A method for predicting and controlling the quality of calcining kiln products

Abstract

The present invention relates to the technical field of product calcination, and specifically discloses a method for predicting and controlling the quality of calcining kiln products, comprising the following steps: S1, combining the correlation analysis between various production parameters and product quality, selecting production parameters with large correlation and no extreme abnormal values ​​as model variable parameters; S2, substituting the model variable parameters into the TCS model training network, performing time series convolution feature extraction, linear output and SVM quality classification prediction of the data; S3, selecting the gas flow rate F of the lower combustion chamber LCCG and lime cooling air flow F LCA The operating parameter variables are used as the control quantity of the intelligent control actuator of the calcining kiln product quality, and the product quality fuzzy controller is constructed. The product quality working condition prediction model is calculated to predict the corresponding product quality category under the parameter working condition. The product quality category under the real-time production parameter condition is predicted through working condition identification, and the main actuators of the production process are controlled to finally achieve the purpose of adjusting product quality.

Description

Technical Field

[0001] The invention relates to the technical field of product calcination, and in particular to a method for predicting and controlling the quality of calcining kiln products. Background Art

[0002] A calcining kiln is a thermal equipment that performs heat treatment on carbon raw materials at high temperatures to improve the performance of the raw materials. The monitoring of variables in the calcining kiln production process mainly relies on the timely and accurate coordination and control of various related systems. It can make timely adjustments to faults or abnormal conditions, and obtain performance index parameters in a timely manner, so that production can run in the best state to a large extent and improve production energy efficiency.

[0003] Yunpeng Fan et al. proposed a quality-relevant kernel neighborhood preserving embedding method (QKNPE) (article source: Yunpeng Fan, Wenyou Du, Yingwei Zhang, Xiaogang Wang. Fault Detection for Multimodal Process Using Quality-Relevant Kernel Neighborhood Preserving Embedding. Mathematical Problems in Engineering. 2015, 2015: 210125. DOI: 10.1155 / 2015 / 210125). The neighborhood preserving embedding technology was introduced to improve the manifold structure and deal with multimodal nonlinear problems. At the same time, the influence of quality variables was considered, and the hidden manifold related to the characteristics of the industrial process was embedded into a low-dimensional space. The identification information of different modes of the monitored system was extracted and applied to the electric fused magnesium furnace (EFMF) process. The experimental results show the effectiveness of this method.

[0004] Chuan Ma et al. used collaborative feature regression (CFR) for abnormal condition detection of multi-view data, corresponded physical variables and image matrices to the working conditions of industrial processes, and established a multi-view data overall feature regression model that weighted sample points based on classification quality (article source: Chuan Ma, Yingwei Zhang, Zhuming Bi, Abnormal Condition Monitoring Based on Multiview Data Collaborative Feature Regression. IEEE Transactions on Instrumentation and Measurement. 2023, 72: 3517710. DOI: 10.1109 / TIM.2023.3277987). Compared with the single-view based method, this method can identify abnormal conditions in a timely and effective manner to ensure the stable operation of the industrial process. The method was applied to the detection of abnormal conditions of electric fused magnesium furnace (EFMF) and its effectiveness was verified.

[0005] In the calcining kiln production system, by using data-driven methods, predictive maintenance can predict faults and convert raw industrial data into key operational insights, thereby proactively solving faults, reducing downtime, and ensuring the service life of key production links such as calcining kiln heating and cooling systems. With the advancement of deep learning and data-driven fault diagnosis and monitoring control methods, the prediction accuracy, reliability and safety of industrial systems can be significantly improved. The integration of predictive maintenance strategies and deep learning algorithms has been effective in optimizing performance, enhancing fault prediction capabilities and improving overall operating efficiency. The use of intelligent control algorithms and computer programs to accurately calculate and predict the operation process of the equipment makes all indicators reach an optimized state and allows thermal energy to be fully utilized. On the basis of realizing computer program control of the production process, the intelligent control of the entire process such as batching, feeding, and discharging is further realized, thereby achieving an increase in product output and quality.

[0006] However, the current prediction and identification control method for calcining furnaces based on deep learning still has the following problems:

[0007] Currently, there are few prediction methods for the quality of calcining furnace products. Most of the research prediction results are parameters of certain or some intermediate links, and most furnace prediction results still need to be interpreted and judged by experts or experienced technicians. Such a working mechanism still has the problem of time lag, making it difficult to detect problems in time and regulate them, and to prevent losses in time, which affects the quality of products being produced.

[0008] Most of the more mature deep learning technologies currently used use a single network stack, but due to the limitations of a single network performance, their diagnostic accuracy and speed cannot meet the requirements to the greatest extent. There is no universally suitable and relatively fixed fusion model for different process flows and different data characteristics. Summary of the invention

[0009] In view of the deficiencies in the prior art, the present invention provides a method for predicting and controlling the quality of calcining kiln products, which solves the problem that the existing control system cannot timely obtain the product quality and perform control under the current operating parameters.

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for predicting and controlling the quality of calcining kiln products, comprising the following steps:

[0011] S1. Model variable parameter selection

[0012] By calculating the grey relative correlation of production parameters, the relative correlation between every two parameters is obtained, and the production parameters with a relative correlation greater than 0.7 with product quality are selected;

[0013] By calculating the Spearman correlation coefficient between production parameters and product quality, production parameters with a Spearman correlation coefficient greater than 0.1 are selected;

[0014] Combined with the box plot of each production parameter and product quality, the gray relative correlation and Spearman value are selected. The data are evenly distributed within the normal value range, the median of the distribution is centered and unskewed, and the production parameters without extreme outliers are used as model variable parameters;

[0015] S2. Product quality prediction

[0016] Substitute the model variable parameters into the TCS model training network to extract the temporal convolution features of the data, perform linear output, and perform SVM quality classification prediction;

[0017] S3. Fuzzy control of calcining kiln product quality

[0018] Select the gas flow rate F in the lower combustion chamber LCCG and lime cooling air flow F LCA The operating parameter variables are used as the control quantity of the intelligent control actuator of the calcining kiln product quality, and the product quality fuzzy controller is constructed. The product quality is set as e and the quality variable is ec. When there is an error in the system, the sequential control method is used to control the gas flow rate F of the lower combustion chamber. LCCG Single fuzzy controller and lime cooling air flow F LCAThe single fuzzy controller matches and controls the gas flow and cooling air flow. The adjusted operating parameters are transmitted to the product quality classifier, that is, the product quality condition prediction model for calculation, predicting the corresponding product quality category under the parameter condition and feeding back to the product quality fuzzy controller.

[0019] Preferably, the specific process of step S2 is as follows:

[0020] S21. Substitute the model variable parameters into the TCS model training network and first preprocess the data;

[0021] S22. Then substitute the preprocessed data into the time series convolution layer for time series convolution feature extraction, specifically: input sequence ,filter , predict the corresponding output , using temporal convolution, expressed as:

[0022] (1);

[0023] Where: d is the expansion coefficient, which is the factorial of 2, d=1,2,4,8,...; k is the convolution kernel size;

[0024] In the residual network ResNet structure, each residual module contains two layers of dilated convolution and nonlinear mapping ReLU. The BN operation uses weight normalization to accelerate the convergence of the network. In addition, random inactivation is used after the activation function. The residual module in the TCN architecture undergoes dilated convolution, weight normalization, activation function, and two rounds of random inactivation as the residual function F(x). When the dimensions of two adjacent layers are the same, H(x)=F(x)+x is used for calculation; when the dimensions of two adjacent layers are different, H(x)=F(x)+W is used. s x is calculated, where W s It is a convolution operation used to adjust the number of dimensions. When the lengths of input and output may be inconsistent in ResNet, the input is added to undergo a one-dimensional convolution 1×1conv, and x is used as a shortcut connection to ensure that the inner product of the elements obtains a tensor of the same size;

[0025] S23. From the hidden layer to the output layer is a multi-category logistic regression, using softmax regression, as shown below:

[0026] (2);

[0027] Among them: k represents the index of the output neuron being calculated; j is the index of the neuron in the group, with a total of J elements; Z represents an array of output neurons;

[0028] The result of the output layer is: Soft max(W2I1+b2) (3);

[0029] Among them, I1 represents the output of the hidden layer; W2 represents the weight; b2 represents the bias;

[0030] Regularization uses the ridge regression method. The objective function of the ridge regression model is It is expressed as:

[0031] (4);

[0032] Among them: x, y are the value of the feature and the label value, m, p are the number of samples and the number of features, β j is the coefficient of the feature quantity, λ is the regularization term coefficient, which controls the model complexity of the regularization term, λ≧0;

[0033] S24.SVM quality classification prediction

[0034] Assume that the training sample set , where the training sample t i Output feature value for the last layer of TCN, training sample t i The regression value f(t i ) and label value y i The maximum deviation between , that is, f(t i ) as the center, and the width of the interval band is , introduce a relaxation factor ξ, and introduce a ξ for each sample in the training set, so that the value of the function distance plus the relaxation factor is greater than or equal to 1, then some data points can be processed so that they can be on the wrong side of the segmentation surface, and the optimization goal can be kept unchanged. The distance from the sample to the hyperplane is:

[0035] (5);

[0036] Then the objective function is:

[0037] (6);

[0038] Where C is the penalty parameter, C>0;

[0039] The radial basis kernel function is used in SVM. The radial basis kernel function uses a vector as an independent variable. When the data point is farther away from the center point, the value will become smaller. The scalar is calculated using the vector distance. The radial basis kernel function is defined as follows:

[0040] (7);

[0041] in: is the squared Euclidean distance between two eigenvectors; σ is a free parameter that determines the scope of the function;

[0042] Solve the hyperplane classification problem:

[0043] For a given N training samples , assuming kernel function K, penalty parameter C (C>0) and interval band half width ε (ε>0), traverse the Lagrange multiplier and Lagrange multiplier optimal solution , minimize:

[0044] (8);

[0045] Constraints:

[0046] ;

[0047] ;

[0048] ;

[0049] SVM performs product quality classification using a one-to-one approach;

[0050] The samples are divided into three categories: I, II, and III. During training, the feature samples corresponding to I, II; I, III or II, III are selected as the training set. The classifier is trained to obtain three results. The test set is used to test the feature samples corresponding to the three results respectively, and then voting is adopted. The voting process is as follows:

[0051] I=II=III=0;

[0052] In the (Ⅰ, Ⅱ) classifier, if Ⅰ wins, then Ⅰ=Ⅰ+1; otherwise, Ⅱ=Ⅱ+1;

[0053] In the (Ⅰ, Ⅲ) classifier, if Ⅰ wins, then Ⅰ=Ⅰ+1; otherwise, Ⅲ=Ⅲ+1;

[0054] In the (Ⅱ, Ⅲ) classifier, if Ⅱ wins, then Ⅱ=Ⅱ+1; otherwise, Ⅲ=Ⅲ+1;

[0055] The result is Max(Ⅰ,Ⅱ,Ⅲ), that is, CQ0\CQ1\CQ2, where CQ0 is the superior grade, CQ1 is the first and second grade, and CQ2 is the third grade and below.

[0056] Preferably, in the step S3, the construction process of the product quality fuzzy controller is as follows:

[0057] The calcining kiln product quality is taken as the controlled object, and the controlled quantity is the gas flow rate F in the lower combustion chamber. LCCG and lime cooling air flow FLCA The actuators are the gas valve of the lower combustion chamber and the lime cooling air fan. The specific execution control quantity includes the opening degree V of the gas valve of the lower combustion chamber Q and lime cooling air fan frequency H Q ;

[0058] Assume that the product quality error is e, take its linguistic variable as E, the domain X = {-2, -1, 0}, and the fuzzy subset on the domain is I i (i=1,2,3), the corresponding language values ​​are {negative large (BP), negative medium (BM), zero (BO)}, which respectively indicate the current quality relative to the set value: poor quality, medium quality, and excellent quality;

[0059] The change in the predicted value of two samplings before and after the product quality error: ec=e t+1 -e t , take its language variable as EC, the domain Y={-3,-2,-1,0}, and the fuzzy subset Q on the domain j (j=1,2,...,4), the corresponding language values ​​are {negative large (OB), negative medium (OM), negative small (OS), zero (Z)}, which respectively indicate that the current quality variable ec is: fast improvement, medium improvement, slow improvement, and unchanged.

[0060] Preferably, the lower combustion chamber gas valve opening V Q The control is as follows:

[0061] Take its language variable as U1, the domain Z1={-3,-2,-1,0,1,2,3}, and the fuzzy subset on the domain is , the corresponding language values ​​are {negative big (MB), negative middle (MM), negative small (MS), zero (O), positive small (PS), positive middle (PM), positive big (PB)}, which respectively represent the control actuator actions as follows: V Q The opening decreases greatly, V Q Opening reduction amount, V Q The opening decreases slightly, V Q Opening unchanged, V Q The opening increase is small, V Q Opening increase, V Q The opening increases significantly.

[0062] Preferably, the lime cooling air fan frequency H Q The control is as follows:

[0063] Take its language variable as U2, the domain Z2={-3,-2,-1,0,1,2,3}, and the fuzzy subset on the domain is , the corresponding language values ​​are {negative big (MB), negative middle (MM), negative small (MS), zero (O), positive small (PS), positive middle (PM), positive big (PB)}, which respectively represent the control actuator actions as follows: H Q The frequency reduction is large, H Q Frequency reduction, H Q The frequency reduction is small, H Q The frequency remains unchanged, H Q The frequency increase is small, H Q Frequency increase, H Q The frequency increase is large.

[0064] The present invention provides a method for predicting and controlling the quality of calcining kiln products, which has the following beneficial effects:

[0065] 1. The present invention selects characteristic parameter variables as a product quality working condition prediction model by combining data set variable analysis, constructs a TCS model training network, predicts the corresponding quality category of production operation parameters, and constructs a fuzzy intelligent control method for product quality prediction. It predicts the product quality category under real-time production parameter conditions through working condition identification, controls the main actuators of the production process, and ultimately achieves the purpose of regulating product quality. It overcomes the defect that the existing control system cannot timely obtain product quality and control under the current operating parameters. It aims to improve non-optimal quality conditions and optimize on the basis of ensuring the stable operation of the production process, thereby improving product quality and improving production quality and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a product quality prediction and control flow chart of the present invention;

[0067] Figure 2 It is the product quality fuzzy control flow chart of the present invention;

[0068] Figure 3 It is a curve diagram of the accuracy change of different model quality prediction experiments of the present invention, wherein a is the change of LSTM model quality prediction accuracy, b is the change of GRU model quality prediction accuracy, c is the change of CNN model quality prediction accuracy, and d is the change of TCN model quality prediction accuracy;

[0069] Figure 4 This is a schematic diagram of comparing the accuracy of the training set and the test set of the quality prediction experiment of different models of the present invention;

[0070] Figure 5 It is a graph showing the change in loss of quality prediction experiments of different models of the present invention, wherein a is the quality prediction loss curve of the LSTM model, b is the quality prediction loss curve of the GRU model, c is the quality prediction loss curve of the CNN model, and d is the quality prediction loss curve of the TCN model;

[0071] Figure 6 Schematic diagram of the comparison of the loss of the training set and the test set in the quality prediction experiment of different models of the present invention. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example

[0073] like Figure 1-6 As shown, an embodiment of the present invention provides a method for predicting and controlling the quality of calcining kiln products, comprising the following steps:

[0074] S1. Model variable parameter selection

[0075] By calculating the grey relative correlation of production parameters, the relative correlation between every two parameters is obtained, and the production parameters with a relative correlation greater than 0.7 with product quality are selected;

[0076] By calculating the Spearman correlation coefficient between production parameters and product quality, production parameters with a Spearman correlation coefficient greater than 0.1 are selected;

[0077] Combined with the box plot of each production parameter and product quality, the gray relative correlation and Spearman value are selected. The data are evenly distributed within the normal value range, the median of the distribution is centered and unskewed, and the production parameters without extreme outliers are used as model variable parameters;

[0078] S2. Product quality prediction

[0079] Substitute the model variable parameters into the TCS model training network to extract the temporal convolution features of the data, perform linear output, and perform SVM quality classification prediction;

[0080] The steps for training the TCS model network for calcining kiln product quality prediction are as follows:

[0081] Output: Quality category CQ0 / CQ1 / CQ2;

[0082] Step 1. Preprocess the dataset

[0083] Input system characteristic behavior sequence Perform standard normalization MinMaxScaler processing (feature range = [0,1]);

[0084] Complete the dimensions and add the third dimension;

[0085] Divide the training set, validation set, and test set according to the proportions in the ratio, and the sum of the proportions is 1;

[0086] Define training set, validation set and test set;

[0087] Save the dataset;

[0088] Load data and reshape and organize the data;

[0089] Define the input array.

[0090] Step 2. Temporal Convolutional Layer

[0091] First layer: dilated convolution (dilation factor = 1);

[0092] Weight normalization;

[0093] The activation function uses the linear rectification function ReLU;

[0094] Random dropout ratio drop_rate=0.2;

[0095] (Two rounds)

[0096] If the input and output lengths are inconsistent;

[0097] Add input one-dimensional convolution 1×1conv;

[0098] Second layer: The network structure is the same as the first layer (expansion coefficient = 2);

[0099] The third layer: The network structure is the same as the first layer (expansion coefficient = 4);

[0100] Fourth layer: The network structure is the same as the first layer (expansion coefficient = 8).

[0101] Step 3. Linear output layer: activation function Softmax;

[0102] Regularized Ridge Regression L2.

[0103] Step 4. Define the training method for training;

[0104] The loss function uses the sparse categorical cross entropy function;

[0105] Gradient optimization algorithm Adam;

[0106] Monitor 'val_loss'.

[0107] Step 5. Extract feature values ​​from the last layer of TCN output.

[0108] Step 6. Load the data and reshape into 2 dimensions.

[0109] Step 7. SVM quality classification prediction.

[0110] OvO Classification:

[0111] The samples were divided into three categories: I, II, and III;

[0112] I=II=III=0;

[0113] In the (Ⅰ, Ⅱ) classifier, if Ⅰ wins, then Ⅰ=Ⅰ+1; otherwise, Ⅱ=Ⅱ+1;

[0114] In the (Ⅰ, Ⅲ) classifier, if Ⅰ wins, then Ⅰ=Ⅰ+1; otherwise, Ⅲ=Ⅲ+1;

[0115] In the (Ⅱ, Ⅲ) classifier, if Ⅱ wins, then Ⅱ=Ⅱ+1; otherwise, Ⅲ=Ⅲ+1;

[0116] The result is Max(Ⅰ,Ⅱ,Ⅲ);

[0117] Kernel function RBF.

[0118] Step 8. Accuracy calculation;

[0119] Training classification accuracy calculation;

[0120] Verify the classification accuracy calculation;

[0121] Test classification accuracy calculation.

[0122] Step 9. Output product quality category CQ0 / CQ1 / CQ2.

[0123] The specific process of the above S2 product quality prediction is as follows:

[0124] S21. Substitute the model variable parameters into the TCS model training network and first preprocess the data;

[0125] S22. Then substitute the preprocessed data into the time series convolution layer for time series convolution feature extraction, specifically: input sequence ,filter , predict the corresponding output , using temporal convolution, expressed as:

[0126] (1);

[0127] Where: d is the expansion coefficient, which is the factorial of 2, d=1,2,4,8,...; k is the convolution kernel size;

[0128] In the residual network ResNet structure, each residual module contains two layers of dilated convolution and nonlinear mapping ReLU. The BN operation uses weight normalization to accelerate the convergence of the network. In addition, random inactivation is used after the activation function. The residual module in the TCN architecture undergoes dilated convolution, weight normalization, and two rounds of random inactivation as the residual function F(x). When the dimensions of two adjacent layers are the same, H(x)=F(x)+x is used for calculation; when the dimensions of two adjacent layers are different, H(x)=F(x)+W is used. s x is calculated, where W s It is a convolution operation used to adjust the number of dimensions. When the lengths of input and output may be inconsistent in ResNet, the input is added to undergo a one-dimensional convolution 1×1conv, and x is used as a shortcut connection to ensure that the inner product of the elements obtains a tensor of the same size;

[0129] S23. From the hidden layer to the output layer is a multi-category logistic regression, using softmax regression, as shown below:

[0130] (2);

[0131] Among them: k represents the index of the output neuron being calculated; j is the index of the neuron in the group, with a total of J elements; Z represents an array of output neurons;

[0132] The result of the output layer is: Soft max(W2I1+b2) (3);

[0133] Among them, I1 represents the output of the hidden layer; W2 represents the weight; b2 represents the bias;

[0134] Regularization uses the ridge regression method. The objective function of the ridge regression model is It is expressed as:

[0135] (4);

[0136] Among them: x, y are the value of the feature and the label value, m, p are the number of samples and the number of features, β j is the coefficient of the feature quantity, λ is the regularization term coefficient, which controls the model complexity of the regularization term, λ≧0;

[0137] S24.SVM quality classification prediction

[0138] Assume that the training sample set , training sample t i (i.e., the output feature value of the last layer of TCN) i ) and label value y i The maximum deviation between , that is, f(ti ) as the center, and the width of the interval band is , introduce a relaxation factor ξ, and introduce a ξ for each sample in the training set, so that the value of the function distance plus the relaxation factor is greater than or equal to 1, then some data points can be processed so that they can be on the wrong side of the segmentation surface, and the optimization goal can be kept unchanged. The distance from the sample to the hyperplane is:

[0139] (5);

[0140] Then the objective function is:

[0141] (6);

[0142] Where C is the penalty parameter, C>0;

[0143] The radial basis kernel function is used in SVM. The radial basis kernel function uses a vector as an independent variable. When the data point is farther away from the center point, the value will become smaller. The scalar is calculated using the vector distance. The radial basis kernel function is defined as follows:

[0144] (7);

[0145] in: is the squared Euclidean distance between two eigenvectors; σ is a free parameter that determines the scope of the function;

[0146] Solve the hyperplane classification problem:

[0147] For a given N training samples , assuming kernel function K, penalty parameter C (C>0) and interval band half width ε (ε>0), traverse the Lagrange multiplier and Lagrange multiplier optimal solution , minimize:

[0148] (8);

[0149] Constraints:

[0150] ;

[0151] ;

[0152] ;

[0153] SVM performs product quality classification using a one-to-one approach;

[0154] The samples are divided into three categories: I, II, and III. During training, the feature samples corresponding to I, II; I, III or II, III are selected as the training set. The classifier is trained to obtain three results. The test set is used to test the feature samples corresponding to the three results respectively, and then voting is adopted. The voting process is as follows:

[0155] I=II=III=0;

[0156] In the (Ⅰ, Ⅱ) classifier, if Ⅰ wins, then Ⅰ=Ⅰ+1; otherwise, Ⅱ=Ⅱ+1;

[0157] In the (Ⅰ, Ⅲ) classifier, if Ⅰ wins, then Ⅰ=Ⅰ+1; otherwise, Ⅲ=Ⅲ+1;

[0158] In the (Ⅱ, Ⅲ) classifier, if Ⅱ wins, then Ⅱ=Ⅱ+1; otherwise, Ⅲ=Ⅲ+1;

[0159] The result is Max(Ⅰ,Ⅱ,Ⅲ), that is, CQ0\CQ1\CQ2, where CQ0 is the superior grade, CQ1 is the first and second grade, and CQ2 is the third grade and below. The quality categories and indicators are shown in Table 1.

[0160] Table 1 Lime quality categories and indicators

[0161]

[0162] S3. Fuzzy control of calcining kiln product quality

[0163] Select the gas flow rate F in the lower combustion chamber LCCG and lime cooling air flow F LCA The operating parameter variables are used as the control quantity of the intelligent control actuator of the calcining kiln product quality, and the product quality fuzzy controller is constructed. The product quality is set as e and the quality variable is ec. When there is an error in the system, the sequential control method is used to control the gas flow rate F of the lower combustion chamber. LCCG Single fuzzy controller and lime cooling air flow F LCA The single fuzzy controller matches and controls the gas flow and cooling air flow, and the adjusted parameters are transmitted to the product quality classifier, that is, the product quality condition prediction model for calculation, predicting the corresponding product quality category under the parameter condition, and feeding back to the product quality fuzzy controller.

[0164] In the above step S3, the construction process of the product quality fuzzy controller is as follows:

[0165] The calcining kiln product quality is taken as the controlled object, and the controlled quantity is the gas flow rate F in the lower combustion chamber. LCCG and lime cooling air flow F LCA The actuators are the gas valve of the lower combustion chamber and the lime cooling air fan. The specific execution control quantity includes the opening degree V of the gas valve of the lower combustion chamberQ and lime cooling air fan frequency H Q ;

[0166] Assume that the product quality error is e, take its linguistic variable as E, the domain X = {-2, -1, 0}, and the fuzzy subset on the domain is I i (i=1,2,3), the corresponding language values ​​are {negative large (BP), negative medium (BM), zero (BO)}, which respectively indicate the current quality relative to the set value: poor quality, medium quality, and excellent quality;

[0167] The change in the predicted value of two samplings before and after the product quality error: ec=e t+1 -e t , take its language variable as EC, the domain Y={-3,-2,-1,0}, and the fuzzy subset Q on the domain j (j=1,2,...,4), the corresponding language values ​​are {negative large (OB), negative medium (OM), negative small (OS), zero (Z)}, which respectively indicate that the current quality variable ec is: fast improvement, medium improvement, slow improvement, and unchanged.

[0168] The opening degree V of the gas valve of the lower combustion chamber Q The control is as follows:

[0169] Take its language variable as U1, the domain Z1={-3,-2,-1,0,1,2,3}, and the fuzzy subset on the domain is , the corresponding language values ​​are {negative big (MB), negative middle (MM), negative small (MS), zero (O), positive small (PS), positive middle (PM), positive big (PB)}, which respectively represent the control actuator actions as follows: V Q The opening decreases greatly, V Q Opening reduction amount, V Q The opening decreases slightly, V Q Opening unchanged, V Q The opening increase is small, V Q Opening increase, V Q The opening increases significantly.

[0170] The above lime cooling air fan frequency H Q The control is as follows:

[0171] Take its language variable as U2, the domain Z2={-3,-2,-1,0,1,2,3}, and the fuzzy subset on the domain is , the corresponding language values ​​are {negative big (MB), negative middle (MM), negative small (MS), zero (O), positive small (PS), positive middle (PM), positive big (PB)}, which respectively represent the control actuator actions as follows: H Q The frequency reduction is large, H QFrequency reduction, H Q The frequency reduction is small, H Q The frequency remains unchanged, H Q The frequency increase is small, H Q Frequency increase, H Q The frequency increase is large.

[0172] The fuzzy control rules of the lower combustion chamber gas valve opening and the lime cooling air fan frequency are shown in Table 2 and Table 3 respectively.

[0173] Table 2 Fuzzy control rules for the opening of the lower combustion chamber gas valve U1

[0174]

[0175] Table 3 Fuzzy control rules for lime cooling air fan frequency U2

[0176]

[0177] The expansion coefficients in the TCS model training network constructed by the present invention are (1, 2, 4, 8), the convolution kernel size is 4, each residual module in the residual network contains two layers of hole convolution and nonlinear mapping ReLU, the BN operation uses weight normalization, and random inactivation is used after the activation function. When the length of the input and output may be inconsistent in ResNet, the input is added to undergo 1×1conv convolution. After the data feature extraction is performed through the TCN architecture, it enters the SVM for product quality category recognition, in which soft intervals and RBF kernel functions are used, and the final classification is performed using the OvO method.

[0178] The TCS model training network constructed by the present invention is compared with the models in the comparative literature. The CNN-SVM model performs SVM classification after extracting the data set features in the last layer of the CNN model, and the TCS model performs SVM classification after extracting the data set features in the last layer of the TCN model. The calculation accuracy and loss change curves of the CNN-SVM and TCS models are the same as those of CNN and TCN, respectively.

[0179] From the experimental results in Table 4, we can see that the models with the highest training set accuracy are CNN-SVM, CNN, TCS, TCN, LSTM and GRU, with values ​​of 96.0199%, 95.8457%, 94.6268%, 91.0952%, 58.61% and 58.24% respectively; while the models with the highest test set accuracy are TCS, TCN, CNN-SVM, CNN, LSTM and GRU, with values ​​of 87.5498%, 87.2146%, 86.5672%, 85.2239%, 56.77% and 56.76% respectively. The training set accuracy of TCS is slightly inferior to that of CNN and CNN-SVM, but the test set accuracy of TCS is the highest, and the gap between the training set and the test set accuracy of TCS is obviously smaller than that of CNN.

[0180] Table 4 Comparison of experimental results of quality condition identification of different models

[0181]

[0182] Combined with different model quality prediction calculation accuracy change curve Figure 3 It can be seen that Figure 3 In the figure, a is the change of quality prediction accuracy of LSTM model, b is the change of quality prediction accuracy of GRU model, c is the change of quality prediction accuracy of CNN model, and d is the change of quality prediction accuracy of TCN model. The LSTM and GRU models have obvious overfitting phenomenon; the accuracy of training set and test set of CNN and TCN models gradually increases and tends to be stable. In comparison, the TCN model has better stability.

[0183] from Figure 4 It can be seen that although the training set accuracy of the CNN-SVM and CNN models exceeds 95%, the test set accuracy does not reach 87%, which is quite different and the overfitting phenomenon is obvious. This is related to the fact that CNN does not capture a small number of non-optimal working conditions when extracting data set features; the training set and test set accuracy of the LSTM and GRU models are slightly different, but both are less than 60%, and the accuracy is low, indicating that these two models are not suitable for the product quality prediction of the heat carrier circulation calcining kiln; the training set and test set accuracy of the TCN and TCS models proposed in the present invention are slightly different, and the values ​​are all above 87%. In comparison, the training set and test set accuracy of the TCS model are higher than those of the TCN model.

[0184] Secondly, let’s observe the loss. From Table 4, we can see that the models with the largest loss in the training set are TCS, TCN, LSTM, GRU, CNN-SVM and CNN, with values ​​of 0.2814, 0.2814, 0.3662, 0.3843, 0.4268 and 0.4268 respectively; the models with the largest loss in the test set are TCS, TCN, GRU, LSTM, CNN-SVM and CNN, with values ​​of 0.2814, 0.2814, 0.3661, 0.3723, 0.4268 and 0.4268 respectively.

[0185] Figure 5 In the figure, a is the quality prediction loss curve of the LSTM model, b is the quality prediction loss curve of the GRU model, c is the quality prediction loss curve of the CNN model, and d is the quality prediction loss curve of the TCN model. Figure 5 and Figure 6 It can be seen that the loss of the training set and the test set of the CNN model differ greatly, and the value is also large, indicating that the model may not only be overfitting, but also encounter bottlenecks in the learning process of the model. In the experiment, the hyperparameters, model structure and data set of the CNN model were debugged in various ways, but no good performance was obtained. This shows that the CNN model used in the experiment has poor learning ability when predicting the product quality of the heat carrier circulation calcining kiln; the loss values ​​of the training set and test set of the GRU and LSTM models are both large, and have not gradually decreased, indicating that the model has overfitting; the loss of the training set and test set of the TCN model has gradually decreased and stabilized, and the difference is relatively small and the value is also small. Combined with different models to calculate the loss change curve of product quality prediction Figure 5 It can be seen that the loss difference between the training set and the test set of the CNN and CNN-SVM models is relatively large and the values ​​are large; the loss difference between the training set and the test set of the GRU and LSTM models is relatively small but the values ​​are large; the loss change between the training set and the test set of the TCN and TCS models constructed in this paper is relatively small and the values ​​are small.

[0186] From the perspective of computing speed, it can be seen from Table 4 that the models from fast to slow are TCS, TCN, GRU, LSTM, CNN-SVM and CNN, with values ​​of 121μs, 121μs, 153μs, 209μs,

[0187] 228μs and 228μs, in terms of computing speed, the TCS and TCN models are obviously more advantageous.

[0188] In summary, the present invention selects characteristic parameter variables as the product quality condition prediction model by combining data set variable analysis, constructs a TCS model training network, predicts the corresponding quality categories of production operation parameters, and constructs a fuzzy intelligent control method for product quality prediction. It predicts the product quality category under real-time production parameter conditions through condition identification, controls the main actuators of the production process, and ultimately achieves the purpose of regulating product quality. It overcomes the defect that the existing control system cannot timely know the product quality and regulate it under the current operating parameters. It aims to improve non-optimal quality conditions and optimize on the basis of ensuring the stable operation of the production process, thereby improving product quality and improving production quality and efficiency.

[0189] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting and controlling the quality of calcining kiln products, characterized in that: The following steps are involved: S1. Model variable parameter selection By calculating the grey relative correlation of production parameters, the relative correlation between every two parameters is obtained, and the production parameters with a relative correlation greater than 0.7 with product quality are selected; By calculating the Spearman correlation coefficient between production parameters and product quality, production parameters with a Spearman correlation coefficient greater than 0.1 are selected; Combined with the box plot of each production parameter and product quality, the gray relative correlation and Spearman value are selected. The data are evenly distributed within the normal value range, the median of the distribution is centered and unskewed, and the production parameters without extreme outliers are used as model variable parameters; S2. Product quality prediction Substitute the model variable parameters into the TCS model training network to extract the temporal convolution features of the data, perform linear output, and perform SVM quality classification prediction; S3. Fuzzy control of calcining kiln product quality Select the gas flow rate F in the lower combustion chamber LCCG and lime cooling air flow F LCA The operating parameter variables are used as the control quantity of the intelligent control actuator of the calcining kiln product quality, and the product quality fuzzy controller is constructed. The product quality is set as e and the quality variable is ec. When there is an error in the system, the sequential control method is used to control the gas flow rate F of the lower combustion chamber. LCCG Single fuzzy controller and lime cooling air flow F LCA The single fuzzy controller matches and controls the gas flow and cooling air flow. The adjusted operating parameters are transmitted to the product quality classifier, that is, the product quality condition prediction model for calculation, predicting the corresponding product quality category under the parameter condition, and feeding back to the product quality fuzzy controller; In the step S3, the construction process of the product quality fuzzy controller is as follows: The calcining kiln product quality is taken as the controlled object, and the controlled quantity is the gas flow rate F in the lower combustion chamber. LCCG and lime cooling air flow F LCA The actuators are the gas valve of the lower combustion chamber and the lime cooling air fan. The specific execution control quantity includes the opening degree V of the gas valve of the lower combustion chamber Q and lime cooling air fan frequency H Q ; Assume that the product quality error is e, take its linguistic variable as E, the domain X = {-2, -1, 0}, and the fuzzy subset on the domain is I i , where i=1,2,3, and the corresponding language values ​​are {negative large (BP), negative medium (BM), zero (BO)}, which respectively indicate that the current quality relative to the set value is: poor quality, medium quality, and excellent quality; The change in the predicted value of two samplings before and after the product quality error: ec=e t+1 -e t , take its language variable as EC, the domain Y={-3,-2,-1,0}, and the fuzzy subset Q on the domain j , where j=1,2,...,4, and the corresponding language values ​​are {negative large (OB), negative medium (OM), negative small (OS), zero (Z)}, which respectively indicate that the current quality variable ec is: fast optimization, medium optimization, slow optimization, and unchanged.

2. The method for predicting and controlling the quality of calcining kiln products according to claim 1, characterized in that: The specific process of step S2 is as follows: S21. Substitute the model variable parameters into the TCS model training network and first preprocess the data; S22. Then substitute the preprocessed data into the time series convolution layer for time series convolution feature extraction, specifically: input sequence ,filter , predict the corresponding output , using temporal convolution, expressed as: （1）； Where: d is the expansion coefficient, which is the factorial of 2, d=1,2,4,8,...; k is the convolution kernel size; In the residual network ResNet structure, each residual module contains two layers of dilated convolution and nonlinear mapping ReLU. The BN operation uses weight normalization to accelerate the convergence of the network. In addition, random inactivation is used after the activation function. The residual module in the TCN architecture undergoes dilated convolution, weight normalization, activation function, and two rounds of random inactivation as the residual function F(x). When the dimensions of two adjacent layers are the same, H(x)=F(x)+x is used for calculation; when the dimensions of two adjacent layers are different, H(x)=F(x)+W is used. s x is calculated, where W s It is a convolution operation used to adjust the number of dimensions. When the lengths of input and output may be inconsistent in ResNet, the input is added to undergo a one-dimensional convolution 1×1conv, and x is used as a shortcut connection to ensure that the inner product of the elements obtains a tensor of the same size; S23. From the hidden layer to the output layer is a multi-category logistic regression, using softmax regression, as shown below: （2）； Among them: k represents the index of the output neuron being calculated; j is the index of the neuron in the group, with a total of J elements; Z represents an array of output neurons; The result of the output layer is: Soft max(W2I1+b2) (3); Among them, I1 represents the output of the hidden layer; W2 represents the weight; b2 represents the bias; Regularization uses the ridge regression method. The objective function of the ridge regression model is It is expressed as: （4）； Among them: x, y are the value of the feature and the label value, m, p are the number of samples and the number of features, β j is the coefficient of the feature quantity, λ is the regularization term coefficient, which controls the model complexity of the regularization term, λ≧0; S24.SVM quality classification prediction Assume that the training sample set , where the training sample t i Output feature value for the last layer of TCN, training sample t i The regression value f(t i ) and label value y i The maximum deviation between , that is, f(t i ) as the center, and the width of the interval band is , introduce a relaxation factor ξ, and introduce a ξ for each sample in the training set, so that the value of the function distance plus the relaxation factor is greater than or equal to 1, then some data points can be processed so that they can be on the wrong side of the segmentation surface, and the optimization goal can be kept unchanged. The distance from the sample to the hyperplane is: （5）； Then the objective function is: （6）； Where C is the penalty parameter, C>0; The radial basis kernel function is used in SVM. The radial basis kernel function uses a vector as an independent variable. When the data point is farther away from the center point, the value will become smaller. The scalar is calculated using the vector distance. The radial basis kernel function is defined as follows: 　　　（7）； in: is the squared Euclidean distance between two eigenvectors; σ is a free parameter that determines the scope of the function; Solve the hyperplane classification problem: For a given N training samples , assuming kernel function K, penalty parameter C, where C>0, interval band half width ε, where ε>0, traversing Lagrange multipliers and Lagrange multiplier optimal solution , minimize: （8）； Constraints: ； ； ； SVM performs product quality classification using a one-to-one approach; The samples are divided into three categories: I, II, and III. During training, the feature samples corresponding to I, II; I, III or II, III are selected as the training set. The classifier is trained to obtain three results. The test set is used to test the feature samples corresponding to the three results respectively, and then voting is adopted. The voting process is as follows: Ⅰ=Ⅱ=Ⅲ=0； In the (Ⅰ, Ⅱ) classifier, if Ⅰ wins, then Ⅰ=Ⅰ+1; otherwise, Ⅱ=Ⅱ+1; In the (Ⅰ, Ⅲ) classifier, if Ⅰ wins, then Ⅰ=Ⅰ+1; otherwise, Ⅲ=Ⅲ+1; In the (Ⅱ, Ⅲ) classifier, if Ⅱ wins, then Ⅱ=Ⅱ+1; otherwise, Ⅲ=Ⅲ+1; The result is Max(Ⅰ,Ⅱ,Ⅲ), that is, CQ0\CQ1\CQ2, where CQ0 is the superior grade, CQ1 is the first and second grade, and CQ2 is the third grade and below.

3. A method for predicting and controlling the quality of calcining kiln products according to claim 2, characterized in that: The lower combustion chamber gas valve opening V Q The control is as follows: Take its language variable as U1, the domain Z1={-3,-2,-1,0,1,2,3}, and the fuzzy subset on the domain is F 1k , where k = 1, 2, ..., 7, and the corresponding language values ​​are {negative large (MB), negative medium (MM), negative small (MS), zero (O), positive small (PS), positive medium (PM), positive large (PB)}, which respectively represent the control actuator actions as follows: V Q The opening decreases greatly, V Q Opening reduction amount, V Q The opening decreases slightly, V Q Opening unchanged, V Q The opening increase is small, V Q Opening increase, V Q The opening increases significantly.

4. The method for predicting and controlling the quality of calcining kiln products according to claim 2, characterized in that: The lime cooling air fan frequency H Q The control is as follows: Take its language variable as U2, the domain Z2={-3,-2,-1,0,1,2,3}, and the fuzzy subset on the domain is F 2k , where k = 1, 2, ..., 7, and the corresponding language values ​​are {negative large (MB), negative medium (MM), negative small (MS), zero (O), positive small (PS), positive medium (PM), positive large (PB)}, which respectively represent the control actuator actions as follows: H Q The frequency reduction is large, H Q Frequency reduction, H Q The frequency reduction is small, H Q The frequency remains unchanged, H Q The frequency increase is small, H Q Frequency increase, H Q The frequency increase is large.

Images