A sintering endpoint prediction model based on spatio-temporal feature fusion and a system thereof

Abstract

This invention discloses a sintering endpoint prediction model and system based on spatiotemporal feature fusion. The model includes: acquiring sintering process data and constructing long-term variable matrices using time series methods. M l and short-term variable matrix M s A sintering endpoint prediction model was established based on the ResNet network and Transformer architecture; the long-term variable matrix was used to... M l and short-term variable matrix M s After matrix reconstruction, the input is fed into the prediction model, where it undergoes feature extraction and reconstruction to obtain a regression input vector. Following iterative training, the predicted value is output. This model divides the input parameters into long-term and short-term variables based on the spatial distribution characteristics of the sintering process. Simultaneously, it constructs a variable matrix based on the temporal characteristics of the variables, fully considering the impact of variables with different temporal and spatial characteristics on the sintering endpoint. The prediction system built based on this model eliminates the need for manual feature extraction, enabling the identification of deeper feature information and significantly improving the accuracy of the prediction model.

Description

Technical Field

[0001] This invention relates to a sintering endpoint prediction model, and more particularly to a sintering endpoint prediction model and system based on spatiotemporal feature fusion, belonging to the field of iron ore sintering production. Background Technology

[0002] The production process employs a long-flow "blast furnace-converter" method, with sinter accounting for over 70% of the blast furnace feed. A rational and stable sintering process plays a crucial role in the efficient and smooth operation of the entire ironmaking process, and the sintering endpoint is the core of stable control throughout the process. The sintering endpoint is dynamically and temporally variable, influenced by numerous factors such as raw material properties and operating parameters, and can only be judged and controlled through manual experience. Advance prediction of the sintering endpoint can provide data support for on-site operators, serving as a prerequisite and foundation for intelligent control of the sintering endpoint. Currently, researchers mainly use a mechanism-based, data-driven model to establish the mapping relationship between process monitoring data and the sintering endpoint. However, limited by traditional modeling methods and computational capabilities, the established models rely on prior induction and summarization of features, making it difficult to comprehensively consider the temporal and spatial relationships between variables. Summary of the Invention

[0003] To address the problems existing in the prior art, the first objective of this invention is to provide a sintering endpoint prediction model based on spatiotemporal feature fusion. This model divides the input parameters into long-term and short-term variables according to the spatial distribution characteristics of the sintering process, and constructs a variable matrix based on the temporal characteristics of the variables. It fully considers the influence of variables with different time and spatial characteristics on the sintering endpoint. Compared with the prior art, this prediction model has the advantages of longer prediction step size and higher accuracy.

[0004] The second objective of this invention is to provide a sintering endpoint prediction system based on spatiotemporal feature fusion, comprising a programmable logic controller, a memory containing the aforementioned prediction model, and a processor. This system, based on convolutional neural networks and Transformers, integrates self-attention mechanisms, residual connections, and other operations to achieve automatic extraction of temporal and spatial data features. It eliminates the need for manual feature extraction, enabling the identification of deeper feature information and significantly improving the accuracy of the prediction model.

[0005] To achieve the above technical objectives, this invention provides a sintering endpoint prediction model based on spatiotemporal feature fusion, comprising:

[0006] Step S1: Obtain sintering process data and construct long-term variable matrices M using time series methods. l and short-term variable matrix M s ;

[0007] Step S2: Establish a sintering endpoint prediction model based on ResNet network and Transformer architecture;

[0008] Step S3: The long-term variable matrix M l and short-term variable matrix M s After matrix reconstruction, the input is fed into the prediction model. The model then undergoes feature extraction and feature reconstruction to obtain the regression input vector. After iterative training, the predicted value of the sintering endpoint is output.

[0009] The matrix obtained by splicing and deformation is a square matrix.

[0010] As a preferred option, the long-term variable matrix M l The sintering process data includes: fuel ratio, fuel feed rate, moisture content of the mixture, feed rate of the mixture, material level in the mixing tank, furnace temperature of the ignition furnace, temperature of the first stage of the holding furnace, and material layer thickness in different channels, all collected in minutes.

[0011] As a preferred embodiment, the short-term variable matrix M s The sintering process data includes: trolley speed in minutes, negative pressure value of No. 1 wind box, negative pressure of No. 24, No. 28 and No. 30 wind boxes, exhaust gas temperature on the north and south sides, exhaust gas temperature and negative pressure of the main flue on the north and south sides, frequency of the main exhaust fan on the north and south sides and its valve opening, as well as TRP, BRP and BTP soft measurement results.

[0012] As a preferred embodiment, the framework of the sintering endpoint prediction model includes: a sequence feature extraction module based on ResNet network, a feature reconstruction module, and a regression prediction module based on Transformer.

[0013] As a preferred embodiment, the feature extraction module consists of a feature extraction branch for the long-term variable matrix and a feature extraction branch for the short-term variable matrix, with each branch using a ResNet network for feature extraction.

[0014] As a preferred embodiment, the stride of the first convolutional layer and the max pooling layer in the ResNet network of the long-term variable feature extraction branch is set to (2,1).

[0015] As a preferred embodiment, the ResNet network in the short-term variable feature extraction branch does not have a max pooling layer.

[0016] As a preferred embodiment, the stride of the 3×3 convolution kernel in the ResNet network for the long-term variable feature extraction branch and the short-term variable feature extraction branch is set to 1, and downsampling is not used.

[0017] As a preferred approach, after feature extraction, the long-term variable matrix Ml and short-term variable matrix M s All converted to a size of ( n*2n ,1) Eigenvectors.

[0018] As a preferred embodiment, the feature reconstruction process is as follows: The long-term variable matrix M... l and short-term variable matrix M s The resulting feature vectors are transformed into a matrix of size (n, 2n), and then concatenated to obtain a regression input matrix of size (2n, 2n).

[0019] Deep learning methods, represented by CNN and Transformer, possess excellent feature extraction capabilities, enabling the sintering endpoint prediction model to shift from point-to-point prediction to sequence-to-point and sequence-to-sequence prediction. Furthermore, by fully exploring the temporal and spatial dependencies within the sintering process variable sequences, and using the long-term variable matrix M... l and short-term variable matrix M s Using it as a training set for a prediction model can significantly improve the accuracy and performance of the prediction model.

[0020] As a preferred embodiment, the regression prediction module consists of a Transformer encoding layer, an attention mechanism weighted layer, and a fully connected layer.

[0021] The present invention also provides a sintering endpoint prediction system based on spatiotemporal feature fusion, which consists of a programmable logic controller (101), a memory (102) containing any of the prediction models described above, and a processor (103).

[0022] As a preferred embodiment, the programmable logic controller (101) is used to collect sintering production process data.

[0023] As a preferred embodiment, the memory (102) is used to store sintering production process data and prediction models.

[0024] As a preferred embodiment, the processor (103) is used to read sintering process production data and execute a prediction model.

[0025] Compared with the prior art, the beneficial technical effects of the technical solution of the present invention are as follows:

[0026] 1) The prediction model provided by this invention divides the input parameters into long-term and short-term variables according to the spatial distribution characteristics of the sintering process. At the same time, it constructs a variable matrix based on the temporal characteristics of the variables, which fully considers the influence of variables with different time and spatial characteristics on the sintering endpoint. Compared with the prior art, this prediction model has the advantages of longer prediction step and higher accuracy.

[0027] 2) The technical solution provided by this invention is based on convolutional neural networks and Transformers, and integrates self-attention mechanism, residual connection and other operations to realize the automatic extraction of temporal and spatial data features. It does not require manual experience to extract data features, and can identify deeper feature information, thereby greatly improving the accuracy of the prediction model. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the prediction model construction process provided in Embodiment 1 of the present invention;

[0029] Figure 2 This is a schematic diagram of the sintering endpoint prediction network in Embodiment 1 of the present invention;

[0030] Figure 3 This is a structural diagram of the prediction system provided in Embodiment 1 of the present invention;

[0031] Figure 4 This is the computer program interface for predicting the sintering endpoint in the prediction system provided in Embodiment 1 of the present invention. Detailed Implementation

[0032] To facilitate understanding of the present invention, a more comprehensive and detailed description of the invention will be provided below in conjunction with the accompanying drawings and preferred embodiments. It should be noted that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0033] Example 1

[0034] This embodiment provides a sintering endpoint prediction model based on spatiotemporal feature fusion:

[0035] S01: Collect 60-minute data for 14 variables, including fuel ratio, fuel feed rate, mixture moisture content, mixture feed rate, mixture tank level, ignition furnace temperature, first stage temperature of the holding furnace, and material layer thickness in six different channels, from minute-level sintering process data. After time alignment, data preprocessing, and data standardization, establish a long-term variable matrix Ml of size 14-60, as shown in Table 1.

[0036] Table 1

[0037]

[0038] S02: Collect 30-minute data points for 22 variables from minute-level sintering process data, including trolley speed, negative pressure value of wind box #1, negative pressure and exhaust gas temperature on the north and south sides of wind boxes #24, #28 and #30, exhaust gas temperature and negative pressure on the north and south main flues, frequency and valve opening of the north and south main exhaust fans, and TRP, BRP and BTP soft measurement results. After time alignment, data preprocessing and data standardization, establish a short-term variable matrix M of size 22×30. s As shown in Table 2.

[0039] Table 2

[0040]

[0041] S03: Input the constructed two-dimensional matrix into the sintering endpoint prediction model. This model is based on the ResNet network and Transformer architecture, and the framework includes: a sequence feature extraction module and a feature reconstruction module based on the ResNet network, and a regression prediction module based on the Transformer, such as... Figure 1 As shown in Table 3, the feature reconstruction module extracts a feature vector of size (2048,1) from the long-term variable matrix and the short-term variable matrix. The feature reconstruction module transforms the two feature vectors into two (32,64) matrices with 64 elements per row, and outputs a 64×64 feature vector that meets the input requirements of the subsequent regression module after concatenation.

[0042] Table 3

[0043]

[0044] The pre-training and fine-tuning process used a PyTorch 2.20 and Python 3.10 training environment. The selected optimizer was Adam, with default parameters and a batch size of 128. Regarding the training strategy, the model primarily employed a two-round training approach. The first round consisted of 200 generations with an initial learning rate of 1e-3. If the loss no longer decreased after 10 consecutive generations, the learning rate was reduced to 0.5 times its original value. The second round consisted of 100 generations. Inheriting the minimum learning rate from the first round, a cosine decay strategy was used to reduce the learning rate to 0.1 times its initial value. Early stopping was also employed to stop the model training promptly. The MSE function was used as the model's loss function. The training set was fed into the network for training, and a grid search method was used to determine the hyperparameters to achieve the desired training effect. The model's metrics are shown in Table 4.

[0045] Table 4

[0046]

[0047] The long-term variable matrix Ml and the short-term variable matrix Ms to be predicted were input into the model. The predicted sintering endpoint after 30 minutes was 28.85. The actual soft measurement result was 29.02, with an error of 0.17.

[0048] Taking the sintering endpoint prediction results over a production period as an example, the comparison between the model predictions and the actual soft-sensor results is as follows: Figure 2 As shown.

[0049] A spatiotemporal feature fusion sintering endpoint prediction system developed based on the above method includes: a programmable logic controller 101, a memory 102, and a processor 103. The memory 102 stores a computer program capable of implementing the above-described sintering endpoint prediction method. The background operations of this computer program are written in Java, and the front-end page is written in JavaScript. The processor 103 is used to read the sintering production process data from the programmable logic controller 101, retrieve and execute the computer program. The prediction system configuration diagram is shown below. Figure 3 As shown, the program interface is as follows: Figure 4 As shown.

Claims

1. A sintering endpoint prediction model based on spatiotemporal feature fusion, characterized in that, include: Step S1: Obtain sintering process data and construct long-term variable matrices using time series methods. M l and short-term variable matrix M s ; Step S2: Establish a sintering endpoint prediction model based on ResNet network and Transformer architecture; Step S3: Calculate the long-term variable matrix. M l and short-term variable matrix M s After matrix reconstruction, the input is fed into the prediction model. The model then undergoes feature extraction and feature reconstruction to obtain the regression input vector. After iterative training, the predicted value of the sintering endpoint is output. The framework of the sintering endpoint prediction model includes: a sequence feature extraction module based on ResNet network, a feature reconstruction module, and a regression prediction module based on Transformer. After feature extraction, the long-term variable matrix M l and short-term variable matrix M s All converted to a size of ( n*2n 1) eigenvectors; the process of reconstructing the features is as follows: converting the long-term variable matrix into eigenvectors. M l and short-term variable matrix M s The resulting feature vector is transformed into a form of size ( n , 2n ) matrix, and then by concatenating the top and bottom parts, we get a matrix of size ( 2n, 2n The regression input matrix; The feature extraction module consists of a feature extraction branch for the long-term variable matrix and a feature extraction branch for the short-term variable matrix, and each branch uses a ResNet network for feature extraction. The stride of the first convolutional layer and the max pooling layer in the ResNet network that performs feature extraction on the long-term variable matrix is ​​set to (2,1); no max pooling layer is set in the ResNet network that performs feature extraction on the short-term variable matrix; the stride of the 3×3 convolutional kernel in the ResNet networks that perform feature extraction on the long-term and short-term variable matrices is set to 1, and downsampling is not used.

2. The sintering endpoint prediction model based on spatiotemporal feature fusion according to claim 1, characterized in that: The long-term variable matrix M l The sintering process data includes: fuel ratio, fuel feed rate, moisture content of the mixture, feed rate of the mixture, material level in the mixing tank, furnace temperature of the ignition furnace, temperature of the first stage of the holding furnace, and material layer thickness in different channels, all collected in minutes.

3. The sintering endpoint prediction model based on spatiotemporal feature fusion according to claim 1, characterized in that: The short-term variable matrix M s The sintering process data includes: trolley speed in minutes, negative pressure value of No. 1 wind box, negative pressure of No. 24, No. 28 and No. 30 wind boxes, exhaust gas temperature on the north and south sides, exhaust gas temperature and negative pressure of the main flue on the north and south sides, frequency of the main exhaust fan on the north and south sides and its valve opening, as well as TRP, BRP and BTP soft measurement results.

4. The sintering endpoint prediction model based on spatiotemporal feature fusion according to claim 1, characterized in that: The regression prediction module consists of a Transformer encoding layer, an attention mechanism weighted layer, and a fully connected layer.

5. A sintering endpoint prediction system based on spatiotemporal feature fusion, characterized in that: It consists of a programmable logic controller (101), a memory (102) containing the prediction model of any one of claims 1 to 4, and a processor (103).

6. The sintering endpoint prediction system based on spatiotemporal feature fusion according to claim 5, characterized in that: The programmable logic controller (101) is used to collect sintering production process data; the memory (102) is used to store sintering production process data and prediction models; the processor (103) is used to read sintering process production data and execute prediction models.

Images