Cold rolling mill rolling force prediction and adaptive control method based on multi-sensor fusion

Abstract

The application provides a cold rolling mill rolling force prediction and adaptive control method based on multi-sensor fusion, relates to the cold rolling mill control technical field, and comprises the following steps: collecting sensor process parameters, constructing a parameter correlation matrix through a graph attention network, reconstructing a working condition characteristic sequence, combining a deformation zone contact arc length change characteristic to predict a rolling force, constructing a control strategy based on a rolling force error value and optimizing control parameters, searching for optimal process parameters in a parameter optimization space and converting the optimal process parameters into control instructions, and the like, so that the rolling force prediction accuracy and control precision can be improved, and the system anti-interference capability is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cold rolling mill control, and particularly relates to a cold rolling mill rolling force prediction and adaptive control method and system based on multi-sensor fusion. BACKGROUND

[0002] In the steel cold rolling production process, the rolling force as a key process parameter has an important influence on product quality and equipment operation stability, and accurate prediction of the rolling force and effective adaptive control are the key to guarantee the cold rolling product quality and improve production efficiency. In the traditional cold rolling production, the rolling force prediction mainly relies on experience, which is difficult to adapt to the complex and changeable working conditions in the actual production process. With the development of industrial automation and intelligent manufacturing, the advanced prediction and control method based on multi-sensor fusion has become an important research direction for cold rolling process optimization.

[0003] The existing rolling force prediction and control mainly analyze the relationship between process parameters and rolling force through experience model and finite element analysis technology, but there are still problems such as difficulty in effectively processing the complex correlation between multi-source heterogeneous sensor data, neglecting the dynamic interaction between different process parameters, leading to insufficient prediction accuracy, control parameters cannot be adjusted in time when facing nonlinear changes and sudden disturbances in the rolling process, which easily causes product quality fluctuations and equipment vibration and other problems, and unable to consider the multi-objective balance of sheet shape quality and equipment operation constraints at the same time, which is difficult to realize global optimal control in actual production, thereby affecting production efficiency and product consistency and other problems. SUMMARY

[0004] The embodiment of the present application provides a cold rolling mill rolling force prediction and adaptive control method based on multi-sensor fusion, which can at least solve some problems in the prior art.

[0005] In a first aspect, the embodiment of the present application provides a cold rolling mill rolling force prediction and adaptive control method based on multi-sensor fusion, comprising:

[0006] Collecting process parameters collected by a cold rolling mill sensor, performing feature extraction on the process parameters to obtain a working condition feature sequence, constructing a dynamic correlation matrix between process parameters through a graph attention network and determining an influence weight, reconstructing the working condition feature sequence based on the influence weight, and calculating a rolling force prediction value in combination with the contact arc length change characteristics of the deformation zone in the rolling process;

[0007] Calculating a rolling force error value between the rolling force prediction value and the actual rolling force, constructing a control strategy based on the rolling force error value and calculating an action evaluation value, optimizing and noise suppressing the rolling mill control parameters based on the action evaluation value to obtain a control compensation value;

[0008] Based on the control compensation value, a parameter optimization space is constructed, in which, in combination with the strip shape quality constraint and the equipment operation constraint, a parameter optimization is performed by a grid search algorithm to obtain optimal process parameters, the optimal process parameters are converted into cold rolling mill control instructions and are issued for execution.

[0009] In an alternative embodiment,

[0010] Process parameters collected by a cold rolling mill sensor are collected, and a working condition feature sequence is obtained by feature extraction on the process parameters. A dynamic correlation matrix between process parameters is constructed by a graph attention network, and an influence weight is determined, including:

[0011] The process parameters are divided into a plurality of parameter sets of types according to physical properties, and a plurality of feature subsequences are obtained by combining the features of each parameter set according to time sequences. The feature subsequences are combined to obtain the working condition feature sequence;

[0012] Based on the working condition feature sequence, a heterogeneous graph structure is constructed, and a meta-path is set in the heterogeneous graph according to the parameter type. For each meta-path, the semantic similarity between different parameter nodes is calculated by a pre-set metric matrix, and the attention score corresponding to the parameter node is calculated based on the attention mechanism;

[0013] Based on the semantic similarity and the attention score, the dynamic correlation matrix is constructed in combination with the edges corresponding to different parameter types in the heterogeneous graph. The degree centrality and feature vector centrality corresponding to each parameter node are calculated based on the dynamic correlation matrix, and the influence weight is solved.

[0014] In an alternative embodiment,

[0015] Based on the influence weight, the working condition feature sequence is reconstructed, and a rolling force prediction value is calculated in combination with the contact arc length change characteristics of the deformation zone in the rolling process, including:

[0016] The material composition and physical properties corresponding to the current processing material are queried in a pre-set material property database, and a material property vector corresponding to the current processing material is constructed. The material property vector and the working condition feature sequence are subjected to Hadamard product operation to obtain a feature correlation sequence. The feature correlation sequence is reconstructed based on the influence weight to obtain a reconstructed feature sequence;

[0017] The working roll radius and deformation amount in the rolling process are collected, and a pre-set multi-layer perception structure is used for processing to obtain the contact arc length change characteristics of the deformation zone and determine the time sequence characteristics corresponding to the contact arc length by a recurrent neural network;

[0018] Calculate the attention weight corresponding to the reconstructed feature sequence based on the timing characteristics, and combine the multi-layer convolution to obtain the rolling force prediction value corresponding to the reconstructed feature sequence.

[0019] In an optional implementation,

[0020] Based on the rolling force error value, a control strategy is constructed and an action evaluation value is calculated, and based on the action evaluation value, the rolling mill control parameters are optimized and noise suppression is performed to obtain a control compensation value, including:

[0021] Based on the rolling force error value and the pre-acquired actual rolling force, a relative rolling force error value is calculated, a timing data is constructed in combination with the pre-acquired historical relative rolling force error value, and a rolling process causal diagram is constructed based on the timing data and the process parameters;

[0022] Calculate the intervention distribution between nodes in the rolling process causal diagram to obtain a causal effect, screen to obtain a key causal path in combination with a pre-set causal effect threshold, determine a control strategy based on the key causal path and divide the control strategy into multiple levels, calculate a hierarchical control amount corresponding to each level based on the causal relationship between different types of nodes in the rolling process causal diagram, and calculate an action evaluation value based on the hierarchical control amount;

[0023] Based on the action evaluation value, the rolling mill control parameters are divided into a fast response parameter group and a steady state compensation parameter group, the first optimization result is obtained by optimizing the fast response parameter group using adaptive weight adjustment based on the causal effect intensity, the second optimization result is obtained by gradually optimizing the steady state compensation parameter group based on the rolling force error accumulation, and the control compensation value is obtained by combining the first optimization result and the second optimization result and performing noise suppression through a multi-layer perception machine.

[0024] In an optional implementation,

[0025] Based on the action evaluation value, the rolling mill control parameters are divided into a fast response parameter group and a steady state compensation parameter group, the first optimization result is obtained by optimizing the fast response parameter group using adaptive weight adjustment based on the causal effect intensity, the second optimization result is obtained by gradually optimizing the steady state compensation parameter group based on the rolling force error accumulation, and the control compensation value is obtained by combining the first optimization result and the second optimization result and performing noise suppression through a multi-layer perception machine.

[0026] Based on the action evaluation value, the rolling mill control parameters are analyzed for responsiveness, the response time is determined, and the rolling mill control parameters with a response time not less than a pre-set response speed threshold are divided into a steady state compensation parameter group, and the rolling mill control parameter group with a response time less than the response speed threshold is divided into a fast response parameter group;

[0027] For the rolling mill control parameters in the fast response parameter group, the sensitivity of each group of control parameters on the key causal path is calculated to obtain a causal effect strength, a weight increment is calculated based on the causal effect strength and the rolling force error value, and adaptive weight updating is performed on the fast response parameter group based on the weight increment to obtain a first optimization result;

[0028] The rolling force error value is integrated over a preset time period to determine a rolling force error accumulation, a compensation amount is calculated in combination with a pre-set gradual optimization coefficient, and gradual optimization is performed on the steady-state compensation parameter group based on the compensation amount to obtain a second optimization result.

[0029] In an optional implementation,

[0030] Based on the control compensation value, a parameter optimization space is constructed, in which, in combination with a plate shape quality constraint and a device operation constraint, a parameter optimization is performed by a grid search algorithm to obtain optimal process parameters, including:

[0031] The temperature distribution and the thermal stress distribution of the rolling area are collected, and the parameter optimization space is constructed in combination with pre-acquired material characteristics and the control compensation value;

[0032] In the parameter optimization space, a plate shape deviation value and a temperature field deviation value are calculated in combination with a plate shape quality constraint and a device operation constraint, and a parameter feasible region is limited to obtain a constrained parameter optimization space;

[0033] Based on the grid search algorithm, a grid is divided in the constrained parameter optimization space, a comprehensive evaluation index corresponding to each grid is calculated, a temperature gradient in the constrained parameter optimization space is calculated based on the comprehensive evaluation index, grids with a temperature gradient greater than a preset gradient threshold are refined, and parameter optimization is performed in the refined grids to obtain candidate process parameters;

[0034] The real-time temperature field is collected, the current thermal deformation amount is calculated in combination with the thermal stress distribution, and the thermal deformation compensation amount is determined based on the current thermal deformation amount. The thermal deformation compensation amount and the candidate process parameters are superimposed to obtain the optimal process parameters.

[0035] In an optional implementation,

[0036] Based on the grid search algorithm, a grid is divided in the constrained parameter optimization space, a comprehensive evaluation index corresponding to each grid is calculated, including:

[0037] In the constrained parameter optimization space, an initial grid is divided based on the grid search algorithm, and an initial position distribution of grid nodes is determined, a topological relationship matrix between the initial grids is determined based on the initial position distribution and production requirements;

[0038] Determine the geometric characteristic index of each initial grid based on the topological relation matrix and calculate the grid quality evaluation value, adjust the grid node position locally based on the grid quality evaluation value, obtain the modified grid and calculate the grid distortion degree and grid orthogonality corresponding to the modified grid;

[0039] Collect the plate shape quality value and temperature uniformity value corresponding to each modified grid, and calculate the comprehensive evaluation index based on the grid distortion degree and the grid orthogonality.

[0040] The second aspect of the embodiment of the application provides a cold rolling mill rolling force prediction and adaptive control system based on multi-sensor fusion, comprising:

[0041] The first unit is used for collecting process parameters collected by a cold rolling mill sensor, extracting features of the process parameters to obtain a working condition feature sequence, constructing a dynamic correlation matrix between the process parameters through a graph attention network and determining an influence weight, reconstructing the working condition feature sequence based on the influence weight, and calculating a rolling force prediction value in combination with a contact arc length change characteristic of a deformation zone in a rolling process;

[0042] The second unit is used for calculating a rolling force error value between the rolling force prediction value and an actual rolling force, constructing a control strategy based on the rolling force error value and calculating an action evaluation value, optimizing and noise suppressing a rolling mill control parameter based on the action evaluation value to obtain a control compensation value;

[0043] The third unit is used for constructing a parameter optimization space based on the control compensation value, in the parameter optimization space, combining a plate shape quality constraint and a device operation constraint, and performing parameter optimization through a grid search algorithm to obtain optimal process parameters, converting the optimal process parameters into a cold rolling mill control instruction and issuing the cold rolling mill control instruction for execution.

[0044] The third aspect of the embodiment of the application provides an electronic device, comprising:

[0045] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0046] The fourth aspect of the embodiment of the application provides a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions are executed by a processor to implement the method described above.

[0047] In the present application, the multi-sensor fusion technology is adopted to construct a dynamic correlation matrix among process parameters in combination with a graph attention network, which can accurately capture the mutual influence relationship among parameters, improve the accuracy and reliability of the rolling force prediction, reduce the negative impact of prediction deviation on control quality, construct a control strategy based on the rolling force error value and calculate the action evaluation value, effectively eliminate system interference through optimization of control parameters and noise suppression technology, improve the stability and robustness of the control system, make the rolling process more stable and controllable, combine the shape quality constraint and equipment operation constraint, and realize adaptive adjustment of the cold rolling mill control parameters through the grid search algorithm for parameter optimization, improve the quality consistency and production efficiency of the rolling products, reduce energy consumption and equipment wear, and prolong the service life of the equipment. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A flowchart of the cold rolling mill rolling force prediction and adaptive control method based on multi-sensor fusion of the present application embodiment is shown in

[0049] Figure 2 A cold rolling process parameter optimization flowchart of the cold rolling mill rolling force prediction and adaptive control method based on multi-sensor fusion of the present application embodiment is shown in DETAILED DESCRIPTION

[0050] To make the purpose, technical scheme and advantages of the present application embodiment clearer, the technical scheme of the present application embodiment will be described clearly and completely below in combination with the drawings of the present application embodiment. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] The technical scheme of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0052] Figure 1 A flowchart of the cold rolling mill rolling force prediction and adaptive control method based on multi-sensor fusion of the present application embodiment is shown in Figure 1 As shown, the method comprises:

[0053] Collecting process parameters collected by the cold rolling mill sensor, performing feature extraction on the process parameters to obtain a working condition feature sequence, constructing a dynamic correlation matrix among process parameters through a graph attention network and determining an influence weight, reconstructing the working condition feature sequence based on the influence weight, and calculating a rolling force prediction value in combination with the contact arc length change characteristics in the deformation zone during the rolling process;

[0054] a rolling force error value between the rolling force prediction value and an actual rolling force, constructing a control strategy based on the rolling force error value and calculating an action evaluation value, and obtaining a control compensation value based on optimization and noise suppression of rolling mill control parameters based on the action evaluation value;

[0055] constructing a parameter optimization space based on the control compensation value, combining plate shape quality constraints and equipment operation constraints in the parameter optimization space, and obtaining optimal process parameters by performing parameter optimization through a grid search algorithm, and converting the optimal process parameters into cold rolling mill control instructions and issuing them for execution.

[0056] In an alternative embodiment,

[0057] Collecting process parameters collected by cold rolling mill sensors, extracting features from the process parameters to obtain a working condition feature sequence, and constructing a dynamic correlation matrix between process parameters through a graph attention network and determining an influence weight include:

[0058] dividing the process parameters into multiple sets of parameters according to physical properties, combining multiple feature subsequences obtained by extracting features from each parameter set according to time sequences, and combining the feature subsequences to obtain the working condition feature sequence;

[0059] constructing a heterogeneous graph structure based on the working condition feature sequence and setting a meta-path in the heterogeneous graph according to parameter types, for each meta-path, calculating semantic similarity between different parameter nodes through a pre-set metric matrix, and calculating an attention score corresponding to the parameter nodes based on an attention mechanism;

[0060] based on the semantic similarity and the attention score, constructing the dynamic correlation matrix in combination with edges corresponding to different parameter types in the heterogeneous graph, calculating a degree centrality and a feature vector centrality corresponding to each parameter node based on the dynamic correlation matrix, and solving to obtain the influence weight.

[0061] The cold rolling mill process parameters can be divided into multiple sets of parameters according to physical properties. For example, the process parameters can be divided into a material parameter set, a device parameter set, and an environment parameter set. The material parameter set includes inlet thickness, outlet thickness, strip width, strip hardness, etc.; the device parameter set includes roll diameter, roll surface roughness, roll temperature, rolling speed, etc.; and the environment parameter set includes environmental temperature, humidity, cooling liquid temperature, etc. Each parameter set is extracted for features, such as the ratio of inlet thickness to outlet thickness, width-to-thickness ratio, etc. for the material parameter set; the ratio of roll diameter to strip thickness, rolling speed change rate, etc. for the device parameter set; and temperature gradient, humidity change rate, etc. for the environment parameter set.

[0062] After feature extraction, the extracted features are combined according to time series to obtain a plurality of feature subsequences. For continuous data with a collection period of 10 milliseconds, data of 30 consecutive sampling points are selected to form a time window, and material feature subsequences, equipment feature subsequences and environment feature subsequences are obtained. For example, if the thickness of the cold rolling mill inlet is 2.5 mm to 2.48 mm and the thickness of the outlet is 1.0 mm in a certain time window, the thickness ratio feature in the time window is 2.5 to 2.48; if the roll diameter is 600 mm, the ratio of the roll diameter to the outlet thickness is 600. These feature subsequences are combined to obtain a complete working condition feature sequence for representing the running state of the cold rolling mill.

[0063] Based on the working condition feature sequence, a heterogeneous graph structure is constructed. In the heterogeneous graph, different types of parameters are used as different types of nodes, and the edges between the nodes represent the correlation between the parameters. For example, there is an edge between the inlet thickness node and the roll diameter node, indicating that there is an influence relationship between the two parameters. According to the parameter type, a meta-path is set in the heterogeneous graph, which defines the connection mode between different types of nodes. For example, material-equipment-material, equipment-environment-equipment, etc. Meta-paths can be defined to capture the indirect association between different types of parameters.

[0064] For each meta-path, the semantic similarity between different parameter nodes is calculated through a pre-set metric matrix. For example, for the two parameters of inlet thickness and roll diameter, the corresponding value in the metric matrix may be 0.7, indicating a high semantic similarity. In actual cases, if the rolling speed increases from 500 m / min to 600 m / min, and the rolling force increases from 8000 kN to 8500 kN, the semantic similarity of the two parameters can be set to 0.8, indicating a strong correlation.

[0065] The attention score corresponding to the parameter node is calculated based on the attention mechanism. The attention mechanism learns the importance relationship between parameters and assigns different weights to different parameters. For example, the feature vector of the parameter node is used as input, and the attention score is calculated through a feedforward neural network. For example, for the strip width parameter, if its importance in predicting the rolling force is high, its attention score may be 0.9; and for the environmental humidity parameter, if its influence is small, the attention score may be 0.2.

[0066] Based on the semantic similarity and the attention score, a dynamic correlation matrix is constructed in combination with the edges corresponding to different parameter types in the heterogeneous graph. The dynamic correlation matrix is a two-dimensional matrix, and the elements in the matrix represent the correlation strength between different parameters. The semantic similarity and the attention score are combined by weighting. For example, if the semantic similarity of the entry thickness and the roll diameter is 0.7, the attention score of the entry thickness is 0.8, and the attention score of the roll diameter is 0.6, then the correlation strength in the dynamic correlation matrix can be calculated as 0.7 x (0.8 + 0.6) / 2 = 0.49.

[0067] The degree centrality and the eigenvector centrality corresponding to each parameter node are calculated based on the dynamic correlation matrix. The degree centrality reflects the number of direct connections of the parameter with other parameters, and the elements in the corresponding row or column of the dynamic correlation matrix are summed. The eigenvector centrality considers that the parameters connected to important parameters are also important, and the principal eigenvector of the dynamic correlation matrix is solved. For example, if the degree centrality of the entry thickness parameter is 3.2 and the eigenvector centrality is 0.85, it indicates that the parameter is in a relatively core position in the entire parameter network.

[0068] The degree centrality and the eigenvector centrality are combined by weighting to solve the influence weight of each parameter. The weight of the degree centrality is set to 0.4, and the weight of the eigenvector centrality is set to 0.6. The influence weight is obtained by weighted summation. For example, if the degree centrality of the entry thickness is 3.2 and the eigenvector centrality is 0.85, then the influence weight is 3.2 x 0.4 + 0.85 x 0.6 = 1.79.

[0069] The key parameters that have the greatest impact on the rolling force of the cold rolling mill are identified by the calculated influence weight, so that the parameters can be adjusted specifically to achieve accurate prediction and adaptive control of the rolling force. For example, if the influence weights of the strip entry thickness, the roll diameter, and the rolling speed are 1.79, 1.56, and 1.42 respectively, then the entry thickness should be adjusted first, followed by the roll diameter and the rolling speed in the control process.

[0070] In this embodiment, by dividing the process parameters into multiple parameter sets according to physical properties and performing feature extraction and time series combination on each parameter set, the correlation characteristics of different types of parameters in the time series evolution process can be effectively preserved, and information loss caused by single processing method can be avoided. The semantic similarity between parameter nodes is calculated by a measurement matrix, and the attention degree of different parameter nodes is allocated by combining the attention mechanism, so that the feature extraction process can highlight the parameter information that is more critical to the working condition change. The dynamic correlation matrix is constructed using the semantic similarity and the attention score, and the degree centrality and the eigenvector centrality of the parameter nodes are calculated on the dynamic correlation matrix, which can comprehensively reflect the degree of action of different parameters in the overall working condition.

[0071] In an alternative embodiment,

[0072] Reconstructing the working condition characteristic sequence based on the influence weight, and calculating the rolling force prediction value in combination with the contact arc length change characteristic of the deformation zone in the rolling process comprises:

[0073] Querying the material composition and physical characteristics corresponding to the current processing material in the pre-set material characteristic database and constructing a material characteristic vector corresponding to the current processing material, performing Hadamard product operation on the material characteristic vector and the working condition characteristic sequence to obtain a characteristic correlation sequence, and reconstructing the characteristic correlation sequence based on the influence weight to obtain a reconstructed characteristic sequence;

[0074] Collecting the work roll radius and deformation in the rolling process, processing through the pre-set multi-layer perception structure to obtain the contact arc length change characteristic of the deformation zone, and determining the time sequence characteristic corresponding to the contact arc length through the recurrent neural network;

[0075] Calculating the attention weight corresponding to the reconstructed characteristic sequence based on the time sequence characteristic, and calculating the rolling force prediction value corresponding to the reconstructed characteristic sequence in combination with the multi-layer convolution.

[0076] The composition and physical characteristic data of various steel, aluminum and other cold rolling materials are stored in the pre-set material characteristic database. The material characteristic database contains the chemical composition of the material, such as carbon content, manganese content, silicon content, etc., and the physical characteristics, such as yield strength, tensile strength, hardness value and other parameters. When the cold rolling mill starts processing a specific material, the relevant data of the current processing material is queried in the database according to the material number or the scanned barcode. For example, for a low carbon steel, the carbon content is 0.1%, the manganese content is 0.5%, the yield strength is 350 MPa, the tensile strength is 450 MPa, and the hardness value is 150 HV. These parameters are constructed into a material characteristic vector, which can be represented as a multi-dimensional numerical vector composed of material composition and physical characteristics.

[0077] Performing Hadamard product operation on the obtained material characteristic vector and the working condition characteristic sequence obtained in the foregoing embodiments, i.e., multiplying the elements at corresponding positions, multiplying each element in the material characteristic vector with the element at the corresponding position in the working condition characteristic sequence to obtain a characteristic correlation sequence considering the influence of material characteristics. Exemplarily, if the working condition characteristic sequence represents an inlet thickness of 2.5 mm in a certain dimension, and the material characteristic vector represents a yield strength of 350 MPa at the corresponding position, then the value at the corresponding position of the characteristic correlation sequence is 2.5 x 350 = 875.

[0078] The feature correlation sequence is reconstructed based on the calculated influence weight to obtain a reconstructed feature sequence. In the reconstruction process, each element in the feature correlation sequence is multiplied by the influence weight of the corresponding parameter, thereby strengthening the influence of important parameters and weakening the influence of secondary parameters. For example, if the influence weight of the inlet thickness is 1.79, the element value 875 in the feature correlation sequence related to the inlet thickness will be adjusted to 875*1.79=1566.25 to obtain the corresponding element value in the reconstructed feature sequence.

[0079] In the rolling process, the working roll radius and deformation data are collected in real time by multiple sensors installed on the cold rolling mill. The working roll radius is usually measured by a displacement sensor, and the deformation is calculated by the difference between the inlet thickness and the outlet thickness. For example, when the working roll radius is 300 mm, the inlet thickness is 2.5 mm, and the outlet thickness is 1.0 mm, the deformation is 1.5 mm. The pre-set multi-layer perception machine structure is used for processing, which includes three hidden layers, each layer has 128, 64, and 32 neurons respectively, and the activation function uses the ReLU function. The input of the multi-layer perception machine is the working roll radius and the deformation, and the output is the contact arc length change characteristic of the deformation zone.

[0080] The contact arc length of the deformation zone calculated by the multi-layer perception machine is 30 mm. As the rolling process proceeds, the contact arc length will change with the change of the working roll radius and the deformation, forming a time series data. This time series data is processed by a recurrent neural network to capture the time sequence features of the contact arc length change. The recurrent neural network uses a long short-term memory network structure, which includes an LSTM layer with 64 hidden units, which is used to learn the long-term dependence of the contact arc length change. For example, if the contact arc length changes from 30 mm to 28 mm and then to 32 mm within 10 seconds, the recurrent neural network can capture the change pattern and generate a time sequence feature vector describing the contact arc length change characteristic.

[0081] Based on the time sequence features obtained by the recurrent neural network, the attention weight corresponding to the reconstructed feature sequence is calculated. The attention mechanism calculates the relationship between the query vector, the key vector, and the value vector to determine the importance of each feature at different times. In this embodiment, the time sequence features are used as the query vector, and each time step of the reconstructed feature sequence is used as the key vector and the value vector. By calculating the similarity between the query vector and the key vector, the attention score of each time step is obtained, and the attention weight is obtained after normalization. For example, for a reconstructed feature sequence with a length of 30, the obtained attention weight distribution is: the weight of the first 10 time steps is 0.02, the weight of the middle 10 time steps is 0.04, and the weight of the last 10 time steps is 0.04, reflecting the importance of different time steps for rolling force prediction.

[0082] The calculated attention weight is combined with the reconstructed feature sequence, and is processed through a multi-layer convolutional neural network to obtain a rolling force prediction value. The multi-layer convolutional neural network includes three convolutional layers, the number of convolutional kernels of each layer is 32, 64 and 128 respectively, the size of the convolutional kernel is 3*3, the step is 1, and the padding is 1. After each layer of convolution, a ReLU activation function and a maximum pooling layer are connected. Finally, two fully connected layers are used, and the output node is 1, that is, the predicted rolling force value. For example, for a low-carbon steel, the inlet thickness is 2.5 mm, the outlet thickness is 1.0 mm, the rolling speed is 500 m / min, and the work roll radius is 300 mm, the predicted rolling force is 8500 kN, which is compared with the actual measured value of 8450 kN, the error is within 1%, which meets the precision requirements of industrial production.

[0083] In the embodiment, by querying the composition and physical properties of the processed material in the material property database, a material property vector is constructed, and a Hadamard product operation is performed with the working condition feature sequence, so that the working condition features and material properties are deeply fused, the influence of material differences on working condition changes is more truly reflected, the feature association sequence is reconstructed based on the influence weight, the key features are highlighted and redundant information is suppressed, the feature expression is more accurate, the work roll radius and deformation in the rolling process are collected, and the time sequence features of the deformation zone contact arc length are extracted by combining a multi-layer perception machine and a recurrent neural network, so that the dynamic influence of geometric changes on mechanical properties in the rolling process can be comprehensively reflected.

[0084] In an alternative embodiment,

[0085] Based on the rolling force error value, a control strategy is constructed and an action evaluation value is calculated, and based on the action evaluation value, the rolling mill control parameters are optimized and noise suppression is performed to obtain a control compensation value, which includes:

[0086] Based on the rolling force error value and the pre-acquired actual rolling force, a relative rolling force error value is calculated, time sequence data is constructed by combining the pre-acquired historical relative rolling force error value, and a rolling process causal diagram is constructed based on the time sequence data and the process parameters.

[0087] The intervention distribution between nodes in the rolling process causal diagram is calculated to obtain a causal effect, a key causal path is screened by combining a pre-set causal effect threshold, the control strategy is determined based on the key causal path, and the control strategy is divided into multiple levels, the hierarchical control amount corresponding to each level is calculated based on the causal relationship between different types of nodes in the rolling process causal diagram, and the action evaluation value is calculated based on the hierarchical control amount.

[0088] The rolling mill control parameters are divided into a fast response parameter group and a steady state compensation parameter group based on the action evaluation value, the fast response parameter group is optimized by adaptive weight adjustment based on the strength of causal effect to obtain a first optimization result, the steady state compensation parameter group is optimized by gradual optimization based on rolling force error accumulation to obtain a second optimization result, and the first optimization result and the second optimization result are combined and noise suppression is performed by a multi-layer perception machine to obtain a control compensation value.

[0089] In the rolling process, the rolling force error value is calculated by comparing the calculated rolling force prediction value with the actually measured rolling force value. For example, if the predicted rolling force is 8500 kN and the actually measured rolling force is 8450 kN, the rolling force error value is 50 kN. The relative rolling force error value is calculated based on the error value and the actual rolling force, i.e. the error value is divided by the actual rolling force value. In this embodiment, the relative error is 50 / 8450≈0.0059, about 0.59%. The calculated relative error is combined with the historically recorded relative error value to construct time series data. For example, the relative error data of the last 100 rolling processes can be saved to form a time series sequence with a length of 100.

[0090] The rolling process causal diagram is constructed in combination with the time series data and the process parameters. The process causal diagram is a directed acyclic graph structure, in which the nodes represent the variables in the rolling process and the edges represent the causal relationship between the variables. In the causal diagram, the process parameters such as the inlet thickness, the outlet thickness, the rolling speed, the work roll radius, etc. are used as parent nodes, and the rolling force is used as a child node. The connection relationship between the nodes is determined by a structure learning algorithm. For example, by analyzing the historical data, it is found that there is a strong causal relationship between the inlet thickness and the rolling force, with a correlation coefficient of 0.85; there is a moderate causal relationship between the rolling speed and the rolling force, with a correlation coefficient of 0.62; and the causal relationship between the environmental temperature and the rolling force is weak, with a correlation coefficient of only 0.15.

[0091] The causal effect is calculated by calculating the intervention distribution between the nodes in the rolling process causal diagram. The intervention distribution is calculated by do-calculus, i.e. the effect of intervention on other nodes is observed by assuming intervention on a certain node. For example, the intervention effect is estimated by Monte Carlo simulation or structural equation modeling. For example, it is found by analysis that when the inlet thickness is intervened and increased from 2.5 mm to 2.6 mm, the rolling force is increased by an average of 150 kN; when the rolling speed is intervened and increased from 500 m / min to 550 m / min, the rolling force is increased by an average of 70 kN.

[0092] The key causal paths are screened in combination with a preset causal effect threshold value. The causal effect threshold value is determined according to engineering experience and historical data statistical analysis. The path with a causal effect intensity greater than 100 kN is defined as a key causal path. Exemplarily, the path from the entry thickness to the rolling force is identified as a key causal path, and the path from the rolling speed to the rolling force is important but does not reach the key path standard. The control strategy is determined based on the key causal path and is divided into multiple levels. For example, the control strategy can be divided into three levels: a fast response level, a medium-term adjustment level and a long-term optimization level. The fast response level mainly adjusts the parameters on the key causal path in real time; the medium-term adjustment level adjusts the parameters on the secondary causal path periodically; and the long-term optimization level focuses on the overall optimization of the process parameters.

[0093] The hierarchical control quantity corresponding to each level is calculated based on the causal relationship between different types of nodes in the rolling process causal diagram. For the fast response level, the adjustment quantity of the entry thickness is calculated according to the causal effect intensity of the entry thickness and the rolling force. Exemplarily, if the rolling force error is 50 kN and the causal effect intensity of the entry thickness is 150 kN / 0.1 mm, the adjustment quantity of the entry thickness is -0.033 mm. For the medium-term adjustment level, the adjustment quantity of the rolling speed is calculated. If the causal effect intensity of the rolling speed is 70 kN / 50 m / min, the adjustment quantity of the rolling speed is -35.7 m / min. For the long-term optimization level, measures such as work roll replacement or wear compensation are involved.

[0094] The action evaluation value is calculated based on the hierarchical control quantity. The action evaluation value reflects the effectiveness and stability of the control action. The evaluation value is calculated by weighted combination of multiple indexes, including control response time, stability index and precision index, etc. For example, for the entry thickness adjustment, if the response time is 100 ms, the stability index is 0.95, and the precision index is 0.98, the corresponding action evaluation value can be calculated as 0.97. For the rolling speed adjustment, if the response time is 500 ms, the stability index is 0.90, and the precision index is 0.92, the corresponding action evaluation value is 0.88.

[0095] The rolling mill control parameters are divided into a fast response parameter group and a steady state compensation parameter group based on the action evaluation value. Parameters with an action evaluation value higher than 0.9 are divided into the fast response parameter group, including the entry thickness, roll gap, etc. Parameters with an action evaluation value between 0.7 and 0.9 are divided into the steady state compensation parameter group, including the rolling speed, cooling liquid flow, etc. The fast response parameter group is optimized by using adaptive weight adjustment based on the causal effect intensity. The control weight is dynamically adjusted according to the causal effect intensity of the parameters. The greater the causal effect intensity, the higher the control weight. For example, if the causal effect intensity of the entry thickness is 150 kN / 0.1 mm and the initial control weight is 1.0, the adjusted control weight can be 1.3, thereby enhancing the control effect of the entry thickness. After optimization, the final adjustment amount of the entry thickness is -0.043 mm, obtaining a first optimization result.

[0096] The steady state compensation parameter group is gradually optimized by using the cumulative amount of rolling force error. The control parameters are gradually adjusted by accumulating the rolling force error in a period of time, which is suitable for parameters with slow response but high stability requirements. For example, if the cumulative amount of rolling force error in 10 seconds is 500 kN·s and the adjustment coefficient of the rolling speed is 0.05 m / min / (kN·s), the adjustment amount of the rolling speed is -25 m / min, obtaining a second optimization result.

[0097] The first optimization result and the second optimization result are combined, and the control compensation value is obtained by noise suppression through a multi-layer perception machine. The multi-layer perception machine includes three hidden layers, each layer having 64, 32 and 16 neurons respectively, and the activation function is LeakyReLU. The input of the network is the combined optimization result, and the output is the control compensation value after noise suppression.

[0098] In this embodiment, the relative error between the rolling force error value and the actual rolling force is calculated, and the time series data is constructed by combining the historical error, which can comprehensively reflect the dynamic change trend of the error in the rolling process. The causal effect between nodes is identified by the causal reasoning method, and the key causal path is obtained by combining the threshold screening, which can effectively eliminate redundant factors, highlight the causal relationship with significant influence on the rolling process, improve the pertinence and reliability of the control decision, and realize the decoupling control of the complex process parameters by dividing the control strategy into different levels and calculating the hierarchical control amount based on the causal relationship, thereby improving the fineness and interpretability of the adjustment.

[0099] In an alternative embodiment,

[0100] dividing the rolling mill control parameters into a fast response parameter group and a steady state compensation parameter group based on the action evaluation value, optimizing the fast response parameter group using adaptive weight adjustment based on the strength of causal effect to obtain a first optimization result, and optimizing the steady state compensation parameter group using gradual optimization based on the rolling force error accumulation to obtain a second optimization result, comprising:

[0101] performing responsiveness analysis on the rolling mill control parameters based on the action evaluation value, determining a response time, and dividing the rolling mill control parameters with a response time not less than a preset response speed threshold into a steady state compensation parameter group, and dividing the rolling mill control parameter group with a response time less than the response speed threshold into a fast response parameter group.

[0102] For the rolling mill control parameters in the fast response parameter group, calculating the sensitivity of each group of control parameters on the key causal path to obtain the strength of causal effect, calculating a weight increment based on the strength of causal effect and the rolling force error value, and performing adaptive weight update on the fast response parameter group based on the weight increment to obtain a first optimization result.

[0103] integrating the rolling force error value within a preset time period to determine a rolling force error accumulation, calculating a compensation amount in combination with a pre-set gradual optimization coefficient, and performing gradual optimization on the steady state compensation parameter group based on the compensation amount to obtain a second optimization result.

[0104] Performing responsiveness analysis on the rolling mill control parameters based on the calculated action evaluation value, determining the response time of each control parameter, which is the time required for the parameter to reach the set value from issuing the control instruction, obtained through experimental measurement or historical data analysis. For example, the response time of roll gap adjustment is 80 milliseconds, the response time of entry tension adjustment is 150 milliseconds, the response time of rolling speed adjustment is 500 milliseconds, and the response time of cooling liquid flow adjustment is 2000 milliseconds. The preset response speed threshold is 300 milliseconds, which is determined according to the actual situation and control requirements of the cold rolling production line. The rolling mill control parameters with a response time not less than 300 milliseconds are divided into a steady state compensation parameter group, including rolling speed, cooling liquid flow, etc.; the rolling mill control parameters with a response time less than 300 milliseconds are divided into a fast response parameter group, including roll gap, entry tension, etc.

[0105] For the rolling mill control parameters in the fast response parameter group, the sensitivity of each group of control parameters on the key causal path is calculated to obtain the causal effect strength. The sensitivity is measured by perturbation analysis method, a small perturbation is applied on the control parameter, and the change amount of rolling force is observed. The sensitivity is determined by calculating the ratio. For example, if the roll gap is increased by 0.01 mm, and the rolling force is observed to decrease by 120 kN, then the sensitivity of the roll gap to the rolling force is 12000 kN / mm, indicating the causal effect strength of the roll gap; if the entry tension is increased by 10 kN, and the rolling force is observed to increase by 30 kN, then the sensitivity of the entry tension to the rolling force is 3, indicating the causal effect strength of the entry tension.

[0106] The weight increment is calculated based on the causal effect strength and the rolling force error value. The calculation of the weight increment considers the size of the causal effect strength and the rolling force error. The product of the causal effect strength and the rolling force error is multiplied by a proportional coefficient, which is set according to the stability requirements of the control system, and is usually a positive number less than 1, for example 0.005. Assuming that the current rolling force error value is 50 kN, the causal effect strength of the roll gap is 12000 kN / mm, and the proportional coefficient is 0.005, then the weight increment of the roll gap parameter is 50x12000x0.005=3000; the causal effect strength of the entry tension is 3, then the weight increment is 50x3x0.005=0.75.

[0107] The adaptive weight update is performed on the fast response parameter group based on the weight increment to obtain the first optimization result. The adaptive weight update refers to adjusting the control weight of the control parameter according to the weight increment. The control weight determines the amplitude of the parameter adjustment. Assuming that the initial control weight of the roll gap is 1.0, and the updated control weight is 1.0+3000x0.0001=1.3; the initial control weight of the entry tension is 1.0, and the updated control weight is 1.0+0.75x0.0001=1.000075. The updated control weight is used to calculate the adjustment amount of the control parameter. For example, if the rolling force error is 50 kN, the sensitivity of the roll gap is 12000 kN / mm, and the updated control weight is 1.3, then the adjustment amount of the roll gap is 50÷12000x1.3=0.0054 mm; the sensitivity of the entry tension is 3, and the updated control weight is 1.000075, then the adjustment amount of the entry tension is 50÷3x1.000075=16.668 kN. The combined adjustment amount obtains the first optimization result.

[0108] The rolling force error values are integrated in a preset time period to determine a rolling force error accumulation. The integration is to accumulate the rolling force errors in a time period to obtain the cumulative effect of the errors. The preset time period is usually several seconds to tens of seconds, which is determined according to the process characteristics of the production line. For example, the time period is set to 10 seconds, the rolling force error is sampled every 100 milliseconds in 10 seconds, a total of 100 points are sampled, and if the average of these error values is 30 kN, the rolling force error accumulation is 30x10=300 kN·s.

[0109] The compensation is calculated in combination with a preset progressive optimization coefficient. The progressive optimization coefficient is determined according to the control characteristics of each parameter in the steady-state compensation parameter group, and reflects the influence degree of parameter adjustment on the rolling force. For example, the progressive optimization coefficient of the rolling speed is 0.02 mm / min / (kN·s), and the progressive optimization coefficient of the cooling liquid flow is 0.005 L / min / (kN·s). The compensation is calculated based on the progressive optimization coefficient and the rolling force error accumulation, and the compensation of the rolling speed is 300x0.02=6 mm / min, and the compensation of the cooling liquid flow is 300x0.005=1.5 L / min.

[0110] The steady-state compensation parameter group is progressively optimized based on the compensation to obtain a second optimization result. The progressive optimization is to adjust the parameters step by step to avoid system instability caused by sudden changes, and the calculated compensation is added or subtracted from the current parameter value to obtain a new parameter setting value. For example, if the current rolling speed is 500 m / min and the calculated compensation is 6 m / min, the rolling speed needs to be reduced considering that the rolling force error is positive, and the new rolling speed setting value is 500-6=494 m / min. If the current cooling liquid flow is 100 L / min and the compensation is 1.5 L / min, it also needs to be reduced, and the new cooling liquid flow setting value is 100-1.5=98.5 L / min, and the second optimization result is determined based on the new parameter setting value.

[0111] In actual application, the first optimization result and the second optimization result are combined and further processed, for example, the control actions of increasing the roll gap by 0.0054 mm, increasing the entry tension by 16.668 kN, reducing the rolling speed to 494 m / min, and reducing the cooling liquid flow to 98.5 L / min are simultaneously performed.

[0112] In the embodiment, by performing responsiveness analysis on the rolling mill control parameters, and dividing the fast response parameter group and the steady-state compensation parameter group according to the response time, the different control parameter action characteristics can be classified and managed, and the control imbalance problem caused by unified adjustment can be avoided. For the fast response parameter group, the sensitivity on the key causal path is calculated, and the adaptive weight update is performed combined with the rolling force error value, so that the control strategy can quickly adjust to the working condition change, effectively improving the instantaneous response ability and dynamic adjustment precision of the system. For the steady-state compensation parameter group, the error accumulation is obtained by integrating the rolling force error value in the preset time period, and the compensation amount is calculated combined with the gradual optimization coefficient, so that the long-term deviation is gradually corrected, and the continuous accumulation of residual error in the system can be effectively suppressed.

[0113] In an alternative embodiment,

[0114] Based on the control compensation value, a parameter optimization space is constructed, in which, combined with the plate shape quality constraint and the equipment operation constraint, the optimal process parameters are obtained by parameter optimization through a grid search algorithm, including:

[0115] The temperature distribution and the thermal stress distribution in the rolling area are collected, and the parameter optimization space is constructed combined with the pre-acquired material characteristics and the control compensation value;

[0116] In the parameter optimization space, the plate shape deviation value and the temperature field deviation value are calculated combined with the plate shape quality constraint and the equipment operation constraint, and the parameter feasible region is limited to obtain a constrained parameter optimization space;

[0117] Based on the grid search algorithm, the grid in the constrained parameter optimization space is divided, the corresponding comprehensive evaluation index of each grid is calculated, the temperature gradient in the constrained parameter optimization space is calculated based on the comprehensive evaluation index, the grid with a temperature gradient greater than a preset gradient threshold is refined, and the candidate process parameters are obtained by parameter optimization in the refined grid;

[0118] The real-time temperature field is collected, the current thermal deformation amount is calculated combined with the thermal stress distribution, and the thermal deformation compensation amount is determined based on the current thermal deformation amount. The thermal deformation compensation amount and the candidate process parameters are superimposed to obtain the optimal process parameters.

[0119] In the cold rolling process, the temperature distribution data of the rolling area is collected by an infrared temperature sensor array installed in the rolling area. The temperature sensor array is uniformly distributed along the rolling direction and the width direction, and the measurement accuracy of each sensor is ±0.5℃, and the sampling frequency is 10Hz. For example, for a four-high cold rolling mill, 24 temperature sensors are installed on the surfaces of the upper and lower work rolls and the backup rolls, and the collected temperature distribution data shows that the surface temperature of the work roll is distributed between 50℃ and 75℃, and increases along the rolling direction, and the temperature at the center of the rolling area is 3℃ to 5℃ higher than that at the edge. The thermal stress distribution data of the rolling area is collected by a strain gauge sensor. The strain gauge sensor is installed at key positions on the surface of the roll, and can measure the thermal stress caused by uneven temperature. For example, the collected thermal stress data shows that the thermal stress on the surface of the roll is distributed between 5MPa and 15MPa, and the thermal stress value at the center of the rolling area is about 30% higher than that at the edge.

[0120] Based on the pre-obtained material properties and the calculated control compensation values, a parameter optimization space is constructed. The material properties include physical parameters such as the elastic modulus, Poisson's ratio, and thermal expansion coefficient of the material. For example, for a common cold-rolled steel plate, the elastic modulus is 210GPa, the Poisson's ratio is 0.3, and the thermal expansion coefficient is 12×10^-6 / ℃. The control compensation values include the roll gap compensation value 0.0054mm, the entry tension compensation value 16.668kN, and the rolling speed compensation value -6m / min. The parameter optimization space is a multi-dimensional space, the dimension of which is determined by the number of process parameters to be optimized, and each dimension corresponds to the value range of a process parameter. In this embodiment, the parameter optimization space includes six parameters: roll gap, rolling speed, entry tension, exit tension, roll bending force, and rolling oil temperature. The value range of each parameter is determined according to the equipment specifications and production requirements.

[0121] In the parameter optimization space, the plate shape deviation value and the temperature field deviation value are calculated in combination with the plate shape quality constraint and the equipment operation constraint, and the parameter feasible region is limited to obtain the constraint parameter optimization space. The plate shape quality constraint refers to the requirements that must be met by the flatness, thickness accuracy and other quality indicators of the finished plate, such as the thickness deviation being not more than ±0.005 millimeters and the flatness I value being not more than 5 units. The equipment operation constraint refers to the parameter range in which the rolling mill equipment can safely and stably operate, such as the rolling force being not more than 12000 kilonewtons, the roll bending force being not more than 1000 kilonewtons, and the rolling speed being not more than 1200 meters / minute. The plate shape deviation value is calculated by comparing the predicted plate shape with the target plate shape, and the temperature field deviation value is calculated by comparing the current temperature field with the ideal temperature field. For example, if the current I value is 7 units and the target I value is 3 units, the plate shape deviation value is 4 units; and if the current roll center temperature is 75℃ and the ideal temperature is 70℃, the temperature field deviation value is 5℃. The parameter feasible region is limited based on the constraints and the deviation values to obtain the constraint parameter optimization space that meets all the constraint conditions.

[0122] Based on the grid search algorithm, the grid is divided in the constraint parameter optimization space, and the performance index corresponding to each grid point is evaluated to find the optimal parameter combination. In this embodiment, the 6-dimensional parameter space is divided into a grid, and each dimension is divided into 10 equal points, obtaining a total of 10^6 grid points. For each grid point, the corresponding comprehensive evaluation index is calculated. The comprehensive evaluation index is the weighted sum of the plate shape quality index, the rolling force stability index and the equipment operation efficiency index. The plate shape quality index includes the thickness accuracy and the flatness, the rolling force stability index reflects the fluctuation degree of the rolling force, and the equipment operation efficiency index reflects the production efficiency and the energy consumption level. For example, for the parameter combination (roll gap = 0.8 millimeters, rolling speed = 600 meters / minute, inlet tension = 800 kilonewtons, outlet tension = 500 kilonewtons, roll bending force = 800 kilonewtons, and rolling oil temperature = 55℃), the calculated plate shape quality index is 0.85, the rolling force stability index is 0.92, the equipment operation efficiency index is 0.78, and the comprehensive evaluation index is 0.85×0.5+0.92×0.3+0.78×0.2=0.853.

[0123] The temperature gradient in the constraint parameter optimization space is calculated based on the comprehensive evaluation index. The temperature gradient is obtained by calculating the rate of change of the comprehensive evaluation index between adjacent grid points. The grid with a temperature gradient greater than a preset gradient threshold is refined. The preset gradient threshold is determined according to the control accuracy requirement, for example, set to 0.02 / °C. If the temperature gradient of a certain grid is 0.03 / °C, which exceeds the threshold, the grid is subdivided into smaller sub-grids, and each dimension is further divided into 5 equal points, obtaining 5^6 sub-grid points. In the refined grid, parameter optimization is continued, and the comprehensive evaluation index of each sub-grid point is calculated. The parameter combination with the highest comprehensive evaluation index is selected as the candidate process parameter. After refinement search, the candidate process parameter found is: roll gap = 0.82 mm, rolling speed = 580 m / min, inlet tension = 820 kN, outlet tension = 510 kN, roll bending force = 830 kN, and rolling oil temperature = 57°C, and the corresponding comprehensive evaluation index is 0.875.

[0124] Real-time temperature field data is collected to obtain the latest temperature distribution during rolling. The real-time temperature field is obtained by the aforementioned infrared temperature sensor array, and the sampling frequency is increased to 20 Hz to capture rapid temperature changes. The current thermal deformation is calculated in combination with the thermal stress distribution data. The thermal deformation is the change in roll shape caused by the temperature gradient, which is calculated by thermal elastic analysis method. The calculation process considers parameters such as the thermal expansion coefficient, elastic modulus, and Poisson's ratio of the roll material. For example, if the center temperature of the roll is 5°C higher than the edge, the roll diameter is 600 mm, and the thermal expansion coefficient is 12×10^-6 / °C, the calculated thermal deformation of the roll is 36 microns, i.e. the center region of the roll expands 36 microns more than the edge region.

[0125] The thermal deformation compensation amount is determined based on the current thermal deformation. The thermal deformation compensation amount is the adjustment amount required for the process parameters to offset the influence of thermal deformation. In this embodiment, the roll bending force and the roll gap are adjusted to compensate for thermal deformation. According to the size and distribution characteristics of the thermal deformation, the corresponding compensation amount is calculated. For example, for a center expansion of 36 microns, the roll bending force needs to be increased by 20 kN and the roll gap needs to be reduced by 0.02 mm. The thermal deformation compensation amount is superimposed on the candidate process parameters to obtain the optimal process parameters: roll gap = 0.80 mm, rolling speed = 580 m / min, inlet tension = 820 kN, outlet tension = 510 kN, roll bending force = 850 kN, and rolling oil temperature = 57°C.

[0126] In the embodiment, by collecting the temperature distribution and thermal stress distribution of the rolling area, combining the material characteristics and control compensation value to construct the parameter optimization space, the material difference, thermodynamic characteristics and control factors can be considered in the process optimization, the comprehensiveness and pertinence of the optimization result are ensured, the plate shape quality constraint and equipment operation constraint are introduced to limit the parameter feasible region, the emergence of unfeasible or unstable working conditions is effectively avoided, the practicality and controllability of the optimization result are improved, by using the grid search algorithm in the constraint parameter optimization space, combining the temperature gradient calculated by the comprehensive evaluation index, the area with large temperature gradient is refined, the resolution of the key parameter area can be improved, the optimization process can more accurately capture the sensitive point affecting the rolling quality, and the optimization precision and efficiency are improved.

[0127] Figure 2 The cold rolling process parameter optimization flowchart of the cold rolling mill rolling force prediction and adaptive control method based on multi-sensor fusion of the embodiment of the application.

[0128] In an optional implementation,

[0129] Based on the grid search algorithm, the grid is divided in the constraint parameter optimization space, and the comprehensive evaluation index corresponding to each grid is calculated, including:

[0130] In the constraint parameter optimization space, the initial grid is divided based on the grid search algorithm, and the initial position distribution of the grid node is determined, the topological relationship matrix between the initial grids is determined based on the initial position distribution and production requirements;

[0131] Based on the topological relationship matrix, the geometric characteristic index of each initial grid is determined, and the grid quality evaluation value is calculated, the grid node position is locally adjusted based on the grid quality evaluation value, the corrected grid is obtained, and the grid distortion degree and grid orthogonality corresponding to the corrected grid are calculated;

[0132] The plate shape quality value and temperature uniformity value corresponding to each corrected grid are collected, and the comprehensive evaluation index is calculated based on the grid distortion degree and the grid orthogonality.

[0133] In the constraint parameter optimization space, the initial grid is divided and the initial position distribution of the grid nodes is determined based on the grid search algorithm. The constraint parameter optimization space is a high-dimensional space containing multiple parameters such as roll gap, rolling speed, inlet tension, outlet tension, roll bending force, rolling oil temperature, etc. The initial grid division adopts a structured grid method, that is, each parameter dimension is equally divided into several segments to form a regular grid structure. For a 6-dimensional parameter space, each dimension is initially divided into 8 equal points, resulting in 8^6 = 262144 grid nodes. The initial position distribution of the grid nodes follows the uniform distribution principle, that is, the spacing between adjacent nodes on each dimension is equal. For example, the value range of the roll gap is 0.5 mm to 1.3 mm, and the node positions of the initial grid on this dimension are 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3 mm; the value range of the rolling speed is 300 to 1000 m / min, and the node positions of the initial grid on this dimension are 300, 400, 500, 600, 700, 800, 900, 1000 m / min.

[0134] Based on the initial position distribution and production requirements, the topological relationship matrix between the initial grid is determined. The topological relationship matrix describes the connection relationship between the grid nodes and is a two-dimensional matrix. The elements in the matrix represent whether two nodes are adjacent. In the structured grid, each internal node is adjacent to 2^d nodes, where d is the dimension of the parameter space, and in this embodiment, d = 6, that is, each internal node is adjacent to at most 64 nodes. The influence of production requirements on the topological relationship is reflected in that some parameter combinations do not meet the production constraints and need to be marked as infeasible in the topological relationship matrix. For example, the combination of high rolling speed and high rolling force may exceed the carrying capacity of the equipment, at which time the corresponding connection relationship in the topological relationship matrix is marked as infeasible. For example, if the connection relationship between node (roll gap = 0.5 mm, rolling speed = 1000 m / min) and node (roll gap = 0.5 mm, rolling speed = 900 m / min) may cause the rolling force to exceed the equipment limit, the corresponding element value in the topological relationship matrix is set to 0, indicating that the two nodes are not directly connected.

[0135] The geometric characteristic indexes of each initial mesh are determined based on the topological relationship matrix, and a mesh quality evaluation value is calculated. The geometric characteristic indexes include the aspect ratio, slope, volume, and other parameters of the mesh, which are used to evaluate the shape quality of the mesh. In a multi-dimensional parameter space, the geometric characteristics of the mesh can be obtained by calculating the projection characteristics of the mesh in each dimension. For example, for a mesh (roll gap = 0.5-0.6 mm, rolling speed = 300-400 m / min, entry tension = 600-700 kN), the aspect ratio can be calculated by the ratio of the range of each dimension, i.e. (0.6-0.5) / (400-300) / (700-600) = 0.1 / 100 / 100 = 1x10-5. The mesh quality evaluation value is a comprehensive evaluation based on the geometric characteristic indexes, calculated by the weighted average method. The weights are determined according to the influence of each index on the optimization result. For example, if the weight of the aspect ratio is 0.5, the weight of the slope is 0.3, and the weight of the volume is 0.2, and the aspect ratio score is 0.8, the slope score is 0.9, and the volume score is 0.7, then the quality evaluation value is 0.8x0.5+0.9x0.3+0.7x0.2 = 0.81.

[0136] Based on the mesh quality evaluation value, the node positions of the mesh are locally adjusted to obtain a corrected mesh. The goal of local adjustment is to improve the mesh quality, especially to increase the mesh density in the parameter sensitive area. The adjustment method uses the spring analogy method, which regards the connection between mesh nodes as a spring, and adjusts the spring stiffness according to the mesh quality evaluation value, so as to realize the optimization of node position. In the mesh area with low quality evaluation value, the spring stiffness is set smaller, so that the node is easy to move; in the mesh area with high quality evaluation value, the spring stiffness is set larger, so that the node position is relatively stable. For example, for a mesh area with a quality evaluation value of 0.81, the spring stiffness between adjacent nodes can be set to 0.9; for a mesh area with a quality evaluation value of 0.6, the spring stiffness can be set to 0.7. Through iterative calculation, the positions of each node are updated until the preset convergence condition or the maximum number of iterations is reached. After local adjustment, the mesh density in the parameter sensitive area increases, and the mesh density in the non-sensitive area decreases, forming a more reasonable mesh distribution. For example, in the area where the roll gap is close to 0.8 mm and the rolling speed is close to 600 m / min, the mesh density may increase by 100%, while in the parameter boundary area, the mesh density may decrease by 30%.

[0137] The grid distortion degree and grid orthogonality corresponding to the corrected grid are calculated. The grid distortion degree measures the degree of deviation of the grid shape from the ideal shape, which is usually calculated by the ratio of the maximum internal angle to the minimum internal angle of the grid. In high-dimensional space, the maximum value of the distortion degree on each dimension projection can be taken as the overall distortion degree. The grid orthogonality measures the degree of perpendicularity of the grid edge to the coordinate axis, which is determined by calculating the minimum value of the cosine of the angle between the grid edge and the coordinate axis. For the corrected grid, if the distortion degrees on each dimension projection are 1.1, 1.2, 1.05, 1.15, 1.08, and 1.1, respectively, the overall distortion degree takes the maximum value of 1.2; if the orthogonality on each dimension projection is 0.98, 0.95, 0.97, 0.96, 0.99, and 0.97, respectively, the orthogonality takes the minimum value of 0.95. The closer the grid distortion degree is to 1, the closer the grid orthogonality is to 1, indicating that the grid quality is higher.

[0138] The plate shape quality value and temperature uniformity value corresponding to each corrected grid are collected. The plate shape quality value is measured by a shape meter installed at the outlet of the cold rolling mill, and the temperature uniformity value is measured by an infrared thermal imager. For each parameter combination corresponding to the corrected grid, actual rolling test or simulation calculation based on historical data and physical model is performed to obtain the corresponding plate shape quality value and temperature uniformity value. For example, for the parameter combination (roll gap = 0.82 mm, rolling speed = 580 m / min, inlet tension = 820 kN, outlet tension = 510 kN, roll bending force = 830 kN, rolling oil temperature = 57°C), the measured plate shape quality value is 0.92 (full score is 1), and the temperature uniformity value is 0.88 (full score is 1).

[0139] The comprehensive evaluation index is calculated by combining the grid distortion degree, grid orthogonality, plate shape quality value, and temperature uniformity value. The comprehensive evaluation index is the weighted average of the four parameters, and the weights are determined according to the actual production requirements. For example, the weight of the plate shape quality value can be set to 0.4, the weight of the temperature uniformity value can be set to 0.3, the weight of the grid distortion degree can be set to 0.15, and the weight of the grid orthogonality can be set to 0.15. If the grid distortion degree is 1.2 and the grid orthogonality is 0.95, the comprehensive evaluation index is calculated as 0.92 x 0.4 + 0.88 x 0.3 + (2 - 1.2) x 0.15 + 0.95 x 0.15 = 0.896. Among them, the grid distortion degree needs to be transformed to make its value range consistent with other indicators, that is, subtract 2 from the distortion degree value, so that the smaller the distortion degree, the higher the evaluation score.

[0140] In this embodiment, by using a grid search algorithm for initial grid division in the constraint parameter optimization space, and combining the production demand to construct the topological relationship matrix between the grids, the parameter space can be reasonably arranged at the beginning of optimization, the balance and structure of the search range are ensured, the geometric characteristic indexes of each grid are calculated based on the topological relationship matrix, and the grid quality evaluation value is obtained, the grid node position is adjusted locally, the uniformity and rationality of grid division can be effectively improved, the optimization error caused by uneven grid distribution is reduced, the stability and numerical calculation accuracy of the grid structure can be quantitatively evaluated by calculating the grid distortion degree and grid orthogonality, the calculation deviation caused by grid distortion or non-orthogonality in the optimization process is avoided, the evaluation index is obtained by combining the corrected grid corresponding to the shape quality value and the temperature uniformity value, and a unified measurement standard between the grid quality and the process target is established, so that the optimization result has numerical calculation stability and can reflect the actual production demand.

[0141] In a second aspect of the embodiment of the present application, a cold rolling mill rolling force prediction and adaptive control system based on multi-sensor fusion is provided, comprising:

[0142] The first unit is used for collecting process parameters collected by the cold rolling mill sensor, extracting features of the process parameters to obtain working condition feature sequences, constructing a dynamic correlation matrix between the process parameters through a graph attention network and determining an influence weight, reconstructing the working condition feature sequences based on the influence weight, and calculating a rolling force prediction value in combination with the contact arc length change characteristics of the deformation zone in the rolling process;

[0143] The second unit is used for calculating a rolling force error value between the rolling force prediction value and the actual rolling force, constructing a control strategy based on the rolling force error value and calculating an action evaluation value, optimizing and noise suppressing the rolling mill control parameters based on the action evaluation value to obtain a control compensation value;

[0144] The third unit is used for constructing a parameter optimization space based on the control compensation value, in which, in combination with the shape quality constraint and the equipment operation constraint, a parameter optimization is performed through a grid search algorithm to obtain optimal process parameters, the optimal process parameters are converted into cold rolling mill control instructions and are executed.

[0145] In a third aspect of the embodiment of the present application, an electronic device is provided, comprising:

[0146] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0147] In a fourth aspect, the present application provides a computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method described above.

[0148] The present application can be a method, an apparatus, a system, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for performing various aspects of the present application.

[0149] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting and adaptively controlling rolling force in a cold rolling mill based on multi-sensor fusion, characterized in that, include: Process parameters collected by sensors in a cold rolling mill are acquired, and feature extraction is performed on the process parameters to obtain a working condition feature sequence. A dynamic correlation matrix between process parameters is constructed using a graph attention network, and the influence weights are determined. The working condition feature sequence is reconstructed based on the influence weights, and the predicted rolling force value is calculated by combining the characteristics of the change in contact arc length in the deformation zone during rolling. Based on the influence weights, the working condition feature sequence is reconstructed, and the predicted rolling force value is calculated by combining the characteristics of the contact arc length change in the deformation zone during rolling, including: The material composition and physical properties of the current processed material are queried in a pre-set material property database, and a material property vector corresponding to the current processed material is constructed. The material property vector is then subjected to a Hadamard product operation with the working condition feature sequence to obtain a feature association sequence. Based on the influence weight, the feature association sequence is reconstructed to obtain a reconstructed feature sequence. The working roller radius and deformation during the rolling process are collected, processed by a pre-set multilayer perceptron structure, and the contact arc length variation characteristics of the deformation zone are obtained. The temporal characteristics corresponding to the contact arc length are then determined by a recurrent neural network. Based on the temporal features, the attention weights corresponding to the reconstructed feature sequence are calculated, and the rolling force prediction value corresponding to the reconstructed feature sequence is obtained by combining multi-layer convolution. Calculate the rolling force error value between the predicted rolling force value and the actual rolling force, construct a control strategy based on the rolling force error value and calculate the action evaluation value, and optimize the rolling mill control parameters and suppress noise based on the action evaluation value to obtain the control compensation value; Based on the control compensation value, a parameter optimization space is constructed. In the parameter optimization space, combined with the plate shape quality constraint and equipment operation constraint, the optimal process parameters are obtained by the grid search algorithm. The optimal process parameters are then converted into cold rolling mill control commands and issued for execution.

2. The method according to claim 1, characterized in that, Process parameters collected by sensors in the cold rolling mill are acquired, and feature extraction is performed on these parameters to obtain a sequence of operating conditions. A dynamic correlation matrix among the process parameters is constructed using a graph attention network, and the influencing weights are determined, including: The process parameters are divided into multiple parameter sets according to their physical properties. After feature extraction of each parameter set, multiple feature subsequences are obtained by combining them according to the time series. The feature subsequences are then combined to obtain the working condition feature sequence. Based on the operating condition feature sequence, a heterogeneous graph structure is constructed and meta-paths are set in the heterogeneous graph according to the parameter type. For each meta-path, the semantic similarity between different parameter nodes is calculated through a pre-set metric matrix, and the attention score corresponding to the parameter node is calculated based on the attention mechanism. Based on the semantic similarity and the attention score, the dynamic association matrix is ​​constructed by combining the edges corresponding to different parameter types in the heterogeneous graph. Based on the dynamic association matrix, the degree centrality and feature vector centrality of each parameter node are calculated, and the influence weight is obtained by solving.

3. The method according to claim 1, characterized in that, A control strategy is constructed based on the rolling force error value, and an action evaluation value is calculated. Based on the action evaluation value, the rolling mill control parameters are optimized and noise is suppressed to obtain control compensation values, including: The relative rolling force error value is calculated based on the rolling force error value and the pre-acquired actual rolling force. Time series data is constructed by combining the pre-acquired historical relative rolling force error values. A cause-effect graph of the rolling process is constructed based on the time series data and the process parameters. The causal effect is obtained by calculating the intervention distribution between nodes in the causal graph of the rolling process. The key causal path is obtained by filtering the key causal path in combination with the pre-set causal effect threshold. The control strategy is determined based on the key causal path and the control strategy is divided into multiple levels. The hierarchical control quantity corresponding to each level is calculated based on the causal relationship between different types of nodes in the causal graph of the rolling process. The action evaluation value is calculated based on the hierarchical control quantity. Based on the action evaluation value, the rolling mill control parameters are divided into a fast response parameter group and a steady-state compensation parameter group. The fast response parameter group is optimized by adaptive weight adjustment based on the intensity of causal effect to obtain a first optimization result. The steady-state compensation parameter group is optimized by asymptotic optimization based on the cumulative amount of rolling force error to obtain a second optimization result. The first optimization result and the second optimization result are combined and noise suppression is performed by a multilayer perceptron to obtain the control compensation value.

4. The method according to claim 3, characterized in that, Based on the action evaluation values, the rolling mill control parameters are divided into a fast response parameter group and a steady-state compensation parameter group. The fast response parameter group is optimized using adaptive weight adjustment based on the intensity of causal effects to obtain a first optimization result. The steady-state compensation parameter group is optimized using asymptotic optimization based on the cumulative amount of rolling force error to obtain a second optimization result, including: Based on the action evaluation value, the responsiveness analysis of the rolling mill control parameters is performed to determine the response time. The rolling mill control parameters with response times not less than the preset response speed threshold are divided into a steady-state compensation parameter group, and the rolling mill control parameters with response times less than the response speed threshold are divided into a fast response parameter group. For the rolling mill control parameters in the fast response parameter group, the sensitivity of each control parameter group on the critical causal path is calculated to obtain the causal effect strength. Based on the causal effect strength and the rolling force error value, the weight increment is calculated, and the fast response parameter group is adaptively updated based on the weight increment to obtain the first optimization result. The cumulative amount of rolling force error is determined by integrating the rolling force error value over a preset time period. The compensation amount is calculated by combining the pre-set progressive optimization coefficients. Based on the compensation amount, the steady-state compensation parameter set is progressively optimized to obtain a second optimization result.

5. The method according to claim 1, characterized in that, Based on the control compensation values, a parameter optimization space is constructed. Within this space, combining plate shape and quality constraints with equipment operation constraints, the optimal process parameters are obtained through a grid search algorithm, including: The temperature and thermal stress distributions in the rolling zone are collected, and the parameter optimization space is constructed by combining the pre-acquired material properties with the control compensation values. In the parameter optimization space, the plate shape deviation value and temperature field deviation value are calculated by combining the plate shape mass constraint and the equipment operation constraint, and the feasible domain of the parameters is limited to obtain the constraint parameter optimization space; The constraint parameter optimization space is divided into grids based on the grid search algorithm. The comprehensive evaluation index corresponding to each grid is calculated. The temperature gradient in the constraint parameter optimization space is calculated based on the comprehensive evaluation index. The grids with temperature gradients greater than a preset gradient threshold are refined. Parameter optimization is performed in the refined grids to obtain candidate process parameters. The real-time temperature field is collected, and the current thermal deformation is calculated based on the thermal stress distribution. The thermal deformation compensation is determined based on the current thermal deformation, and the thermal deformation compensation is superimposed with the candidate process parameters to obtain the optimal process parameters.

6. The method according to claim 5, characterized in that, Based on the grid search algorithm, the constraint parameter optimization space is divided into grids, and the comprehensive evaluation index corresponding to each grid is calculated, including: In the constraint parameter optimization space, an initial grid is divided based on a grid search algorithm, and the initial position distribution of the grid nodes is determined. Based on the initial position distribution and production requirements, the topological relationship matrix between the initial grids is determined. The geometric characteristic index of each initial grid is determined based on the topological relationship matrix and the grid quality evaluation value is calculated. The grid node positions are locally adjusted based on the grid quality evaluation value to obtain the corrected grid and the grid distortion degree and grid orthogonality corresponding to the corrected grid are calculated. The plate shape quality value and temperature uniformity value corresponding to each corrected grid are collected, and a comprehensive evaluation index is calculated by combining the grid distortion degree and the grid orthogonality.

7. A cold rolling mill rolling force prediction and adaptive control system based on multi-sensor fusion, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to collect process parameters collected by sensors in the cold rolling mill, extract features from the process parameters to obtain a working condition feature sequence, construct a dynamic correlation matrix between process parameters through a graph attention network and determine the influence weights, reconstruct the working condition feature sequence based on the influence weights, and calculate the predicted rolling force value by combining the contact arc length variation characteristics of the deformation zone during rolling. The second unit is used to calculate the rolling force error value between the predicted rolling force value and the actual rolling force, construct a control strategy based on the rolling force error value and calculate the action evaluation value, and optimize the rolling mill control parameters and suppress noise based on the action evaluation value to obtain the control compensation value. The third unit is used to construct a parameter optimization space based on the control compensation value. In the parameter optimization space, the optimal process parameters are obtained by combining the plate shape quality constraints and equipment operation constraints through a grid search algorithm. The optimal process parameters are then converted into cold rolling mill control commands and issued for execution.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images