Sintering end point prediction model based on spatio-temporal feature fusion and system thereof

Abstract

The invention discloses a time-space feature fusion-based burn-through point prediction model and a system thereof. The model comprises the following steps: acquiring sintering process data, and respectively constructing a long-term variable matrix M1 and a short-term variable matrix Ms by adopting a time sequence method; a sintering end point prediction model is established based on a ResNet network and a Transform architecture; the long-term variable matrix M1 and the short-term variable matrix Ms are input into a prediction model after matrix reconstruction, a regression input vector is obtained through feature extraction and feature reconstruction in sequence, and a prediction value is output through iterative training. According to the model, input parameters are divided into long-term variables and short-term variables according to spatial distribution characteristics of a sintering process, and meanwhile, a variable matrix is constructed according to time sequence characteristics of the variables, so that the influence of the variables with different time and spatial characteristics on a sintering end point is fully considered; a prediction system constructed based on the model does not need to adopt artificial experience to carry out data feature extraction, and deeper-level feature information can be identified, so that the accuracy of the prediction model is greatly improved.

Description

Technical Field

[0001] The invention relates to a sintering endpoint prediction model, in particular to a sintering endpoint prediction model based on spatiotemporal feature fusion and a system thereof, and belongs to the field of iron ore sintering production. Background Art

[0002] The steel industry is a basic industry of the national economy and one of the important indicators to measure the level of a country's national strength. my country's steel industry still uses the "blast furnace-converter" long-process production process, and the proportion of sintered ore in the blast furnace exceeds 70%. The rationality and stability of the sintering process plays a key role in the efficient and smooth operation of the entire ironmaking process, and the sintering end state is the core of the stable control of the entire sintering process. The sintering end point is affected by many factors such as the properties of raw materials and operating parameters. It has dynamic time-varying properties and can only be judged and controlled by manual experience. The advance prediction of the sintering end point can provide data support for on-site operators and is the premise and basis for realizing intelligent control of the sintering end point. At present, researchers mainly use a combination of mechanism and data-driven models to establish a mapping relationship between process detection data and the sintering end point, but due to the limitations of traditional modeling methods and computing power, the established model relies on the previous induction and summary of features, and the temporal and spatial correlations between variables are difficult to comprehensively consider. Summary of the invention

[0003] In response to the problems existing in the prior art, the first purpose of the present invention is to provide a sintering endpoint prediction model based on the fusion of time and space features. The model divides the input parameters into long-term variables and short-term variables according to the spatial distribution characteristics of the sintering process, and constructs a variable matrix according to the time series characteristics of the variables. It fully considers the influence of variables with different time and space characteristics on the sintering endpoint. Compared with the prior art, this prediction model has the advantages of longer prediction step and higher accuracy.

[0004] The second object of the present invention is to provide a sintering endpoint prediction system based on spatiotemporal feature fusion, which is composed of a programmable logic controller, a memory containing the above prediction model, and a processor. The system is based on convolutional neural networks and Transformer, and integrates self-attention mechanisms, residual connections and other operations to achieve 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.

[0005] In order to achieve the above technical objectives, the present invention provides a sintering endpoint prediction model based on spatiotemporal feature fusion, comprising: Step S1, obtain sintering process data, and use time series method to construct long-term variable matrix M l and the short-term variable matrixM s ; Step S2, establishing a sintering endpoint prediction model based on the ResNet network and the Transformer architecture; Step S3: Long-term variable matrix M l and the short-term variable matrix M s After matrix reconstruction, it is input into the prediction model, and then feature extraction and feature reconstruction are performed in sequence to obtain the regression input vector, and then iterative training is performed to output the sintering endpoint prediction value; The matrix obtained by splicing and deformation is a square matrix.

[0006] As a preferred solution, the long-term variable matrix M l The sintering process data include: fuel ratio, fuel feeding amount, mixture moisture, mixture feeding amount, mixture trough material level, ignition furnace hearth temperature, holding furnace section temperature and material layer thickness in different channels collected in minutes.

[0007] As a preferred solution, the short-term variable matrix M s The sintering process data include: trolley speed in minutes, negative pressure value of No. 1 wind box, negative pressure of No. 24, No. 28 and No. 30 wind boxes, north and south exhaust gas temperatures, north and south flue exhaust gas temperatures and negative pressures, north and south main exhaust fan frequencies and valve openings, and TRP, BRP and BTP soft measurement results.

[0008] As a preferred solution, the framework of the sintering endpoint prediction model includes: a sequence feature extraction module based on a ResNet network, a feature reconstruction module, and a regression prediction module based on a Transformer.

[0009] As a preferred solution, the feature extraction module is composed of a feature extraction branch for a long-term variable matrix and a feature extraction branch for a short-term variable matrix, and each branch uses a ResNet network for feature extraction.

[0010] As a preferred solution, the step size of the first convolutional layer and the maximum pooling layer in the ResNet network of the long-term variable feature extraction branch is set to (2,1).

[0011] As a preferred solution, no maximum pooling layer is set in the ResNet network of the short-term variable feature extraction branch.

[0012] As a preferred solution, the step size of the 3´3 convolution kernel in the ResNet network of the long-term variable feature extraction branch and the short-term variable feature extraction branch is set to 1, and downsampling is not used.

[0013] As a preferred solution, after the feature extraction, the long-term variable matrix M l and the short-term variable matrix M s All are converted to size ( A , 1) Feature vector.

[0014] As a preferred solution, the feature reconstruction process is: convert the long-term variable matrix M l and the short-term variable matrix M s The resulting eigenvector is transformed into a size of ( n , 2n ) matrix, and then by splicing up and down, we get a size of ( 2n, 2n ) is the regression input matrix.

[0015] Deep learning methods represented by CNN and Transformer have good feature information extraction capabilities, which can realize the transformation of the sintering endpoint prediction model from the original point-to-point prediction method to sequence-to-point and sequence-to-sequence prediction. Furthermore, the temporal and spatial dependencies in the sintering process variable sequence are fully explored, and the long-term variable matrix is ​​used to M l and the short-term variable matrix M s As a training set for the prediction model, it can significantly improve the accuracy and performance of the prediction model.

[0016] As a preferred solution, the regression prediction module consists of a Transformer encoding layer, an attention mechanism weighted layer and a fully connected layer.

[0017] As a preferred solution, the attention mechanism weighted layer performs weighted fusion on the vectors according to Formula 1 and outputs the regression vector Z ; Formula 1: ; In formula 1: x i is a matrix row vector; is the trainable weight matrix of the fully connected layer, where r is a hyperparameter, r=8; The fully connected layer weight matrix can be trained for the fully connected layer, and the Sigmoid activation function is used to provide attention weights.

[0018] The present invention also provides a sintering endpoint prediction system based on spatiotemporal feature fusion, which comprises a programmable logic controller (101), a memory (102) containing any one of the prediction models described above, and a processor (103).

[0019] As a preferred solution, the programmable logic controller (101) is used to collect sintering production process data.

[0020] As a preferred solution, the memory (102) is used to store sintering production process data and prediction models.

[0021] As a preferred solution, the processor (103) is used to read the sintering process production data and execute the prediction model.

[0022] Compared with the prior art, the beneficial technical effects of the technical solution of the present invention are: 1) The prediction model provided by the present invention divides the input parameters into long-term variables and short-term variables according to the spatial distribution characteristics of the sintering process, and constructs a variable matrix according to the temporal characteristics of the variables, fully considering the influence of variables with different time and spatial characteristics on the sintering endpoint. Compared with the existing technology, the prediction model has the advantages of longer prediction step and higher accuracy.

[0023] 2) In the technical solution provided by the present invention, based on convolutional neural networks and Transformer, self-attention mechanism, residual connection and other operations are integrated to realize the automatic extraction of temporal and spatial data features. There is no need to use manual experience for data feature extraction, and deeper feature information can be identified, thereby greatly improving the accuracy of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a schematic diagram of the prediction model construction process provided in Example 1 of the present invention; Figure 2 Schematic diagram of the structure of the sintering endpoint prediction network in Example 1 of the present invention; Figure 3 This is a structural diagram of the prediction system provided in Example 1 of the present invention; Figure 4 It is the computer program interface for predicting the sintering endpoint in the prediction system provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0025] In order to facilitate the understanding of the present invention, the present invention will be described in more comprehensive and detailed manner in conjunction with the accompanying drawings and preferred embodiments of the specification. It should be noted that the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Example 1 This embodiment provides a sintering endpoint prediction model based on spatiotemporal feature fusion: S01: 60-minute data of 14 variables, including fuel ratio, fuel feeding amount, mixture moisture, mixture feeding amount, mixing tank material level, ignition furnace hearth temperature, holding furnace section temperature and material layer thickness of six different channels, are collected from the minute-level sintering process data. After time alignment, data preprocessing and data standardization, a long-term variable matrix Ml of size 14´60 is established, as shown in Table 1.

[0027]

[0028] S02: Collect 30-minute data of 22 variables, including trolley speed, negative pressure value of No. 1 wind box, negative pressure of No. 24, No. 28 and No. 30 wind boxes and exhaust gas temperature on the north and south sides, exhaust gas temperature and negative pressure of the south and north flues, frequency of main exhaust fans on the south and north sides and their valve opening, and soft measurement results of TRP, BRP and BTP, from the minute-level sintering process data. After time alignment, data preprocessing and data standardization, a short-term variable matrix of size 22´30 is established. M s , as shown in Table 2.

[0029]

[0030] S03: Input the constructed two-dimensional matrix into the sintering endpoint prediction model, which is based on the ResNet network and the Transformer architecture. The framework includes: a sequence feature extraction module based on the ResNet network, a feature reconstruction module, 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 splicing them up and down, as shown in Table 3.

[0031]

[0032] The pre-training and fine-tuning process uses the model training environment of pytorch=2.20 and python=3.10. The optimizer selected for training is Adam, the parameters are set to default, and the batch size is 128. In terms of training strategy, the model is mainly trained in two rounds. The first round of training generations is 200 generations, and the initial learning rate is 1e-3. If the training loss does not decrease after 10 consecutive generations, the learning rate is reduced to 0.5 times the original; the second round of training is carried out for a total of 100 rounds. On the basis of inheriting the minimum learning rate of the first round of training, the cosine decay strategy is adopted to decay the learning rate to 0.1 times the initial learning rate of this round, and the early stopping method is used to stop the model training in time; the MSE function is used as the loss function of the model, the training set is sent to the network training, and the grid search method is used to determine the hyperparameters to achieve the training effect. The various indicators of the model are shown in Table 4.

[0033]

[0034] The long-term variable matrix M1 and short-term variable matrix Ms to be predicted are input into the model, and the predicted sintering endpoint after 30 minutes is: 28.85. The actual soft measurement result is 29.02, with an error of 0.17.

[0035] Taking the sintering endpoint prediction results in a production period as an example, the model prediction value is compared with the actual soft measurement result. Figure 2 shown.

[0036] A sintering endpoint prediction system based on the above method and integrating spatiotemporal features is developed. The system includes: a programmable logic controller 101, a memory 102, and a processor 103. The memory 102 stores a computer program that can implement the above sintering endpoint prediction method. The computer program background business is compiled in Java language, and the front-end page is compiled in JavaScript. The processor 103 is used to read the sintering production process data of the programmable logic controller 101, retrieve and execute the computer program. The prediction system composition diagram is shown in FIG. Figure 3 As shown in the figure, the program interface is as follows Figure 4 shown.

Claims

1. A sintering endpoint prediction model based on spatiotemporal feature fusion, characterized in that: include: Step S1, obtain sintering process data, and use time series method to construct long-term variable matrix M l and the short-term variable matrix M s ; Step S2, establishing a sintering endpoint prediction model based on the ResNet network and the Transformer architecture; Step S3: Long-term variable matrix M l and the short-term variable matrix M s After matrix reconstruction, it is input into the prediction model, and then feature extraction and feature reconstruction are performed in sequence to obtain the regression input vector, and then iterative training is performed to output the sintering endpoint prediction value; The matrix obtained by splicing and deformation is a square matrix.

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 include: fuel ratio, fuel feeding amount, mixture moisture, mixture feeding amount, mixture trough material level, ignition furnace hearth temperature, holding furnace section temperature and material layer thickness in different channels 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 include: trolley speed in minutes, negative pressure value of No. 1 wind box, negative pressure of No. 24, No. 28 and No. 30 wind boxes, north and south exhaust gas temperatures, north and south flue exhaust gas temperatures and negative pressures, north and south main exhaust fan frequencies and valve openings, and 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 framework of the sintering endpoint prediction model includes: a sequence feature extraction module based on a ResNet network, a feature reconstruction module and a regression prediction module based on a Transformer.

5. The sintering endpoint prediction model based on spatiotemporal feature fusion according to claim 4 is characterized in that: The feature extraction module is composed of a feature extraction branch for a long-term variable matrix and a feature extraction branch for a short-term variable matrix, and each branch uses a ResNet network for feature extraction.

6. The sintering endpoint prediction model based on spatiotemporal feature fusion according to claim 5, characterized in that: The step size of the first convolution layer and the maximum pooling layer in the ResNet network of the long-term variable feature extraction branch is set to (2,1); the maximum pooling layer is not set in the ResNet network of the short-term variable feature extraction branch; the step size of the 3´3 convolution kernel in the ResNet networks of the long-term variable feature extraction branch and the short-term variable feature extraction branch is set to 1, and downsampling is not used.

7. The sintering endpoint prediction model based on spatiotemporal feature fusion according to claim 1, characterized in that: After the feature extraction, the long-term variable matrix M l and the short-term variable matrix M s All are converted to size ( A , 1) feature vector; the feature reconstruction process is: the long-term variable matrix M l and the short-term variable matrix M s The resulting eigenvector is transformed into a size of ( n , 2n ) matrix, and then by splicing up and down, we get a size of ( 2n, 2n ) is the regression input matrix.

8. The sintering endpoint prediction model based on spatiotemporal feature fusion according to claim 4, characterized in that: The regression prediction module consists of a Transformer encoding layer, an attention mechanism weighted layer, and a fully connected layer; the attention mechanism weighted layer performs weighted fusion on the vector according to formula 1 and outputs a regression vector Z ; Formula 1: ; In formula 1: x i is a matrix row vector; is the trainable weight matrix of the fully connected layer, where r is a hyperparameter, r=8; The fully connected layer weight matrix can be trained for the fully connected layer, and the Sigmoid activation function is used to provide attention weights.

9. 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 described in any one of claims 1 to 8, and a processor (103).

10. The sintering endpoint prediction system based on spatiotemporal feature fusion according to claim 9, 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 a prediction model; and the processor (103) is used to read the sintering process production data and execute the prediction model.

Images