A control-oriented sintering endpoint modeling method based on coupled mapping framework

Abstract

The application discloses a control-oriented sintering endpoint modeling method based on a coupling mapping framework, and belongs to the technical field of metallurgical industrial process modeling and process control. First, sintering endpoint time series data and multiple sintering data time series sequences are collected, and a sintering endpoint modeling model containing an attention gate cycle unit module and a bidirectional gate cycle unit is constructed; the sintering data is divided into manipulated variables, coupling disturbance variables and disturbance variables; after the various time series data are divided into time series segments, the coupling disturbance variables and the manipulated variables are respectively subjected to the attention gate cycle unit module to obtain state vector prediction values, and then the state vector prediction values are input into the bidirectional gate cycle unit together with the divided manipulated variables, disturbance variables and coupling disturbance variables to obtain sintering endpoint single-step prediction values; further, a loss function is calculated according to true values, and sintering endpoint modeling model parameters are iteratively updated. The method guarantees a simple model structure, and takes into account prediction accuracy and control availability.

Description

Technical Field

[0001] This invention belongs to the field of metallurgical industrial process modeling and process control technology, specifically relating to a control-guided sintering endpoint modeling method based on a coupled mapping framework. Technical Background In modern steel production, the sintering process is a crucial preliminary step in blast furnace ironmaking, and its production stability directly affects the blast furnace's operating status and the quality of molten iron. The Burning Through Point (BTP) is an important state variable reflecting whether the sintering process is sufficient; deviation from its position will lead to a decrease in sinter quality, an increase in energy consumption, and unstable equipment operation.

[0002] Because the sintering process is characterized by many variables, strong nonlinearity, complex coupling relationships, and significant time lag, the sintering endpoint is difficult to measure directly through on-site sensors. It can usually only be indirectly judged by the wind box temperature or human experience after sintering, which is difficult to meet the requirements of modern sintering processes for advance prediction and closed-loop control.

[0003] Existing data-driven sintering endpoint modeling methods mostly focus on improving prediction accuracy while neglecting the impact of changes in manipulated variables on the internal state evolution of the system. When the manipulated variables are adjusted, the model often exhibits prediction jumps or discontinuous responses, making it difficult to directly embed it into control systems. Therefore, there is an urgent need for a sintering endpoint modeling method that can explicitly characterize the coupling relationship between manipulated variables and system states and possesses control-oriented characteristics. Summary of the Invention

[0004] To address the issues of insufficient stability and poor control adaptability of existing sintering endpoint prediction models under varying manipulated variables, this invention proposes a control-oriented sintering endpoint modeling method based on a coupled mapping framework.

[0005] S1. Collect sintering endpoint time series data and various sintering variable time series sequences used to predict the sintering endpoint during the metallurgical sintering process; construct a sintering endpoint modeling model including an attention-gated loop unit module and a bidirectional gated loop unit. S2, the sintering trolley speed variable in the sintering variables is recorded as the manipulated variable, and the remaining sintering variables are recorded as the disturbance variables; combined with the Pearson correlation coefficient and high and low frequency mutual information, the disturbance variables are further divided into coupled disturbance variables and disturbance variables; S3, divides the sintering endpoint timing data, the timing sequence of manipulated variables, the timing sequence of coupled interference variables, and the timing sequence of disturbance variables into timing segments; S4. The time series sequences of the coupled interference variables and the time series sequences of the manipulated variables are input into two attention-gated loop unit modules respectively, and then merged to obtain the state vector prediction value. The time series sequences of the manipulated variables, the time series sequences of the disturbance variables, and the time series sequences of the coupled interference variables are further combined and then passed through a bidirectional gated loop unit to obtain the single-step prediction value of the sintering endpoint. S5. The loss function is calculated based on the single-step prediction value and the divided sintering endpoint time sequence. All parameters in the sintering endpoint modeling model are updated according to the loss function and S4-S5 are repeated until the loss function converges. At this time, the trained sintering endpoint modeling model is obtained. S6 collects various sintering variables in real time during the metallurgical sintering process and forms corresponding time series sequences. After repeating S2-S3, the sequences are uniformly input into the trained sintering endpoint model to obtain the predicted sintering endpoint.

[0006] Furthermore, in S2, the sintering variables include ignition temperature, temperature of each wind box, sintering trolley speed, fuel flow rate, quicklime flow rate, return ore flow rate, iron ore flow rate, gas pressure, gas temperature, plate feed speed, roller feed speed, fuel pressure, ignition pressure, ignition gas pressure, ignition gas flow rate, ignition gas temperature, sintering monitoring point temperature, fuel gas temperature, and ring cooling temperature.

[0007] Furthermore, in S2, the interference variables are further divided into coupled interference variables and disturbance variables. Specifically, the weighted coupling degree between the time series corresponding to each interference variable and the time series corresponding to the manipulation variable is calculated sequentially. Interference variables with a weighted coupling degree higher than a threshold are recorded as coupled interference variables, while those with a weighted coupling degree lower than a threshold are recorded as disturbance variables.

[0008] Furthermore, the weighted coupling degree is specifically as follows: ; in, and It is the scaling factor. and These represent the low-frequency mutual information and the high-frequency mutual information of the interference variables, respectively. The Pearson correlation coefficient is used. This represents the weighted coupling degree of the interference variables.

[0009] Furthermore, the low-frequency and high-frequency mutual information of the interference variables are specifically as follows: for each interference variable, the interference variable and the manipulation variable are first normalized separately; then, according to the preset binning, the high-frequency and low-frequency mutual information between the normalized interference variable and the manipulation variable are calculated respectively. ; ; in, Indicates that after normalization, it simultaneously belongs to The nth sub-box and The The number of samples per bin, i.e. and The number of samples that fall within the same numerical range after normalization; express The number of samples in the nth bin. express No. The number of samples per bin and These are the time series corresponding to the disturbance variable and the time series corresponding to the manipulation variable, respectively. and These represent the number of cells under low-frequency mutual information and the number of cells under high-frequency mutual information, respectively.

[0010] Furthermore, in S3, the time segment division adopts the sliding time window method.

[0011] Furthermore, S4 specifically includes: S401, input the partitioned coupling interference variable time series and the partitioned manipulation variable time series into the attention gating recurrent unit module respectively to obtain the first context variable and the second context variable; S402, after unifying the dimensions of the first and second context variables respectively, the predicted value of the state vector is obtained by element-wise multiplication; S403, the partitioned time series of manipulated variables, the partitioned time series of disturbance variables, the partitioned time series of coupled disturbance variables and the predicted value of the state vector are concatenated and input into the bidirectional gated recurrent unit to obtain the forward hidden variables and the backward hidden variables; S404, multiply the forward and backward hidden variables element-wise to obtain the single-step prediction value of the sintering endpoint.

[0012] Furthermore, in S402, the dimensions are unified, specifically as follows: ; ; in, All are learnable weight matrices. , and All are learnable bias matrices. and These are the first context variable and the second context variable, respectively. and These are the first and second context variables after dimension unification, respectively, with T in the upper right corner representing the transpose.

[0013] Furthermore, in S5, the loss function is specifically: ; in, For loss function, for, This represents the actual sintering endpoint value at the current moment. This is a single-step prediction of the sintering endpoint. The current state vector is obtained by concatenating the time series sequences of the manipulated variables, the time series sequences of the perturbation variables, the time series sequences of the coupled disturbance variables, and the predicted state vector values. Let be the true state vector at time t. For the dimension of coupling interference variables.

[0014] The beneficial effects of this invention are: 1. By explicitly characterizing the nonlinear coupling relationship between the manipulated variable and the system state through the coupling mapping mechanism, the stability of the model under the condition of changing manipulated variables is improved.

[0015] 2. The model output is continuous and controllable, making it suitable for online deployment of control systems.

[0016] 3. While ensuring prediction accuracy, it also takes into account the simplicity of the model structure, and has good engineering application value. Attached Figure Description

[0017] Figure 1 This is an overall framework diagram of the sintering endpoint modeling method based on the coupling mapping framework of the present invention; Figure 2 This is a schematic diagram of the iron ore sintering process; Figure 3 Graph showing the results of variable selection; Figure 4 A schematic diagram illustrating the construction of a sliding time window for time series data; Figure 5 A detailed flowchart of the coupling mapping framework constructed for this invention; Figure 6 A comparison of the sintering endpoint prediction results between the coupling mapping model constructed for this invention and other models; Figure 7 The diagram shows the comparison of the controlled effects of the coupling mapping model constructed in this invention with other models. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0019] Sintering refers to the process of transforming a powdered mixture into a solid mass by placing it in a high-temperature heating environment. The specific operation is as follows: First, iron ore powder, flux, fuel, and recycled ore are mixed according to a specific ratio. After adding water and granulating, the mixture is conveyed to a mixing silo for temporary storage via a belt conveyor. Then, the mixture is evenly spread onto a belt sintering machine using a spreading device. Under a set negative pressure environment, the mixture is ignited by an igniter, allowing the fuel in the mixture to burn completely. The flue gas generated during combustion is exhausted from top to bottom by an exhaust fan. As the sintering trolley continues to move, the mixture gradually undergoes melting and combustion reactions from top to bottom. The large amount of heat released by the fuel combustion melts the surface of the powdered iron ore, generating a certain amount of liquid phase. This liquid phase wets the unmelted ore particles, thereby achieving the bonding of the sintered ore particles into lumps. After the sintering reaction is complete, the finished sintered ore falls to the tail end of the machine as it travels along. After being crushed by a single-roll crusher, it is conveyed to the finished product granulation system. The bottom material and return ore generated during granulation are recycled for the sintering process, while the remaining qualified product is conveyed to the blast furnace as finished sintered ore for subsequent use. A schematic diagram of the sintering process is shown below. Figure 2 As shown.

[0020] Based on the above background, this invention proposes a control-guided sintering endpoint modeling method based on a coupled mapping framework, the overall process of which is as follows: Figure 1 As shown, specifically: (1) By integrating the variable screening method of linear correlation and nonlinear mutual information, the manipulated variables, disturbance variables and state variables in the sintering process are systematically screened to construct a model input feature set with physical consistency and information sufficiency.

[0021] First, historical time-series data of various sintering variables were collected during the sintering process, including but not limited to ignition temperature, temperature of each wind box, and sintering trolley speed, as well as the corresponding actual sintering endpoint time-series data. Among these, the sintering trolley speed variable was used as the manipulated variable, and the other variables were recorded as disturbance variables.

[0022] Secondly, using a combination of Pearson correlation coefficient and high- and low-frequency mutual information, interference variables are screened and further divided into coupled interference variables and disturbance variables, specifically: Let the length be Two variables and Their Pearson correlation coefficient Defined as: in and Sequences and The average value.

[0023] In mutual information computation, the number of bins often affects the resolution and sparsity of the data. (The last sentence appears to be incomplete and possibly refers to a separate topic: "Conducting sequence...") and Normalized to the interval [0, 1], assuming the calculation of low-frequency mutual information... High-frequency mutual information The number of boxes at each time are respectively and Therefore: in Indicates that after normalization, it simultaneously belongs to The nth sub-box and The The number of samples per bin, i.e. and The number of samples that are within the same numerical range after normalization. Indicates belonging to The number of samples in the nth bin. Indicates belonging to The The number of samples in each bin. Therefore, define the variable. and Weighted coupling degree between for: in and It is the scaling factor.

[0024] Using the method described above, the historical time series corresponding to each interference variable is traversed. According to its relationship with manipulated variables of The disturbance variables are classified into coupler variables X and disturbance variables D: in To select a threshold.

[0025] (2) Using the existing input features, the sliding time window method is used to fragment the timing data of the sintering process and construct a set of key process variables.

[0026] Using the sliding window approach for segmentation sampling has two advantages: it directly solves the problems of small data files and short acquisition time; and it adapts to the modeling characteristics of multivariable timing systems such as sintering processes.

[0027] A sliding window is used to divide the historical time series data of both the coupling disturbance variable and the manipulation variable into multiple overlapping time series segments. The divided time series data are as follows: Figure 4 As shown (in a specific embodiment of the present invention, the time width of the sliding window is 40 minutes). At a certain moment... The partitioned historical coupling interference variable sequence With manipulated variable sequence They are constructed into the following formats respectively: in To couple the dimensions of the interference variables, Time window length The same sliding window method is used to process the observed actual sintering endpoint distance values, so that each fragmented input sample corresponds to a sintering endpoint output. This constructs a sequence dataset, which is used as input for the subsequent sintering endpoint modeling model.

[0028] (3) Construct a sintering endpoint model based on the coupling mapping framework, perform time-series encoding on the manipulation variable sequence and the perturbation variable sequence respectively, and realize the coupling mapping between the two in the low-dimensional latent space by element-wise multiplication, so as to explicitly characterize the modulation effect of the change of the manipulation variable on the system state evolution process; based on the latent state after coupling mapping, use the state decoding model to predict the sintering endpoint.

[0029] The overall framework of the sintering endpoint modeling model based on the coupling mapping framework is as follows: Figure 5 As shown, it includes two attention-gated loop unit modules and one bidirectional gated loop unit. The processing flow of the sintering endpoint modeling model is divided into two steps: Step 3.1: Capture the current system state by coupling the historical sequence of disturbance variables with the sequence of manipulated variables. The historical coupled interference variable sequence and the manipulated variable sequence are input into different attention-gated recurrent unit modules. Taking the historical coupled interference variable sequence as an example, from the input variables... To the hidden layer The transmission process is as follows: in, and These represent the reset gate variable and the update gate variable at time t, respectively. This represents the intermediate variable after the gating calculation at time t. , , The weight matrix to be learned. This is the sigmoid function.

[0030] Attention mechanisms enable the model to automatically focus on The most important part. Attention weights. and context vector The calculation method is as follows: in and Here are the weight and bias matrices to be learned, with T in the upper right corner representing the transpose. For the manipulated variable sequence... Perform the same operation to obtain the corresponding context vector. The hidden layer variables obtained from the historical coupling interference variable sequence and the manipulated variable sequence are transformed into... after dimensional unification. and Then, the predicted value of the current state vector is obtained by element-wise multiplication. : in These are all weight matrices to be learned. , and These are all bias matrices to be learned. This is for element-wise multiplication.

[0031] Step 3.2: Obtain the predicted value of the sintering endpoint at the next moment by decoding the output of the bidirectional gated loop unit.

[0032] At the current moment , as well as and spliced ​​together A unified input bidirectional gated loop unit is used to obtain forward and backward hidden variables, which are then multiplied element-wise to calculate the single-step prediction value of the sintering endpoint. The specific calculation method is as follows: in and Here are the weight matrix and bias matrix to be learned. For forward hiding variables, For backward hiding variables, It is a bidirectional gated loop unit.

[0033] Step 3.3: Model Training and Update Based on real-time data from the sintering process, the parameters of the coupled mapping model are adjusted in real time, continuously optimized and iterated, repeating steps 3.1 and 3.2 to make the model more robust. The training objective is to find the mean squared error between the model output and the current true state value and the true index. Minimize: in The vector represents the true state at time t, which is directly read from the dataset.

[0034] The model is updated in reverse according to L until L converges. At this point, the model training is considered complete, the iteration stops, and the trained sintering endpoint model is obtained.

[0035] During the online operation phase, based on the real-time data collected from the sintering process, the dataset of key process variables is input into the trained sintering endpoint model to obtain the sintering endpoint prediction result, thus achieving online prediction of the sintering endpoint.

[0036] To verify the effectiveness of the sintering endpoint model constructed in this invention, further experiments were conducted on the No. 2 sintering machine of a steel group in the eastern coastal region of China. A total of 9936 valid samples were collected, with a sampling interval of 1 minute. After data preprocessing, sample fragments were obtained using a sliding window, and the data fragments were divided into 8000 training samples and 1936 test samples. In addition, combining mechanistic analysis, literature review, and statistical data analysis, this experiment determined the weighted coupling coefficient between the temperature of the sixth wind box and the speed of the trolley in the No. 2 sintering machine. As a selection threshold The final variable selection results are as follows: Figure 3 As shown in Table 1, the classification results of the preferred variables are as follows.

[0037] Table 1 Variable Classification Results Since tuning deep learning models is quite complicated, the model parameters were set experimentally as shown in Table 2 below.

[0038] Table 2. Hyperparameters of the model To compare the performance of the established models, the comparative models used here are two deep learning algorithms based on recurrent neural networks: LSTM and GRU networks, and two soft measurement algorithms for process industry indicators: Autoformer and BTPNet networks. The evaluation metric is the coefficient of determination. Root mean square error and mean absolute error As shown in the formula below: Where i is the i-th sample, and n is the number of samples. Indicates the predicted value. This represents the mean of the true values. Represents the actual value.

[0039] Table 3 and Figure 6 The prediction performance of the sintering endpoint prediction model based on the coupling mapping framework is demonstrated. It can be seen that the prediction performance of the two deep learning algorithms based on recurrent neural networks, LSTM and GRU, is relatively poor. All values ​​were below 0.85, indicating that this type of model has certain limitations for complex industrial processes. However, the Autoformer network and BTPNet networks, two soft measurement algorithms for process industry indicators, showed significant improvements in performance. All are above 0.88. However, for stable industrial applications, the required accuracy has not yet been met. In contrast, the proposed sintering endpoint prediction model based on the coupled mapping framework significantly improves the accuracy of sintering endpoint prediction, achieving precise modeling and characterization of the iron ore sintering process.

[0040] Table 3 Comparison of Single-Step Prediction Results for Sintering Endpoint Table 3 Comparison of Single-Step In addition, since the coupled mapping framework can track the current system state, it can be embedded as a whole into a closed-loop control framework and used as the controlled object for control algorithm experiments, which has significant scientific research and industrial application value. The discrete incremental PID control algorithm was used to conduct control experiments on the controlled objects constructed by the above five modeling methods. The expression of the discrete incremental PID control algorithm is as follows: in for The error between the set value and the actual feedback value at any given time. , and To control parameters, set , , .

[0041] The results of the controlled experiment are as follows Figure 7 As shown, LSTM, GRU, Autoformer, and BTPNet all lack a system state update module, making them unsuitable for online control experiments as dynamic systems. The coupling relationship between the system output and the manipulated variables cannot be accurately reflected, resulting in a significant deviation between the controlled curve and the setpoint. In contrast, the coupling mapping framework enhances the role of the system output, maintaining the average error between the controlled sintering endpoint and the setpoint at approximately 0.2m. This demonstrates the superiority of the sintering endpoint modeling method proposed in this invention in closed-loop control experiments.

[0042] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A control-guided sintering endpoint modeling method based on a coupled mapping framework, characterized in that, include: S1. Collect sintering endpoint time series data and various sintering variable time series sequences used to predict the sintering endpoint during the metallurgical sintering process; construct a sintering endpoint modeling model including an attention-gated loop unit module and a bidirectional gated loop unit. S2, the sintering trolley speed variable in the sintering variables is recorded as the manipulated variable, and the remaining sintering variables are recorded as the disturbance variables; combined with the Pearson correlation coefficient and high and low frequency mutual information, the disturbance variables are further divided into coupled disturbance variables and disturbance variables; S3, divides the sintering endpoint timing data, manipulated variable timing sequence, coupled interference variable timing sequence, and disturbance variable timing sequence into timing segments; S4. The time series sequences of the coupled interference variables and the time series sequences of the manipulated variables are input into two attention-gated loop unit modules respectively, and then merged to obtain the state vector prediction value. The time series sequences of the manipulated variables, the time series sequences of the disturbance variables, and the time series sequences of the coupled interference variables are further combined and then passed through a bidirectional gated loop unit to obtain the single-step prediction value of the sintering endpoint. S5. The loss function is calculated based on the single-step prediction value and the divided sintering endpoint time sequence. All parameters in the sintering endpoint modeling model are updated according to the loss function and S4-S5 are repeated until the loss function converges. At this time, the trained sintering endpoint modeling model is obtained. S6 collects various sintering variables in real time during the metallurgical sintering process and forms corresponding time series sequences. After repeating S2-S3, the sequences are uniformly input into the trained sintering endpoint model to obtain the predicted sintering endpoint.

2. The control-guided sintering endpoint modeling method based on a coupled mapping framework according to claim 1, characterized in that, In S2, the sintering variables include ignition temperature, temperature of each wind box, sintering trolley speed, fuel flow rate, quicklime flow rate, return ore flow rate, iron ore flow rate, gas pressure, gas temperature, plate feed speed, roller feed speed, fuel pressure, ignition pressure, ignition gas pressure, ignition gas flow rate, ignition gas temperature, sintering monitoring point temperature, fuel gas temperature, and ring cooling temperature.

3. The control-guided sintering endpoint modeling method based on a coupled mapping framework according to claim 1, characterized in that, In step S2, the interference variables are further divided into coupled interference variables and disturbance variables. Specifically, the weighted coupling degree between the time series corresponding to each interference variable and the time series corresponding to the manipulation variable is calculated sequentially. Interference variables with a weighted coupling degree higher than a threshold are recorded as coupled interference variables, while those with a weighted coupling degree lower than a threshold are recorded as disturbance variables.

4. The control-guided sintering endpoint modeling method based on a coupled mapping framework according to claim 3, characterized in that, The weighted coupling degree is specifically: ； in, and It is the scaling factor. and These represent the low-frequency mutual information and the high-frequency mutual information of the interference variables, respectively. The Pearson correlation coefficient is used. This represents the weighted coupling degree of the interference variables.

5. The control-guided sintering endpoint modeling method based on a coupled mapping framework according to claim 4, characterized in that, The low-frequency and high-frequency mutual information of the interference variables are as follows: For each interference variable, the interference variable and the manipulation variable are first normalized separately; then, according to the preset binning, the high-frequency and low-frequency mutual information between the normalized interference variable and the manipulation variable are calculated respectively. ； ； in, Indicates that after normalization, it simultaneously belongs to The nth sub-box and The The number of samples per bin, i.e. and The number of samples that fall within the same numerical range after normalization; express The number of samples in the nth bin. express No. The number of samples per bin and These are the time series corresponding to the disturbance variable and the time series corresponding to the manipulation variable, respectively. and These represent the number of cells under low-frequency mutual information and the number of cells under high-frequency mutual information, respectively.

6. The control-guided sintering endpoint modeling method based on a coupled mapping framework according to claim 1, characterized in that, In S3, the time segment division adopts the sliding time window method.

7. The control-guided sintering endpoint modeling method based on a coupled mapping framework according to claim 1, characterized in that, Specifically, S4 is: S401, input the partitioned coupling interference variable time series and the partitioned manipulation variable time series into the attention gating recurrent unit module respectively to obtain the first context variable and the second context variable; S402, after unifying the dimensions of the first and second context variables respectively, the predicted value of the state vector is obtained by element-wise multiplication; S403, the partitioned time series of manipulated variables, the partitioned time series of disturbance variables, the partitioned time series of coupled disturbance variables and the predicted value of the state vector are concatenated and input into the bidirectional gated recurrent unit to obtain the forward hidden variables and the backward hidden variables; S404, multiply the forward and backward hidden variables element-wise to obtain the single-step prediction value of the sintering endpoint.

8. The control-guided sintering endpoint modeling method based on a coupled mapping framework according to claim 7, characterized in that, In S402, the dimensions are unified, specifically as follows: ； ； in, All are learnable weight matrices. , and All are learnable bias matrices. and These are the first context variable and the second context variable, respectively. and These are the first and second context variables after dimension unification, respectively, with T in the upper right corner representing the transpose.

9. The control-guided sintering endpoint modeling method based on a coupled mapping framework according to claim 1, characterized in that, In S5, the loss function is specifically as follows: ； in, For loss function, for, This represents the actual sintering endpoint value at the current moment. This is a single-step prediction of the sintering endpoint. The current state vector is obtained by concatenating the time series sequences of the manipulated variables, the time series sequences of the perturbation variables, the time series sequences of the coupled disturbance variables, and the predicted state vector values. Let be the true state vector at time t. For the dimension of coupling interference variables.

Images