Sinter feo prediction method based on recursive coordinate mapping

Abstract

The application discloses a sinter FeO prediction method based on recursive coordinate mapping and belongs to the field of industrial process monitoring, modeling and simulation. By regarding operation variables as serialized input data, derivative analysis of orthogonal basis functions maps the input to a high-dimensional coordinate space. Then, in the high-dimensional coordinate space, a transpose attention mechanism is adopted to independently calculate attention scores in different degrees of coordinate space, dynamically weight, and reconstruct data to the time domain by the coordinate space. The method has a good fitting effect on time series data and can be widely applied to the prediction and optimization of industrial systems with long-term dependence, strong nonlinearity and strong dynamics.

Description

Technical Field

[0001] This invention belongs to the field of industrial process monitoring, modeling and simulation, and specifically relates to an improved method for predicting technical indicators of sintering ironmaking production. Background Technology

[0002] Ferrous oxide (FeO) content is an important performance indicator of sinter, directly affecting its strength, particle size, and metallurgical properties, and is also related to the blast furnace coke ratio in subsequent processes. Generally, an increase in FeO content means an increase in the content of poorly reducible silicates or calcium iron olivine in the sinter, which is considered to increase the drum index and decrease the reducibility of the sinter. Practical production experience shows that for every 1% increase in FeO content in sinter, the blast furnace coke ratio will increase by 1.5%, while the output will decrease by 1.5%. The ironmaking process industry should determine appropriate and stable FeO control indicators for sinter based on actual production needs, while ensuring the strength of the sinter. This is of great significance for stabilizing blast furnace conditions and reducing energy consumption.

[0003] The field-phase coupling and inherent nonlinear coupling of sintering systems are exceptionally complex, involving multiple complex solid-phase and liquid-phase reactions, making it difficult to detect key parameters characterizing operational performance online. Currently, most ironmaking plants predict FeO indicators by estimating them based on operator experience. However, this empirical method, heavily reliant on worker experience and actual plant conditions, is not only unsuitable for widespread application but also suffers from poor prediction accuracy, rendering it unsuitable for subsequent large-scale ironmaking system control. Therefore, developing FeO prediction algorithms for the sintering process and improving the informatization and intelligentization of sintering system production is crucial for large-scale ironmaking systems in my country.

[0004] Existing machine learning and statistical learning methods, such as partial least squares, Kalman filtering, neural networks, support vector machines, and subspace identification methods, either theoretically require the system to have strictly linear time-invariant properties, or require the collection of large amounts of training data and complex hyperparameter optimization processes. These prior conditions limit the application of such methods in real sintering machines. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method for predicting FeO in sintered ore based on recursive coordinate mapping.

[0006] A method for predicting FeO in sintered ore based on recursive coordinate mapping includes the following steps:

[0007] 1) Determine FeO-related variables from the sintering plant database as input features, and perform sliding window encapsulation preprocessing on the data;

[0008] 2) Construct a soft measurement model for FeO content, a quality index of sintered ore, based on a recursive coordinate coding network. The input data is mapped to a high-dimensional coordinate space through recursive coordinate mapping, and then the coordinate space data is mapped back to the time domain through nonlinear mapping transformation to realize the soft measurement of FeO. The recursive coordinate coding network structure includes input data u, output data y, system transition matrix A, system input matrix B, and inverse nonlinear mapping module BNLM.

[0009] In step 1), the variables related to FeO are: neutralization bin flow rate, light burning bin flow rate, pulverized coal bin flow rate, coke bin flow rate, return ore bin flow rate, quicklime bin flow rate, first mixed water flow rate, second mixed water flow rate, sinter thickness, main flue negative pressure, main flue exhaust gas temperature, main fan frequency, No. 21 wind box temperature, No. 22 wind box temperature, No. 23 wind box temperature, No. 24 wind box temperature, Section I annular cooler temperature, Section II annular cooler temperature, and Section III annular cooler temperature. The model will extract high-level latent space features through input variables to achieve FeO soft measurement.

[0010] In step 1), the sliding window encapsulation preprocessing refers to the following processing of the filtered input data during the deployment of the sintering soft measurement model:

[0011]

[0012]

[0013] in Represents the packaged first Window data, Represents the original data. This represents indexing the data. Representing FeO data, This represents the window size.

[0014] In step 2), the recursive coordinate mapping involves performing the following operations on the encapsulated filtered data:

[0015]

[0016]

[0017]

[0018] in Represents the first-order derivative operator. Represents the moment The next The order coordinates can be taken as follows: , , Consistent with the definition in step 2).

[0019] In step 2), the inverse nonlinear mapping change refers to the operation performed on the output after the aforementioned coordinate mapping during the sintering process:

[0020]

[0021] in Represents the input vector. , Represents a linear transformation matrix. , Is with Vectors with a consistent shape.

[0022] The beneficial effects of this invention are as follows:

[0023] 1. The proposed method abstracts the problem of sintered ore product quality into a time series prediction problem, cleverly avoiding the complex solid-liquid phase chemical modeling process and avoiding dependence on worker experience, thus having greater versatility.

[0024] 2. The proposed method considers the time dependence and nonlinearity of the data separately, and adopts an innovative coordinate encoding method for modeling, which greatly improves the accuracy of model prediction.

[0025] 3. This method fully considers the relative position information of the time series, avoids the noise caused by embedded coding algorithms such as coordinate coding or time coding, and makes the model more robust. Attached Figure Description

[0026] Figure 1 The diagram shows the structure of the FeO prediction method for sintered ore based on recursive coordinate mapping.

[0027] Figure 2 The image shows the impulse response diagram of the FeO prediction method for sintered ore based on recursive coordinate mapping.

[0028] Figure 3 The diagram shows the data flow for model training optimization and prediction.

[0029] Figure 4 The image shown is a demonstration of the data model effect in a real ironworks. Detailed Implementation

[0030] The mechanism of this invention will be explained first.

[0031] A method for predicting FeO in sintered ore based on recursive coordinate mapping (RCM-BNLM) is proposed. This method extends the nonlinear mapping architecture of deep learning based on state space and retains the state transition matrix to characterize the dynamic characteristics of the system. The difference between RCM-BNLM and explicit network architectures such as gated recurrent neural networks lies in solving the long-term dependency problem through implicit initialization. Recursive coordinate mapping (RCM) represents the latent space as the coefficients of orthogonal basis functions, which allows the input data to be encoded into a high-dimensional feature space with minimal information loss, effectively realizing multi-step prediction and having greater utility and practical significance.

[0032] The Recursive Coordinate Mapping (RCM) expression is as follows:

[0033]

[0034] For coordinate variables in coordinate space, For real-world input variables, For time variables, For the derivative operator that takes the first derivative with respect to the time variable, This is the coordinate weight matrix. For the input weight vector, For vector content indexing, For indexing matrix contents, Let the order be the global coordinate space. The index is the subscript of the coordinate order.

[0035] The coordinate weight matrix and input weight vector expressions are as follows:

[0036]

[0037]

[0038] The obtained output is then back mapped to the time domain using the Inverse Nonlinear Mapping (BNLM) module, completing the closed-loop operation of the recursive coordinate mapping. The BNLM formula is as follows:

[0039]

[0040] The maximum value operator acts as an activation function in the expression, adding nonlinear modeling capability to the model. and As a matrix, it acts as a linear operator in the expression, enhancing the model's expressive power and enabling optimization in the coordinate space. and It is the corresponding bias vector.

[0041] Given that the sintering process is a process industry with strong operational dependence and long time delay, the quality of the finished product is heavily dependent on the quality of the preceding processes, such as the proportion of neutralized ore, fuel, iron, water content in the raw material silo, the thickness of the sintering material layer, the speed of the sintering machine, and the real-time power of the bellows. The time interval from material preparation to finished product output is about three hours. Traditional machine learning and statistical learning methods cannot solve such a long time delay problem.

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention, and not to limit the present invention.

[0043] Figure 1 The diagram shows the specific structure of the FeO prediction method for sintered ore based on recursive coordinate mapping. It can be seen from the perspective of the conceptual diagram that RCM is essentially no different from general state-space model-based algorithms. The construction of dynamic data is completed through iterative calculation of the state space, while the construction of nonlinear data is completed through the inverse nonlinear mapping (BNLM) module.

[0044] The purpose of mapping the sintering process variable data collected by the sintering plant to a high-dimensional latent space is to...

[0045] 1) Achieving data dimensionality upgrade. Traditional machine learning methods such as kernel functions have revealed that low-dimensional linearly inseparable data can be linearly separable in high-dimensional space. Dimensionality upgrade can effectively improve the model's modeling ability and reduce the difficulty of data fitting.

[0046] 2) Retain original data information to the greatest extent possible. The success of statistical analysis methods such as principal component analysis and canonical correlation analysis shows that when performing feature engineering and dimensional operations, it is important to retain the effective information in the original data, and the less information loss, the better.

[0047] To achieve data dimensionality enhancement with minimal information loss, a loss function is used. As a metric for information loss, among which and These represent the reconstructed and actual sintering plant process parameter data, respectively. (Subscripts are also provided.) and These represent the starting and ending points of the calculation, respectively.

[0048] It can be seen that the loss is small when the reconstructed data is close to the real data. No information loss is considered to have occurred if and only if the reconstructed data is completely identical to the real data; in this case, there is no loss whatsoever. This analysis is consistent with real-world experience.

[0049] The proposed method achieves data mapping through orthogonal basis function decomposition. This requires selecting a set of orthogonal functions as the basis of the coordinate space. Here, Legendre polynomials are chosen as the orthogonal basis functions. Therefore, the expression for the nth-order basis function is:

[0050]

[0051] in Represents the Legendre polynomial basis function family. This represents the input data. The inner product operation is used to decompose the data, splitting the original input data into several mutually orthogonal coordinates in the coordinate space.

[0052]

[0053] in These are real data input from a sintering plant, such as data from the bellows temperature sensor. These are the coordinate results obtained after the input data is decomposed using the i-th order basis function. It is the measurement weight function corresponding to the Legendre polynomial as an orthogonal basis, which is assumed to be... .

[0054] Input data can be reconstructed using coordinate information.

[0055]

[0056] in Through coordinates This is obtained by reconstructing the data using a normalized Legendre basis function. The purpose of normalization is to eliminate the overall linear bias of the data.

[0057]

[0058] in This represents the family of basis functions for the normalized Legendre polynomials.

[0059] Theories related to function approximation show that input data decomposed and reconstructed by orthogonal polynomials can achieve arbitrary precision reconstruction and approximation without limiting the size of the coordinate space. Based on this, we can map and store arbitrary sintering process data in a coordinate space that not only contains more detailed frequency division information of the process data but also has the ability to reconstruct with minimal information.

[0060] However, real-world process data is updated in a streaming manner, rather than statically. To reflect the actual conditions of sintering production, a recursive update method needs to be modeled. This requires analyzing the dynamic characteristics of the coordinate decomposition step and introducing additional time variables. Used to represent past time steps, still in Represents the current time step.

[0061] At this point, the sintering process data accumulated through historical experience can be used... Therefore, the dynamic characteristic analysis of equation (6) can be obtained as follows:

[0062]

[0063] in and These are the expressions for the normalized Legendre function and the measurement weight function in the new bivariate coordinate domain, respectively.

[0064]

[0065] Defined as That is, variables only in the past moment Variables existing at the current time The weight is 1 when the window is 1 in size, and 0 otherwise.

[0066] Due to the closed-loop representation of Legendre polynomial functions, i.e., their derivatives can be expressed as linear combinations of functions themselves,

[0067]

[0068] right The following impulse function is obtained by differentiation analysis. linear combination

[0069]

[0070] Finally, combining equations (12) and (11), the theoretical analysis result of equation (9) can be obtained as follows:

[0071]

[0072]

[0073] Equation (14) is obtained after simplification.

[0074]

[0075] By summarizing it into a general representation of a state-space model, we can obtain...

[0076]

[0077]

[0078]

[0079] Using equations (16), (17), and (18), a state-space model can now be used to recursively map the sintering process data updated in a streaming manner, achieving coordinate mapping with minimal information loss. The coordinate data can then be back-mapped to the time domain using the subsequent BNLM module, enabling the model to model nonlinear relationships.

[0080] Figure 2 A schematic diagram of the impulse response for predicting FeO in sintered ore using the processed recursive coordinate mapping is provided. In state-space model-based methods, the impulse response is crucial, reflecting the model's intrinsic processing logic for time-series data. It can be seen that all coordinate components rapidly decay to 0 near the x-coordinate of 1, which is consistent with the analysis of the measurement weight function in the proposed method, indirectly proving the reliability of the proposed method.

[0081] The present invention will be further explained below with reference to specific examples.

[0082] Application Examples

[0083] (1) Introduction to the sintering process mechanism

[0084] This experiment was conducted on sintering machine No. 360 of a steel group in South China. Sintering is a crucial pretreatment step in steel production, involving a series of complex and precise operations. First, raw materials such as iron ore powder, recycled materials, coke powder, and limestone are mixed with fuel and flux. This mixing process not only involves uniform mixing of materials but also requires the addition of a certain amount of water for preheating. The mixture is then carefully spread on the conveyor belt of the sintering machine, requiring coarse particles at the bottom and fine particles on top. This process needs to be uniform and precise to ensure that the mixture layer is evenly heated throughout the sintering process. After being ignited, the mixture moves with the sintering machine, undergoing complex physicochemical reactions internally, where iron powder particles gradually combine into dense, porous small blocks. Finally, the sintered ore, after combustion and heat treatment, is unloaded, cooled, and crushed to prepare for subsequent blast furnace smelting.

[0085] (2) Selection of variables in the sintering process

[0086] The sintering process is a complex, dynamic, and time-varying process. Due to its numerous intricate steps and physicochemical reactions, it exhibits strong coupling, large time lags, and strong nonlinearity, significantly hindering the analysis using mathematical and statistical models. To better understand the relationship between the sintering process and variables, characteristic engineering of the sintering process is necessary. Through a review of sintering variables, the following 19 variables with the strongest correlation to FeO content were identified, as shown in Table 1.

[0087] Table 1 Selection of Sintering Input Variables

[0088] <![CDATA[ x 1 ]]> Zhonghe Warehouse Flow kilograms per meter <![CDATA[ x 2 ]]> Light burning chamber flow kilograms per meter <![CDATA[ x 3 ]]> Pulverized coal silo flow kilograms per meter <![CDATA[ x 4 ]]> Coke powder silo flow kilograms per meter <![CDATA[ x 5 ]]> Return to mining warehouse flow kilograms per meter <![CDATA[ x 6 ]]> quicklime silo flow rate kilograms per meter <![CDATA[ x 7 ]]> Mixed water flow rate cubic meters per hour <![CDATA[ x 8 ]]> Secondary mixed water flow rate cubic meters per hour <![CDATA[ x 9 ]]> Sinter thickness millimeters <![CDATA[ x 10 ]]> Large flue negative pressure kPa <![CDATA[ x 11 ]]> Large flue gas temperature Celsius <![CDATA[ x 12 ]]> Main fan frequency hertz <![CDATA[ x 13 ]]> Temperature of No. 21 bellows Celsius <![CDATA[ x 14 ]]> Temperature of No. 22 bellows Celsius <![CDATA[ x 15 ]]> Temperature of No. 23 bellows Celsius <![CDATA[ x 16 ]]> Temperature of No. 24 bellows Celsius <![CDATA[ x 17 ]]> I-stage annular cooler temperature Celsius <![CDATA[ x 18 ]]> II-stage annular cooler temperature Celsius <![CDATA[ x 19 ]]> III-stage annular cooler temperature Celsius y FeO content percentage

[0089] (3) Data preprocessing methods

[0090] Time series modeling assumes that the dynamic characteristics of the input-output relationship are concentrated within a certain time window. This idea is reflected in typical time series analysis methods such as ARIMA and LSTM. Therefore, a sliding window is used to encapsulate the collected raw sintering process variable data.

[0091]

[0092]

[0093] in Represents window size. Represents the packaged first Window data, Represents the original data. This represents indexing the data. This represents FeO data. The encapsulated data exists in matrix form, which not only facilitates subsequent calculations and operations but also reduces the computational burden, allowing the model to avoid traversing all accumulated variables, thus greatly saving computational resources.

[0094] (4) Model training and deployment

[0095] Step 1: Offline modeling;

[0096] Step 1.1: Through the aforementioned sintering process variable selection, collect the 19 variables that are most correlated with FeO content, store them in the database, and perform sliding window preprocessing on the database data.

[0097] Step 1.2: Read data from the database and divide it into training, validation, and test sets. The training and validation sets are used to train and optimize model parameters, while the test set is used to finally evaluate model performance as the final evaluation metric.

[0098] Step 1.3: Extract training sample data from the training set, input the training sample data into the aforementioned coordinate mapping unit, and obtain the coordinate output. Note that the coordinate mapping unit can only process one-dimensional data input, so in actual operation, the 19 input variables are input separately to obtain the coordinate output, and all coordinate outputs are concatenated together.

[0099] Step 1.4: Output the stitched coordinates to the time domain using the BNLM module. As can be seen from the above coordinate acquisition and reconstruction process, the essence of the reconstruction operation is a weighted sum of the coordinates and the Legendre basis function. Therefore, linear operators in BNLM can be used for approximation. Considering the strong coupling and strong nonlinearity of the sintering process, [further details are needed]. The nonlinear features introduced by activation functions enable models to model nonlinear data relationships.

[0100] Step 1.4: By comparing the FeO predicted content output by BNLM with the labels of the training samples, the model output function is calculated, and the internal parameters of the model are updated according to the model output function to achieve model training and optimization.

[0101] Step 2: Online detection;

[0102] Step 2.1: As Figure 3 As shown, after the model completes offline training and optimization, an input variable matrix is ​​constructed using data from a real factory database. After data preprocessing and encapsulation, the matrix is ​​input into the model, and the online detection function is achieved through the model output.

[0103] (5) Performance testing

[0104] To verify the model's effectiveness, 2514 actual production samples were collected from the sintering plant. Based on this data, a Partial Minimum Regression (PLS) model and a Deep Long Short-Term Memory (LSTM) network were used as comparative models in experiments. The evaluation metrics were the coefficient of determination and the root mean square error, as shown in the following formula:

[0105]

[0106]

[0107] in Represents the coefficient of determination. Represents the root mean square error. Represents the total number of samples involved in the calculation. The FeO content represented by the model output This represents the mean of the FeO labels in the actual output.

[0108] Table 2 Comparison of experimental results

[0109] RCM-BNLM 0. 000134 0.9799 LSTM 0.001073 0.8388 PLS 0.002391 0.6389

[0110] Table 2 and Figure 4 Detailed results of the comparative experiments are presented, demonstrating that the RCM-BNLM method of this invention achieves optimal measurement results. While PLS is a classic regression modeling method, it overemphasizes the construction of correlation functions, neglecting the dynamic and nonlinear problems introduced by long-term time dependencies during the sintering process, thus exhibiting the lowest performance. LSTM employs a gating structure, mitigating gradient explosion or vanishing caused by long-term dependencies and accelerating model convergence. However, LSTM's explicit gating structure modeling is inefficient, and long-term dependency issues remain a concern. RCM-BNLM utilizes a recursive coordinate mapping structure to achieve data dimensionality enhancement, operating on the data in a higher dimension, effectively improving model performance, accelerating convergence, and providing accurate FeO soft measurement results.

[0111] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for predicting FeO in sintered ore based on recursive coordinate mapping, characterized in that, Includes the following steps: 1) Determine FeO-related variables from the sintering plant database as input features, and perform sliding window encapsulation preprocessing on the data; 2) Construct a soft measurement model for FeO content, a quality index of sintered ore, based on a recursive coordinate coding network. The input data is mapped to a high-dimensional coordinate space through recursive coordinate mapping, and then the coordinate space data is mapped back to the time domain through nonlinear mapping transformation to realize the soft measurement of FeO. The recursive coordinate coding network structure includes input data u, output data y, system transition matrix A, system input matrix B, and inverse nonlinear mapping module BNLM. In step 1), the sliding window encapsulation preprocessing refers to the following processing of the filtered input data during the deployment of the sintering soft measurement model: ； ； in Represents the packaged first Window data, Represents the original data. This represents indexing the data. Representing FeO data, Represents window size; In step 2), the recursive coordinate mapping involves performing the following operations on the encapsulated filtered data: ； ； ； in Represents the first-order derivative operator. Represents the moment The next The order coordinates can be taken as follows: , , Consistent with the definition in step 2), Let the order be the global coordinate space. The index is the subscript of the coordinate order; In step 2), the inverse nonlinear mapping change refers to the operation performed on the output after the aforementioned coordinate mapping during the sintering process: ； in Represents the input vector. , Represents a linear transformation matrix. , Is with Vectors with a consistent shape.

2. The method according to claim 1, characterized in that, In step 1), the variables related to FeO are: neutralization bin flow rate, light burning bin flow rate, pulverized coal bin flow rate, coke bin flow rate, return ore bin flow rate, quicklime bin flow rate, first mixed water flow rate, second mixed water flow rate, sinter thickness, main flue negative pressure, main flue exhaust gas temperature, main fan frequency, No. 21 wind box temperature, No. 22 wind box temperature, No. 23 wind box temperature, No. 24 wind box temperature, Section I annular cooler temperature, Section II annular cooler temperature, and Section III annular cooler temperature. The model will extract high-level latent space features through input variables to achieve FeO soft measurement.

Images