A rotary kiln sintering temperature probability interval prediction method, system and device

Abstract

The present application provides a rotary kiln sintering temperature probability interval prediction method, system and device, which introduces a parallel multi-head self-attention mechanism. By fusing the multi-head self-attention feature with the context vector of the sintering temperature prediction model, more abundant information is provided for the decoder, thereby improving the understanding ability of the complex dynamic relationship in the input sequence and enhancing the modeling ability of the long-term dependence relationship. In addition, the present application also adopts a Gaussian process regression (GPR) method to predict the probability interval of the sintering temperature. GPR can infer the posterior distribution and establish a probability prediction model according to the prior distribution of the prediction and the existing data set. Compared with the deterministic method, the GPR prediction contains the uncertainty of parameter estimation, provides a probability prediction interval, and better guides the process control decision.

Description

Technical Field

[0001] This invention belongs to the field of rotary kiln industrial control technology, and in particular relates to a rotary kiln sintering temperature probability range prediction system and method. Background Technology

[0002] Rotary kilns are large-scale, high-energy-consuming thermal equipment widely used in cement, metallurgical, and other production processes. During sintering, a key issue is controlling the temperature within an ideal range, which is closely related to product quality and directly determines material and energy consumption. Sintering temperature, referring to the highest temperature in the sintering zone of the rotary kiln, is a crucial indicator of the sintering process and is of great significance for its monitoring and control. However, due to the complexity and nonlinearity of rotary kilns, direct measurement of sintering temperature is very difficult, requiring temperature probes made of high-temperature refractory materials. These probes are susceptible to wear, blockage, and interference, leading to inaccurate or unreliable measurements. In actual production, optimizing and controlling the rotary kiln sintering process is a challenging task, posing a significant challenge to the level of industrial process control. Therefore, establishing a sintering temperature prediction model is of great importance.

[0003] Existing sintering temperature prediction models primarily rely on industrial data and theoretical methods such as machine learning to construct mathematical models between easily detectable and difficult-to-detect thermal variables, thereby addressing the challenge of predicting key process or quality indicators. For example, image processing techniques combined with deep learning algorithms can be used to identify dynamic features in blurred flame images to infer the temperature state of the rotary kiln. Alternatively, RBF neural network soft-sensor modeling and biogeographic optimization algorithms can be employed to solve the modeling and optimization problems of rotary kiln pellet quality indicators.

[0004] However, the following problems still exist: On the one hand, modeling methods such as RBF are all shallow neural networks, which are difficult to effectively mine the deep structural information hidden in the rotary kiln sintering process data, resulting in limited model generalization ability. On the other hand, since the rotary kiln combustion process is a complex nonlinear dynamic system, the thermodynamic data collected by the process sensors is a multivariate time series with typical strong coupling nonlinear dynamic characteristics. Traditional static network structures cannot describe the dynamic change law of the sintering process.

[0005] Furthermore, significant uncertainties exist in actual industrial processes, with complex factors such as raw material quality and external disturbances all impacting the process. Existing spatiotemporal depth soft-sensing dynamic modeling methods can only provide deterministic predictions at a single scale and cannot quantify the uncertainties surrounding future sintering temperatures. Summary of the Invention

[0006] This invention addresses the problems existing in the prior art by providing a rotary kiln sintering temperature probability range prediction system and method.

[0007] This invention provides a method for predicting the probabilistic range of sintering temperature in a rotary kiln. It employs a rotary kiln sintering temperature prediction model based on parallel multi-head self-attention DA-LSTM-GPR, and includes the following steps:

[0008] S1: Determine the input and output of the sintering temperature prediction model;

[0009] S2: Construct a sintering temperature prediction model based on parallel multi-head self-attention DA-LSTM-GP;

[0010] S3: The sintering temperature prediction model is pre-trained using the training set data, and the parameters of the sintering temperature prediction model are adjusted and optimized.

[0011] S4: Real-time prediction of sintering temperature is performed using a pre-trained parallel multi-head self-attention DA-LSTM-GP prediction model.

[0012] Preferably, step S1 above further includes the following steps:

[0013] S11: Based on the thermal data in the rotary kiln industrial control computer database, select thermal variables and construct a dataset;

[0014] S12: Normalize the data in the dataset.

[0015] Preferably, the selected thermal variables include: coal feed rate, kiln head temperature, kiln tail temperature, main engine load, cooler load, slurry flow rate, blower flow rate, and kiln head negative pressure;

[0016] Preferably, in step S2 above, during the process of constructing the sintering temperature prediction model, LSTM with input attention mechanism, multi-head self-attention mechanism, and LSTM with time attention mechanism are introduced.

[0017] Preferably, step S2 above further includes establishing a prediction model for GPR: collecting a training set D = {(x i ,y i )|i=1,2,...,k}=(X,Y),Y=f(X)+ε where X is the output of the PMA_DA_LSTM network, Y is the observation, and noise Then we can obtain the prior distribution of the observed value Y and the joint prior distribution of the observed value Y and the predicted value y.

[0018]

[0019] Where K(X,X) is a positive definite covariance matrix, K(X,x) * ) represents the test point x * The covariance matrix between the training input X and the training input X, K(x)* ,x * Let be the covariance of the test points; the squared exponential kernel is selected as the kernel function, and the formula for the squared exponential covariance function is as follows:

[0020]

[0021] And the posterior distribution of the predicted value y.

[0022]

[0023]

[0024]

[0025] This is the prediction result of GPR, and the interval prediction result corresponding to the 95% confidence level is...

[0026] Preferably, the following evaluation index is used for step S3 above:

[0027]

[0028]

[0029]

[0030] Among them, y i y' is the actual sintering temperature. i is the predicted value, and N is the number of samples in the validation set.

[0031] Meanwhile, the present invention provides a rotary kiln sintering temperature probability range prediction system, including an input-output determination module for determining the input and output of the sintering temperature prediction model;

[0032] The model building module is used to build a sintering temperature prediction model based on parallel multi-head self-attention DA_LSTM_GPR.

[0033] The model training module is used to pre-train the sintering temperature prediction model using training set data and to adjust and optimize the parameters of the sintering temperature prediction model.

[0034] The real-time prediction module is used to make real-time predictions using a pre-trained parallel multi-head self-attention DA_LSTM_GPR sintering temperature prediction model.

[0035] The beneficial effects of this invention are:

[0036] This invention addresses the problem of low accuracy in sintering temperature detection in traditional neural networks by proposing a parallel multi-head autonomous DA-LSTM-GPR network. This network can distinguish the importance of different input features and time steps, and utilizes GPR to provide an estimate of the uncertainty of the prediction results, rather than just a single predicted value, thereby improving the accuracy and reliability of the prediction. Attached Figure Description

[0037] Figure 1 This is a probability prediction diagram of sintering temperature for the PMA_DA_LSTM_GPR method of the present invention.

[0038] Figure 2 This graph compares the test results of the sintering temperature prediction method of this invention with other methods, namely ELM, CNN, LSTM, CNN-LSTM, Seq-seq, DARNN, ​​and PMA-DA-LSTM. The dashed lines in the graph represent the true values, and the solid lines represent the predicted values.

[0039] Figure 3 This is a comparison chart of the prediction errors of the sintering temperature prediction method of the present invention with other methods, namely ELM, CNN, LSTM, CNN_LSTM, Seq_seq, DARNN, ​​and PMA_DA_LSTM.

[0040] Figure 4 This is a flowchart of the sintering temperature prediction method of the present invention. Detailed Implementation

[0041] The techniques described below can be modified in various ways and have multiple embodiments, which are described in detail below with reference to the accompanying drawings. However, this does not mean that the techniques described below are limited to the specific embodiments. It should be understood that the present invention includes all similar modifications, equivalents, and substitutions without departing from the spirit and scope of the techniques described below.

[0042] The improvement of this invention lies in the encoder section, where a parallel multi-head self-attention mechanism is introduced. This mechanism allows the model to dynamically focus on different parts of the input sequence, more comprehensively capturing the complex relationships between multiple variables. By fusing the multi-head self-attention features with the context vector of the sintering temperature prediction model, richer information is provided to the decoder, thereby improving the understanding of complex dynamic relationships in the input sequence and enhancing the ability to model long-term dependencies.

[0043] Furthermore, existing spatiotemporal depth soft-sensing dynamic modeling methods can only provide deterministic predictions at a single scale, failing to quantify the uncertainty of future sintering temperatures. In contrast, probabilistic predictions expressed as prediction intervals and probability density functions can help operators make more informed process control decisions in uncertain environments. Therefore, this invention also employs Gaussian process regression (GPR) to predict the probabilistic interval of sintering temperatures. GPR can infer the posterior distribution and establish a probabilistic prediction model based on the predicted prior distribution and existing datasets. Compared to deterministic methods, GPR predictions incorporate the uncertainty of parameter estimation, providing a probabilistic prediction interval and better guiding process control decisions.

[0044] This invention provides a method for predicting the probabilistic range of sintering temperature in a rotary kiln, specifically a method based on parallel multi-head self-attention DA-LSTM-GPR (PMA_DA_LSTM_GPR), comprising the following steps:

[0045] S1: Determine the input and output of the sintering temperature prediction model;

[0046] S2: Construct a sintering temperature prediction model based on parallel multi-head self-attention DA_LSTM_GPR;

[0047] S3: The sintering temperature prediction model is pre-trained using the training set data, and the parameters of the sintering temperature prediction model are adjusted and optimized.

[0048] S4: Real-time prediction of sintering temperature is performed using a pre-trained parallel multi-head self-attention DA_LSTM_GPR prediction model.

[0049] Specifically, in step S1: a large amount of thermal data in the rotary kiln industrial control computer database, combined with the practical experience of on-site workers, selected eight variables: coal feed rate, kiln head temperature, kiln tail temperature, main unit load, cooler load, slurry flow rate, blower flow rate, and kiln head negative pressure, with a sampling interval of 1-2 minutes.

[0050] The dataset is constructed as follows:

[0051]

[0052] Y(t)=[y(t),y(t-1),...,y(tT)] (2)

[0053] Where X(t) is a thermal variable, as shown in Table 1, y(t) is the moisture content at the outlet of the drying machine, the time interval T is the sample selection interval length, and m is the number of thermal variables.

[0054] Table 1 Thermal variables of rotary kiln

[0055]

[0056] In actual production, data acquisition systems inevitably contain random noise and outliers, affecting model performance. To improve the quality of modeling data, it is necessary to preprocess the raw data from the sintering process. The values ​​of various thermal variables differ significantly in dimension; to reduce their impact on model convergence performance, normalization is required, as shown in the following equation:

[0057]

[0058] Where x represents the value of the variable before normalization, and x′ represents the value after normalization. min x max These represent the minimum and maximum values ​​of the variable before normalization.

[0059] Furthermore, the sintering process involves numerous production state parameters and control variables, such as kiln head temperature and airflow, which are thermal variables. The sintering temperature is influenced by a multitude of parameters. Introducing redundant information can easily increase the complexity of the network structure and reduce the model's prediction accuracy and timeliness. Therefore, previously, feature selection was often necessary to remove redundant variables that had a weak impact on the output.

[0060]

[0061] Where I(X) i ,Y) and I(X) i ,X j ) represent individual features X i Mutual information and features X between target Y and target Y i With X j Mutual information between them. The mRMR was used to screen the initial features of the thermal variables, and 5-fold cross-validation was used to determine the optimal number of features. Four variables were selected as input variables: kiln head temperature, kiln tail temperature, main unit load, and blower flow rate X2, X3, X4, and X7.

[0062] In step S2, during the construction of the sintering temperature prediction model, LSTM with input attention mechanism, multi-head self-attention mechanism, and LSTM with time attention mechanism are introduced.

[0063] Among them, the LSTM with input attention mechanism: at time t, the feature vector Attention weights are calculated using a single-layer neural network.

[0064]

[0065]

[0066] Where V tT W e U e These are the training weights, input to the attention layer, and output. Input the weight coefficients of the features at time t, and then... t The input is fed into the Softmax layer, resulting in all attention weights summing to 1. This represents the weight of the k-th input feature at time t. Using these weights, sequence feature values ​​highly correlated with the quality variable can be adaptively extracted.

[0067]

[0068] The input features are

[0069]

[0070] This is a nonlinear operation of the LSTM unit, and the calculation process is as follows:

[0071]

[0072] By utilizing the proposed input attention mechanism, encoder can selectively focus on specific driving sequences, rather than treating all input driving sequences equally.

[0073] Multi-head self-attention mechanism: The multi-head attention mechanism focuses on learning the dependencies between the vectors of the input feature sequence itself, paying attention to different parts of the input, thereby better capturing the relationships between the inputs. Multi-head self-attention is calculated as follows:

[0074]

[0075] Where n represents the number of self-attention modules, D K W is a factor used to scale attention weights. i Q W i K W i V Let head represent the weight matrix of the i-th self-attention module. i This is the output of the i-th self-attention module. Finally, the outputs of all self-attention modules are concatenated and mapped to the final output vector through a linear transformation.

[0076] LSTM with temporal attention mechanism: Each encoder hidden state is assigned a temporal attention value. Then, an adaptively weighted context vector is obtained as input to the LSTM in decoding. The temporal attention weights for the hidden state at time t are calculated as follows:

[0077]

[0078]

[0079] Where d t-1 ,s′ t-1 This decodes the hidden state and cell state of the LSTM at the previous time step. Note the weights. It is the i-th encoded hidden state at time t. The importance of prediction. Because each encoded hidden state The temporal weights are mapped to the input, so the attention mechanism will map the context vector c. t Calculate for all encoded outputs The weighted sum, such as:

[0080]

[0081] Obtain the context vector c t Combining it with the target sequence yields:

[0082]

[0083] Utilizing new Update the hidden state of decode at time t:

[0084]

[0085] Where f2 represents the nonlinear operation of the LSTM unit, calculated as follows:

[0086]

[0087] Finally, calculate the predicted output.

[0088]

[0089] Where [d] T ;c T ] represents the hidden state d decoded at time T. T and context information c T V y and b v Weight matrix and bias vector.

[0090] Preferably, step S2 above further includes establishing a prediction model for GPR: collecting a training set D = {(x i ,y i )|i=1,2,...,k}=(X,Y),Y=f(X)+ε where X is the thermal variable, Y is the observed value, and noise Then we can obtain the prior distribution of the observed value Y and the joint prior distribution of the observed value Y and the predicted value y.

[0091]

[0092]

[0093] Where K(X,X) is a positive definite covariance matrix, K(X,x) * ) represents the test point x * The covariance matrix between the training input X and the training input X, K(x) * ,x * Let be the covariance of the test points; the squared exponential kernel is selected as the kernel function, and the formula for the squared exponential covariance function is as follows:

[0094]

[0095] And the posterior distribution of the predicted value y.

[0096]

[0097] This is the prediction result of GPR, and the interval prediction result corresponding to the 95% confidence level is...

[0098]

[0099] Preferably, the following evaluation index is used for step S3 above:

[0100]

[0101] Among them, y i y' is the actual sintering temperature. i is the predicted value, and N is the number of samples in the validation set.

[0102] To verify the effectiveness of the parallel multi-head attention and probabilistic prediction model, the proposed algorithm was tested using data from a thermal analyzer at the No. 2 rotary kiln of an aluminum company. The test results demonstrate the effectiveness and feasibility of the proposed method, providing operators with reliable information on production status.

[0103] Table 2. Experimental results of sintering temperature prediction using various algorithms

[0104]

[0105]

[0106] As can be seen from the table, this invention has the best prediction accuracy, with the smallest root mean square error (RMSE) and the lowest regression evaluation index (R²).2 The maximum value indicates a strong correlation between the actual and predicted values. Further, from... Figure 2-3 It can be seen that, compared with other algorithms and other models, the error between the predicted sintering temperature and the actual value obtained by the model of this invention is within a very close range, except for a few singularities. The predicted temperature change curve has a higher fitting degree with the actual temperature change curve, which further demonstrates the reliability of the method proposed in this invention.

[0107] Meanwhile, the present invention provides a rotary kiln sintering temperature probability range prediction system, including an input-output determination module for determining the input and output of the sintering temperature prediction model;

[0108] The model building module is used to build a sintering temperature prediction model based on parallel multi-head self-attention DA_LSTM_GPR.

[0109] The model training module is used to pre-train the sintering temperature prediction model using training set data and to adjust and optimize the parameters of the sintering temperature prediction model.

[0110] The real-time prediction module is used to perform real-time predictions using a prediction model based on parallel multi-head self-attention DA_LSTM_GPR sintering temperature.

[0111] The present invention also provides a rotary kiln sintering temperature probability range prediction device, including a processor and a memory;

[0112] Memory, used to store computer programs;

[0113] When a processor executes a program stored in memory, it implements the method steps as described in the above embodiments.

[0114] The present invention further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method steps of the above embodiments.

[0115] To address the challenge of accurately predicting the sintering temperature of rotary kilns, this invention proposes a parallel multi-head autonomous attention DA-LSTM-GPR method for rotary kiln sintering temperature prediction. First, preliminary feature selection of thermal variables is performed, and 5x cross-validation is used to determine the optimal number of features, reducing feature dimensionality and simplifying model complexity. Second, the weights of the input sequence are dynamically adjusted using an input attention mechanism to better capture key information in the sequence. Simultaneously, a parallel multi-head attention enhanced encoder is constructed to fully extract features and dependencies between variables. Then, a concatenation layer fuses the two parallel feature information into a deep spatial representation, which is input into a temporal attention decoder network to capture long-term temporal dependencies in the input data. This representation is then input into a fully connected layer network to predict the outlet sintering temperature. Finally, GPR is used to correct and optimize the sintering temperature prediction results of the parallel multi-head attention DA-LSTM model, thereby improving the accuracy and stability of the prediction results. Experimental results verify the feasibility and effectiveness of the proposed algorithm. It automatically assigns weights to thermal variables to accurately predict the trend of sintering temperature changes.

[0116] Although the present invention has been described in detail above with general descriptions and specific embodiments, some modifications or improvements can be made to it. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Other changes and modifications made by those skilled in the art without departing from the spirit and scope of the present invention are still included within the scope of protection of the present invention.

Claims

1. A method for predicting the probability interval of sintering temperature in a rotary kiln, characterized in that: A rotary kiln sintering temperature prediction model based on parallel multi-head self-attention DA-LSTM-GPR (PMA_DA_LSTM_GPR) is adopted, including the following steps: S1: Determine the input and output of the sintering temperature prediction model; S2: Construct a sintering temperature prediction model based on parallel multi-head self-attention DA-LSTM-GPR; S3: The sintering temperature prediction model is pre-trained using the training set data, and the parameters of the sintering temperature prediction model are adjusted and optimized. S4: Real-time prediction of sintering temperature is performed using a pre-trained parallel multi-head self-attention DA-LSTM-GPR prediction model. Step S1 further includes preliminary feature screening of thermal variables using mRMR, and determining the optimal number of features using 5-fold cross-validation, selecting four variables as input variables: kiln head temperature, kiln tail temperature, main unit load, and blower flow rate. ； in and Each represents a single feature Mutual information and features between the target Y and the target Y and Mutual information between them; In step S2, during the construction of the sintering temperature prediction model, LSTM with input attention mechanism, multi-head self-attention mechanism, and time attention mechanism are used simultaneously. Step S2 also includes building a prediction model for GPR: collecting a training set. ; ； Where X is the output of the PMA_DA_LSTM network, Y is the observation, and noise ε: Then we obtain the prior distribution of the observed value Y and the joint prior distribution of the observed value Y and the predicted value y. ； ； Where K(X,X) is a positive definite covariance matrix. Indicates test point The covariance matrix between the training input X and the training input X Let be the covariance of the test points; a squared exponential kernel is chosen as the kernel function, and the formula for the squared exponential covariance function is as follows: ； And the posterior distribution of the predicted value y: ； ； ； These are the GPR prediction results, and the corresponding interval prediction results at the 95% confidence level are:

2. The method for predicting the probability range of rotary kiln sintering temperature according to claim 1, characterized in that, Step S1 further includes the following steps: S11: Based on the thermal data in the rotary kiln industrial control computer database, select thermal variables and construct a dataset; S12: Normalize the data in the dataset.

3. The method for predicting the probability range of rotary kiln sintering temperature according to claim 1, characterized in that, The selected thermal variables include: coal feed rate, kiln head temperature, kiln tail temperature, main engine load, cooler load, slurry flow rate, blower flow rate, and kiln head negative pressure.

4. The method for predicting the probability interval of rotary kiln sintering temperature according to claim 1, characterized in that, The following evaluation metrics will be used for step S3: ； ； ； in, This is the actual sintering temperature. is the predicted value, and N is the number of samples in the validation set.

5. The method for predicting the probability range of rotary kiln sintering temperature according to claim 1, characterized in that, By dynamically adjusting the weights of the input sequence using the input attention mechanism, a parallel multi-head attention-enhanced encoder is constructed to extract features and dependencies between variables. The two parallel feature information are fused into a deep spatial representation through a concatenation layer and then input into a temporal attention-based decode network to capture the long-term temporal dependency features of the input data.

6. A rotary kiln sintering temperature probability interval prediction system, used to implement the rotary kiln sintering temperature probability interval prediction method according to any one of claims 1-5, characterized in that, include: The input / output determination module is used to determine the input and output of the sintering temperature prediction model. The model building module is used to build a sintering temperature prediction model based on parallel multi-head self-attention DA_LSTM_GPR. The model training module is used to pre-train the sintering temperature prediction model using training set data and to adjust and optimize the parameters of the sintering temperature prediction model. The real-time prediction module is used to make real-time predictions using a pre-trained parallel multi-head self-attention DA_LSTM_GPR sintering temperature prediction model.

7. A rotary kiln sintering temperature probability range prediction device, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the rotary kiln sintering temperature probability interval prediction method according to any one of claims 1-5.

Images