Sintered ore drum strength prediction model based on improved Transform network and system thereof

Abstract

The invention discloses a sinter drum strength prediction model based on an improved Transform network and a system of the sinter drum strength prediction model. The model firstly obtains continuous state data of a sintering process, constructs a sintering process variable matrix M1 and a variable variation trend matrix M2 after preprocessing, constructs an improved Transform network model composed of a variable sequence feature extraction branch-regression block and a variable sequence variation trend feature extraction branch-regression block by adopting an Encoder structure of Transform as a basic framework, and finally constructs a variable sequence variation trend feature extraction branch-regression block. And inputting the matrix M1 and the matrix M2 into an improved Transform network model, carrying out weighted fusion on feature variables extracted from the two branches, inputting a fused variable Z as a regression block, and outputting a predicted value. According to the model, a deep learning network is adopted, and data features are extracted through operations such as a self-attention mechanism and residual connection, so that process data change trend information loss is avoided, the prediction result accuracy and the prediction speed are greatly improved, and the technical problems of time lag, manual subjectivity and the like in an existing detection method are effectively solved.

Description

Technical Field

[0001] The invention relates to a sintering drum lightness prediction model, in particular to a sintering drum strength prediction model based on an improved Transformer network and a system thereof, and belongs to the technical field of iron ore sintering production. Background Art

[0002] The quality indicators of sintered ore mainly include chemical composition, physical properties and metallurgical properties. Among them, drum strength is an important indicator of the physical properties of sintered ore, reflecting the ability of sintered ore to resist crushing and wear during transportation, loading and unloading and blast furnace smelting. Therefore, drum strength is a comprehensive indicator to measure the quality of sintered ore and the level of production operation. Sintered ore with high drum strength can maintain a good particle size distribution in the blast furnace, reduce pulverization and melting, and ensure the stable operation of the blast furnace. At present, the drum strength of sintered ore is generally sampled and tested on the finished product belt by a machine gripper. Although this method is simple and direct, and has high accuracy, due to the particularity of the equipment and the time lag of the sintering production process, the equipment sampling cycle is generally more than 2 hours, and the detection time is more than 2 hours after the sintering completion time. The above problems seriously affect the perception and control of the drum strength of sintered ore at the production site.

[0003] The sintering process is a semi-closed high-temperature reaction process. Key variables such as the windbox exhaust gas temperature can roughly reflect the heat level of the sintering process. Experienced sintering personnel can estimate the drum strength level of this batch of sintered ore based on the sintering process variables such as temperature and their changing trends. Therefore, by developing a sinter drum strength prediction method and system based on an improved Transformer network to replace manual observation, the sinter drum strength can be accurately measured, solving the shortcomings of the traditional drum strength detection method, which is of great significance to improving the sintering production operation level. Summary of the invention

[0004] In view of the problems existing in the prior art, the first purpose of the present invention is to provide a sinter drum strength prediction model based on an improved Transformer network. The model adopts the Encoder structure of the Transformer in the deep learning model as the basic framework, can directly use the sintering production process data sequence as input, extract data features through self-attention mechanism, residual connection and other operations, effectively avoid the loss of process data change trend information, and greatly improve the prediction speed while improving the accuracy of the prediction results, thereby effectively solving the technical problems such as time lag and artificial subjectivity existing in the existing detection methods.

[0005] The second purpose of the present invention is to provide a sinter drum strength prediction system based on an improved Transformer network. Based on the excellent accuracy and prediction speed of the model, the prediction system formed by the model can replace manual observation. After testing, the prediction system using the model has an absolute value of ≤0.4% of the absolute error in the continuous 32h prediction process compared with the test results, and the average absolute value of the absolute error is only 0.24%, showing excellent prediction accuracy.

[0006] To achieve the above technical objectives, the present invention provides a sinter drum strength prediction model based on an improved Transformer network, comprising:

[0007] Step S1, obtaining continuous state data of the sintering process, and constructing a sintering process variable matrix M1 and a variable change trend matrix M2 after preprocessing;

[0008] Step S2, using the Transformer Encoder structure as the basic framework, constructing an improved Transformer network model consisting of a variable sequence feature extraction branch-regression block and a variable sequence change trend feature extraction branch-regression block;

[0009] Step S3, input the sintering process variable matrix M1 and the variable change trend matrix M2 into the improved Transformer network model, use the attention weighted layer to weightedly fuse the feature variables extracted from the two branches, and use the fused variable Z as the regression block input to output the predicted value of the sinter drum strength;

[0010] The variable sequence feature extraction branch also has a placeholder embedding layer before the first Transformer encoding layer.

[0011] As a preferred scheme, the continuous state data of the sintering process includes dolomite ratio, quicklime ratio, fuel ratio, return ore ratio, mixture moisture, mixture ratio, mixture tank position, sintering machine speed, mixture layer thickness, 18~30 bellows exhaust gas temperature, 18~30 bellows negative pressure, 18~30 bellows damper opening and ring cooling inlet hood exhaust gas temperature.

[0012] As a preferred solution, the preprocessing process is: performing outlier elimination, cubic spline interpolation and data standardization operations on the continuous state data of the sintering process obtained, thereby obtaining.

[0013] As a preferred solution, the construction process of the sintering process variable matrix M1 is: read the continuous state data of the sintering process at the minute level, use the time step of the state data as the matrix row vector, and use the variable sequence of the state data as the matrix column vector.

[0014] As a preferred solution, the construction process of the variable change trend matrix M2 is: use a sliding window with a size of 5 and a step size of 1 to perform a variable feature trend extraction operation based on least squares fitting on each column of variables in M1, obtain the variable change trend m, and after traversal calculation, obtain the variable change trend matrix M2.

[0015] As a preferred solution, the calculation process of the variable change trend m is:

[0016] Formula 1: ;

[0017] In formula 1: is the variable subsequence in the window, s is the time step, and y is the normalized variable value.

[0018] As a preferred solution, the variable sequence feature extraction branch structure and the variable sequence change trend feature extraction branch are both composed of a position embedding layer, a position encoding layer, a Transformer encoding layer, a Transformer encoding layer and a Transformer encoding layer; the regression block is composed of an attention mechanism weighted layer and a fully connected layer.

[0019] As a preferred solution, the attention mechanism function is:

[0020] Formula 2: ;

[0021] In formula 2: are query, key, and value vectors respectively, X is the input vector, and is a trainable projection matrix, is the dimension of the key vector.

[0022] As a preferred solution, the construction process of the variable sequence feature extraction branch-regression block is: use the Encoder layer of the three-layer Transformer to extract the feature variable sequence features, and at the same time add a row of parameter trainable row vectors above the matrix M1, and extract the first row vector x1 of the output feature matrix as the input feature of the subsequent regression block.

[0023] As a preferred solution, the construction process of the variable sequence change trend feature extraction branch-regression block is: using the Encoder layer of the three-layer Transformer to extract the feature variable sequence features, and extracting the last row vector x2 of the output matrix as the input feature of the subsequent regression block.

[0024] As a preferred solution, the weighted fusion process is:

[0025] Formula 3: ;

[0026] In formula 3: is the fused variable, is the trainable fully connected layer weight matrix of the fully connected layer, is the trainable weight matrix of the fully connected layer, r is a hyperparameter, is the row vector of input features of the regression block in the two branches.

[0027] The present invention also provides a sintered ore drum strength prediction system based on an improved Transformer network, comprising: a programmable logic controller (101), a memory (102) and a processor (103); the memory stores a computer program that can implement the prediction model described in any one of claims 1 to 9; the processor is used to read the sintering production process data in the programmable logic controller, and retrieve and execute the computer program that implements the prediction model in the memory.

[0028] Compared with the prior art, the beneficial technical effects of the technical solution of the present invention are:

[0029] 1) The prediction model provided by the present invention adopts the Encoder structure of the Transformer in the deep learning model as the basic framework, and can directly take the sintering production process data sequence as input, extract data features through operations such as self-attention mechanism and residual connection, effectively avoiding the loss of information on the trend of process data changes, and greatly improving the prediction speed while improving the accuracy of the prediction results, thereby effectively solving the technical problems such as time lag and human subjectivity existing in existing detection methods.

[0030] 2) In the technical solution provided by the present invention, both extraction branches only use the Encoder layer of the three-layer Transformer to extract the features of the feature variable sequence, which effectively improves the calculation efficiency of the model while reducing the risk of model overfitting. In addition, by adding a row of parameter trainable row vectors above the sintering process variable matrix M1 in the variable sequence feature extraction branch, the variable information of itself and the entire current matrix can be obtained, thereby further improving the prediction accuracy of the model. After testing, the prediction system using this model has an absolute value of ≤0.4% of the absolute error during the continuous 32h prediction process compared with the test results, and the average absolute value of the absolute error is only 0.24%, showing excellent prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a schematic diagram of a process variable sequence construction method in Example 1 of the present invention;

[0032] Figure 2 This is a schematic diagram of the structure of the improved Transformer network provided in Example 1 of the present invention;

[0033] Figure 3 This is a diagram showing the structure of a sinter drum strength prediction system based on an improved Transformer network in Example 1 of the present invention;

[0034] Figure 4 It is the computer program interface of the sinter drum strength prediction system based on the improved Transformer network in Example 1 of the present invention. DETAILED DESCRIPTION

[0035] 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.

[0036] Example 1

[0037] The present invention provides a sinter drum strength prediction model based on an improved Transformer network, and the specific process is as follows:

[0038] Step S1, read the dolomite ratio, quicklime ratio, fuel ratio, return ore ratio, mixture moisture, mixture ratio, mixture slot, sintering machine speed, mixture layer thickness, 18~30 bellows exhaust gas temperature, 18~30 bellows negative pressure, 18~30 bellows air door opening and 40 minutes of data of 49 variables of the ring cooling inlet hood from the minute-level sintering production process data, as shown in Table 1, use Lagrange interpolation method to correct the outliers, use data standardization to scale the process data, splice the variables into a matrix to obtain a sintering process variable sequence matrix M1 with a size of 40×49, as shown in Table 2, each row of the matrix is ​​a time step, and each column is a variable sequence; use a sliding window with a size of 5 and a step size of 1 to perform a variable feature trend extraction operation based on least squares fitting on each column of variables in M1, and set the variable subsequence in the window to be , s is the time step, y is the normalized variable value, and the calculation method of the variable change trend m is as shown in Formula 1:

[0039] Formula 1: ;

[0040] According to formula 1, each column of variables is traversed and calculated to obtain the variable change trend matrix M2 with a size of 36×49, as shown in Table 3. The construction process of M1 and M2 matrices is as follows: Figure 1 shown.

[0041]

[0042]

[0043]

[0044] Step S2, select the Transformer Encoder structure as the basic framework of the model, which consists of a variable sequence feature extraction branch, a variable sequence trend feature extraction branch, and a regression block. The branch structure is similar, consisting of a position embedding layer, a position encoding layer, a Transformer encoding layer, a Transformer encoding layer, and a Transformer encoding layer. The variable sequence feature extraction branch also has a placeholder embedding layer before the first Transformer encoding layer; the regression block consists of an attention mechanism weighted layer and a fully connected layer. The self-attention function of the Transformer is shown in Formula 2:

[0045] Formula 2: ;

[0046] In formula 2: are query, key, and value vectors respectively, X is the input vector, and is a trainable projection matrix, is the dimension of the key vector. The role of Softmax is to normalize all correlation scores so that the sum is 1 and generate attention weights. In the sequence variable feature extraction branch, the placeholder embedding layer generates a parameter-trainable row vector a1 in the first row of the matrix M1, forming a 41×49 matrix. When the branch is finally output, the row vector can obtain its information with the current entire matrix variable.

[0047] The attention weighted layer is used to perform weighted fusion on the feature variables extracted by the two branches. The first row vector x1 of the variable sequence feature extraction branch output matrix is ​​used as the input feature of the final regression block, and the last row vector x2 of the variable trend change feature extraction branch output matrix is ​​used as the input feature of the final regression block. The sizes of x1 and x2 vectors are both [1,49]. The attention weighted layer is used to perform weighted fusion on the feature variables extracted by the two branches. The fusion formula is shown in Equation 3:

[0048] Formula 3: ;

[0049] In formula 3: is the trainable weight matrix of the fully connected layer, where r is a hyperparameter and r=4 is selected. The fully connected layer weight matrix can be trained for the fully connected layer; the Sigmoid activation function is selected to provide attention weights.

[0050] The fused variable Z is used as the regression block input to output the predicted value of sinter drum strength. The model structure is as follows: Figure 2 shown.

[0051] Historical data was collected and a data set was made to train the improved Transformer network model to obtain the sinter drum strength prediction model. The process variable sequence data set was randomly divided into a training set and a test set at a ratio of 8:2. Adam was used as the optimizer of the model, and Huber loss with δ=0.5 was used as the loss function of the model. LeakyReLU (LeakyRectified Linear Unithtt) and Relu were selected as the activation functions of the network. The He Initialization method was used to initialize the weight matrix of each layer of the model. The training set was sent to the network for training. The method of preheating the training model was used to accelerate the convergence of the model. The experimental results are shown in Table 4.

[0052]

[0053] Step S3, respectively input the process variable sequence matrix M1 and the variable change trend sequence matrix M2 into the model to obtain the sintered ore drum strength prediction result; input the sintering process data sequence to be predicted into the sintered ore drum strength prediction model to obtain the sintered ore drum strength prediction result. Taking the actual operation results of an actual production period as an example, the comparison between the model prediction value and the inspection and test results is shown in Table 5.

[0054]

[0055] It can be seen from Table 5 that the prediction system using this model has an absolute value of ≤0.4% and an average absolute value of the absolute error of only 0.24% compared with the test results during the continuous 32-h prediction process, showing excellent prediction accuracy.

[0056] The present invention also provides a sinter drum strength prediction system based on an improved Transformer network, the system comprising: a programmable logic controller 101, a memory 102, and a processor 103; the memory 102 stores a computer program that can implement the above-mentioned sinter drum strength prediction model, 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, and retrieve and execute the computer program; the prediction system is composed as follows Figure 3 As shown, the program interface is as follows Figure 4 shown.

Claims

1. A sinter drum strength prediction model based on an improved Transformer network, characterized in that: include: Step S1, obtaining continuous state data of the sintering process, and constructing a sintering process variable matrix M1 and a variable change trend matrix M2 after preprocessing; Step S2, using the Transformer Encoder structure as the basic framework, constructing an improved Transformer network model consisting of a variable sequence feature extraction branch-regression block and a variable sequence change trend feature extraction branch-regression block; Step S3, input the sintering process variable matrix M1 and the variable change trend matrix M2 into the improved Transformer network model, use the attention weighted layer to weightedly fuse the feature variables extracted from the two branches, and use the fused variable Z as the regression block input to output the predicted value of the sinter drum strength; The variable sequence feature extraction branch also has a placeholder embedding layer before the first Transformer encoding layer.

2. The sinter drum strength prediction model based on the improved Transformer network according to claim 1 is characterized in that: The continuous state data of the sintering process include dolomite ratio, quicklime ratio, fuel ratio, return ore ratio, mixture moisture, mixture ratio, mixture tank position, sintering machine speed, mixture layer thickness, 18~30 bellows exhaust gas temperature, 18~30 bellows negative pressure, 18~30 bellows damper opening and ring cooling inlet hood exhaust gas temperature.

3. The sinter drum strength prediction model based on the improved Transformer network according to claim 1, characterized in that: The preprocessing process is: performing outlier elimination, cubic spline interpolation and data standardization operations on the continuous state data of the sintering process obtained, thereby obtaining the result.

4. The sinter drum strength prediction model based on the improved Transformer network according to claim 1, characterized in that: The construction process of the sintering process variable matrix M1 is as follows: read the continuous state data of the sintering process at the minute level, use the time step of the state data as the matrix row vector, and use the variable sequence of the state data as the matrix column vector.

5. The sinter drum strength prediction model based on improved Transformer network according to claim 1, characterized in that: The construction process of the variable change trend matrix M2 is as follows: a sliding window with a size of 5 and a step size of 1 is used to perform a variable feature trend extraction operation based on least squares fitting on each column variable in M1 to obtain the variable change trend m, and after traversal calculation, the variable change trend matrix M2 is obtained; The calculation process of the variable change trend m is: Formula 1: ; In formula 1: is the variable subsequence in the window, s is the time step, and y is the normalized variable value.

6. The sinter drum strength prediction model based on the improved Transformer network according to claim 1, characterized in that: The variable sequence feature extraction branch structure and the variable sequence change trend feature extraction branch are both composed of a position embedding layer, a position encoding layer, a Transformer encoding layer, a Transformer encoding layer, and a Transformer encoding layer; the regression block is composed of an attention mechanism weighted layer and a fully connected layer; The attention mechanism function is: Formula 2: ; In formula 2: , , are query, key, and value vectors respectively, X is the input vector, and is a trainable projection matrix, is the dimension of the key vector.

7. The sinter drum strength prediction model based on improved Transformer network according to claim 1, characterized in that: The construction process of the variable sequence feature extraction branch-regression block is as follows: the Encoder layer of the three-layer Transformer is used to extract the feature variable sequence features, and at the same time, a row of parameter trainable row vectors is added above the matrix M1, and the first row vector x1 of the output feature matrix is ​​extracted as the input feature of the subsequent regression block.

8. The sinter drum strength prediction model based on improved Transformer network according to claim 1, characterized in that: The construction process of the variable sequence change trend feature extraction branch-regression block is: using the Encoder layer of the three-layer Transformer to extract the feature variable sequence features, and extracting the last row vector x2 of the output matrix as the input feature of the subsequent regression block.

9. The sinter drum strength prediction model based on improved Transformer network according to claim 1, characterized in that: The weighted fusion process is as follows: Formula 3: ; In formula 3: is the fused variable, is the trainable fully connected layer weight matrix of the fully connected layer, is the trainable weight matrix of the fully connected layer, r is a hyperparameter, is the row vector of input features of the regression block in the two branches.

10. A sinter drum strength prediction system based on an improved Transformer network, characterized in that: include: A programmable logic controller (101), a memory (102) and a processor (103); the memory stores a computer program that can implement the prediction model described in any one of claims 1 to 9; the processor is used to read the sintering production process data in the programmable logic controller, and retrieve and execute the computer program that implements the prediction model in the memory.

Images