Method and system for predicting internal temperature of cement rotary kiln and thickness of kiln skin kiln brick

Abstract

This invention provides a method and system for predicting the internal temperature and kiln lining / brick thickness of a cement rotary kiln. The method includes: collecting external surface temperature data using an infrared cylinder scanner, inputting it into an RKT-Net model containing ResNet and Transformer structures to obtain the axial temperature curve of the internal surface; and inputting two sets of temperature data based on preset time intervals into an RKH-Net model containing dual convolutional channels and a CBAM attention mechanism to output the kiln lining and brick thickness distribution matrix. The two models are trained respectively using a training set generated from a heat transfer mechanism model, a voxel dataset, and a global simulated heat conduction database to ensure prediction accuracy. This invention combines deep learning and simulation technology to solve the problems of underconstraint and data scarcity, achieving accurate predictions and possessing strong industrial applicability.

Description

Technical Field

[0001] This invention relates to the field of cement rotary kiln monitoring technology, specifically to a method and system for predicting the internal temperature and kiln lining / brick thickness of a cement rotary kiln. Background Technology

[0002] The rotary kiln is a crucial piece of equipment in cement production. Its internal temperature distribution, kiln lining thickness, and kiln brick thickness directly determine cement production efficiency, product quality, and equipment operational safety. Kiln lining shedding and kiln brick wear can lead to overheating and damage to the kiln body. Since the internal temperature cannot be directly measured, accurately predicting these parameters is essential for the safe and stable operation of the rotary kiln.

[0003] In existing technologies, internal temperature is often estimated by combining infrared detection of external surface temperature with simple empirical models, resulting in low accuracy and weak anti-interference capabilities. Prediction of kiln lining and brick thickness relies heavily on manual inspections or single sensor data, leading to under-constraints and preventing precise quantification. Furthermore, real-world temperature-thickness pairing data from industrial sites is scarce, traditional deep learning models suffer from insufficient training data and poor generalization. Moreover, existing models do not specifically enhance the correlation between temperature and thickness, further reducing prediction accuracy and failing to meet the actual monitoring needs of industry. Summary of the Invention

[0004] This invention provides a method for predicting the internal temperature of a cement rotary kiln, comprising: collecting cylindrical scanning data of the outer surface temperature of the cement rotary kiln using an infrared cylindrical scanner; averaging the preprocessed outer surface temperature cylindrical scanning data along the axial direction to obtain an axial average temperature curve of the outer surface; inputting the axial average temperature curve of the outer surface into an RKT-Net model to obtain an axial temperature curve of the inner surface; wherein, the RKT-Net model is constructed based on training data and test data, and the training data and the test data are paired samples consisting of the preprocessed axial average temperature curve of the outer surface and the axial temperature curve of the inner surface generated by the heat transfer mechanism model.

[0005] According to one embodiment of the present invention, the method further includes constructing a heat transfer mechanism model, specifically including: constructing a heat transfer mechanism model based on classical heat and mass transfer theory and combined with process parameters under normal operating conditions of a cement rotary kiln using mathematical modeling; the process parameters include at least one of the following: kiln inner and outer diameter, kiln length, kiln rotation speed, fuel calorific value, kiln air mass flow rate, kiln tail material mass flow rate, kiln tail material initial temperature, kiln tail gas temperature, thermal conductivity of the kiln metal shell, thermal conductivity of refractory bricks, and thermal conductivity of the kiln skin; the cement rotary kiln is divided into segments according to a preset length, and a multi-field coupled energy equation set covering radiative heat transfer, convective heat transfer, conductive heat transfer, combustion reaction heat, and chemical reaction heat is established for each segment; by solving the energy equation set for each segment in sequence, the inner wall temperature of each segment is obtained, and the integrated result outputs an axial temperature curve of the inner surface covering the entire kiln axis.

[0006] According to one embodiment of the present invention, the method further includes constructing a training dataset for the heat transfer mechanism model, specifically including: fitting and downsampling the axial temperature curve of the inner surface, and downsampling the axial average temperature curve of the outer surface to unify their dimensions to a preset length; using the dimension-unified axial average temperature curve of the outer surface and the corresponding axial temperature curve of the inner surface as a set of training samples, and generating a rotary kiln internal temperature prediction training set in batches based on infrared cylinder scanning data over continuous time periods; and dividing the prediction training set into a training subset and a prediction subset according to a preset ratio.

[0007] According to one embodiment of the present invention, the method further includes constructing an RKT-Net model, specifically comprising: the RKT-Net model including a ResNet residual structure and a Transformer structure; the ResNet residual structure comprising a preset number of groups of residual units, each group of residual units being configured with a main path and a residual path; the main path sequentially deploying a convolutional layer, a normalization layer, an activation function, another convolutional layer, and a normalization layer, with the two normalization layers respectively cooperating with the corresponding convolutional layers to complete feature standardization, thereby improving the stability of feature extraction; the residual path directly passes the input features to the end of the main path, and performs element-wise addition with the features processed by the main path to achieve fusion, and then outputs the results after passing through an activation function; multiple groups of residual units work together to progressively extract the outer surface temperature curve. The local features are mapped to the input dimension sequentially from the third preset dimension to 64, 128, and 256 dimensions. The Transformer structure has a first preset number of attention heads, a second preset number of network layers, and a third preset number of feature dimensions. It is used to construct the global dependency relationship of temperature features, receive the 256-dimensional feature vector output by the ResNet residual structure, and perform global feature fusion and decoding. The RKT-Net model also includes a fully connected layer, which is used to restore the feature vector output by the Transformer structure to a preset length, consistent with the dimension of the inner surface axial temperature curve. The RKT-Net model takes the third preset dimension of the outer surface axial average temperature curve as input and outputs the inner surface axial temperature curve.

[0008] According to one embodiment of the present invention, the training of the RKT-Net model further includes: conducting supervised training using a training subset of the prediction training set, wherein the prediction subset of the prediction training set is used to verify the training accuracy of the RKT-Net model; performing normalization processing on the input data of the training subset and the prediction subset; performing restoration processing on the output data of the RKT-Net model; setting the adaptive learning rate to a first preset learning rate, adopting the ReduceLROnPlateau strategy, setting the factor to a first preset factor, setting the patience to a first preset patience, setting the batch size to a first preset batch, setting the maximum number of iterations to a first preset number, confirming that the validation set loss has not decreased for a consecutive first preset number of rounds, stopping the training, and retaining the best network weights; after training is completed, using the prediction subset to verify the accuracy, which requires that the average absolute value between the predicted temperature inside the cement rotary kiln and the true value is less than a preset temperature, and the degree of fit between the predicted temperature inside the cement rotary kiln and the true value is greater than a preset fitting value.

[0009] According to one embodiment of the present invention, the method further includes preprocessing the outer surface cylindrical scan data, specifically including: splitting the outer surface temperature cylindrical scan data according to the sampling time to obtain an outer surface temperature field matrix of a first preset dimension; verifying the data integrity through thermal visualization and removing abnormal missing data; dynamically adjusting the sampling frequency of the outer surface temperature cylindrical scan data according to the rotary kiln rotation speed to ensure that the spatial overlap rate of adjacent sampling points is greater than a preset percentage; correcting the sampling start position deviation of each outer surface temperature field matrix based on the rotation period of the rotary kiln, so that the row index of all the outer surface temperature field matrices is aligned with the axial physical position of the rotary kiln; averaging the temperature field matrices along the axial direction after alignment to obtain the original outer surface axial temperature curve of a second preset dimension; and performing a downsampling operation on the original outer surface axial temperature curve to unify it into an outer surface axial average temperature curve of a third preset dimension.

[0010] This invention also provides a method for predicting the internal temperature of a cement rotary kiln according to an embodiment of the invention, which includes: collecting two sets of operating data of the cement rotary kiln based on a preset time interval, each set of data including the external surface temperature field matrix generated by preprocessing the external surface cylindrical sweep data, and the corresponding internal surface axial temperature curve output by the RKT-Net model; inputting the data pairs consisting of the two sets of external surface temperature field matrices and the corresponding internal surface axial temperature curves into the RKT-Net model, and outputting the kiln lining thickness distribution matrix and the kiln brick thickness distribution matrix along the entire kiln axis of the cement rotary kiln; wherein, the RKT-Net model is trained using a quadruple consisting of the external surface temperature field matrix, the corresponding internal surface axial temperature curve, the corresponding kiln lining thickness distribution matrix, and the corresponding kiln brick thickness distribution matrix at two different time points as samples, and constructing constraints based on the two sets of temperature data at different time points to solve the under-constraint problem in predicting the kiln lining and kiln brick thickness.

[0011] According to one embodiment of the present invention, the method further includes constructing an RKH-Net model, specifically comprising: the RKH-Net model including a dual convolutional input channel, an Encoder module, a Transformer module, a multiple Decoder module, and a CBAM attention mechanism; the dual convolutional input channel is used to extract local features of two sets of temperature data respectively and superimpose them; the CBAM attention mechanism strengthens the response of key temperature regions where the changes in kiln skin and kiln brick thickness are strongly correlated; the RKH-Net model inputs two sets of data pairs consisting of an outer surface temperature field matrix and an inner surface axial temperature curve at different time points, and outputs a kiln skin thickness distribution matrix and a kiln brick thickness distribution matrix; the output results must satisfy that the average absolute value between the predicted and actual values ​​of the kiln skin thickness is less than a first preset thickness, and the average absolute value between the predicted and actual values ​​of the kiln brick thickness is less than a second preset thickness.

[0012] According to one embodiment of the present invention, the training of the RKH-Net model is further included, specifically comprising: constructing a training set for the RKH-Net model based on a voxel simulation heat conduction database and a global simulation heat conduction database; wherein the samples of the training set are quadruples consisting of two sets of external surface temperature field matrices, corresponding internal surface axial temperature curves, corresponding kiln lining thickness distribution matrices, and corresponding kiln brick thickness distribution matrices at two different time points; dividing the training set into a training subset and a prediction subset according to a preset ratio; performing normalization processing on the external surface temperature field matrices and the internal surface axial temperature curves of the training subset and the prediction subset; and using the Adam optimizer, setting the adaptive learning rate to a second preset learning rate, and using Redu... The ceLROnPlateau strategy is used, with the factor set to the second preset factor, patience set to the second preset patience, batch size set to the second preset batch, and maximum iteration count set to the second preset number. The validation set loss is continuously monitored, and training is stopped if the validation set loss does not decrease for the second preset number of consecutive iterations, while retaining the optimal network weights. The output data of the RKH-Net model is then processed to map the abstract numerical values ​​output by the model to kiln lining thickness and kiln brick thickness values ​​with actual physical meaning. After training, the accuracy is verified using the predicted subset, requiring that the average absolute value between the predicted and actual kiln lining thickness values ​​is less than the first preset thickness, and the average absolute value between the predicted and actual kiln brick thickness values ​​is less than the second preset thickness.

[0013] This invention also provides a system for predicting the internal temperature and kiln lining / brick thickness of a cement rotary kiln, comprising: an infrared data acquisition module for acquiring cylinder scan data of the external surface temperature of the cement rotary kiln using an infrared cylinder scanner, and transmitting the acquired cylinder scan data to a data preprocessing module; and a data preprocessing module for performing preprocessing operations on the external surface temperature cylinder scan data, including splitting, removing abnormal missing data, dynamically adjusting the sampling frequency, correcting the sampling start position deviation, calculating the axial mean, and downsampling, generating an external surface axial average temperature curve with a third preset dimension, and transmitting the external surface axial average temperature curve to RKT-Net. The model module divides the cement rotary kiln into segments of a preset length and establishes a multi-field coupled energy equation system covering radiative heat transfer, convective heat transfer, conductive heat transfer, combustion reaction heat, and chemical reaction heat. After solving the equations, it integrates and outputs the axial temperature curve of the inner surface of the entire kiln along the entire axis, which is then transmitted to the data storage module. The RKT-Net model module is used to input the average axial temperature curve of the outer surface and output the axial temperature curve of the inner surface. The RKT-Net model module has a built-in RKT-Net model containing a ResNet residual structure, a Transformer structure, and a fully connected layer. It uses the rotary kiln's internal temperature prediction training set and... The training process includes normalization, restoration processing, the ReduceLROnPlateau learning rate strategy, and an early stopping mechanism. The training accuracy meets the requirement that the average absolute value between the predicted temperature inside the cement rotary kiln and the actual temperature is less than a preset temperature, and the degree of fit is greater than a preset fitted value. The data storage module stores the cylindrical sweep data of the outer surface temperature, the data preprocessed by the data preprocessing module, the training set of the RKT-Net model module, and the weights of the trained model. The result output module receives the axial temperature curve of the inner surface output by the RKT-Net model module and converts it into a visual chart or numerical report.

[0014] This invention achieves a significant technological breakthrough in the prediction of key parameters for cement rotary kilns by deeply integrating deep learning and simulation technologies. In temperature prediction, the RKT-Net model, through the synergistic effect of the ResNet residual structure and the Transformer structure, efficiently extracts local features of the outer surface temperature curve and constructs global dependencies. Combined with multi-step precise data preprocessing, it significantly improves the accuracy and reliability of inner surface temperature prediction, ensuring that the absolute average value and fitting degree of the prediction results meet stringent industrial-grade requirements. In thickness prediction, the dual-convolutional input channels of the RKH-Net model can fully exploit the correlation features between two sets of temperature data at different time points. The CBAM attention mechanism further strengthens the temperature region response closely related to changes in kiln lining and kiln brick thickness, effectively solving the under-constraint problem faced by traditional methods and achieving accurate quantitative output of the axial thickness distribution of the entire kiln.

[0015] Addressing the industry pain point of scarce on-site industrial measurement data, this invention generates temperature prediction training data through a heat transfer mechanism model. It then constructs the four-tuple samples required for thickness prediction based on a voxel and overall simulated heat conduction database, completely resolving the poor generalization problem caused by insufficient training data in traditional deep learning models. The system adopts a modular design; the infrared data acquisition module and preprocessing workflow can flexibly adapt to the dynamic operating conditions of the rotary kiln. Data quality is ensured through dynamic adjustments to the sampling frequency and correction of sampling start-point deviations. The output of visual charts and numerical reports facilitates maintenance personnel's rapid understanding of equipment status, providing strong support for the safe and stable operation of the rotary kiln.

[0016] Both models employ scientific training strategies. Normalization and restoration processes ensure data dimensionality consistency and physical meaning restoration. The ReduceLROnPlateau learning rate adjustment strategy, combined with an early stopping mechanism, guarantees both the convergence speed of model training and improves long-term stability and anti-interference capabilities. The overall solution does not rely on complex on-site modifications or additional sensor deployments, and can be quickly integrated into existing cement production processes. While improving production efficiency and ensuring product quality, it reduces the risk of kiln overheating damage, demonstrating strong industrial applicability and promotional value. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the method for predicting the internal temperature of a cement rotary kiln provided by the present invention.

[0019] Figure 2 This is a schematic diagram of the process for constructing a heat transfer mechanism model provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the process for constructing the training dataset for the heat transfer mechanism model provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the process for constructing the RKT-Net model provided by the present invention.

[0022] Figure 5 This is a schematic diagram of the training process for the RKT-Net model provided by the present invention.

[0023] Figure 6 This is a schematic diagram of the process for preprocessing external surface cylindrical scan data provided by the present invention.

[0024] Figure 7 This is a flowchart illustrating the method for predicting the thickness of kiln lining and kiln bricks in a cement rotary kiln provided by the present invention.

[0025] Figure 8 This is a schematic diagram of the process for constructing the RKH-Net model provided by the present invention.

[0026] Figure 9 This is a schematic diagram of the training process for the RKH-Net model provided by the present invention.

[0027] Figure 10 This is a block diagram of the prediction system for the internal temperature and kiln lining / brick thickness of a cement rotary kiln provided by the present invention.

[0028] Figure label: 100: Cement rotary kiln internal temperature and kiln lining / brick thickness prediction system; 110: Infrared data acquisition module; 120: Data preprocessing module; 130: RKT-Net model module; 140: RKH-Net model module; 150: Data storage module; 160: Result output module. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] The following is combined with Figures 1 to 10 The present invention describes a method for predicting the internal temperature of a cement rotary kiln, a method for predicting the thickness of the kiln lining and kiln bricks, and a system for predicting the internal temperature and the thickness of the kiln lining and kiln bricks of a cement rotary kiln.

[0031] Figure 1 This is a schematic flowchart of the method for predicting the internal temperature of a cement rotary kiln provided by the present invention. Figure 1 As shown, the present invention provides a method for predicting the internal temperature of a cement rotary kiln, comprising: in step S100, collecting cylindrical scanning data of the outer surface temperature of the cement rotary kiln using an infrared cylindrical scanner, and averaging the preprocessed outer surface cylindrical scanning data along the axial direction to obtain the outer surface axial average temperature curve; in step S200, inputting the outer surface axial average temperature curve into an RKT-Net model to obtain the inner surface axial temperature curve; wherein, the RKT-Net model is constructed based on training data and test data, and the training data and test data are paired samples consisting of the preprocessed outer surface axial average temperature curve and the inner surface axial temperature curve generated by the heat transfer mechanism model.

[0032] Specifically, step S100 requires the accurate acquisition and standardized preprocessing of external surface temperature data to provide a high-quality, highly consistent data foundation for model input. During the data acquisition phase, an infrared cylinder scanner is used to continuously monitor the temperature of the cement rotary kiln's external surface. The scanner's sampling frequency is fixed at 10Hz. This frequency accurately captures dynamic fluctuations in surface temperature caused by kiln skin crust formation and localized wear, while avoiding data redundancy and storage pressure caused by high-frequency sampling. In actual operation, the sampling frequency needs to be dynamically fine-tuned according to the real-time rotational speed of the rotary kiln (0.5~4 rpm). The requirement is to ensure that the spatial overlap between adjacent sampling points is not less than 50%, thereby guaranteeing the continuity and integrity of temperature data along the kiln's axial direction.

[0033] After data acquisition, the preprocessing stage begins. The first step involves splitting the raw cylindrical scan data into multiple independent external surface temperature field matrices based on the sampling time. In this embodiment, all matrices are uniformly 200*100 pixels in size. Thermal visualization allows for intuitive verification of the data integrity of each matrix group, promptly removing invalid data with abnormal missing values ​​or abrupt changes. The second step addresses the issue of inconsistent starting positions of the temperature field matrices at different sampling times by performing position alignment. The alignment method is achieved using the following formula: in, This represents the vertical resolution of the temperature field matrix, which is set to 100 in this embodiment. For the first The sampling start time of the outer surface temperature field matrix; The rotation cycle of the rotary kiln; and These represent the row indices of the matrix before and after alignment, respectively. This calculation ensures that the row indices of all temperature field matrices are precisely matched with the axial physical positions of the kiln body, eliminating feature misalignment caused by initial sampling bias. The third step calculates the mean along the axial direction of the aligned temperature field matrix, obtaining an original axial temperature curve of the outer surface with a length of 200. The fourth step performs a downsampling operation on this original curve, uniformly scaling its dimension to 50, ultimately forming the average axial temperature curve of the outer surface that meets the input requirements of the RKT-Net model.

[0034] The construction of the RKT-Net model requires the generation of training and testing data based on a heat transfer mechanism model and measured data. Both training and testing data consist of paired samples comprised of a pre-processed 50-dimensional axial average temperature curve of the outer surface and an axial temperature curve of the inner surface generated by the heat transfer mechanism model. The inner surface axial temperature curve is obtained by solving a multi-field coupled energy equation system from the heat transfer mechanism model. Its original dimension was 79, which was smoothed by polynomial fitting and then downsampled and scaled to 50 dimensions to maintain consistency with the outer surface curve's dimension. Based on 843 sets of valid infrared cylinder scan data from four consecutive months, the above paired samples were generated in batches to form a complete dataset, which was then divided into a training subset and a testing subset in a 7:3 ratio for supervised training and accuracy verification of the RKT-Net model, respectively.

[0035] In step S200, the preprocessed 50-dimensional outer surface axial average temperature curve is input into the trained RKT-Net model. The model operation must follow fixed mathematical logic and parameter configurations. First, the input curve is normalized using the formula... Temperature data is mapped to a reasonable range to eliminate the interference of numerical range differences on model training performance. The data then enters the model's feature extraction and mapping stage. The forward propagation process is as follows: first, local temperature features are extracted step-by-step through three sets of one-dimensional residual structures, mapping the input 50-dimensional vector sequentially to 64-dimensional and 128-dimensional vectors, ultimately outputting a 256-dimensional feature vector. This feature vector is then fed into the Transformer module, which has 8 attention heads, 5 network layers, and 512 feature dimensions. Through a self-attention mechanism, the temperature correlation between any two positions in the sequence is directly calculated, completing global feature fusion and decoding, while maintaining the 256-dimensional output. Finally, a fully connected layer restores the 256-dimensional feature vector to 50 dimensions, and then... (The sentence is incomplete and requires further context to translate accurately.) The output data is restored to obtain an internal surface axial temperature curve with practical physical meaning. This curve has a length of 50, which is consistent with the dimension of the external surface axial average temperature curve. It can fully reflect the internal temperature distribution of the entire kiln along the axis, providing key data support for subsequent prediction of kiln lining and kiln brick thickness.

[0036] Figure 2 This is a schematic diagram of the process for constructing a heat transfer mechanism model provided by the present invention. For example... Figure 2As shown, according to an embodiment of the present invention, the implementation further includes constructing a heat transfer mechanism model, specifically including: in step S210, based on classical heat and mass transfer theory and combined with process parameters under normal operating conditions of the cement rotary kiln, a heat transfer mechanism model is constructed using mathematical modeling; the process parameters include at least one of the following: kiln inner and outer diameter, kiln length, kiln rotation speed, fuel calorific value, kiln air mass flow rate, kiln tail material mass flow rate, kiln tail material initial temperature, kiln tail gas temperature, kiln metal shell thermal conductivity, refractory brick thermal conductivity, and kiln skin thermal conductivity; in step S220, the cement rotary kiln is divided into segments according to a preset length, and a multi-field coupled energy equation set covering radiative heat transfer, convective heat transfer, conductive heat transfer, combustion reaction heat, and chemical reaction heat is established for each segment; in step S230, by solving the energy equation set of each segment in sequence, the kiln inner wall temperature of each segment is obtained, and after integration, an inner surface axial temperature curve covering the entire kiln axis is output.

[0037] Specifically, step S210 is based on classical heat and mass transfer theory, combined with key process parameters under normal operating conditions of a cement rotary kiln, and a heat transfer mechanism model is constructed using mathematical modeling. The selected process parameters are all derived from actual data directly obtainable from the industrial site, including a kiln inner diameter of 4.8m, an outer diameter of 5.6m, a total kiln length of 79m, a kiln rotation speed range of 0.5 to 4 rpm, a fuel calorific value of 29308 kJ / kg, and an air mass flow rate of 5000–8000 m³ / kg. 3 The kiln tail inflow mass flow rate is 150–250 t / h, the initial temperature of the kiln tail inflow material is 20–80℃, the kiln tail gas temperature is 800–1100℃, the thermal conductivity of the kiln body metal shell is 45–55 W / (m·K), the thermal conductivity of the refractory brick is 6 W / (m·K), and the thermal conductivity of the kiln lining is 1.3 W / (m·K). The kiln air mass flow rate is used to calculate the convective heat transfer coefficient between the gas and the material layer and inner wall; the kiln tail inflow mass flow rate and initial temperature serve as the initial boundary conditions for the material layer energy balance equation; the kiln tail gas temperature serves as the inlet boundary condition for the gas energy balance equation; and the thermal conductivity of the kiln body metal shell is used to correct the boundary heat loss term in the kiln wall conduction heat transfer calculation. During the modeling process, these process parameters are organically combined with heat transfer laws to clarify the quantitative relationship between the kiln structure, material movement, and heat transfer, enabling the model to realistically reflect the heat transfer characteristics under actual operating conditions and providing a solid framework for the subsequent establishment and solution of the energy equation set.

[0038] Step S220 requires first completing the reasonable segmentation of the kiln body, and then constructing a multi-field coupled energy equation set that comprehensively covers the heat transfer and generation process for each segment. To adapt to the differences in heat distribution, kiln brick material characteristics, and wear rate in different sections, this embodiment further divides the kiln body into 7 segments along the axial direction according to the kiln brick material. The specific segment range and parameter values ​​are as follows: Segment 1 (0-1.8m), Segment 2 (1.8-5m, preheating zone), Segment 3 (5-11m), Segment 4 (11-22m), Segment 5 (22-27.6m), Segment 6 (27.6-37.8m), and Segment 7 (37.8-78m). Each segment has a defined range for its internal surface temperature, kiln lining thickness, and kiln brick thickness. For example, segment 2 (preheating zone) has an internal surface temperature of 800–1200℃, a kiln lining thickness of 0–500 mm, and a kiln brick thickness of 100–220 mm. Subsequent voxel simulations and heat transfer mechanism model solutions are performed according to these segment parameters to ensure the industrial adaptability of the technical solution. The kiln body is divided into 79 independent segments along the axial direction, with a preset length of 1 m. This ensures relatively uniform heat transfer conditions within each segment, effectively reducing the difficulty of solving the equations. For each segment, a multi-field coupled energy equation system is established, centered on the energy balance of the material layer and the gas. Specifically, it includes the material layer energy balance equation, the gas energy balance equation, and the kiln wall energy balance equation. The material layer energy balance equation is as follows: , The gas energy balance equation is , The energy balance equation for the kiln wall is: .

[0039] All heat transfer and generation terms covered by the equations are determined through explicit calculation methods. The calculation of radiative heat transfer between the gas and the material layer and the inner wall follows the following: Radiative heat transfer between the gas and the inner wall follows the principle of: This is the Stefan-Boltzmann constant. The radiation area of ​​the gas and the material layer. This represents the radiation area between the gas and the inner wall. For gas emissivity, For the emissivity of the material layer, The emissivity of the inner wall. For gas absorption rate, The temperature of the material bed. The inner wall temperature The gas temperature is used. Convective heat transfer between the gas and the material bed is achieved using: The convective heat transfer between the gas and the inner wall is achieved by: The convective heat transfer coefficient between the gas and the material layer. The convective heat transfer area between the gas and the material layer. The convective heat transfer coefficient between the gas and the inner wall is denoted as . Let be the convective heat transfer area between the gas and the inner wall. The conductive heat transfer between the inner and outer walls is calculated strictly according to the thermal conductivity laws of a cylindrical wall, using the following formula: The thermal conductivity of refractory materials, The length of the segmented kiln body, The temperature of the outer wall of the kiln. The thickness of the refractory lining. The inner diameter of the kiln is given. The heat of combustion and the heat of chemical reaction are calculated based on the fuel characteristics and material reaction laws under the corresponding operating conditions. The types and relationships of all heat terms are determined based on the heat and mass transfer theory and the actual heat transfer conditions inside the kiln, ensuring that the equations can comprehensively and accurately describe the complex energy balance relationship inside the kiln.

[0040] In step S230, the multi-field coupled energy equations for each segment are solved sequentially along the axial axis. The solution result of the previous segment is used as the boundary condition for the next segment to ensure the continuity of heat transfer between segments and avoid abrupt temperature changes that do not conform to actual heat transfer laws. During the solution process, mature numerical calculation methods are used, combined with the process parameters and heat transfer characteristics of each segment, to accurately calculate the kiln wall temperature corresponding to each segment. After all segmented solutions are completed, the 79 kiln inner wall temperature data points are systematically integrated to eliminate minor temperature deviations at segment boundaries, ensuring smooth connection between the temperature data points and ultimately forming an axial temperature curve covering the entire kiln axis. This curve, with a length of 79, fully reflects the axial temperature distribution characteristics inside the kiln, providing high-quality temperature data support for training the RKT-Net model and directly validating the accuracy of subsequent model predictions, thus ensuring the reliability of the entire prediction scheme.

[0041] Figure 3 This is a schematic diagram illustrating the process of constructing the training dataset for the heat transfer mechanism model provided by this invention. For example... Figure 3As shown, according to an embodiment of the present invention, the method further includes constructing a training dataset for a heat transfer mechanism model, specifically including: in step S310, fitting and downsampling the axial temperature curve of the inner surface and downsampling the axial average temperature curve of the outer surface, so that the dimensions of the two are unified to a preset length; in step S320, using the dimension-unified axial average temperature curve of the outer surface and the corresponding axial temperature curve of the inner surface as a set of training samples, and generating a batch of rotary kiln internal temperature prediction training set based on infrared cylinder scanning data of continuous time periods; in step S330, dividing the prediction training set into a training subset and a prediction subset according to a preset ratio.

[0042] Specifically, the goal of step S310 is to unify the dimensions of the input and label data, providing consistent basic data for model training. The axial temperature curve of the inner surface is generated by solving the heat transfer mechanism model. The original length is consistent with the number of kiln segments, which is 79. First, the curve is smoothed by polynomial fitting to eliminate the small temperature fluctuations that may occur during the segmented solution process, ensuring the continuity and physical rationality of the curve. After fitting, a downsampling operation is performed, and the original 79-dimensional curve is scaled down to 50 dimensions by sampling data points at equal intervals. This allows the curve to retain the main temperature distribution characteristics while reducing the computational complexity of the model.

[0043] The original length of the outer surface axial average temperature curve is 200. Without additional fitting, it directly employs the same downsampling strategy as the inner surface axial temperature curve, extracting data points at equal intervals and scaling the dimension to a uniform 50. The preset length of 50 here balances data feature preservation with model computational efficiency and perfectly matches the input-output dimension requirements of the subsequent RKT-Net model. This ensures that the processed outer surface axial average temperature curve is dimensionally identical to the inner surface axial temperature curve, satisfying the input-label pairing requirements in supervised training.

[0044] Step S320 requires generating a batch of rotary kiln internal temperature prediction training sets with sufficient sample size based on high-quality data from continuous time periods. First, the pairing rules for the samples are clarified. Each training sample consists of a 50-dimensional outer surface axial average temperature curve and a corresponding 50-dimensional inner surface axial temperature curve, with a one-to-one correspondence between the two and originating from the kiln state data at the same sampling time, ensuring the correlation and authenticity of the samples.

[0045] During the data acquisition phase, infrared cylinder scanning data from four consecutive months were selected. 843 sets of valid data were filtered out through equal-interval sampling to avoid temporal concentration bias in the dataset. Each set of selected valid data underwent dimensional standardization processing according to step S310, resulting in 843 sets of formatted training samples. This ultimately constructed a training set of 843 samples for predicting the internal temperature of the rotary kiln. The batch-generated training set covers the temperature characteristics of the kiln under different operating conditions and during different operating periods, providing rich learning samples for the RKT-Net model and effectively improving the model's generalization ability and prediction accuracy.

[0046] Step S330 provides independent data support for model training and accuracy verification by scientifically dividing the data subset. The preset ratio here is set to 7:3, and the prediction training set with a capacity of 843 is split into a training subset and a prediction subset by random partitioning to ensure that the two sets of data are consistent in temperature feature distribution and to avoid model training bias caused by improper partitioning.

[0047] After partitioning, the training subset contains 590 samples (843 × 70% ≈ 590), primarily used for parameter learning and iterative optimization of the RKT-Net model, providing sufficient training data to explore the mapping relationship between the outer and inner surface temperatures. The prediction subset contains 253 samples (843 × 30% ≈ 253), used to independently verify the model's training effect. By comparing the model's prediction results with the actual inner surface temperature curve, the model's prediction accuracy is quantitatively evaluated. This partitioning method conforms to conventional practices in deep learning model training, ensuring sufficient training data while objectively reflecting the model's generalization performance through an independent prediction subset, providing a reliable evaluation basis for subsequent model parameter optimization.

[0048] Figure 4 This is a schematic diagram illustrating the process of constructing the RKT-Net model provided by this invention. For example... Figure 4As shown, according to an embodiment of the present invention, the implementation further includes constructing an RKT-Net model, specifically including: in step S410, the RKT-Net model includes a ResNet residual structure and a Transformer structure; in step S420, the ResNet residual structure includes a preset number of groups of residual units, each group of residual units is configured with a main path and a residual path; the main path sequentially deploys a convolutional layer, a normalization layer, an activation function, another convolutional layer, and a normalization layer, the two normalization layers respectively cooperate with the corresponding convolutional layers to complete feature standardization, improving the stability of feature extraction; the residual path directly passes the input features to the end of the main path, and performs element-wise addition with the features processed by the main path to achieve fusion, and then outputs through the activation function; multiple groups of residual units work together to gradually extract local features of the outer surface temperature curve, and map the input dimension from the third preset dimension to 64 dimensions, 128 dimensions, and 256 dimensions sequentially; in step S43 In step S440, the Transformer structure has a first preset number of attention heads, a second preset number of network layers, and a third preset number of feature dimensions. It is used to construct the global dependency relationship of temperature features, receive the 256-dimensional feature vector output by the ResNet residual structure, and perform global feature fusion and decoding. In step S450, the RKT-Net model also includes a fully connected layer, which is used to restore the feature vector output by the Transformer structure to a preset length, consistent with the dimension of the inner surface axial temperature curve. In step S450, the RKT-Net model inputs the third preset dimension of the outer surface axial average temperature curve and outputs the inner surface axial temperature curve.

[0049] Specifically, step S410 clarifies the architectural composition of the RKT-Net model, achieving an organic combination of local temperature feature extraction and global dependency construction through the synergistic cooperation of the ResNet residual structure and the Transformer structure. This architectural design leverages the stability of ResNet in capturing local features while utilizing the global correlation mining capability of Transformer for long-sequence data to accurately adapt to the mapping requirements from the outer surface temperature curve to the inner surface temperature curve, providing structural support for the model's prediction accuracy.

[0050] Step S420 focuses on the detailed construction of the ResNet residual structure, which includes a predetermined number of 10 groups of residual units (5×2 groups). Each group of residual units is designed in parallel with the main path and the residual path to balance the depth and effectiveness of feature extraction. The main path sequentially deploys a convolutional layer, a normalization layer, an activation function, another convolutional layer, and a normalization layer. The first convolutional layer is responsible for dimensionality transformation and preliminary extraction of the input features, and the normalization layer eliminates feature distribution differences, improving training stability. The activation function introduces nonlinear transformation to enhance the model's ability to fit complex temperature features. The second convolutional layer further deepens feature extraction, and the output features are normalized again by the normalization layer. The residual path directly passes the input features to the end of the main path, and the features processed by the main path are element-wise added to achieve fusion, effectively alleviating the gradient vanishing problem caused by the increase in network depth. The extraction results of this group of residual units are then output by the activation function. When multiple residual units work together, they follow a fixed dimension mapping rule to map the input third preset dimension (50 dimensions) outer surface axial average temperature curve to 64 dimensions and 128 dimensions in sequence, and finally output a 256-dimensional local feature vector. This not only preserves the key local fluctuation features of the temperature curve, but also achieves the orderly improvement of feature dimensions, laying the foundation for subsequent global feature processing.

[0051] Step S430 configures the parameters and implements the functionality of the Transformer structure. Specifically, it has 8 attention heads (first preset number), 5 network layers (second preset number), and a feature dimension of 512 (third preset number). This structure receives the 256-dimensional feature vector output from the ResNet residual structure and maps the feature dimension to 512 dimensions through an embedding layer to meet the processing requirements of the Transformer. The 8 attention heads are computed in parallel, focusing on the correlation between different positions in the feature sequence, using a self-attention mechanism formula: in, To achieve the dimensionality of the attention head (512 / 8=64), the association weight between any two temperature feature points in the sequence is directly calculated, realizing the interaction and fusion of feature information on a global scale. The five-layer network is progressively advanced, gradually deepening the decoding process of global features, and finally outputting a global fusion feature vector with a dimension of 256. This makes up for the deficiency of ResNet, which is only good at local feature extraction, and fully captures the global distribution pattern of the temperature curve along the axis.

[0052] In step S440, the fully connected layer performs dimensionality reduction and feature mapping. Through linear transformation, it reduces the 256-dimensional globally fused feature vector output by the Transformer structure to a preset length of 50 dimensions. This preset length perfectly matches the dimension of the inner surface axial temperature curve, ensuring that the model output directly corresponds to a temperature sequence in actual physical terms. The weight parameters of the fully connected layer are iteratively optimized through subsequent training processes to accurately learn the nonlinear mapping relationship between the globally fused features and the inner surface temperature curve, guaranteeing the accuracy of the output results.

[0053] Step S450 clarifies the input and output specifications of the RKT-Net model. The input data is the axial average temperature curve of the outer surface in the third preset dimension (50-dimensional). This curve has undergone preprocessing and dimensionality unification, possessing the characteristics of standardized format and complete features. The model, through the collaborative operation of the ResNet residual structure and the Transformer structure, as well as the dimensionality restoration of the fully connected layer, finally outputs a 50-dimensional axial temperature curve of the inner surface. This output curve fully reflects the internal temperature distribution of the kiln body along the axis, maintaining consistency with the label dimension of the training dataset, and providing accurate temperature data input for subsequent prediction of kiln lining and kiln brick thickness, forming a complete technical chain.

[0054] Figure 5 This is a schematic diagram illustrating the training process of the RKT-Net model provided by this invention. For example... Figure 5 As shown, according to an embodiment of the present invention, the method further includes training the RKT-Net model, specifically including: in step S510, supervised training is performed using a training subset of the prediction training set, and the prediction subset of the prediction training set is used to verify the training accuracy of the RKT-Net model; in step S520, normalization processing is performed on the input data of the training subset and the prediction subset; in step S530, restoration processing is performed on the output data of the RKT-Net model; in step S540, the adaptive learning rate is set to a first preset learning rate, the ReduceLROnPlateau strategy is adopted, the factor is set to a first preset factor, and the learning rate is set to... The first preset patience is set, the batch size is set to the first preset batch, and the maximum number of iterations is set to the first preset number. If the validation set loss does not decrease for the first preset number of consecutive rounds, training is stopped, and the best network weights are retained. In step S550, after training is completed, the accuracy is verified using the predicted subset. It is required that the average absolute value between the predicted temperature and the actual temperature inside the cement rotary kiln is less than the preset temperature. The average absolute value is the mean absolute error (MAE), which is defined as the arithmetic mean of the absolute values ​​of the differences between the predicted temperature and the actual temperature of all samples. It is used to intuitively reflect the overall degree of deviation of the prediction results. The degree of fit between the predicted temperature and the actual temperature inside the cement rotary kiln is greater than the preset fit value.

[0055] Specifically, step S510 clarifies the division of labor between model training and accuracy verification, ensuring the effectiveness of the training process and the objectivity of the verification results. Based on the previously constructed training set of 843 rotary kiln temperature prediction sets, after a 7:3 ratio, 590 sets constitute the training subset, and 253 sets serve as the prediction subset. Supervised training uses the preprocessed, dimensionally uniform axial average temperature curve of the outer surface as input and the corresponding axial temperature curve of the inner surface as the label. Backpropagation is performed using the difference between the model output and the true label to iteratively optimize the various network parameters of the RKT-Net model. The prediction subset remains independent throughout the process, not participating in any training process, and is only used for real-time monitoring of the model's performance on unseen data, avoiding accuracy evaluation bias caused by data leakage, and providing a reliable basis for determining subsequent training termination conditions.

[0056] The normalization process in step S520 is used to eliminate dimensional differences in the input data, improving the stability and convergence speed of model training. Specifically, the maximum and minimum values ​​of the axial average temperature of the outer surface in the training subset are first statistically analyzed. Considering the operating conditions of a cement rotary kiln, the outer surface temperature typically ranges from 100 to 300°C. Based on this, a unified normalization mapping rule is established, scaling all input data from both the training and prediction subsets to the range of 0 to 1. The normalization of the prediction subset strictly follows the statistical benchmark of the training subset, rather than its own extreme values, ensuring the consistency of the input data distribution and avoiding systematic biases in model predictions due to benchmark differences, thus guaranteeing the model's generalization performance.

[0057] Step S530, the restoration process, is the inverse operation of normalization. Its purpose is to map the abstract numerical values ​​output by the model to internal surface temperature values ​​with actual physical meaning. Training uses the mean squared error (MSE) loss function to quantify the difference between the predicted internal surface temperature curve and the true curve, as shown in the formula: in, For the first The true internal surface temperature value of each sample These are the model's predicted values. The total number of samples is represented by this function. The squared penalty applied to the error prioritizes correcting significant temperature deviations, meeting the stringent requirements of industry for temperature prediction accuracy. This complements the MAE metric used in the subsequent validation phase, comprehensively ensuring the model training effect.

[0058] Because the input data is normalized and scaled during model training, the output results are also in the 0 to 1 range, which cannot directly reflect the true temperature. Therefore, based on the extreme values ​​of the inner surface temperature of the training subset recorded in step S520 (the inner surface temperature range is typically 0 to 1000℃), a restoration mapping rule needs to be established to inversely scale the abstract values ​​output by the model to the actual temperature range. This processing step ensures that the model output results can be directly used for subsequent practical scenarios such as kiln operation status assessment and kiln lining / brick thickness prediction, completing the complete link of data preprocessing, model training, and result restoration.

[0059] Step S540 balances model training effectiveness and computational resource consumption by finely setting training parameters and termination strategies. Considering the task characteristics of rotary kiln temperature prediction, the values ​​of each preset parameter are set with reference to industry conventions and experimental verification results. The first preset learning rate for adaptive learning is set to 0.001, balancing convergence speed and parameter optimization accuracy. The ReduceLROnPlateau strategy, commonly used in deep learning, is adopted to dynamically adjust the learning rate, with the first preset factor set to 0.5 and the first preset patience set to 10, meaning that learning rate adjustment is triggered when the validation set loss does not decrease for 10 consecutive rounds. The first preset batch size is set to 32, achieving a balance between single-round training efficiency and gradient stability. The first preset number of iterations, i.e., the maximum number of iterations, is set to 500 to avoid overfitting due to overtraining. The training termination condition is clearly defined as the validation set loss not decreasing for 10 consecutive rounds (usually 10). At this point, training is immediately stopped, and the optimal network weights at the previous point of minimum loss are automatically retained, avoiding ineffective iterations that consume resources while ensuring the model has optimal generalization performance.

[0060] Step S550 is the acceptance test for model training effectiveness, ensuring the model meets industrial application requirements through dual precision indicators. The average absolute error (ADE) is the absolute value between the predicted temperature inside the cement rotary kiln and the actual temperature. Its preset temperature, based on industrial precision requirements, is set to 5℃. This indicator directly reflects the overall deviation of the model's predictions, ensuring the accuracy of the prediction results. The coefficient of determination (COD) is the goodness of fit between the predicted temperature inside the cement rotary kiln and the actual temperature. Its preset fit value is set to 0.95. This indicator reflects the degree of agreement between the predicted curve and the actual temperature curve, ensuring the model can accurately capture the axial distribution pattern of temperature inside the kiln. If the model meets both indicator requirements, the training is considered successful; otherwise, the process returns to step S540 to adjust the training parameters and retrain until the indicators meet the standards, ensuring the final output model has reliable industrial application value.

[0061] Figure 6 This is a schematic diagram of the process for preprocessing external surface cylindrical scan data provided by the present invention. For example... Figure 6As shown, according to an embodiment of the present invention, the method further includes preprocessing the external surface cylindrical scanning data, specifically including: in step S610, splitting the external surface temperature cylindrical scanning data according to the sampling time to obtain an external surface temperature field matrix of a first preset dimension, verifying the data integrity through thermal visualization, and removing abnormal missing data; in step S620, dynamically adjusting the sampling frequency of the external surface temperature cylindrical scanning data according to the rotary kiln rotation speed to ensure that the spatial overlap rate of adjacent sampling points is greater than a preset percentage; in step S630, correcting the sampling start position deviation of each external surface temperature field matrix based on the rotation period of the rotary kiln, so that the row index of all external surface temperature field matrices is aligned with the axial physical position of the rotary kiln; in step S640, averaging the temperature field matrix along the axial direction after position alignment to obtain the original external surface axial temperature curve of the second preset dimension; and performing a downsampling operation on the original external surface axial temperature curve to unify it into an external surface axial average temperature curve of the third preset dimension.

[0062] Specifically, step S610 is the fundamental step in data preprocessing, aiming to split the raw data and filter out valid information. The outer surface temperature data collected by the infrared cylinder scanner is continuous time-series data. After being split according to the sampling time, it forms multiple independent outer surface temperature field matrices. The first preset dimension is set to 200×100, with the rows of the matrix corresponding to the axial position of the kiln body and the columns corresponding to the circumferential position. This dimension can not only completely preserve the spatial distribution information of the kiln body surface temperature, but also control the data volume to avoid redundancy. After splitting, the numerical distribution of each temperature field matrix is ​​visually presented through a heatmap visualization method, quickly identifying abnormal data with problems such as large-area missing data and local numerical abrupt changes. Such data that cannot reflect the true temperature state of the kiln body is directly removed, ensuring the integrity and reliability of the data in subsequent processing.

[0063] Step S620 ensures the spatial continuity of temperature data by dynamically adapting the sampling frequency. The real-time rotational speed of the rotary kiln is within a dynamic range of 0.5 to 4 rpm. Changes in rotational speed can cause differences in the coverage area of ​​adjacent sampling points on the kiln surface. If the sampling frequency is fixed, the interval between sampling points may be too large, leading to missed temperature information, or the interval may be too small, resulting in data duplication. Therefore, the sampling frequency is flexibly adjusted according to the real-time rotational speed. When the rotational speed increases, the sampling frequency is appropriately increased, and when the rotational speed decreases, the sampling frequency is correspondingly decreased. The goal is to ensure that the spatial overlap rate of adjacent sampling points is always greater than 50%. Sufficient spatial overlap allows continuously acquired temperature data to form an effective connection, avoiding recording gaps in the temperature distribution on the kiln surface due to rotational speed fluctuations, and providing a coherent data foundation for subsequent position alignment and feature extraction.

[0064] Step S630 focuses on correcting the deviation of the sampling start position to achieve accurate matching between temperature data and physical location. The sampling start points of temperature field matrices at different sampling times may deviate due to differences in equipment triggering timing and kiln rotation phase. Direct use of these matrices would result in row indices of different matrices not corresponding to the same axial physical position of the kiln, affecting the consistency of temperature characteristics. Correction is made based on the rotation period of the rotary kiln, which is calculated from the real-time rotation speed; the faster the speed, the shorter the period, and vice versa. By calculating the difference between the sampling start time of each temperature field matrix and the baseline start time, and combining this with the rotation period, the row index order of the matrices is adjusted to ensure that the row indices of all temperature field matrices accurately correspond to the axial physical position of the kiln from one end to the other. This completely eliminates the feature misalignment problem caused by the starting position deviation, ensuring that temperature data at different times have a comparable basis.

[0065] Step S640 standardizes and simplifies the data dimensions through mean calculation and downsampling. First, the mean of the aligned temperature field matrix is ​​calculated along the axial direction; that is, for each axial position (each row) in the matrix, the average temperature of all circumferential positions (each column) is calculated, resulting in the original axial temperature curve of the outer surface that only reflects axial temperature changes. The second preset dimension is 200, consistent with the number of axial rows in the temperature field matrix, ensuring the original curve fully preserves the axial temperature distribution trend of the kiln body. Since the original curve has a high dimension, which would increase the computational burden of subsequent models, a downsampling operation is performed, using equally spaced data points to uniformly scale the 200-dimensional original curve to a third preset dimension of 50. This dimension preserves the characteristics of the axial temperature distribution while significantly reducing the data volume, ensuring the final outer surface axial average temperature curve fully meets the input requirements of the RKT-Net model, providing efficient and standardized data input for model computation.

[0066] Figure 7 This is a flowchart illustrating the method for predicting the thickness of kiln lining and kiln bricks in a cement rotary kiln provided by the present invention. The present invention also provides a method for predicting the thickness of kiln lining and kiln bricks based on the above-described method for predicting the internal temperature of a cement rotary kiln, such as... Figure 7As shown, the process includes: in step S700, two sets of operating data of a cement rotary kiln are collected based on a preset time interval. Each set of data includes an outer surface temperature field matrix generated by preprocessing the outer surface cylindrical sweep data, and the corresponding inner surface axial temperature curve output by the RKT-Net model; in step S800, the data pairs consisting of the two sets of outer surface temperature field matrices and the corresponding inner surface axial temperature curves are input into the RKT-Net model, which outputs the kiln skin thickness distribution matrix and the kiln brick thickness distribution matrix along the entire kiln axis; wherein, the RKT-Net model is trained using a quadruple consisting of two sets of outer surface temperature field matrices at different time points, the corresponding inner surface axial temperature curves, the corresponding kiln skin thickness distribution matrix, and the corresponding kiln brick thickness distribution matrix as samples, and the constraint conditions are constructed based on the two sets of temperature data at different time points to solve the under-constraint problem of kiln skin and kiln brick thickness prediction.

[0067] Specifically, step S700 captures the temperature change characteristics of the kiln body through time-series data acquisition, providing effective input for thickness prediction. This method uses "dual sampling points" to overcome the under-constraint problem in thickness prediction, assuming that the thickness distribution changes of the kiln lining and kiln bricks are negligible within a preset 2-hour time interval, and that in industrial practice, the wear rate of the kiln lining and the loss rate of the kiln bricks are far lower than the temperature fluctuations at this time scale. Based on this assumption, the two sets of time-series data satisfy the heat transfer law: in, This represents the heat transfer mapping relationship (including constant parameters such as material thermal conductivity and metal layer thickness). The number of constraints and unknowns in the two sets of equations are as follows: , With consistent quantities, a unique solution can be obtained, completely solving the under-constraint problem of traditional methods.

[0068] The preset time interval is set based on the actual industrial patterns of kiln lining and brick thickness variations, and is set to 2 hours. This interval ensures a detectable thickness-related temperature difference between the two sets of data, while avoiding data redundancy due to too short an interval or missing key changes due to too long an interval. Each set of running data contains two types of information: one is the outer surface temperature field matrix after preprocessing in steps S610 to S630, with a dimension of 200×100, which fully preserves the axial and circumferential temperature spatial distribution of the kiln body; the other is the inner surface axial temperature curve output in real time by the trained RKT-Net model, with a dimension of 50, accurately reflecting the internal axial temperature gradient of the kiln body. The two types of data correspond one-to-one, both originating from the kiln operating status at the same sampling time, ensuring the correlation and authenticity between the data, laying the foundation for subsequent inference of thickness changes through temperature differences.

[0069] The training samples used to construct the RKH-Net model are in a four-tuple structure, specifically including two sets of external surface temperature field matrices at different time points, two corresponding sets of internal surface axial temperature curves, a corresponding kiln lining thickness distribution matrix, and a corresponding kiln brick thickness distribution matrix. During model training, constraints are constructed based on the two sets of temperature data at different time points. Utilizing the assumption that the change in kiln lining and kiln brick thickness within a preset 2-hour time interval is negligible, the two sets of temperature data correspond to the same set of thickness parameters, forming two independent heat transfer constraint equations. This ensures that the two unknowns—kiln lining thickness and kiln brick thickness—are precisely matched with the number of constraint equations, fundamentally solving the under-constraint problem in kiln lining and kiln brick thickness prediction and providing a mathematical basis for finding the unique solution for the thickness.

[0070] Step S800 achieves accurate mapping from temperature data to thickness distribution, relying on the feature extraction and association learning capabilities of the RKH-Net model. First, the rules for data pair composition are clarified: two sets of data collected at two different time points are combined into one input data pair. This pair includes the outer surface temperature field matrix and the inner surface axial temperature curve at the first time point, as well as the outer surface temperature field matrix and the inner surface axial temperature curve at the second time point. These four types of data collaboratively constitute the complete input to the model. This input format can fully preserve the spatial distribution differences of the temperature field and the changes in the axial temperature gradient at different time points, providing rich information for the model to explore the dynamic correlation between temperature and thickness.

[0071] After the data is input into the RKH-Net model, the model extracts local features from two sets of temperature data through dual convolutional input channels and superimposes them. The CBAM attention mechanism specifically enhances the response of key temperature regions strongly correlated with thickness changes. The Encoder and Transformer modules work together to construct the global dependencies of temperature features, while multiple Decoder modules decode the thickness information of the kiln lining and kiln bricks respectively. The final output kiln lining thickness distribution matrix and kiln brick thickness distribution matrix have the same dimensions (200×100) as the input outer surface temperature field matrix. The rows of the matrix correspond to the axial position of the kiln body, and the columns correspond to the circumferential position, which can completely present the thickness distribution of the entire kiln in the axial and circumferential directions, achieving accurate quantification of thickness parameters. The output results not only meet the refined requirements of industrial site monitoring of the thickness of different parts of the kiln body, but also provide intuitive and reliable data support for subsequent equipment maintenance decisions.

[0072] Figure 8 This is a schematic diagram of the process for constructing the RKH-Net model provided by this invention. For example... Figure 8As shown, according to an embodiment of the present invention, the implementation further includes constructing an RKH-Net model, specifically including: in step S810, the RKH-Net model includes a dual convolutional input channel, an Encoder module, a Transformer module, a multiple Decoder module, and a CBAM attention mechanism; in step S820, the dual convolutional input channel is used to extract local features of two sets of temperature data respectively and superimpose them; the CBAM attention mechanism strengthens the response of key temperature regions where the changes in kiln skin and kiln brick thickness are strongly correlated; in step S830, the RKH-Net model inputs two sets of data pairs consisting of the outer surface temperature field matrix and the inner surface axial temperature curve at different time points, and outputs the kiln skin thickness distribution matrix and the kiln brick thickness distribution matrix; in step S840, the output results must satisfy that the average absolute value between the predicted value and the true value of the kiln skin thickness is less than a first preset thickness, and the average absolute value between the predicted value and the true value of the kiln brick thickness is less than a second preset thickness.

[0073] Specifically, step S810 clarifies the architecture of the RKH-Net model, using TransUnet as the main framework. Through the coordinated operation of dual convolutional input channels, an Encoder module, a Transformer module, multiple Decoder modules, and a CBAM attention mechanism, it achieves accurate mapping of temperature features to thickness distribution. This architecture retains the efficient extraction capability of CNNs for local spatial features while leveraging the Transformer module to mine global correlation information. Simultaneously, the multiple Decoder module enables the simultaneous calculation of kiln skin and kiln brick thicknesses, specifically addressing the under-constraint problem in thickness prediction. Each module performs its specific function while remaining closely integrated. The dual convolutional input channels handle initial feature extraction and temporal data fusion; the Encoder and Transformer modules deepen feature representation; the CBAM attention mechanism enhances the response to key information; and the multiple Decoder module completes the final thickness decoding. This forms a complete technical chain: first extracting temporal temperature features, then fusing cross-time dimension information, then enhancing the key temperature response, and finally completing the thickness mapping. This ensures that every step from input data to output results accurately adapts to the thickness prediction requirements.

[0074] Step S820 focuses on the functional implementation of the dual-convolutional input channel and the CBAM attention mechanism, ensuring effective extraction of temperature features and enhancement of key information. The dual-convolutional input channel corresponds to two sets of temperature data spaced 2 hours apart. Each set of input data includes a 200×100 outer surface temperature field matrix and a 50-dimensional inner surface axial temperature curve. Through dimensionality adaptation, the inner surface temperature curve is expanded into a 200×100 matrix form and fused with the outer surface temperature field matrix into a unified input data format. Each channel contains a convolutional layer, a normalization layer, and an activation function, independently extracting local temperature features at the corresponding time point and outputting a 32×50×100 feature vector. The two feature vectors are then superimposed to form a 64×50×100 fused feature vector, fully preserving the differences and correlations between the two sets of time-series data. The CBAM attention mechanism is embedded after feature overlay. Through the synergistic effect of channel attention and spatial attention, it automatically identifies and enhances the temperature region response that is strongly correlated with the thickness variation of kiln skin and kiln bricks, suppresses irrelevant background noise, and makes the model more focused on key temperature features, thereby improving the pertinence and accuracy of thickness prediction.

[0075] Step S830 clarifies the input and output specifications of the model, ensuring consistency between data format and physical meaning. The model input consists of two sets of temperature data pairs at different time points. Each set of data includes a preprocessed 200×100 outer surface temperature field matrix and a corresponding 50-dimensional inner surface axial temperature curve. Utilizing the negligible variation in kiln lining and brick thickness within a 2-hour interval, constraints are constructed to address the under-constraint problem. The input data, after feature extraction via dual convolution channels and superposition, is sequentially fed into the Encoder and Transformer modules: The Encoder module progressively deepens local feature extraction through multi-layer convolution and pooling operations, mapping the 64×50×100 fused feature vector sequentially into feature maps of 128×25×50, 256×13×26, and 512×7×14; the Transformer module constructs global dependencies of the features, uncovering global patterns in temperature distribution, while maintaining an output dimension of 512×7×14. RKH-Net uses TransUNet as its main framework, preserving the skip connection structure between the encoder and decoder: high-resolution local temperature features extracted by the CNN encoder (such as 128×25×50 and 256×13×26 feature maps) are directly fused with the global correlation features output by the Transformer decoder (512×7×14 feature map) through skip connections. This design can compensate for the loss of detail during the upsampling process, allowing local temperature fluctuation features and global temperature distribution patterns to complement each other, improving the accuracy of thickness prediction. The multi-decoder module adopts the upsampling structure corresponding to the encoder module. Through four upsampling steps and three skip connections, the high-dimensional feature map is gradually restored to a 200×100 dimension. The two decoders independently decode the thickness information of the kiln skin and kiln bricks, and finally output two sets of 200×100 thickness distribution matrices. The rows of the matrix correspond to the axial position of the kiln body, and the columns correspond to the circumferential position, fully presenting the thickness distribution of the entire kiln.

[0076] Step S840 sets the accuracy standards for the model output to ensure that the prediction results meet the needs of industrial applications. The first preset thickness is set to 40mm, meaning the average absolute value between the predicted and actual kiln lining thickness must be less than 40mm; the second preset thickness is set to 25mm, meaning the average absolute value between the predicted and actual kiln brick thickness must be less than 25mm. This accuracy standard is based on the actual industrial operation of cement rotary kilns, taking into account both the reasonable error range of on-site measurements and ensuring the reliability of operation and maintenance decisions. The prediction accuracy of kiln lining thickness can meet the timely identification of abnormal states such as kiln lining ring formation and detachment, while the prediction accuracy of kiln brick thickness can provide a precise basis for kiln brick wear and replacement, avoiding kiln overheating damage or production capacity loss due to thickness judgment errors, ensuring that the model has strong industrial applicability.

[0077] Figure 9This is a schematic diagram illustrating the training process of the RKH-Net model provided by this invention. For example... Figure 9 As shown, according to an embodiment of the present invention, the training of the RKH-Net model is further included, specifically: in step S910, a training set of the RKH-Net model is constructed based on the voxel simulation heat conduction database and the overall simulation heat conduction database. The samples of the training set are quadruples consisting of two sets of external surface temperature field matrices, corresponding internal surface axial temperature curves, corresponding kiln skin thickness distribution matrices, and corresponding kiln brick thickness distribution matrices at two different time points. The training set is divided into a training subset and a prediction subset according to a preset ratio. In step S920, the external surface temperature field matrices and internal surface axial temperature curves of the training subset and the prediction subset are normalized. In step S930, the Adam optimizer is used, the adaptive learning rate is set to a second preset learning rate, the ReduceLROnPlateau strategy is used, the factor is set to a second preset factor, and the patience is set to a second preset tolerance. In step S940, the output data of the RKH-Net model is restored to map the abstract numerical values ​​output by the model to kiln skin thickness and kiln brick thickness values ​​with actual physical meaning. In step S950, after training is completed, the accuracy is verified using a predicted subset. The average absolute value between the predicted kiln skin thickness and the actual kiln brick thickness must be less than the first preset thickness, and the average absolute value between the predicted kiln brick thickness and the actual kiln brick thickness must be less than the second preset thickness.

[0078] Specifically, step S910 focuses on the construction and partitioning of the training set, providing high-quality and sufficient sample support for model training. The training set is jointly constructed based on the voxel simulation heat conduction database and the overall simulation heat conduction database. The two databases generate temperature-thickness paired data in batches through industrial simulation technology, effectively solving the pain point of scarce on-site industrial measurement data. The construction of the voxel simulation database is performed according to the above 7 segments. Each segment uses an exhaustive method to traverse all possible parameter combinations: the inner surface temperature is taken at 10℃ intervals, the kiln lining thickness is taken at 20mm intervals, and the kiln brick thickness is taken at 10mm intervals. Taking segment 2 (preheating zone) as an example, 41 values ​​can be obtained for the inner surface temperature (800~1200℃), 26 values ​​can be obtained for the kiln lining thickness (0~500mm), and 13 values ​​can be obtained for the kiln brick thickness (100~220mm), forming a total of 41×26×13=13858 working conditions. After each operating condition is simulated by Ansys steady-state heat conduction, the data is stored as tabular data in the format of [inner surface temperature, outer surface temperature, kiln skin thickness, kiln brick thickness], which ultimately forms a dedicated voxel simulation heat conduction database for each segment. The overall simulation heat conduction database is constructed based on a voxel database using a lookup table method: First, prototype data of the inner surface temperature curve (10 sets of differentiated data extracted from the temperature prediction training set and fitted upsampled to 200 dimensions), kiln skin thickness field (32 sets artificially constructed), and kiln brick thickness field (20 sets artificially constructed) are generated and cross-matched to form 20×32×20=12800 basic combinations; three types of noise enhancement are added to the thickness field and temperature curve of each combination, namely wave noise, constant superposition, and thickness anomaly point, expanding to 51200 combinations; finally, for each unit area of ​​the complete kiln wall, according to its segment, preset thickness, and inner surface temperature, the closest data is found from the corresponding segment voxel database to match the outer surface temperature, and finally a quadruplet training sample containing two sets of time-series temperature data, kiln skin thickness matrix, and kiln brick thickness matrix is ​​formed, with a total capacity of 115200 sets. The training set consists of a single sample in a quadruple structure, containing two sets of external surface temperature field matrices at different time points (each with a dimension of 200×100), two corresponding sets of internal surface axial temperature curves (each with a dimension of 50), a corresponding kiln lining thickness distribution matrix (dimension 200×100), and a corresponding kiln brick thickness distribution matrix (dimension 200×100), fully covering all data types of model input and output. The total capacity of the training set is 115,200 sets, randomly divided into a training subset and a prediction subset according to a preset ratio of 7:3. After the division, the training subset contains 80,640 sets of samples, used for model parameter learning and iterative optimization; the prediction subset contains 34,560 sets of samples, which are completely independent of the training process and are used only for objectively verifying the model's generalization performance, ensuring that the temperature-thickness feature distributions of the two sets of data are consistent after the division, and avoiding the impact of data bias on the training effect.

[0079] Step S920 eliminates dimensional differences in the input data through normalization, ensuring the stability and convergence speed of model training. The training uses the mean squared error (MSE) loss function to simultaneously quantify the prediction errors of kiln lining thickness and kiln brick thickness, as shown in the formula: in, For the first The actual thickness value of each sample (including kiln lining and kiln bricks). To predict the thickness value for the model, The total number of samples is denoted as . The function strengthens the correction for larger deviations through squared error, which directly corresponds to the MAE accuracy requirements in the subsequent validation phase, ensuring that the training direction is accurately matched with the needs of industrial operation and maintenance.

[0080] The normalization process targets the outer surface temperature field matrix and the inner surface axial temperature curve of both the training and prediction subsets. The processing rules are based on the temperature extremes of the training subset: the outer surface temperature has a minimum of 100℃ and a maximum of 300℃, while the inner surface temperature has a minimum of 0℃ and a maximum of 1000℃. All input data are uniformly scaled to the range of 0 to 1. The normalization of the prediction subset strictly follows the statistical baseline of the training subset, without calculating its own extreme values. This ensures the consistency of the input data distribution and avoids systematic biases in model predictions due to baseline differences, laying a data foundation for the model to learn the mapping relationship between temperature and thickness.

[0081] Step S930 balances model training effectiveness and computational resource consumption by scientifically configuring training parameters and termination strategies. The Adam optimizer is used for training, which has adaptive learning rate adjustment capabilities, effectively improving training efficiency and convergence stability. The second preset learning rate is set to 0.001, balancing parameter optimization accuracy and convergence speed. The ReduceLROnPlateau strategy is used to dynamically adjust the learning rate, with the second preset factor set to 0.5 and the second preset patience set to 5. This means that if the validation set loss does not decrease for five consecutive rounds, the learning rate is automatically reduced by 50% to overcome training bottlenecks. The second preset batch size is set to 32, achieving a balance between single-round training efficiency and gradient stability. The second preset number of iterations, i.e., the maximum number of iterations, is set to 1000 to avoid overfitting due to overtraining. During training, the validation set loss is continuously monitored. When the validation set loss does not decrease for two consecutive preset rounds (20 rounds), training is immediately stopped, and the optimal network weights at the point of minimum loss are automatically retained. This avoids ineffective iterations that consume resources while ensuring the model has optimal generalization performance.

[0082] The restoration process in step S940 is the inverse operation of normalization, mapping the abstract numerical values ​​output by the model to thickness values ​​with actual physical meaning. Since the input data is normalized and scaled during model training, the output thickness-related values ​​are also in the 0-1 range, making them unsuitable for direct application in industrial scenarios. Therefore, a restoration mapping rule is established based on the actual thickness range of the kiln lining and kiln bricks: the kiln lining thickness is restored to the 0-500mm range, and the kiln brick thickness is restored to the 100-220mm range. Through linear mapping, the abstract values ​​are transformed into specific thickness quantification results, ensuring that the output data directly reflects the actual state of the kiln body and meets the actual needs of operation and maintenance decisions.

[0083] Step S950 is the acceptance test for the model training effect, ensuring that the model meets the requirements of industrial applications through clear accuracy indicators. The accuracy standard in this embodiment directly matches the needs of industrial operation and maintenance: the ideal kiln lining thickness ranges from 200 to 500 mm. A thickness <200 mm accelerates kiln brick wear, while >500 mm reduces heat transfer efficiency, leading to an 8%–12% increase in coal consumption. A MAE accuracy of 40 mm can accurately identify these abnormal states. The kiln brick thickness varies from 100 to 220 mm depending on the section. A MAE accuracy of 25 mm can provide timely warnings of excessive kiln brick wear risks, preventing a sudden rise in kiln surface temperature to above 400°C and causing a "red kiln" accident, ensuring that the prediction results directly serve equipment maintenance decisions. The first preset thickness is set to 40 mm, meaning the absolute average between the predicted and actual kiln lining thickness values ​​must be less than 40 mm; the second preset thickness is set to 25 mm, meaning the absolute average between the predicted and actual kiln brick thickness values ​​must be less than 25 mm. This accuracy standard is formulated based on the actual industrial operation of cement rotary kilns, taking into account both the reasonable error range of on-site measurements and the reliability of operation and maintenance decisions. The accuracy of kiln lining thickness prediction can promptly identify abnormal conditions such as ring formation and detachment, while the accuracy of kiln brick thickness prediction provides a precise basis for wear and replacement, avoiding overheating damage or production loss of the kiln body due to thickness judgment errors. If the model simultaneously meets both accuracy indicators, the training is deemed successful; if not, the process returns to step S930 to adjust the training parameters and retrain until the indicators meet the standards, ensuring that the final output model has strong industrial applicability.

[0084] The present invention also provides a system for predicting the internal temperature and kiln lining / brick thickness of a cement rotary kiln. Figure 10 This is a block diagram of the cement rotary kiln internal temperature and kiln lining / brick thickness prediction system provided by the present invention. Figure 10As shown, the cement rotary kiln internal temperature and kiln lining / brick thickness prediction system 100 includes: an infrared data acquisition module 110, used to acquire cylinder scan data of the cement rotary kiln's external surface temperature using an infrared cylinder scanner, and transmit the acquired cylinder scan data to a data preprocessing module; and a data preprocessing module 120, used to perform preprocessing operations on the external surface temperature cylinder scan data, including splitting, removing abnormal missing data, dynamically adjusting the sampling frequency, correcting the sampling start position deviation, calculating the axial mean, and downsampling, generating an external surface temperature field matrix of a first preset dimension and an external surface axial average temperature curve of a third preset dimension, and transmitting the external surface temperature field matrix to the RKH-Net model module, and transmitting the external surface axial average temperature curve to the RKT-Net. Model modules: RKT-Net model module 130, used to input the axial average temperature curve of the outer surface and output the axial temperature curve of the inner surface; the RKT-Net model module has a built-in RKT-Net model containing ResNet residual structure, Transformer structure and fully connected layer, using the rotary kiln internal temperature prediction training set, and completes training through normalization, reduction processing and ReduceLROnPlateau learning rate strategy and early stopping mechanism. The training accuracy meets the requirement that the absolute average value between the predicted temperature inside the cement rotary kiln and the actual value is less than the preset temperature, and the fitting degree is greater than the preset fitting value; RKH-Net model module 140, used to input two sets of data pairs consisting of the outer surface temperature field matrix and the inner surface axial temperature curve at different time points, and output the kiln skin thickness distribution matrix and the kiln brick thickness distribution matrix; the RKH-Net model module has a built-in RKH-Net model containing dual convolution input channels, Encoder module, Transformer module, multiple Decoder module and CBAM attention mechanism, based on the four-tuple training set constructed by the voxel simulation heat conduction database and the overall simulation heat conduction database, and uses Adam The optimizer, ReduceLROnPlateau learning rate strategy, early stopping mechanism, normalization, and restoration processing complete the training. The training accuracy meets the requirement that the average absolute value between the predicted and actual values ​​of kiln skin thickness is less than a first preset thickness, and the average absolute value between the predicted and actual values ​​of kiln brick thickness is less than a second preset thickness. The data storage module 150 is used to store the outer surface temperature cylinder scan data, the data preprocessed by the data preprocessing module, the training set of the RKT-Net model module, and the trained model weights. The result output module 160 is used to receive the inner surface axial temperature curve output by the RKT-Net model module and convert it into a visual chart or numerical report output.

[0085] The cement rotary kiln internal temperature and kiln lining / kiln brick thickness prediction system provided by this invention corresponds one-to-one with the above-mentioned cement rotary kiln internal temperature prediction method and kiln lining / kiln brick thickness prediction method. The six modules of the system are precisely matched with each step of the method. The technical logic, parameter configuration and accuracy requirements of the two are completely consistent and can be used for mutual reference. The relevant details will not be repeated here.

[0086] The specific correspondences are as follows: the infrared data acquisition module corresponds to the external surface temperature cylinder scan data acquisition stage in the method; the data preprocessing module corresponds to the external surface cylinder scan data preprocessing process in the method; the RKT-Net model module and the RKH-Net model module correspond to the construction, training, and prediction processes of the RKT-Net model and the RKH-Net model, respectively, in the method; the data storage module is used to store various types of data generated during the implementation of the method; and the result output module corresponds to the output stages of the inner surface axial temperature curve and the kiln skin and kiln brick thickness distribution matrix in the method.

[0087] All parameters, process rules, and accuracy standards specified in the method are applicable to the corresponding system modules. The system, as a physical implementation of the method, completely adopts the method's technical solution and requires no additional explanation. Those skilled in the art can understand the system's working principle and implementation through the detailed description of the method.

[0088] This invention constructs a complete technical system covering data acquisition, preprocessing, model computation, and result output through the collaborative innovation of heat transfer mechanism models and deep learning models, and the complementary support of voxel simulation and overall simulation database. This method and system require no additional sensor deployment, successfully addressing pain points such as the scarcity of temperature-thickness paired data in industrial settings and the lack of constraints in kiln lining and brick thickness prediction. The prediction accuracy fully meets the industrial-grade requirements for cement rotary kiln operation and maintenance. Furthermore, the entire solution possesses strong adaptability and scalability, integrating into existing production processes without complex on-site modifications. It can prevent safety risks such as kiln overheating and kiln lining detachment through precise monitoring, and improve heat transfer efficiency and reduce energy consumption by optimizing kiln lining conditions, providing reliable technical support for the safe, efficient, and low-cost operation of cement production.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the internal temperature of a cement rotary kiln, characterized in that, include: The outer surface temperature data of the cement rotary kiln is collected by an infrared cylinder scanner. The average value of the pre-processed outer surface temperature cylinder scanner data is calculated along the axial direction to obtain the axial average temperature curve of the outer surface. The axial average temperature curve of the outer surface is input into the RKT-Net model to obtain the axial temperature curve of the inner surface. The RKT-Net model is constructed based on training data and test data. The training data and the test data are paired samples consisting of the preprocessed axial average temperature curve of the outer surface and the axial temperature curve of the inner surface generated by the heat transfer mechanism model.

2. The method according to claim 1, characterized in that, This also includes constructing a heat transfer mechanism model, specifically including: Based on classical heat and mass transfer theory, and combined with the process parameters under normal operating conditions of cement rotary kilns, a heat transfer mechanism model is constructed using mathematical modeling. The process parameters include at least one of the following: kiln inner and outer diameter, kiln length, kiln rotation speed, fuel calorific value, kiln air mass flow rate, kiln tail material mass flow rate, kiln tail material initial temperature, kiln tail gas temperature, thermal conductivity of kiln metal shell, thermal conductivity of refractory bricks, and thermal conductivity of kiln lining. The cement rotary kiln is divided into segments according to a preset length, and a multi-field coupled energy equation set covering radiative heat transfer, convective heat transfer, conductive heat transfer, combustion reaction heat and chemical reaction heat is established for each segment. By solving the energy equations for each segment in sequence, the temperature of the inner wall of each kiln segment is obtained, and the integrated results output the axial temperature curve of the inner surface covering the entire kiln axis.

3. The method according to claim 1 or 2, characterized in that, It also includes the training dataset for constructing the heat transfer mechanism model, specifically including: The axial temperature curve of the inner surface is fitted and downsampled, and the axial average temperature curve of the outer surface is downsampled to unify the dimensions of the two to a preset length. The axial average temperature curve of the outer surface after dimension unification and the corresponding axial temperature curve of the inner surface are used as a set of training samples. Based on the infrared cylinder scanning data of continuous time periods, a training set for predicting the internal temperature of the rotary kiln is generated in batches. The prediction training set is divided into a training subset and a prediction subset according to a preset ratio.

4. The method according to claim 1, characterized in that, It also includes building the RKT-Net model, specifically including: The RKT-Net model includes a ResNet residual structure and a Transformer structure; The ResNet residual structure includes a preset number of residual units, each set with a main path and a residual path. The main path sequentially deploys a convolutional layer, a normalization layer, an activation function, another convolutional layer, and a normalization layer. The two normalization layers work in conjunction with the corresponding convolutional layers to standardize features, improving the stability of feature extraction. The residual path directly passes the input features to the end of the main path, where they are element-wise added to the features processed by the main path to achieve fusion, and then output after passing through the activation function. Multiple sets of residual units work together to gradually extract local features of the outer surface temperature curve, and the input dimension is sequentially mapped from a third preset dimension to 64, 128, and 256 dimensions. The Transformer structure has a first preset number of attention heads, a second preset number of network layers, and a third preset number of feature dimensions. It is used to construct the global dependency relationship of temperature features, receive the 256-dimensional feature vector output by the ResNet residual structure, and perform global feature fusion and decoding. The RKT-Net model also includes a fully connected layer, which is used to restore the feature vector output by the Transformer structure to a preset length, so as to keep the dimension consistent with the axial temperature curve of the inner surface. The RKT-Net model takes the third preset dimension of the outer surface axial average temperature curve as input and outputs the inner surface axial temperature curve.

5. The method according to claim 3, characterized in that, It also includes training the RKT-Net model, specifically including: Supervised training is carried out using a training subset of the predicted training set, and the predicted subset of the predicted training set is used to verify the training accuracy of the RKT-Net model. Normalization is performed on the input data of the training subset and the prediction subset; Perform restoration processing on the output data of the RKT-Net model; The adaptive learning rate is set to the first preset learning rate. The ReduceLROnPlateau strategy is adopted. The factor is set to the first preset factor, the patience is set to the first preset patience, the batch size is set to the first preset batch, and the maximum number of iterations is the first preset number. If the validation set loss does not decrease for the first preset number of consecutive rounds, the training is stopped and the best network weights are retained. After training, the accuracy is verified using the predicted subset. The absolute average value between the predicted temperature inside the cement rotary kiln and the actual temperature must be less than the preset temperature, and the degree of fit between the predicted temperature inside the cement rotary kiln and the actual temperature must be greater than the preset fitting value.

6. The method according to claim 1, characterized in that, It also includes preprocessing the outer surface cylindrical scan data, specifically including: The outer surface temperature cylinder scan data is split according to the sampling time to obtain the outer surface temperature field matrix of the first preset dimension. The integrity of the data is verified by thermal visualization, and abnormal missing data is removed. The sampling frequency of the outer surface temperature cylinder scan data is dynamically adjusted according to the rotary kiln rotation speed to ensure that the spatial overlap rate of adjacent sampling points is greater than a preset percentage. Based on the rotation period of the rotary kiln, the sampling start position deviation of each of the outer surface temperature field matrices is corrected so that the row index of all the outer surface temperature field matrices is aligned with the axial physical position of the rotary kiln. The average value of the temperature field matrix after position alignment along the axis is obtained to obtain the original axial temperature curve of the outer surface in the second preset dimension; the original axial temperature curve of the outer surface is downsampled to be unified into the average axial temperature curve of the outer surface in the third preset dimension.

7. A method for predicting the thickness of kiln lining and kiln bricks according to any one of claims 1-6, characterized in that, include: Two sets of operating data of the cement rotary kiln are collected based on a preset time interval. Each set of data includes the outer surface temperature field matrix generated by preprocessing the outer surface cylinder sweep data, and the corresponding inner surface axial temperature curve output by the RKT-Net model. The data pairs consisting of the two sets of external surface temperature field matrices and the corresponding internal surface axial temperature curves are input into the RKH-Net model, and the kiln skin thickness distribution matrix and kiln brick thickness distribution matrix of the entire kiln axis of the cement rotary kiln are output. The RKH-Net model is trained using a quadruple consisting of the outer surface temperature field matrix, the corresponding inner surface axial temperature curve, the corresponding kiln skin thickness distribution matrix, and the corresponding kiln brick thickness distribution matrix at two different time points as samples. Furthermore, it constructs constraints based on the temperature data sets at the two different time points to solve the under-constraint problem in predicting the thickness of the kiln skin and kiln bricks.

8. The method according to claim 7, characterized in that, It also includes building the RKH-Net model, specifically including: The RKH-Net model includes dual convolutional input channels, an Encoder module, a Transformer module, multiple Decoder modules, and a CBAM attention mechanism. The dual convolutional input channel is used to extract local features from the two sets of temperature data and then superimpose them; the CBAM attention mechanism enhances the response of key temperature regions that are strongly correlated with changes in kiln skin and kiln brick thickness. The RKH-Net model takes two sets of data pairs consisting of the outer surface temperature field matrix and the inner surface axial temperature curve at different time points as inputs, and outputs the kiln skin thickness distribution matrix and the kiln brick thickness distribution matrix. The output results must satisfy the following conditions: the average absolute value between the predicted and actual values ​​of the kiln lining thickness is less than the first preset thickness, and the average absolute value between the predicted and actual values ​​of the kiln brick thickness is less than the second preset thickness.

9. The method according to claim 7, characterized in that, It also includes training the RKH-Net model, specifically including: The training set of the RKH-Net model is constructed based on the voxel simulation heat conduction database and the overall simulation heat conduction database. The samples of the training set are four-tuple samples consisting of the outer surface temperature field matrix, the corresponding inner surface axial temperature curve, the corresponding kiln skin thickness distribution matrix, and the corresponding kiln brick thickness distribution matrix at two different time points. The training set is divided into a training subset and a prediction subset according to a preset ratio. Normalization is performed on the outer surface temperature field matrix and the inner surface axial temperature curve of the training subset and the prediction subset; The Adam optimizer is used, the adaptive learning rate is set to the second preset learning rate, the ReduceLROnPlateau strategy is adopted, the factor is set to the second preset factor, the patience is set to the second preset patience, the batch size is set to the second preset batch, the maximum number of iterations is the second preset number, the validation set loss is continuously monitored, and if the validation set loss does not decrease for the second preset number of consecutive rounds, the training is stopped and the best network weights are retained. The output data of the RKH-Net model is restored to map the abstract numerical values ​​output by the model into kiln skin thickness and kiln brick thickness values ​​with actual physical meaning. After training, the accuracy is verified using the predicted subset. The absolute average value between the predicted and actual values ​​of the kiln lining thickness must be less than the first preset thickness, and the absolute average value between the predicted and actual values ​​of the kiln brick thickness must be less than the second preset thickness.

10. A system for predicting the internal temperature and kiln lining / brick thickness of a cement rotary kiln, characterized in that, include: The infrared data acquisition module is used to collect the temperature data of the outer surface of the cement rotary kiln using an infrared cylinder scanner, and transmit the collected cylinder scan data to the data preprocessing module. The data preprocessing module is used to perform preprocessing operations on the outer surface temperature cylinder scan data, including splitting, removing abnormal missing data, dynamically adjusting the sampling frequency, correcting the sampling start position deviation, calculating the axial mean, and downsampling, to generate an outer surface temperature field matrix of a first preset dimension and an outer surface axial average temperature curve of a third preset dimension. The outer surface temperature field matrix is ​​then transmitted to the RKH-Net model module, and the outer surface axial average temperature curve is transmitted to the RKT-Net model module. The RKT-Net model module is used to input the axial average temperature curve of the outer surface and output the axial temperature curve of the inner surface. The RKT-Net model module has a built-in RKT-Net model containing a ResNet residual structure, a Transformer structure and a fully connected layer. It uses the rotary kiln internal temperature prediction training set and completes the training through normalization, restoration processing and ReduceLROnPlateau learning rate strategy and early stopping mechanism. The training accuracy meets the requirement that the absolute average value between the predicted temperature inside the cement rotary kiln and the actual temperature is less than the preset temperature and the fitting degree is greater than the preset fitting value. The RKH-Net model module is used to input two sets of data pairs consisting of the outer surface temperature field matrix and the inner surface axial temperature curve at different time points, and output the kiln skin thickness distribution matrix and the kiln brick thickness distribution matrix. The RKH-Net model module has a built-in RKH-Net model containing dual convolution input channels, an Encoder module, a Transformer module, a multi-Decoder module and a CBAM attention mechanism. It is based on a quadruplet training set constructed from a voxel simulation heat conduction database and a global simulation heat conduction database. The Adam optimizer, ReduceLROnPlateau learning rate strategy, early stopping mechanism and normalization and restoration processing are used to complete the training. The training accuracy meets the requirement that the absolute average value between the predicted value and the true value of the kiln skin thickness is less than a first preset thickness, and the absolute average value between the predicted value and the true value of the kiln brick thickness is less than a second preset thickness. The data storage module is used to store the outer surface temperature cylinder scan data, the data preprocessed by the data preprocessing module, the training sets of the RKT-Net model module and the RKH-Net model module, and the trained model weights. The result output module is used to receive the axial temperature curve of the inner surface output by the RKT-Net model module and the kiln skin thickness distribution matrix and the kiln brick thickness distribution matrix output by the RKH-Net model module, and convert them into visual charts or numerical reports.

Images