Intelligent carbon potential control method and system based on multi-source data perception

Abstract

The invention relates to an intelligent carbon potential control method and system based on multi-source data perception, and the method comprises the steps: collecting parameter vectors corresponding to M parameter types, comprehensively obtaining the multi-source information related to the carbon potential, carrying out the reasonable splicing and processing of the parameter vectors, and obtaining the multi-source information related to the carbon potential. The carbon potential corresponding to the preset target time point can be accurately predicted through the trained time sequence offset prediction model, the offset mapping model and the parameter prediction model, and compared with a traditional prediction mode depending on a single parameter or a small number of parameters, potential relations and time sequence characteristics among different parameters are fully mined, and the prediction accuracy is improved. The accuracy of carbon potential prediction is greatly improved, and a reliable basis is provided for subsequent accurate control. The multi-source data sensing mode can comprehensively reflect complex and changeable working conditions in actual production, the model can learn and adapt to the relation between parameters and carbon potential under different working conditions, high robustness and adaptability are achieved, and the application range of the method in different production environments is widened.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon potential control, and in particular to an intelligent carbon potential control method and system based on multi-source data perception. Background Art

[0002] With the development of heat treatment process technology, carbon potential control plays a key role in ensuring product quality and performance. Traditional carbon potential control methods often rely on a single data source or simple mathematical models to regulate carbon potential. For example, some methods only indirectly infer carbon potential based on temperature data within the furnace. However, in actual production environments, there are many factors that affect carbon potential, such as gas flow rate, component ratio, and workpiece material properties. A single data source cannot fully and accurately reflect the dynamic changes of carbon potential.

[0003] Some methods, while taking multiple parameters into account, fail to fully exploit the inherent correlations and timing characteristics between different parameters. Faced with complex and changing working conditions, these traditional methods are unable to predict carbon potential in a timely and accurate manner, resulting in poor carbon potential control precision and prone to problems such as insufficient or excessive carburizing. Insufficient carburizing will result in the workpiece surface hardness and wear resistance failing to meet requirements, affecting the product's service life; excessive carburizing may cause the workpiece to deform and increase brittleness, also reducing product quality. At the same time, the inability to achieve precise control often requires more energy and raw materials to compensate for control errors, increasing production costs.

[0004] Therefore, how to improve the accuracy of carbon potential control has become an urgent problem to be solved. Summary of the Invention

[0005] In order to achieve the purpose of the present invention, the technical solution adopted by the present invention is: an intelligent carbon potential control method based on multi-source data perception, the intelligent carbon potential control method based on multi-source data perception comprises the following steps: S101, collecting parameter vectors corresponding to M parameter types in a preset time period, where M is a positive integer; S102, using a parameter vector of a first preset parameter type as a basic vector, and other parameter vectors as reference vectors; S103, obtaining an initial parameter matrix by concatenating the basis vector and each reference vector; S104, inputting the initial parameter matrix into the trained timing offset prediction model to obtain the timing offset corresponding to each reference vector; S105, jointly inputting the initial parameter matrix and the timing offsets corresponding to each reference vector into the trained offset mapping model to obtain an intermediate parameter matrix; S106, adjusting the intermediate parameter matrix according to a preset mapping function to obtain a target parameter matrix; S107, extracting a target sub-matrix from the target parameter matrix, inputting the target sub-matrix into the trained parameter prediction model to obtain a target prediction vector corresponding to a preset target time point; S108, determining the predicted carbon potential corresponding to the target time point based on the target prediction vector; S109: If the predicted carbon potential corresponding to the target time point is less than the preset first carbon potential threshold, the first carbon potential control process is executed; if the predicted carbon potential corresponding to the target time point is greater than the preset second carbon potential threshold, the second carbon potential control process is executed.

[0006] Compared with the prior art, the present invention has the following beneficial effects: 1. By collecting parameter vectors corresponding to M parameter types, we comprehensively acquire multi-source information related to carbon potential. These parameter vectors are rationally spliced ​​and processed. Through the trained timing offset prediction model, offset mapping model, and parameter prediction model, we can accurately predict the carbon potential corresponding to the preset target time point. Compared with traditional prediction methods that rely on a single or a small number of parameters, this model fully explores the potential relationships and timing characteristics between different parameters, greatly improving the accuracy of carbon potential prediction and providing a reliable basis for subsequent precise control.

[0007] 2. The multi-source data perception method can fully reflect the complex and changeable working conditions in actual production. The model can learn and adapt to the relationship between parameters and carbon potential under different working conditions. Whether in complex situations such as equipment aging, environmental changes, or fluctuations in raw material characteristics, it can stably and accurately predict and control carbon potential. It has strong robustness and adaptability, which broadens the application scope of this method in different production environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a flow chart of an intelligent carbon potential control method based on multi-source data perception in the first embodiment of the present invention; Figure 2 This is a structural diagram of an intelligent carbon potential control system based on multi-source data perception in Example 2 of the present invention. DETAILED DESCRIPTION

[0009] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments 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 making creative efforts are within the scope of protection of the present invention.

[0010] Example 1: like Figure 1As shown, the present invention provides a technical solution: an intelligent carbon potential control method based on multi-source data perception, characterized in that the intelligent carbon potential control method based on multi-source data perception includes the following steps: S101 , collecting parameter vectors corresponding to M parameter types in a preset time period.

[0011] Wherein, M is a positive integer, the preset time period may include several preset time points, the time interval between adjacent preset time points is fixed to T, and a single parameter vector may include parameter values ​​collected by sensors of corresponding parameter types at each preset time point.

[0012] S102: Using a parameter vector of a first preset parameter type as a basic vector and other parameter vectors as reference vectors.

[0013] The first preset type may refer to a parameter type directly related to carbon potential.

[0014] Specifically, in this embodiment, there is one basis vector and M-1 reference vectors.

[0015] In a specific embodiment, the parameter types include at least oxygen probe signal parameters, carbon monoxide concentration signal parameters, furnace temperature signal parameters, and furnace pressure parameters; The first preset type is the oxygen probe signal parameter.

[0016] Among them, parameter types can also include enriched gas flow, furnace circulation fan current, workpiece surface infrared temperature measurement, etc. The implementer can set the parameter type according to actual needs.

[0017] The oxygen probe signal parameters are the type of parameters that are directly related to carbon potential.

[0018] S103, an initial parameter matrix is ​​obtained by concatenating the basic vector and each reference vector.

[0019] Among them, the basic vector is used as the first row vector and is spliced ​​with other reference vectors row by row to obtain the initial parameter matrix.

[0020] S104: Input the initial parameter matrix into the trained timing offset prediction model to obtain the timing offset corresponding to each reference vector.

[0021] Among them, the timing offset defaults to an integer. The architecture of the timing offset prediction model can adopt a convolution module and a fully connected layer. The output of the convolution module is connected to the input of the fully connected layer. The convolution module is used to extract the feature information of the input initial parameter matrix, and the fully connected layer is used to map the feature information extracted by the convolution module to the timing offset corresponding to each reference vector.

[0022] It should be noted that in the embodiment, the output dimension of the full connection layer is set to be the same as the number of columns of the initial parameter matrix, so as to better correspond the time offset to the parameter type. Since the parameter vector of the first preset type is a basis vector and does not undergo offset processing, the time offset corresponding to the parameter vector of the first preset type is zero by default. In this step, only the time offset corresponding to each reference vector is described.

[0023] In a specific embodiment, the training process of the time offset prediction model includes the following steps: Obtain M historical vectors corresponding to M parameter types, and take the historical vector of the first preset type as a first vector and the other historical vectors as second vectors; Splice the M historical vectors corresponding to the M parameter types to obtain a parameter sample matrix; For any second vector, obtain a preset initial offset of the parameter type corresponding to the second vector, take the preset initial offset as a temporary offset, and perform a shift operation on the second vector according to the temporary offset to obtain a shifted vector corresponding to the second vector; Splice the first vector and the shifted vector corresponding to the second vector to obtain a temporary matrix; According to the preset initial offset, extract a temporary sub-matrix from the temporary matrix; Input the temporary sub-matrix into the trained parameter prediction model to obtain temporary prediction vectors corresponding to a plurality of historical time points; Obtain temporary true vectors corresponding to the plurality of historical time points, and calculate a prediction accuracy according to the temporary prediction vectors corresponding to the plurality of historical time points and the temporary true vectors corresponding to the plurality of historical time points; If the prediction accuracy is less than a preset accuracy threshold, update the temporary offset by using a preset optimization algorithm until the prediction accuracy corresponding to the temporary offset meets an optimization target, and take the finally updated temporary offset as a target offset corresponding to the second vector; Splice the first preset value and the target offsets corresponding to the second vectors to obtain offset label data; Calculate a first training loss according to the parameter sample matrix, the offset label data, and a preset first training loss function, train the time offset prediction model according to the first training loss until the first training loss converges, and obtain a trained time offset prediction model.

[0024] The historical vector can be formed by parameter values collected by a sensor of a corresponding parameter type in a historical time period.

[0025] The preset initial offset of the parameter type can be set by the implementer or can be the offset used in the historical inference process. The purpose of introducing the preset initial offset is to provide a data basis for optimizing the temporary offset when the temporary offset needs to be optimized, so as to minimize the calculation amount and time consumption of the optimization and improve the optimization efficiency.

[0026] Specifically, when the second vector is shifted according to the temporary offset, the embodiment uses a zero padding operation to ensure that the shift vector corresponding to the second vector is aligned with the first vector.

[0027] The first vector is taken as the first row, and the shift vector corresponding to the second vector is spliced by row to obtain a temporary matrix.

[0028] Under the premise that the preset initial offset is known, the time interval covered by the first vector and the shift vector corresponding to the second vector can be determined, that is, the effective columns in the temporary matrix are determined, so that the temporary sub-matrix is extracted from the temporary matrix according to the effective columns.

[0029] The temporary sub-matrix is input into the trained parameter prediction model, and iterative prediction is performed by column to obtain temporary prediction vectors corresponding to respective historical time points. The temporary prediction vector is a column vector. It should be noted that, in order to ensure that the output size of the temporary sub-matrix and the parameter prediction model is consistent, the first vector, the shift vector corresponding to the second vector, and the zero vector corresponding to each of the other second vectors can be spliced by row to obtain a temporary matrix. Since the columns of the temporary sub-matrix can represent time sequences, iterative prediction is performed by column. For details of the iterative prediction process, refer to the description of the parameter prediction model.

[0030] It can be known that each historical time point has actually performed parameter collection, so the temporary true vectors corresponding to respective historical time points can be directly obtained. According to the temporary prediction vector and the temporary true vector corresponding to each historical time point, the sub-accuracy corresponding to each historical time point can be calculated by using a cosine similarity or other measurement method. The prediction accuracy can be obtained by performing mean value calculation on the sub-accuracy corresponding to each historical time point.

[0031] In the embodiment, the preset accuracy threshold can be 0.8. The implementer can adjust the accuracy threshold according to the actual prediction accuracy requirement. The higher the actual prediction accuracy requirement, the closer the accuracy threshold should be to 1.

[0032] The preset optimization algorithm uses the hill climbing method in the embodiment. After updating the temporary offset, the prediction accuracy is recalculated until the prediction accuracy corresponding to the temporary offset meets the optimization target. The finally updated temporary offset is taken as the target offset corresponding to the second vector.

[0033] The first preset value can be set to 0, and the first training loss function can adopt a mean square error loss function.

[0034] In a specific embodiment, the optimization goal is that the prediction accuracy corresponding to the temporary offset is a local maximum; After the prediction accuracy corresponding to the temporary offset meets the optimization goal, it also includes: The offset confidence of the parameter type corresponding to the second vector is determined according to the prediction accuracy corresponding to the temporary offset.

[0035] Among them, since this embodiment performs optimization based on a preset initial offset, the optimization process can be simplified, so the hill climbing method is used for optimization, and the local maximum is used as the optimization target. Of course, the implementer can also use an optimization method such as a simulated annealing algorithm, and use the global maximum as the optimization target.

[0036] The offset confidence can be used to indicate the updated timing offset in the subsequent prediction process.

[0037] Specifically, in this embodiment, the prediction accuracy rates are normalized, so the prediction accuracy rate corresponding to the temporary offset can be directly used as the offset confidence level of the corresponding parameter type.

[0038] S105 , jointly inputting the initial parameter matrix and the timing offsets corresponding to each reference vector into the trained offset mapping model to obtain an intermediate parameter matrix.

[0039] Among them, the first preset value corresponding to the basic vector and the timing offset corresponding to each reference vector can be spliced ​​into a one-dimensional vector with the same number of columns as the initial parameter matrix. By splicing the initial parameter matrix and the one-dimensional vector by row, the input matrix of the trained offset mapping model can be obtained. The size of the intermediate parameter matrix is ​​consistent with the size of the initial parameter matrix.

[0040] The offset mapping model can adopt a network structure similar to U-Net, the only difference is that the input size of the encoder and the output size of the decoder are different.

[0041] Specifically, the offset mapping model is only used to perform mapping of the offset processing to ensure the differentiability of the offset processing, thereby facilitating the subsequent reverse update of the timing offset. Therefore, the implementer can randomly generate training samples including matrices and offsets, determine the training labels based on the shift processing, and train the offset mapping model based on the training samples and training labels in combination with the mean square error loss function to obtain a trained offset mapping model.

[0042] S106: Adjust the intermediate parameter matrix according to a preset mapping function to obtain a target parameter matrix.

[0043] The preset mapping function is used to extract the effective columns of the intermediate parameter matrix and ensure the differentiability of the extraction process.

[0044] In a specific embodiment, the intermediate parameter matrix is ​​adjusted according to a preset mapping function to obtain a target parameter matrix, including: Calculate the minimum value of each column in the intermediate parameter matrix, and add the minimum value of each column and the second preset value as the indicator value of each column; For any element value in the intermediate parameter matrix, mapping is performed to obtain a mapping value corresponding to the element value according to the element value, the indication value of the column to which the element value belongs, and a preset mapping function; The target parameter matrix is ​​determined according to the mapping values ​​corresponding to all element values.

[0045] The second preset value c can be a minimum value, which can be set to 10 in this embodiment. -6 .

[0046] In a specific implementation, the preset mapping function is specifically: y=exp(-a / b j )×d ij ; Among them, d ij is the element value of the i-th row and j-th column in the intermediate parameter matrix, i is an integer in the range [1, M], j is an integer in the range [1, K], K is the total number of columns in the intermediate parameter matrix, b j is the indicator value of the j-th column, a is the third preset value, and exp() is the exponential function.

[0047] The third preset value is also a minimum value, but the exponential magnitude of the third preset value is greater than the exponential magnitude of the second preset value. In this embodiment, the third preset value can be 10 -3 .

[0048] Specifically, when there is an element value of zero in the jth column, it can be considered that the column is not a valid column, that is, not all shift vectors have corresponding collection values ​​in the column, then b j =10 -6 , -a / b j =-10 3 , exp(-a / b j ) approaches 0, so that any element in the column is mapped to 0.

[0049] When there is no zero element value in the jth column, the column can be considered as a valid column, that is, all shift vectors have corresponding collection values ​​in the column, then b j =min j +c,min j The minimum element value in the jth column, the default is bj Much larger than the third preset value, b j Can be approximately regarded as min j If they are the same, then -a / b j Approaching 0, exp(-a / b j ) approaches 1, so that any element in the column is mapped to itself.

[0050] S107 , extracting a target sub-matrix from the target parameter matrix, inputting the target sub-matrix into the trained parameter prediction model, and obtaining a target prediction vector corresponding to a preset target time point.

[0051] Among them, the parameter prediction model can adopt a time series prediction model, such as a recurrent neural network model, a long short-term memory network model, and a time domain convolutional network model.

[0052] In this embodiment, the parameter prediction model adopts a time domain convolutional network model. The architecture of the time domain convolutional network model will not be described in detail here. The training of the time domain convolutional network model can be directly performed based on the parameter sample matrix obtained by splicing historical vectors. Since this embodiment improves the accuracy of the prediction process by adjusting the temporal relationship of the parameter vectors, there is no need for additional training of the parameter prediction model, and conventional training methods can be used.

[0053] In a specific embodiment, a target submatrix is ​​extracted from a target parameter matrix, and the target submatrix is ​​input into a trained parameter prediction model to obtain a target prediction vector corresponding to a preset target time point, including: Extracting a target submatrix from the target parameter matrix, where the number of columns of the target submatrix is ​​N, and N is a positive integer; Aligning the left edge of the preset sliding window with the first column of the target submatrix, and extracting the first intermediate matrix from the target submatrix using the preset sliding window, wherein the length of the preset sliding window is L, where L is a positive integer less than N; Input the first intermediate matrix into the trained parameter prediction model to obtain LN intermediate prediction vectors; Calculate the prediction reliability based on the column vectors corresponding to the L+1th column to the Nth column in the target submatrix and the LN intermediate prediction vectors; If the prediction reliability is greater than or equal to the preset threshold, aligning the right edge of the preset sliding window with the last column of the target submatrix, and extracting a second intermediate matrix from the target submatrix using the preset sliding window; Input the second intermediate matrix into the trained parameter prediction model to obtain the target prediction vector corresponding to the target time point; If the prediction reliability is less than the preset threshold, the timing offsets corresponding to the respective reference vectors are optimized according to the offset confidences corresponding to the respective reference vectors, and the process returns to step S105 .

[0054] The preset sliding window width is M, which is the same as the number of rows of the target submatrix.

[0055] Specifically, the size of the target submatrix is ​​N×M, corresponding to the target parameter vectors from the 1st to the Nth preset time points, the time interval between adjacent preset time points is fixed to T, the first intermediate matrix corresponds to the target parameter vectors from the 1st to the Lth preset time points, the first intermediate matrix is ​​input into the trained parameter prediction model to obtain the intermediate prediction vector of the L+1th preset time point, and then the target parameter vectors from the 2nd to the Lth preset time points and the intermediate prediction vector of the L+1th preset time point are input into the trained parameter prediction model to obtain the intermediate prediction vector of the L+2th preset time point, and then the target parameter vectors from the 3rd to the Lth preset time point and the intermediate prediction vector from the L+1th preset time point to the L+2th intermediate point are input into the trained parameter prediction model to obtain the intermediate prediction vector of the L+3th preset time point, and so on, iterative prediction is performed to obtain LN intermediate prediction vectors.

[0056] According to any preset time point from the L+1th preset time point to the Nth preset time point, the sub-reliability is calculated using the intermediate prediction vector and the column vector in the target sub-matrix corresponding to the preset time point through metrics such as cosine similarity, and then the mean of all sub-reliabilities is calculated to obtain the prediction reliability.

[0057] The preset degree threshold is set to 0.8 in this embodiment. Similarly, the implementer can adjust the preset degree threshold according to the actual prediction accuracy requirement. The higher the actual prediction accuracy requirement, the closer the preset degree threshold should be to 1.

[0058] When the prediction reliability is greater than or equal to the preset threshold, it means that the timing offset is relatively reliable in this case, and there is no need to adjust the timing offset. The second intermediate matrix can be directly extracted for iterative prediction. The target time point can refer to the Rth preset time point, where R is an integer greater than N. The target prediction vector corresponding to the target time point can be obtained by iterative prediction based on the second intermediate matrix. The specific iterative prediction process will not be repeated here.

[0059] When the prediction reliability is less than the preset threshold, it means that the timing offset is unreliable in this case. The reason is that when the timing offset is determined, the local maximum is used as the optimization target. The local maximum may not meet the prediction accuracy threshold. Therefore, it is necessary to optimize the timing offset corresponding to each reference vector. The optimization method can use the reverse gradient descent method. The learning rate of the timing offset update corresponding to the reference vector with different offset confidence is different. The higher the offset confidence, the lower the learning rate when the timing offset of the corresponding reference vector is updated.

[0060] S108: Determine the predicted carbon potential corresponding to the target time point according to the target prediction vector.

[0061] Among them, the first row elements in the target prediction vector are extracted to obtain the oxygen probe signal parameters corresponding to the target time point.

[0062] Based on the oxygen probe signal parameters and the prior relationship between oxygen partial pressure and carbon potential, the predicted carbon potential corresponding to the target time point can be calculated.

[0063] S109: If the predicted carbon potential corresponding to the target time point is less than the preset first carbon potential threshold, the first carbon potential control process is executed; if the predicted carbon potential corresponding to the target time point is greater than the preset second carbon potential threshold, the second carbon potential control process is executed.

[0064] Among them, the implementer can set the carbon potential range interval. When the predicted carbon potential corresponding to the target time point is within the carbon potential range interval, no regulation is required. The left boundary value of the carbon potential range interval can be the first carbon potential threshold, and the right boundary value can be the second carbon potential threshold.

[0065] Specifically, when the predicted carbon potential corresponding to the target time point is less than the preset first carbon potential threshold, the first carbon potential control process is executed. The first carbon potential control process may include increasing the carbon supply, reducing the exhaust valve opening, etc. When the predicted carbon potential corresponding to the target time point is greater than the preset second carbon potential threshold, the second carbon potential control process is executed. The second carbon potential control process may include reducing the enriched gas flow rate, appropriately increasing the furnace circulation fan speed, etc.

[0066] It should be noted that the implementer of the regulation object corresponding to the specific carbon potential control process can set it according to actual conditions, and the specific regulation amount of the regulation object can be determined by using PID control, expert preset rules, reinforcement learning and other methods in existing technologies.

[0067] This embodiment 1 comprehensively obtains multi-source information related to carbon potential by collecting parameter vectors corresponding to M parameter types, reasonably splices and processes these parameter vectors, and accurately predicts the carbon potential corresponding to the preset target time point through the trained timing offset prediction model, offset mapping model, and parameter prediction model. Compared with the traditional prediction method that relies on a single or a small number of parameters, it fully explores the potential relationship and timing characteristics between different parameters, greatly improves the accuracy of carbon potential prediction, and provides a reliable basis for subsequent precise control. The multi-source data perception method can fully reflect the complex and changeable working conditions in actual production. The model can learn and adapt to the relationship between parameters and carbon potential under different working conditions. It has strong robustness and adaptability, and broadens the application scope of this method in different production environments.

[0068] Example 2: like Figure 2 As shown, the present invention provides a technical solution: an intelligent carbon potential control system based on multi-source data perception, characterized in that the intelligent carbon potential control system based on multi-source data perception includes: The parameter collection module 201 is used to collect parameter vectors corresponding to M parameter types in a preset time period, where M is a positive integer; A vector classification module 202, configured to use parameter vectors of a first preset parameter type as basic vectors and other parameter vectors as reference vectors; A vector concatenation module 203 is configured to concatenate the basic vectors and the reference vectors to obtain an initial parameter matrix; The offset prediction module 204 is used to input the initial parameter matrix into the trained timing offset prediction model to obtain the timing offset corresponding to each reference vector; The matrix mapping module 205 is used to input the initial parameter matrix and the timing offsets corresponding to each reference vector into the trained offset mapping model to obtain an intermediate parameter matrix; The matrix adjustment module 206 is used to adjust the intermediate parameter matrix according to a preset mapping function to obtain a target parameter matrix; The parameter prediction module 207 is used to extract a target sub-matrix from the target parameter matrix, input the target sub-matrix into the trained parameter prediction model, and obtain a target prediction vector corresponding to a preset target time point; The carbon potential prediction module 208 is used to determine the predicted carbon potential corresponding to the target time point based on the target prediction vector; The carbon potential control module 209 is used to execute the first carbon potential control process if the predicted carbon potential corresponding to the target time point is less than the preset first carbon potential threshold, and to execute the second carbon potential control process if the predicted carbon potential corresponding to the target time point is greater than the preset second carbon potential threshold.

[0069] It should be noted that the specific definition of an intelligent carbon potential control system based on multi-source data perception can be found in the definition of an intelligent carbon potential control method based on multi-source data perception above, and will not be repeated here. The information interaction and execution process between the above modules, etc., are based on the same concept as the method embodiments of the present invention. Their specific functions and technical effects can be found in the method embodiments section and will not be repeated here.

[0070] The embodiments disclosed in the present invention are preferred embodiments, but are not limited to them. Ordinary technicians in this field can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. As long as they do not deviate from the spirit of the present invention, they are all within the scope of protection of the present invention.

Claims

1. An intelligent carbon potential control method based on multi-source data perception, characterized in that: The intelligent carbon potential control method based on multi-source data perception includes the following steps: S101, collecting parameter vectors corresponding to M parameter types in a preset time period, where M is a positive integer; S102, using a parameter vector of a first preset parameter type as a basic vector, and other parameter vectors as reference vectors; S103, obtaining an initial parameter matrix by concatenating the basic vector and each reference vector; S104, inputting the initial parameter matrix into the trained timing offset prediction model to obtain the timing offset corresponding to each reference vector; S105, jointly inputting the initial parameter matrix and the timing offsets corresponding to each reference vector into a trained offset mapping model to obtain an intermediate parameter matrix; S106, adjusting the intermediate parameter matrix according to a preset mapping function to obtain a target parameter matrix; S107, extracting a target sub-matrix from the target parameter matrix, inputting the target sub-matrix into a trained parameter prediction model to obtain a target prediction vector corresponding to a preset target time point; S108, determining the predicted carbon potential corresponding to the target time point according to the target prediction vector; S109, if the predicted carbon potential corresponding to the target time point is less than the preset first carbon potential threshold, the first carbon potential control process is executed; if the predicted carbon potential corresponding to the target time point is greater than the preset second carbon potential threshold, the second carbon potential control process is executed.

2. The intelligent carbon potential control method based on multi-source data perception according to claim 1 is characterized in that: The parameter types include at least oxygen probe signal parameters, carbon monoxide concentration signal parameters, furnace temperature signal parameters, and furnace pressure parameters; The first preset type is the oxygen probe signal parameter.

3. The intelligent carbon potential control method based on multi-source data perception according to claim 1 is characterized in that: The training process of the timing offset prediction model includes the following steps: Obtaining historical vectors corresponding to M parameter types, taking the historical vector whose parameter type is the first preset type as the first vector, and the other historical vectors as the second vector; The parameter sample matrix is ​​obtained by concatenating the historical vectors corresponding to the M parameter types; For any second vector, obtain a preset initial offset of a parameter type corresponding to the second vector, use the preset initial offset as a temporary offset, and perform shift processing on the second vector according to the temporary offset to obtain a shift vector corresponding to the second vector; Concatenate the first vector and the shift vector corresponding to the second vector to obtain a temporary matrix; extracting a temporary sub-matrix from the temporary matrix according to the preset initial offset; Inputting the temporary submatrix into the trained parameter prediction model to obtain temporary prediction vectors corresponding to several historical time points; Obtaining the temporary true vector corresponding to each historical time point, and calculating the prediction accuracy based on the temporary prediction vector corresponding to each historical time point and the temporary true vector corresponding to each historical time point; If the prediction accuracy is less than a preset accuracy threshold, the temporary offset is updated using a preset optimization algorithm until the prediction accuracy corresponding to the temporary offset meets the optimization target, and the final updated temporary offset is used as the target offset corresponding to the second vector; The offset label data is obtained by concatenating the first preset value and the target offset corresponding to each second vector; According to the parameter sample matrix, the offset label data and the preset first training loss function, a first training loss is calculated, and the timing offset prediction model is trained according to the first training loss until the first training loss converges to obtain the trained timing offset prediction model.

4. The intelligent carbon potential control method based on multi-source data perception according to claim 3 is characterized in that: The optimization goal is that the prediction accuracy corresponding to the temporary offset is a local maximum; After the prediction accuracy corresponding to the temporary offset satisfies the optimization target, the method further includes: The offset confidence of the parameter type corresponding to the second vector is determined according to the prediction accuracy corresponding to the temporary offset.

5. The intelligent carbon potential control method based on multi-source data perception according to claim 4 is characterized in that: Extracting a target submatrix from the target parameter matrix, inputting the target submatrix into a trained parameter prediction model to obtain a target prediction vector corresponding to a preset target time point, including: Extracting a target submatrix from the target parameter matrix, wherein the number of columns of the target submatrix is ​​N, where N is a positive integer; Aligning a left edge of a preset sliding window with a first column of the target submatrix, and extracting a first intermediate matrix from the target submatrix using the preset sliding window, wherein a length of the preset sliding window is L, where L is a positive integer less than N; Inputting the first intermediate matrix into the trained parameter prediction model to obtain LN intermediate prediction vectors; Calculating the prediction reliability based on the column vectors corresponding to the L+1th column to the Nth column in the target submatrix and the LN intermediate prediction vectors; If the prediction reliability is greater than or equal to a preset threshold, aligning the right edge of the preset sliding window with the last column of the target submatrix, and extracting a second intermediate matrix from the target submatrix using the preset sliding window; Inputting the second intermediate matrix into the trained parameter prediction model to obtain a target prediction vector corresponding to the target time point; If the prediction reliability is less than the preset threshold, the timing offsets corresponding to the respective reference vectors are optimized according to the offset confidences corresponding to the respective reference vectors, and the process returns to step S105.

6. The intelligent carbon potential control method based on multi-source data perception according to claim 1 is characterized in that: The step of adjusting the intermediate parameter matrix according to a preset mapping function to obtain a target parameter matrix includes: Calculating the minimum value of each column in the intermediate parameter matrix, and adding the minimum value of each column and a second preset value as an indication value of each column; For any element value in the intermediate parameter matrix, mapping is performed to obtain a mapping value corresponding to the element value according to the element value, the indication value of the column to which the element value belongs, and the preset mapping function; The target parameter matrix is ​​determined according to the mapping values ​​corresponding to all element values.

7. The intelligent carbon potential control method based on multi-source data perception according to claim 6 is characterized in that: The preset mapping function is specifically: y=exp(-a / b j )×d ij ; Among them, d ij is the element value of the i-th row and j-th column in the intermediate parameter matrix, i is an integer in the range of [1, M], j is an integer in the range of [1, K], K is the total number of columns in the intermediate parameter matrix, and b j is the indicator value of the j-th column, a is the third preset value, and exp() is the exponential function.

8. An intelligent carbon potential control system based on multi-source data perception, characterized in that: The intelligent carbon potential control system based on multi-source data perception includes: A parameter collection module is used to collect parameter vectors corresponding to M parameter types in a preset time period, where M is a positive integer; a vector classification module, configured to use a parameter vector of a first preset parameter type as a basic vector and other parameter vectors as reference vectors; A vector splicing module, configured to obtain an initial parameter matrix by splicing the basic vector and each reference vector; An offset prediction module, configured to input the initial parameter matrix into a trained timing offset prediction model to obtain timing offsets corresponding to respective reference vectors; A matrix mapping module is used to input the initial parameter matrix and the timing offsets corresponding to each reference vector into a trained offset mapping model to obtain an intermediate parameter matrix; A matrix adjustment module is used to adjust the intermediate parameter matrix according to a preset mapping function to obtain a target parameter matrix; A parameter prediction module is used to extract a target submatrix from the target parameter matrix, input the target submatrix into a trained parameter prediction model, and obtain a target prediction vector corresponding to a preset target time point; a carbon potential prediction module, configured to determine the predicted carbon potential corresponding to the target time point based on the target prediction vector; The carbon potential control module is used to execute the first carbon potential control process if the predicted carbon potential corresponding to the target time point is less than the preset first carbon potential threshold, and to execute the second carbon potential control process if the predicted carbon potential corresponding to the target time point is greater than the preset second carbon potential threshold.

Images