CN-121765283-B - Cold rolling friction coefficient prediction method based on physical graph topology and graph convolution neural network

Abstract

The cold rolling friction coefficient prediction method based on the physical graph topology and the graph rolling neural network comprises the steps of defining a building rule of a graph topology structure for friction coefficient analysis in plate and strip cold continuous rolling, selecting analysis parameters according to plate and strip cold continuous rolling process characteristics and production experience, determining node sets, physical relation function sets and edge sets in the graph topology structure, constructing an adjacent matrix and a corresponding self-adjacent matrix of the graph rolling network with physical weights, performing normalization processing on the self-adjacent matrix by an introduction degree matrix to generate the normalized self-adjacent matrix, constructing the graph rolling neural network based on the physical characteristics according to the setting and the basic form of the graph rolling network, selecting cold rolling product samples, obtaining analysis parameters and friction coefficient priori values, training the graph rolling neural network based on the physical characteristics, and evaluating the performance of the graph rolling neural network based on the physical characteristics by adopting decision coefficients, mean square errors, average absolute errors and average absolute percentage errors.

Inventors

• LI XU
• JIN SHUREN
• HAN YUEJIAO
• ZHANG XIN

Assignees

• 东北大学

Dates

Publication Date

20260512

Application Date

20260302

Claims (5)

1. 1. The cold rolling friction coefficient prediction method based on the physical graph topology and the graph convolution neural network is characterized by comprising the following steps of: Step 1, defining a rule for establishing a graph topological structure for friction coefficient analysis in plate and strip cold continuous rolling based on a heuristic search theory; step 2, selecting analysis parameters according to the plate and strip cold continuous rolling process characteristics and production experience, and determining a node set, a physical relationship function set and an edge set in a topological structure of the graph; step 3, constructing an adjacency matrix and a corresponding self-adjacency matrix of a graph rolling network with physical weights based on a Reynolds equation; Step 4, introducing a degree matrix, performing normalization processing on the self-adjacent matrix to meet the operation requirement of the graph rolling network, and generating a normalized self-adjacent matrix; step 5, constructing a physical characteristic diagram convolutional neural network based on the cold rolling friction coefficient according to the basic forms of the map convolution network and the settings of the steps 2 to 4; Step 6, selecting a cold-rolled product sample, acquiring analysis parameters and friction coefficient priori values, taking the analysis parameters as input based on a physical characteristic diagram convolutional neural network, taking the friction coefficient priori values as output, and training the convolutional neural network based on the physical characteristic diagram; step 7, evaluating the performance of the neural network based on the physical characteristic graph by adopting a decision coefficient, a mean square error, an average absolute error and an average absolute percentage error; The step 3 specifically comprises the following steps: The reynolds equation for the lubrication process in the cold continuous rolled strip is: (6) In the formula, The thickness of the oil film is instantaneous, and the unit is mm; The unit is mm 2 /s, which is the initial viscosity of the lubricant; Is the viscosity-pressure coefficient; Rolling pressure is unit MPa; The rolling speed is integrated, and the unit is m/s; Is a differential operator; Discretizing the formula (6) to obtain edges Physical weight coefficient under corresponding process Represented as; (7) In the formula, Is a side Corresponding to the instantaneous oil film thickness in mm under the process; Is that Rolling pressure difference under the corresponding process is unit MPa; Is a discretization step length; Defining the physical weight coefficient according to the basic principle of the graph neural network Adjacent matrix of (a) The method comprises the following steps: (8) Adjacent matrix The modification is expressed as a self-adjacency matrix : (9) In the formula, Is a unit matrix, by introducing the unit matrix to make And the self-environment protection certificate node self-feature propagation is increased.
2. 2. The method for predicting the cold rolling friction coefficient based on the physical graph topology and the graph rolling neural network according to claim 1, wherein the step 1 is specifically: let the node set in the graph structure be And is expressed as: (1) In the node Representing an input feature, total number of nodes The number of the input feature categories is the same as that of the input feature categories; make the physical relation function set as : (2) In the formula, Is composed of nodes Related nodes formed for independent variables Is a function of (a) and (b), Is the number of physical relationship functions; Let the edges in the graph structure be gathered as : (3) In the formula, Is a node And (3) with By a function of An established edge; based on formulas (1) - (3), the graph structure is defined The following are provided: (4) In the formula, Is based on a set of functions In (3) by means of edge sets Defining a node set Directed edges between nodes in the (a).
3. 3. The method for predicting the cold rolling friction coefficient based on the physical graph topology and the graph rolling neural network according to claim 2, wherein the step 2 is specifically: Selecting rolling force RF, working roll flattening radius RA, oil film thickness OL, working roll surface roughness RO, rolling mileage RD, rolling speed RS, tensile stress TS, deformation resistance DR, rolling reduction RE, temperature K and emulsion viscosity LV as analysis parameters, collecting nodes The definition is as follows: (5) Based on the functional relation among the parameters, the existing mechanism analysis result and the production experience, the method determines And 。
4. 4. The method for predicting the cold rolling friction coefficient based on the physical graph topology and the graph rolling neural network according to claim 1, wherein the step 4 is specifically: Definition matrix The method comprises the following steps: (10) by means of For a pair of The following normalization process was performed: (11) In the formula, To normalize the self-adjacency matrix.
5. 5. The method for predicting cold rolling friction coefficient based on physical graph topology and graph rolling neural network according to claim 4, wherein the operation flow based on physical feature graph rolling neural network in step 5 is as follows: For graph rolling network, neighbor aggregation result of graph rolling operation The method comprises the following steps: (12) In the formula, An input feature matrix composed of analysis parameters; is a learnable weight; Middle (f) The specific calculation mode of the row is as follows: (13) In the formula, For normalizing elements of row i and column j in the adjacency matrix; Is a node Is used for the degree of discharge of the steel wire, Is a node Is a degree of departure of (2); Is a node Corresponding feature vectors; Is the first to input the feature matrix A column; The output layer is removed, and the propagation architecture for a graph rolling network with one hidden layer is: (14) In the formula, As a learnable weight of the first layer, An output matrix for the first layer; global pooling of the final hidden layer to aggregate features of all nodes into graph-embedded vectors : (15) In the formula, The predicted output of the convolutional neural network P-GCN based on the physical feature map is expressed as: (16) In the formula, Is a predictive vector; For linear weighting of output layers Is a rank of the transition; Is a bias vector.

Description

Cold rolling friction coefficient prediction method based on physical graph topology and graph convolution neural network Technical Field The invention belongs to the technical field of metallurgical rolling, and relates to a cold rolling friction coefficient prediction method based on a physical graph topology and a graph convolution neural network. Background The cold continuous rolling process of the plate and the strip is a core process for realizing accurate regulation and control of the performance of the high-end plate and efficient production in the steel industry, the control precision of the process directly determines the mechanical property, the dimensional precision and the surface quality of a finished product, and the process is a guarantee for the reliability and the competitiveness of key materials in modern manufacturing. The friction coefficient is used as a key medium of stress transmission and a dominant factor of neutral point deviation, and accurate control of the friction coefficient directly influences rolling stability and is a core parameter of process optimization. First, the coefficient of friction directly affects the rolling force energy parameter. A higher friction coefficient means that more rolling force and energy consumption are required to achieve the same reduction, increasing the equipment load and power consumption. Secondly, it concerns the shape and surface quality of the board. The high friction coefficient is easy to cause the defects of vibration lines, scratch and the like on the surface of the strip steel, while the low friction coefficient can influence the rolling stability due to the increase of the forward slip value, even cause slip and damage the plate shape. In addition, the uniform and stable friction coefficient is a precondition for ensuring the consistent deformation of the strip steel along the width direction and obtaining good plate shape. The current friction coefficient research mainly comprises two methods of theoretical model and experimental determination. However, the simplification of the actual process by theoretical models and physical simulation experiments leads to distortion of the characterization of the coefficient of friction, which is manifested by deviations of critical rolling parameters (such as set values from measured values). In addition, although the adaptive correction method is adopted, the method is limited by the deficiency of an effective reference standard and the deficiency of correlation analysis of the physical characteristics of data by the adaptive method, and the actual and effective friction coefficient is still difficult to obtain. At present, most manufacturers still rely on manual intervention to correct friction coefficients and abnormal rolling characteristics caused by the friction coefficients. Therefore, achieving high-precision autonomous calculation of friction coefficients is a critical issue to be solved. Disclosure of Invention In order to solve the technical problems, the invention aims to provide a cold rolling friction coefficient prediction method based on a physical graph topology and a graph convolution neural network. The invention provides a cold rolling friction coefficient prediction method based on a physical graph topology and a graph convolution neural network, which comprises the following steps: Step 1, defining a rule for establishing a graph topological structure for friction coefficient analysis in plate and strip cold continuous rolling based on a heuristic search theory; step 2, selecting analysis parameters according to the plate and strip cold continuous rolling process characteristics and production experience, and determining a node set, a physical relationship function set and an edge set in a topological structure of the graph; step 3, constructing an adjacency matrix and a corresponding self-adjacency matrix of a graph rolling network with physical weights based on a Reynolds equation; Step 4, introducing a degree matrix, performing normalization processing on the self-adjacent matrix to meet the operation requirement of the graph rolling network, and generating a normalized self-adjacent matrix; step 5, constructing a physical characteristic diagram convolutional neural network based on the cold rolling friction coefficient according to the basic forms of the map convolution network and the settings of the steps 2 to 4; Step 6, selecting a cold-rolled product sample, acquiring analysis parameters and friction coefficient priori values, taking the analysis parameters as input based on a physical characteristic diagram convolutional neural network, taking the friction coefficient priori values as output, and training the convolutional neural network based on the physical characteristic diagram; and 7, evaluating the performance of the neural network based on the physical characteristic map by adopting a decision coefficient, a mean square error, an average absolute error an