Sintering process physical quantity prediction method based on multi-model fusion

Abstract

The invention discloses a sintering process physical quantity prediction method based on multi-model fusion, and belongs to the technical field of sintering process optimization. The method comprises the following steps: firstly, acquiring unstable heat distribution data under different carbon distribution working conditions through fluid simulation, extracting dynamic heat characteristics from the unstable heat distribution data, and constructing mixed characteristics by combining static coke distribution characteristics; a multi-model fusion neural network composed of a convolutional neural network, a bidirectional long-short-term memory network and Transform is adopted to carry out spatial-temporal feature joint learning, and prediction of future time step heat distribution is realized. And when actually measured data are missing, predicting initial heat data based on static characteristics through a trained initial step generator, and iteratively updating input by adopting an autoregression mode to complete full-time-domain heat evolution prediction. According to the method, the precision and generalization ability of heat prediction in the sintering process are effectively improved, and process optimization, energy conservation and emission reduction are supported.

Description

Technical Field

[0001] This invention relates to the field of sintering process optimization technology, and in particular to a method for predicting physical quantities in sintering processes based on multi-model fusion. Background Technology

[0002] Sintering is a crucial process in steel production, and its heat distribution directly impacts product quality and energy consumption. Traditional methods, relying primarily on empirical formulas and single models for heat prediction, suffer from the following problems: 1. Low prediction accuracy, making it difficult to adapt to complex working conditions; 2. The model has poor generalization ability and suffers from severe overfitting; 3. Lack of joint modeling of static process parameters and dynamic spatiotemporal characteristics; 4. It cannot effectively handle complex relationships between high-dimensional features.

[0003] In existing technologies, single machine learning models such as support vector machines and random forests are widely used in sintering process prediction. However, these methods suffer from limited feature representation capabilities and difficulty in capturing nonlinear relationships. While deep learning models such as neural networks have powerful feature learning capabilities, they are prone to overfitting and require a large amount of data.

[0004] Therefore, there is an urgent need for a method for predicting the heat of sintering processes that can combine the advantages of multiple models and improve prediction accuracy and generalization ability. Summary of the Invention

[0005] To address the above problems, this invention provides a method for predicting physical quantities in sintering processes based on multi-model fusion. It aims to achieve accurate and rapid prediction of the multiphase thermal field in the sintering process by integrating convolutional neural networks, long short-term memory networks, and attention mechanisms. It can also provide reliable initial input when measured data is missing, thus supporting process optimization and energy conservation and emission reduction.

[0006] This invention provides a method for predicting physical quantities in sintering processes based on multi-model fusion, comprising: S1. Transient simulation of iron ore sintering process under different carbonization conditions is performed using fluid simulation software to obtain unsteady heat distribution data at different depths inside the material layer at each time step. S2, extract dynamic heat characteristics from unsteady heat distribution data, and construct static coke distribution characteristics based on the ratio of coke to material in the depth direction of the material layer under the working conditions; S3, the dynamic heat characteristics are spliced ​​with the static coke distribution characteristics to obtain mixed characteristics, and heat prediction data is obtained based on a neural network model.

[0007] Furthermore, the neural network model consists of a convolutional neural network, a bidirectional long short-term memory network, and a Transformer model. When training the neural network model, the Adam optimizer is used, the mean squared error is used as the loss function, and a Dropout layer is added to the Transformer model to prevent overfitting.

[0008] By employing convolutional neural networks (CNNs) to extract spatial features, bidirectional long short-term memory networks (Bi-LSTM) to capture temporal evolution patterns, and combining these with a Transformer model for feature fusion, the model's ability to model complex spatiotemporal coupling relationships during sintering is significantly improved. Training with the Adam optimizer and mean squared error loss function effectively accelerates convergence and optimizes prediction accuracy. Adding a Dropout layer to the Transformer model significantly suppresses overfitting and enhances its generalization performance and robustness under different operating conditions.

[0009] Furthermore, S3 specifically includes: The hybrid features are input into a convolutional neural network to extract spatial features while preserving their temporal structure, thus obtaining the first feature. The hybrid temporal features are input into a bidirectional long short-term memory network to learn the temporal evolution of the hybrid temporal features and output a second feature enriched with the heat information of the current time window. The second feature and the static coke distribution feature are input into the Transformer model. The static coke distribution feature is used as the query vector, and the second feature is used as the key vector and value vector. The third feature is obtained by weighted fusion through the attention mechanism. The third feature is input into a fully connected layer for mapping, and the heat prediction data for several future time steps is output.

[0010] Through a phased feature extraction and fusion mechanism, the CNN first extracts local spatial features while maintaining the temporal structure, then the Bi-LSTM learns the temporal evolution law, and finally the attention mechanism of the Transformer explicitly fuses the static working condition features, enabling the model to more accurately understand the impact of static carbon distribution parameters on dynamic heat evolution, thereby improving the prediction accuracy and physical consistency of the heat field at multiple time steps.

[0011] Furthermore, the method also includes: When measured data is missing, the dynamic thermal characteristics of the first few time steps are predicted by a pre-trained initial step generator based on the static coke distribution characteristics, and used as the initial input data for the neural network model. An autoregressive approach is used to iteratively update the input of the neural network model, enabling the prediction of thermal evolution across the entire time domain.

[0012] Introducing an initial step generator based on static features provides reliable initial input for the neural network model when measured data is missing or highly noisy, significantly improving the system's applicability and stability in real-world industrial scenarios. Combined with an autoregressive iterative prediction mechanism, continuous prediction of thermal evolution across the entire time domain is achieved, enhancing the model's consistency and reliability in long-term prediction tasks.

[0013] Furthermore, the method of iteratively updating the input of the neural network model using an autoregressive approach to achieve prediction of the full-time-domain thermal evolution specifically includes: S41, the initial step generator generates the predicted dynamic thermal features for the first N time steps; S42, The predicted dynamic heat characteristics and static coke distribution characteristics are concatenated and input into the neural network model to obtain the predicted heat characteristics for the subsequent K time steps; S43, the predicted heat features of the K time steps are concatenated with the predicted dynamic heat features of the next NK time steps input in the previous step to form new predicted dynamic heat features. S44. Repeat S42 and S43 until the desired full-time-domain thermal characteristic prediction results are obtained.

[0014] By employing a specially designed autoregressive iterative process, the input time window is gradually updated, effectively reducing the problem of error accumulation in long-term predictions and improving the stability and controllability of the prediction process. This mechanism ensures the continuity of prediction results in the time dimension, better conforms to the physical evolution law of the sintering process, and is suitable for real-time optimization and digital twin systems.

[0015] Furthermore, the input data construction process of the initial step generator is as follows: Extract static coke distribution characteristics, including the depth of different layers in the sintering material layer, the proportion of coke in each layer, and the mass of added coke; The extracted dynamic thermal features are encoded using location features, and normalized deep features are defined. The larger the value, the deeper the location of the collected heat feature; The extracted dynamic thermal features are encoded using time features, and normalized time step features are defined. The smaller the value, the earlier the data collection time step; Combining normalized deep features Features of normalized time steps Data augmentation is performed using power functions, exponential functions, and trigonometric functions to obtain enhanced dynamic features. The enhanced dynamic features are combined with the static coke distribution features to generate the input data for the initial step generator.

[0016] By performing structured encoding and data augmentation on static parameters and dynamic features, rich and physically meaningful feature inputs were constructed, significantly improving the prediction accuracy of the initial step generator. The introduction of normalized depth and time features enhanced the model's ability to model sintering lag and timing advance effects, thus enabling the generation of high-quality initial thermal field data even without measured data.

[0017] Furthermore, between S2 and S3, the following is also included: The dynamic heat characteristics and static coke distribution characteristics were standardized using the Z-Score method, transforming all characteristics into standard normal distribution data with a mean of 0 and a standard deviation of 1.

[0018] The Z-Score normalization method is used to process the input features, eliminating the influence of different physical dimensions on model training and improving the model's convergence speed and training stability. The normalized feature distribution is more consistent with the data assumptions of neural networks, which helps to improve the model's adaptability to various working conditions and the final prediction accuracy.

[0019] Furthermore, the extraction of dynamic heat features from unsteady heat distribution data specifically includes: A fixed-length time window is extracted from the unsteady heat distribution data, and four different types of heat data are extracted from this window as dynamic heat features, including solid-phase source term heat, gas-phase reaction heat, solid-phase reaction heat, and gas-solid convection heat transfer.

[0020] By clearly extracting four key heat types—solid-phase source heat, gas-phase reaction heat, solid-phase reaction heat, and gas-solid convection heat transfer—as dynamic features, the model is able to comprehensively capture the multiphase and multi-mechanism heat transfer and transformation behavior during sintering, thereby improving the physical rationality and engineering applicability of the prediction results.

[0021] Compared with existing technologies, the beneficial effects of this invention are as follows: by integrating convolutional neural networks, bidirectional long short-term memory networks, and Transformer attention mechanisms, the accuracy of thermal field prediction and spatiotemporal modeling capabilities in the sintering process are significantly improved; the introduction of an initial step generator based on static features effectively solves the prediction startup problem under conditions of missing or noisy measured data, enhancing the practicality and robustness of the model; the overall method combines physical consistency and real-time performance, providing reliable technical support for sintering process optimization, energy consumption reduction, and digital twin systems. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this drawing 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 only some embodiments of this drawing. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the entire model; Figure 2 This is a schematic diagram of the sintering material layer depth scale; Figure 3 This is a graph showing the model loss during the training phase. Figure 4 Bland-Altman plot of model predictions versus true values ​​during the inference phase; Figure 5 This is a graph showing the change of autoregressive error during the inference phase with the number of iterations. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0025] This invention provides a method for predicting physical quantities in sintering processes based on multi-model fusion, specifically including the following steps: S1 uses fluid simulation software to perform transient simulation of the iron ore sintering process under different carbonization conditions, and obtains the unsteady heat distribution data at different depths inside the material layer at each time step.

[0026] S2 extracts dynamic heat characteristics from unsteady heat distribution data and constructs static coke distribution characteristics based on the ratio of coke to material in the depth direction of the material layer under the working conditions.

[0027] Specifically, a fixed-length time window is extracted from the unsteady heat distribution data, and then four different types of heat data are extracted from this window as dynamic heat features. Including solid-phase heat source terms Heat of gas-phase reaction Heat of solid-state reaction Gas-solid convection heat transfer Dynamic thermal characteristics It consists of a depth dimension and a time step dimension.

[0028] The process of constructing static coke distribution characteristics is as follows: For a sintered material layer with a depth of 800 mm, depth position measuring points were evenly set at 10 mm intervals along its depth direction, resulting in a total of 80 measuring points. These measuring points were used to collect heat data at the corresponding locations, including solid phase source heat, gas phase reaction heat, solid phase reaction heat, and gas-solid convection heat transfer.

[0029] To characterize the fuel distribution at different depths, each measuring point is assigned a characteristic value, namely the mass ratio of coke to material at that location (coke ratio). In actual sintering processes, a layered coke distribution strategy is often adopted, with different coke ratios set for different depth layers, such as... Figure 2 As shown, the darker color in the upper layer represents a higher specific gravity (coke ratio) of fuel (coke) to material; while the lighter color in the lower layer represents a lower coke ratio at that location. Arranging these 80 coke ratio values ​​in order of their corresponding depth, a one-dimensional vector is constructed, namely the static coke distribution feature vector.

[0030] Furthermore, to address the issue of missing measured data, this invention designs an initial step generator, such as... Figure 1 As shown, this initial step generator is used to generate data based on static coke distribution characteristics. Complete the dynamic thermal characteristics of the initial few time steps The initial step generator generates predictions to produce initial input data that can be used by neural network models. Furthermore, if the measured data is not missing, the initial step generator can also be used to monitor the measured data and improve the noise level of the data.

[0031] The initial step generator is an XGBoost model, and the process of constructing its input data is as follows: Static coke distribution characteristics related to operating conditions were extracted, including the depth of different layers in the sintering material, the proportion of coke in each layer, and the mass of added coke. .

[0032] Simultaneously, dynamic thermal characteristics were collected at various depth locations and time steps. Perform feature encoding: Location feature encoding: Defining normalized deep features The larger the value, the deeper the location of the collected heat feature, and the later the time when that location enters the high-temperature region during the sintering process; Time feature encoding: Defining normalized time step features The smaller the value, the earlier the heat data was collected in a given time step. This characteristic is consistent with the aforementioned normalized depth feature. Together, they indicate whether the heat data at the collected location has entered or exited the high-temperature sintering zone.

[0033] Combining normalized deep features Features of normalized time steps Data augmentation is performed using power functions, exponential functions, and trigonometric functions to obtain enhanced dynamic features. ; Will enhance dynamic features Characteristics of static coke distribution The data is stitched together to generate the input data for the initial step generator, ensuring that the heat data sampled at each depth location and time step has complete feature support.

[0034] Furthermore, the initial step generator can also be used to validate its output heat prediction data, as follows: The initial step generator is based on normalized deep features. Normalized time characteristics The system predicts the static coke distribution characteristics under corresponding operating conditions, outputting calorific data for a specific time step. This time step is compared with the normalized time features in the input. Correspondingly, therefore, it can be based on the same Values ​​and different depth positions The value is then used to call the initial step generator again to predict the heat data at that moment for subsequent comparison.

[0035] During the validation process, the prediction result of the neural network model at a certain time step is denoted as Q1, and the prediction result of the initial step generator at the same time step is denoted as Q2. If it is assumed that the initial step generator has higher prediction reliability, then Q2 is used as the reference truth value, and Q1 is subjected to statistical testing. For example, a t-test can be used to calculate the statistic. If the statistic exceeds the preset confidence interval, it is considered that the prediction error of the neural network model at that time step is large, meaning that errors may accumulate during its iteration process, and the reliability of subsequent prediction results will decrease. In this case, it is recommended to pause the prediction, wait for new sensor monitoring data input, and restart the iteration process.

[0036] Furthermore, if measured data is available, a statistical test can be performed between the measured data Q3 and the output Q2 of the initial step generator to assess whether the measured data is affected by noise. If the test shows that Q3 and Q2 are significantly different, a Kalman filter algorithm can be considered to correct the new input data.

[0037] Furthermore, the Z-Score method is used to standardize the dynamic heat characteristics, static coke distribution characteristics, and enhanced dynamic characteristics, transforming all characteristics into standard normal distribution data with a mean of 0 and a standard deviation of 1, thus eliminating the influence caused by the different dimensions of different physical quantities. The Z-Score formula is: ; Where z represents the standardized data; The mean of the sample; This is the population mean; The population standard deviation; n is the sample size.

[0038] S3, the dynamic heat characteristics are spliced ​​with the static coke distribution characteristics to obtain mixed characteristics, and heat prediction data is obtained based on a neural network model.

[0039] Specifically, the neural network model consists of convolutional neural networks, bidirectional long short-term memory networks, and Transformer models, such as... Figure 1 As shown, the data prediction process includes: The hybrid features are input into a convolutional neural network to extract spatial features while preserving their temporal structure, thus obtaining the first feature. The hybrid temporal features are input into a bidirectional long short-term memory network to learn the temporal evolution of the hybrid temporal features and output a second feature enriched with the heat information of the current time window. The second feature and the static coke distribution feature are input into the Transformer model. The static coke distribution feature is used as the query vector, and the second feature is used as the key vector and value vector. The third feature is obtained by weighted fusion through the attention mechanism. The third feature is input into a fully connected layer for mapping, and the heat prediction data for several future time steps is output.

[0040] The training of a neural network model uses mean squared error (MSE) as the loss function. The model loss during training is as follows: Figure 3 As shown, the definition of the MSE function is as follows: ; Where N is the total number of samples; This represents the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0041] To improve the model's generalization performance and robustness, the neural network model used in this invention adds a Dropout layer to the Transformer model to prevent overfitting and ensure prediction performance. The Dropout layer is typically placed after the activation function. Taking a network with M neurons as an example, the Dropout layer is constructed as follows: ; in, Let M be a Bernoulli random mask vector of length. p is the retention probability; Bernoulli follows the Bernoulli distribution; This is the output matrix of the layer preceding the Dropout layer; L is the number of samples in the batch; This is the output matrix after applying Dropout.

[0042] For each neuron, a Bernoulli vector r with values ​​of 0 or 1 is randomly generated according to the retention probability p, and the neuron's output is randomly set to zero.

[0043] The neural network model is trained using the Adam optimizer, which uses an exponentially weighted moving average to estimate the momentum and second moment of the gradient, i.e.: ; in, The first moment estimate of the gradient at the current time step; The first moment estimate of the gradient at the current time step; This is the second moment estimate of the gradient at the current time step; Second moment estimate of the gradient of the previous time step , The exponential decay rate hyperparameter is typically taken as... ; This represents the gradient at the current time step.

[0044] To prevent large initialization deviations, the above state variables are normalized: ; in, This is the first-order moment estimate after bias correction; This is the second-order moment estimate after bias correction.

[0045] Update gradient: ; in, The updated gradient; The values ​​of the model parameters at the current time step; The values ​​of the model parameters at the previous time step; The learning rate; is the numerical stability constant.

[0046] The initial learning rate η selected during training is 5 × 10⁻⁶. -4 Furthermore, the validation set loss is monitored during training, and an early stopping mechanism is triggered when there is no improvement for 10 consecutive rounds, causing the model training to stop.

[0047] Cosine annealing is used to train neural network models. In the early stages of training, the model needs to iterate and update itself rapidly, thus requiring a large learning rate. However, as the number of training epochs increases, the model gradually converges and stabilizes. At this point, a learning rate that is too large can easily disrupt the learned structure and make the model unstable. Therefore, a variable learning rate method can be used to balance rapid convergence in the early stages with fine-grained optimization in the later stages. A common method is to change the step size according to the cosine annealing pattern.

[0048] When training the initial step generator XGBoost model, a fixed learning rate of 0.03 is used because tree models are less sensitive to learning rates, and using a fixed learning rate is more efficient and convenient.

[0049] Error calibration and performance evaluation of neural network models: Every 20 time steps, the model output is calibrated using the "true heat values ​​from the training set" to reduce the accumulation of errors in long-term time-series predictions. Performance is evaluated from two perspectives: goodness of fit and correlation. The specific calculation methods for these metrics are as follows: R² coefficient of determination: ; in, The coefficient of determination; The sample mean of the true values; n is the number of samples.

[0050] Pearson correlation coefficient: ; Where r is the Pearson correlation coefficient; The standard deviation of the true value; The standard deviation of the predicted values; Cov represents covariance.

[0051] S4 uses an autoregressive approach to iteratively update the input of the neural network model, thereby achieving prediction of the full-time-domain thermal evolution. The specific process is as follows: S41, the initial step generator generates the predicted dynamic thermal features for the first N time steps; S42, The predicted dynamic heat characteristics and static coke distribution characteristics are concatenated and input into the neural network model to obtain the predicted heat characteristics for the subsequent K time steps; S43, the predicted heat features of the K time steps are concatenated with the predicted dynamic heat features of the next NK time steps input in the previous step to form new predicted dynamic heat features. S44. Repeat S42 and S43 until the desired full-time-domain thermal characteristic prediction results are obtained.

[0052] The trained neural network model was tested on a test set. First, an initial step generator was used to generate the thermal feature prediction results for the initial N time steps based on the operating conditions. Then, the neural network model was used to iterate continuously using its autoregressive inference function, and finally the model inference prediction results for the entire time domain were obtained.

[0053] Compare the predicted result with the true values ​​on the test set and draw its Bland-Altman diagram, as shown below. Figure 4 As shown, the prediction error of the neural network model changes with the number of iterations during inference, as illustrated in the figure. Figure 5 As shown, it can be observed that in the initial few iterations, the autoregressive error of the model is small; however, as the number of iterations increases, the prediction error of the neural network model begins to grow rapidly and fluctuate significantly; the more iterations there are, the more unstable the prediction model becomes, and at this point the prediction results are close to divergent and have low reliability.

[0054] It should be noted that the present invention is not limited to the above-described embodiments. The above embodiments are merely examples, and any embodiments that have the same structure and perform the same effects as the technical concept within the scope of the present invention are included within the scope of the present invention. Furthermore, various modifications that can be conceived by those skilled in the art to the embodiments, and other ways of constructing by combining some of the constituent elements of the embodiments, without departing from the spirit of the present invention, are also included within the scope of the present invention.

Claims

1. A method for predicting physical quantities in sintering processes based on multi-model fusion, characterized in that, include: S1. Transient simulation of iron ore sintering process under different carbonization conditions is performed using fluid simulation software to obtain unsteady heat distribution data at different depths inside the material layer at each time step. S2, extract dynamic heat characteristics from unsteady heat distribution data, and construct static coke distribution characteristics based on the ratio of coke to material in the depth direction of the material layer under the working conditions; S3, the dynamic heat characteristics are spliced ​​with the static coke distribution characteristics to obtain mixed characteristics, and heat prediction data is obtained based on a neural network model.

2. The method for predicting physical quantities of sintering process based on multi-model fusion as described in claim 1, characterized in that, The neural network model consists of a convolutional neural network, a bidirectional long short-term memory network, and a Transformer model. The Adam optimizer is used when training the neural network model, the mean squared error is used as the loss function, and a Dropout layer is added to the Transformer model to prevent overfitting.

3. The method for predicting physical quantities of sintering process based on multi-model fusion as described in claim 2, characterized in that, S3 specifically includes: The hybrid features are input into a convolutional neural network to extract spatial features while preserving their temporal structure, thus obtaining the first feature. The hybrid temporal features are input into a bidirectional long short-term memory network to learn the temporal evolution of the hybrid temporal features and output a second feature enriched with the heat information of the current time window. The second feature and the static coke distribution feature are input into the Transformer model. The static coke distribution feature is used as the query vector, and the second feature is used as the key vector and value vector. The third feature is obtained by weighted fusion through the attention mechanism. The third feature is input into a fully connected layer for mapping, and the heat prediction data for several future time steps is output.

4. The method for predicting physical quantities of sintering processes based on multi-model fusion as described in claim 1, characterized in that, The method further includes: When measured data is missing, the dynamic thermal characteristics of the first few time steps are predicted by a pre-trained initial step generator based on the static coke distribution characteristics, and used as the initial input data for the neural network model. An autoregressive approach is used to iteratively update the input of the neural network model, enabling the prediction of thermal evolution across the entire time domain.

5. The method for predicting physical quantities of sintering process based on multi-model fusion as described in claim 4, characterized in that, The method of iteratively updating the input of the neural network model using an autoregressive approach to predict the evolution of heat over the entire time domain specifically includes: S41, the initial step generator generates the predicted dynamic thermal features for the first N time steps; S42, The predicted dynamic heat characteristics and static coke distribution characteristics are concatenated and input into the neural network model to obtain the predicted heat characteristics for the subsequent K time steps; S43, the predicted heat features of the K time steps are concatenated with the predicted dynamic heat features of the next NK time steps input in the previous step to form new predicted dynamic heat features. S44. Repeat S42 and S43 until the desired full-time-domain thermal characteristic prediction results are obtained.

6. The method for predicting physical quantities of sintering process based on multi-model fusion as described in claim 4, characterized in that, The input data construction process for the initial step generator is as follows: Extract static coke distribution characteristics, including the depth of different layers in the sintering material layer, the proportion of coke in each layer, and the mass of added coke; The extracted dynamic thermal features are encoded using location features, and normalized deep features are defined. The larger the value, the deeper the location of the collected heat feature; The extracted dynamic thermal features are encoded using time features, and normalized time step features are defined. The smaller the value, the earlier the data collection time step; Combining normalized deep features Features of normalized time steps Data augmentation is performed using power functions, exponential functions, and trigonometric functions to obtain enhanced dynamic features. The enhanced dynamic features are combined with the static coke distribution features to generate the input data for the initial step generator.

7. The method for predicting physical quantities of sintering processes based on multi-model fusion as described in claim 1, characterized in that, Between S2 and S3, the following is also included: The dynamic heat characteristics and static coke distribution characteristics were standardized using the Z-Score method, transforming all characteristics into standard normal distribution data with a mean of 0 and a standard deviation of 1.

8. The method for predicting physical quantities of sintering process based on multi-model fusion as described in claim 1, characterized in that, The extraction of dynamic heat features from unsteady heat distribution data specifically includes: A fixed-length time window is extracted from the unsteady heat distribution data, and four different types of heat data are extracted from this window as dynamic heat features, including solid-phase source term heat, gas-phase reaction heat, solid-phase reaction heat, and gas-solid convection heat transfer.

Images