the problem of the auxiliary variables of the time period corresponding to the cement clinker f-CaO time point, perform time sequence matching according to the cement process, intercept effective data segments, and extract effective features of the auxiliary variables of multiple time sequences through the grouped convolution fusion attention network, and finally input into the LSTM, thereby meeting the actual needs of accuracy and rapidity, and guiding the production in real time. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 To intercept auxiliary variable data according to scheme 1;

[0042] Figure 2 Optimal proportion coefficient prediction effect diagram. DETAILED DESCRIPTION

[0043] In this embodiment, a cement clinker f-CaO prediction method considering multiple time sequences is performed according to the following steps:

[0044] Step 1, according to the process mechanism of cement clinker f-Cao, select n auxiliary variables with larger correlation coefficient with cement clinker f-Cao, let the sampling period of cement clinker f-Cao data be T1, let the sampling period of auxiliary variables be T2, and T1 >> T2. Obtain p auxiliary variables and cement clinker f-Cao data with a sampling period of T1 from the DCS system, and perform preprocessing and column normalization to obtain an auxiliary variable matrix denoted as wherein, denotes transposition, A i denotes the auxiliary variable data obtained in the i-th sampling period T1, and wherein denotes the time series length, wherein A yf , A hzy , A blj respectively denote the auxiliary variable data obtained from the preheating and decomposition system, the rotary kiln, and the grate cooler; and the obtained cement clinker f-Cao data is denoted as a vector b i denotes the cement clinker f-Cao data in the i-th sampling period T1;

[0045] Specifically, taking the data of a new dry process cement production line in a certain cement plant as an example, according to the process mechanism of cement clinker f-Cao, 11 variables including the outlet temperature of the decomposition furnace, the coal feeding amount of the decomposition furnace, the raw material feeding amount, the tertiary air temperature, the kiln current, the kiln tail flue gas chamber temperature, the kiln head cover temperature, the kiln speed, the kiln head coal feeding amount, the kiln head negative pressure, and the first stage grate under pressure are selected as auxiliary variables, and 276 continuous cement clinker f-CaO data with a sampling period of 2h are obtained from the laboratory; the corresponding auxiliary variable data of the cement clinker f-Cao for two hours are obtained from the DCS, such as collecting the cement clinker f-Cao data at 8 o'clock, collecting the auxiliary variable data in the time period from 6 o'clock to 8 o'clock, the auxiliary variable data with a sampling period of 10s, 2h, a total of 720 auxiliary variable data, the collected auxiliary variable data shape is (276, 720, 11). Wherein 2h = 7200s, 7200 / 10 = 720.

[0046] There are a small amount of abnormal data in the collected data, which are processed by two methods. In the first case, the f-Cao data is missing or there is a shutdown data, the auxiliary variable data and the cement clinker f-Cao data in the same sampling period of 2h are regarded as a group, and the corresponding group data is directly deleted at the same time. In another case, a variable suddenly changes and then returns to normal, in which case the 3σ rule is used to find the abnormal data points, and these abnormal data are removed and the corresponding missing data are filled with the last normal data point. After processing the abnormal data, there are 231 groups of input and output, i.e. the shape of the cement clinker f-CaO data is (231, 1), and the shape of the auxiliary variable data is (231, 720, 11).

[0047] Step 2, timing matching:

[0048] Step 2.1, according to the empirical formula, the residence time of the cement material in the preheating and decomposing system, the rotary kiln and the grate cooler is T yf , T hzy and T blj respectively; and T yf <<T hzy <T blj , T>T yf +T hzy +T blj ;

[0049] In this embodiment, the residence time of the cement material in the preheating and decomposing system is 40s, the residence time of the cement material in the rotary kiln is 30min, and the residence time of the cement material in the grate cooler is 40min.

[0050] Step 2.2, let the starting time of the i-th sampling period T i be and the ending time of the i-th sampling period T i be T e i nd ; let the matching time period of the auxiliary variable data A of the grate cooler in the i-th sampling period T i be T blj , and the data segment from to the residence time T blj be intercepted in the auxiliary variable data A of the grate cooler in the i-th sampling period T i ;

[0051] let the matching time period of the auxiliary variable data A hzy of the rotary kiln in the i-th sampling period T i be bl1·T hzy , and the data segment from to the residence time T be intercepted in the auxiliary variable data A of the rotary kiln in the i-th sampling period T i ;

[0052] let the matching time period of the auxiliary variable data A yf of the preheating and decomposing system in the i-th sampling period T i be bl2·T yf , and the data segment from to the residence time T be intercepted in the auxiliary variable data A of the preheating and decomposing system in the i-th sampling period T iAuxiliary variable data of inner preheating decomposition system Data intercepted from to , wherein bl1 and bl2 represent duration coefficients, and bl1 and bl2>1;

[0053] In this embodiment, bl1=1.2 and bl2=4 are taken as examples, the preheating decomposition system has four variables, which are the outlet temperature of the decomposition furnace, the coal feeding amount of the decomposition furnace, the raw material feeding amount, and the tertiary air temperature, the time period selected is [48min, 51min20s] with a length of 160s, and the number of time series data points is 16; the rotary kiln has six variables, which are the kiln current, the temperature of the kiln tail smoke chamber, the temperature of the kiln head cover, the kiln speed, the kiln head coal feeding amount, and the kiln head negative six, the time period selected is [47min, 83min] with a length of 36min, and the number of time series data points is 156; the grate cooler has one variable, which is the one-stage grate underpressure, the time period selected is [80min, 120min] with a length of 40min, and the number of time series data points is 240, and the interception process is as shown in Figure 1 ;

[0054] Step 2.3, after the auxiliary variable data of the grate cooler in p sampling periods are intercepted respectively according to the process of step 2.2, the p grate cooler data intercepted are compressed into a grate cooler matrix with a length of q by the piecewise mean value method; similarly, a rotary kiln matrix and a preheating decomposition system matrix with a length of q are obtained; thus, the auxiliary variable single-channel matrix A' with a dimension of p×q×n×1 is formed by the grate cooler matrix, the rotary kiln matrix, and the preheating decomposition system matrix;

[0055] In this embodiment, the variables in the rotary kiln and the grate cooler are compressed by mean value, the variables in the rotary kiln take an average value every 10(156 / 16) data points, and the variables in the grate cooler take an average value every 15(240 / 16) data points, to obtain matrices with the same dimension, and reconstruct an input matrix with a dimension of (231, 16, 11, 1).

[0056] Step 3, an auxiliary variable feature extraction network based on device fusion attention is constructed, which sequentially includes a time point data weighting module, a channel conversion layer, a device weighting module, and an LSTM network;

[0057] Step 3.1, the time point data weighting module sequentially includes a two-dimensional convolution layer and a time point data weighting layer;

[0058] The auxiliary variable single-channel matrix A' is input into the auxiliary variable feature extraction network, and after being processed by the two-dimensional convolution layer with a convolution kernel size of , a feature map Q with a dimension of p×q'×n×1 is obtained 1A , wherein q' represents the length of the time period after interception and compression.

[0059] The time point data weighting layer utilizes three convolution kernels with a 1×1 convolution layer to process the feature map Q 1A , and correspondingly obtain three feature maps Q 1B , Q 1C , and Q 1D . Then, Q 1B , Q 1C , and Q 1D are respectively reconstructed in the auxiliary variable number dimension and the time period dimension to obtain three feature maps Q 1B,reshape , Q 1C,reshape , and Q 1D,reshape with a dimension of p×N×1.

[0060] The correlation degree of the kth time point data Q 1B,reshape in Q k and the jth time point data Q 1B,reshape in Q 1C,reshape is obtained by using formula (1), so as to obtain the correlation degree matrix of the auxiliary variables.

[0061]

[0062] In formula (1), N represents the number of time point data of all auxiliary variables, and N=q′×n.

[0063] The transpose of Q 1D,reshape and Q 1S is multiplied to obtain a feature map matrix Q 1F,reshape . Then, Q 1F,reshape is reconstructed to obtain a time point weight feature map Q 1A with the same dimension as Q 1F . Finally, Q 1A and Q 1F are added to obtain a weighted auxiliary variable feature map Q 1 .

[0064] Step 3.2, the channel conversion layer utilizes the transpose function to exchange the channel dimension and the auxiliary variable number dimension of Q 1 to obtain a multivariable channel feature map Q 2 with a shape of p×q′×1×n.

[0065] ​In this embodiment, a Python machine learning framework using TensorFlow as the underlying library is employed. Conv2D is used to extract auxiliary variable features, with the convolution kernel size set to (4, 1) and the number of output channels set to 1. After the dimensions are swapped using the transpose function, the output data has dimensions of (2^31, 1^3, 1, 11) multi-channel data, where 11 channels represent 11 variables. Data near each pixel in an image typically exhibits similar characteristics. In time series data, the data for each variable at the current moment has similar characteristics to data before the current moment, and the influence between variables is highly uncertain. Assigning a channel to each variable extracts its temporal features while avoiding mutual interference between variables.

[0066] Step 3.3: The device weighting module includes, in sequence, a group fusion module and a device attention mechanism module;

[0067] The group fusion module includes: a grouped convolutional layer and a depthwise separable convolutional layer;

[0068] The grouped convolutional layers process the multivariable channel feature map Q. 2 The preheating decomposition system is divided into sections to obtain a characteristic map. Rotary Kiln General Map Characteristic diagram of a grate cooler

[0069] The depth-separable convolutional layer utilizes pointwise convolution to... The process is performed to obtain the characteristic map of the fused preheated decomposition system. General map of the integrated rotary kiln Feature diagram of the fused grate cooler

[0070] Will The feature maps of any u-th device and v-th device are denoted as Q. u 3C and Q v 3C The device attention mechanism module respectively Q u 3C and Q v 3C After transposition and reconstruction, the feature map matrix Q of the u-th device is obtained. u 3D and the feature map matrix Q of the v-th device v 3D Therefore, the influence Q of the u-th device on the v-th device can be obtained using equation (2). uv 3X This leads to the acquisition of the equipment influence feature map.

[0071] (2)

[0072] The weighted feature map Q of the u-th device is obtained by using formula (3) u 3E , thereby obtaining the weighted time sequence feature map Q of the device 3E :

[0073]

[0074] In formula (3), Q 3β represents a scale coefficient;

[0075] In this embodiment, the preheating and decomposition system sequentially includes the outlet temperature of the decomposition furnace, the coal feeding amount of the decomposition furnace, the raw material feeding amount, the tertiary air temperature, the kiln current, the kiln tail flue gas chamber temperature, the kiln head cover temperature, the kiln speed, the kiln head coal feeding amount, and the kiln head negative pressure in the kiln grouping, and the grate cooler grouping includes the first-stage grate underpressure.

[0076] Step 3.4, the time sequence feature map Q 3E is sent into the LSTM network for prediction to obtain a prediction result of the auxiliary variable with a dimension of p x 1;

[0077] In this embodiment, the number of layers of the LSTM is set to 50, the Dropout parameter is set to 0.05, and the number of layers of the last dense is set to 1.

[0078] Step 4, determination of the optimal duration and prediction of the cement clinker f-CaO;

[0079] The auxiliary variable matrix A and the cement clinker f-CaO vector b are input into the model auxiliary variable feature extraction network, different bl1 and bl2 are selected, the network is trained through the Adam optimizer, the network parameters are updated through the calculation of the mean square error loss function, and the training is stopped until the maximum training number is reached, thereby obtaining the optimal model auxiliary variable feature extraction network and the optimal duration coefficients bl1 and bl2 for predicting the cement clinker f-CaO from the auxiliary variable matrix. * *

[0080] In this embodiment, an electronic device includes a memory and a processor, the memory is used to store a program supporting the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0081] In this embodiment, a computer readable storage medium stores a computer program on the computer readable storage medium, and the computer program is executed by the processor to perform the steps of the above method.

[0082] ​​In this embodiment, bl1 takes {1, 1.1, 1.2, 1.3, 1.4} respectively, bl2 takes {1, 2, 3, 4, 5} respectively, bl3 = 0.6, the loss function is set to mean square error, and the training number is set to 100; by selecting bl1 and bl2 multiple times and recording the average value of the mean square error as shown in Table 1:

[0083] Table 1 is a comparison of mean square errors of different duration ratio coefficients;

[0084] [bl1\bl2] 1 2 3 4 5 1.1 0.153 0.126 0.098 0.085 0.096 1.2 0.121 0.104 0.079 0.067 0.073 1.3 0.151 0.126 0.115 0.075 0.113 1.4 0.168 0.145 0.137 0.099 0.126 1.5 0.205 0.196 0.159 0.124 0.154

[0085] Finally, the optimal duration is found Save the model parameters at this time, and the prediction effect diagram is as shown in Figure 2 Compared with the ordinary model, the accuracy of the method is greatly improved.

Claims

1. A method for predicting f-CaO in cement clinker considering multiple time series, applied to production equipment consisting of a preheating decomposition system, a rotary kiln, and a grate cooler, characterized in that... The method for predicting f-CaO in cement clinker is carried out according to the following steps: Step 1: Based on the process mechanism of f-Cao generation in cement clinker, select the components with a high correlation coefficient with f-Cao in cement clinker. There are two auxiliary variables, let the sampling period of the cement clinker f-Cao data be [missing information]. According to the sampling period p data points of cement clinker f-Cao and their corresponding auxiliary variables are obtained from the DCS system, and preprocessed and normalized by column to obtain the corresponding cement clinker f-Cao vector. and auxiliary variable matrix ,in, Indicates transpose. Indicates the i-th sampling period Data on cement clinker f-Cao within the range; Indicates the i-th sampling period Auxiliary variable data within, and , , , Representing the i-th sampling period Auxiliary variable data collected from the preheating decomposition system, rotary kiln, and grate cooler; Step 2, Processing the model input: Step 2.1: Based on empirical formulas, the residence times of cement materials in the preheating and decomposition system, rotary kiln, and grate cooler are respectively... , and ;and << < , > + + ; Step 2.2: Let the i-th sampling period be... The start time is The i-th sampling period The end time is ;make In the i-th sampling period Auxiliary variable data of internal grate cooler The time period matched in is And in the i-th sampling period Auxiliary variable data of internal grate cooler Extract from Start to stay time The data segment up to this point; make In the i-th sampling period Auxiliary variable data of internal rotary kiln The time period matched in is And in the i-th sampling period Auxiliary variable data of internal rotary kiln Extract from Start to stay time The data segment up to this point; make In the i-th sampling period Auxiliary variable data of the internal preheating decomposition system The time period matched in is And in the i-th sampling period Auxiliary variable data of the internal preheating decomposition system Extract from to The data up to date, of which, , Represents the duration coefficient, and , >1; Step 2.3: Following the process in Step 2.2, the auxiliary variable data of the grate cooler for each of the p sampling periods are extracted. Then, the extracted p grate cooler data are compressed into a grate cooler matrix of length q using the segmented averaging method. Similarly, a rotary kiln matrix and a preheating decomposition system matrix of length q are obtained. Thus, the grate cooler matrix, rotary kiln matrix, and preheating decomposition system matrix constitute a matrix with dimension q. Auxiliary variable single-channel matrix ; Step 3: Construct an auxiliary variable feature extraction network based on device-specific fusion attention, which includes: a time-point data weighting module, a channel conversion layer, a device-specific weighting module, and an LSTM network. Step 3.1: The time-point data weighting module includes, in sequence: a two-dimensional convolutional layer and a time-point data weighting layer; The auxiliary variable single-channel matrix The input is given to the auxiliary variable feature extraction network and processed by a convolution kernel with a size of [missing value]. After processing by a two-dimensional convolutional layer, the resulting product has a dimension of Feature map ,in, This indicates the length of the compressed time segment. The time-point data weighting layer uses three convolutional kernels to... The convolutional layers respectively process the feature maps After processing, three feature maps are obtained. , , Then separately , , Reconstructing the dimensions based on the number of auxiliary variables and the time period, we obtain the following dimensions: Three feature maps , , ; Using equation (1) to obtain Data at the k-th time point and Data at the j-th time point degree of correlation Thus, the correlation matrix of the auxiliary variables is obtained. ; (1) In equation (1), This represents the number of time point data points for all auxiliary variables, and ; Will and Multiplying by the transpose yields the feature map matrix. Then to After reconstruction, the result is the same as Time-point weighted feature maps with the same dimension Ultimately and The feature map of the weighted auxiliary variable is obtained by adding them together. ; Step 3.2: The channel conversion layer uses the transpose function to... By swapping the channel dimension and the number of auxiliary variables, we obtain a shape of... Multivariable channel feature map ; Step 3.3: The device weighting module includes, in sequence, a group fusion module and a device attention mechanism module; The group fusion module includes: a grouped convolutional layer and a depthwise separable convolutional layer; The grouped convolutional layers process multivariable channel feature maps The preheating decomposition system is divided into sections to obtain a characteristic map. Rotary kiln general map Characteristic diagram of a grate cooler ; The depth-separable convolutional layer utilizes pointwise convolution to... , , The process is performed to obtain the characteristic map of the fused preheated decomposition system. General map of the integrated rotary kiln Feature diagram of the fused grate cooler ; Will , , The feature maps of any u-th device and v-th device are denoted as follows: and The device attention mechanism module will respectively and After transposition and reconstruction, the feature map matrix of the u-th device is obtained. and the feature map matrix of the v-th device Therefore, the influence of the u-th device on the v-th device can be obtained using equation (2). This leads to the acquisition of the equipment influence feature map. : (2) The weighted feature map of the u-th device is obtained using equation (3). Thus, the device-weighted time-series feature map is obtained. : (3) In equation (3), Indicates the scaling factor; Step 3.4: Extract the time series feature map. The data is fed into an LSTM network for prediction, resulting in a dimension of... The prediction results of auxiliary variables; Step 4: Determination of the optimal duration and prediction of f-CaO in cement clinker; auxiliary variable matrix and cement clinker f-Cao vector Input the auxiliary variable feature extraction network into the model and select different , Then, the network is trained using the Adam optimizer, and the mean squared error loss function is calculated to update the network parameters until the maximum number of training iterations is reached, thereby obtaining the optimal model auxiliary variable feature extraction network and the optimal duration coefficients. and It is used to predict cement clinker f-Cao from the auxiliary variable matrix.

2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the cement clinker f-CaO prediction method of claim 1, and the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the cement clinker f-CaO prediction method of claim 1.

Images