A cold rolling industrial process monitoring method based on graph neural network

By using graph neural networks to construct graph structure data sets and graph convolution neural network models in the cold rolling industry, the shortcomings of traditional methods in monitoring complex systems are solved, and efficient and accurate monitoring of cold rolling industrial processes is achieved, which improves the Product quality and production stability.

CN116560277BActive Publication Date: 2024-11-01UNIV OF SCI & TECH BEIJING +1
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
CN202310564327.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-05-17
Publication Date
2024-11-01
Estimated Expiration
2043-05-17

AI Technical Summary

Technical Problem

In the cold rolling industry, traditional multivariate statistical analysis methods have long running time and low calculation accuracy, making it difficult to effectively monitor abnormalities and failures in complex systems, resulting in product quality and output affecting.

Method used

The process monitoring method based on graph neural network is adopted, and by constructing graph structure data sets and graph convolution neural network models, the output is predicted and faults are monitored, and problems are detected and dealt with in advance to avoid interfering with normal production processes.

Benefits of technology

The fit and calculation speed of the monitoring model are improved, accurate monitoring of the cold rolling industrial process is achieved, the interpretability of the data model is enhanced, and the stability of product quality and output is ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116560277B_ABST
    Figure CN116560277B_ABST
Patent Text Reader

Abstract

The present invention discloses a cold rolling industrial process monitoring method based on a graph neural network, comprising: obtaining industrial data, processing the industrial data, and establishing a row unit data set; based on the row unit data set, establishing a graph structure data set of the cold rolling process according to the spatiotemporal graph of the strip rolling process combined with an expert experience knowledge base; constructing a graph neural network process monitoring model, determining a calculation method for a statistical threshold according to data characteristics and calculating an abnormal threshold; comparing the obtained value of the monitoring model with the abnormal threshold, and monitoring the cold rolling industrial process through the comparison result. The present invention improves the model accuracy by constructing a graph structure that is consistent with the real mechanism model, and has a certain interpretability for the data model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of cold-rolled steel strip, and in particular to a cold-rolling industrial process monitoring method based on graph neural network. Background Art

[0002] As an important part of my country's industry, the cold rolling industry is the cornerstone for my country to maintain rapid economic development. Cold-rolled strip steel has achieved a high degree of automation, and computer data storage technology and data processing technology have made continuous breakthroughs, which has greatly reduced the cost of data storage and analysis. Data-driven cold rolling industrial process monitoring has flourished, promoting industrial digitalization and intelligent transformation.

[0003] The most important parameter in cold-rolled strip data is the flatness deviation, which can reflect the unevenness of the surface of the final product. The following formula is the calculation formula for the flatness deviation:

[0004]

[0005] Among them, i represents the strip being divided into i strips horizontally, λ(i) represents the elongation difference of the strip after the i-th division, ΔL represents the length difference between the i-th strip and the reference strip, L represents the reference length of the strip, and IU value is usually used as the unit of elongation difference. The physical meaning of 1IU is that after the residual stress of the strip is eliminated, there is an elongation difference of 10μm per meter. In the actual production process, IU is often used to represent the plate shape deviation.

[0006] With the increasing market demand for cold-rolled product output and the continuous improvement of quality and precision requirements, increasing output while ensuring that the product has sufficient quality and precision has become the current research goal. The cold-rolling production process has gradually developed in the direction of large-scale, integrated, and intelligent. Nowadays, the industrial production process is becoming increasingly complex. The multi-condition, multi-system, and mutually coupled production system makes the cold-rolling industrial production process very prone to abnormalities and failures. Once it deviates from the normal working conditions, it will seriously affect the product quality and output, and cause major economic losses, waste of resources, and even personal safety and other major production accidents. In view of this situation, process monitoring in the cold-rolling production process has gradually become a research hotspot. In the field of traditional process monitoring, the process monitoring model based on the mechanism model is very difficult to understand the relationship between the variables in the production process for modern multi-system, multi-variable production processes with complex mechanism models, resulting in a gradual decline in its effect. Now the data-driven process monitoring model is gradually developing. The data-driven process monitoring model regards the cold rolling industrial production system as a black box. It does not analyze the internal mechanism and directly models it based on the relationship between the input and output data in the research object. The model has strong online correction capability and can be applied to highly nonlinear and severely uncertain systems, thus providing an effective way to solve the modeling problem of complex systems.

[0007] Graph neural network aims to overcome the problem of graph structure data that traditional neural networks cannot process. Graph neural network is a neural network that processes non-standardized data structures. It has been widely used in fields including applied physics, molecular chemistry and knowledge graphs. Graph neural network is a neural network that calculates on graphs. Graph is a universal model that can efficiently express the complex relationships between a large number of things. It is a data representation form that describes complex transactions composed of nodes and edges. Mathematically, it is generally expressed as a graph, where G = (V, E), V (Vertical) represents the node, and characterizes each runtime parameter during the cold rolling process. E (Edge) represents the edge, which describes the relationship between the two parameters. Currently, the popular graph neural networks include graph convolution networks (GCN), graph attention networks (GAT) and GraphSAGE. GCN can learn the characteristics of nodes and structures at the same time, and GCN is highly practical and can be used for network models with arbitrary topological structures. It is currently the most in-depth mainstream application method for graph neural network research.

[0008] The cold rolling industrial process is gradually developing towards large-scale and complex development, and its process variables are gradually increasing. The traditional multivariate statistical analysis method has the problems of long running time and low calculation accuracy in actual production application. The neural network algorithm that has been widely studied at present has a good fit after training with a large amount of data. The trained model has the advantages of fast calculation speed and high calculation accuracy. Summary of the invention

[0009] In order to solve the above technical problems, the present invention proposes a cold rolling industrial process monitoring method based on graph neural network, establishes a data-driven process monitoring model, predicts output and monitors faults based on graph neural network, so as to detect and deal with problems early and in time to avoid interfering with normal production process and product quality.

[0010] To achieve the above object, the present invention provides a cold rolling industrial process monitoring method based on graph neural network, comprising:

[0011] Acquire industrial data, process the industrial data, and establish a row-based data set;

[0012] Based on the row unit data set, a graph structure data set of the cold rolling process is established according to the spatiotemporal graph of the strip rolling process and the expert experience knowledge base;

[0013] Construct a graph neural network process monitoring model, determine the calculation method of the statistical threshold according to the data characteristics and calculate the abnormal threshold; compare the value obtained by the monitoring model with the abnormal threshold, and monitor the cold rolling industrial process through the comparison result.

[0014] Preferably, processing the industrial data includes:

[0015] Cleaning the industrial data to remove reference values ​​and duplicate values;

[0016] Feature selection is performed on the industrial data to remove low-quality features that are irrelevant to the target.

[0017] Preferably, the Pearson correlation coefficient is used as a standard to perform feature selection on the industrial data, and the calculation method of the Pearson correlation coefficient is:

[0018]

[0019] Among them, X and Y are variables, and N is the sample size.

[0020] Preferably, establishing the graph structure dataset includes:

[0021] The original data set in units of rows or columns is converted into a graph structure in units of graphs, and the graph structure data set is obtained.

[0022] Preferably, establishing a graph structure data set of the cold rolling process includes:

[0023] Based on the spatiotemporal graph of the cold rolling process and the sharing of expert experience knowledge base, a graph structure dataset of the cold rolling process is established.

[0024] Preferably, before constructing the graph neural network process monitoring model, the following steps are included:

[0025] Generate node representations and determine activation functions for nonlinear operations, including;

[0026] According to the GCN calculation paradigm, node features are generated, and the graph convolutional neural network results are input into the LeakyReLU activation function to transform the model that originally fits linear mapping into a model that can fit nonlinear results.

[0027] Preferably, the method according to the GCN calculation paradigm is:

[0028]

[0029] Among them, H l is the input feature of the lth layer, H l+1 is the output feature of the lth layer; w l is the linear transformation matrix; σ(·) is the nonlinear activation function; is the self-connected adjacency matrix; is the degree matrix of the self-connected matrix;

[0030] The LeakyReLU activation function is:

[0031]

[0032] Among them, x is the activation value and a is the hyperparameter.

[0033] Preferably, constructing the graph neural network process monitoring model includes:

[0034] Based on the determination method of statistical threshold in traditional multivariate statistical methods, the 3σ rule is used to determine the abnormal IU value threshold of cold-rolled strip.

[0035] A GCN process prediction model is established to estimate the flatness IU value after any two moments. If the flatness IU estimated value exceeds the threshold, it is considered that an abnormality has occurred, and then the graph neural network process monitoring model is constructed.

[0036] Compared with the prior art, the present invention has the following advantages and technical effects:

[0037] Compared with the traditional multivariate statistical analysis method used in traditional fault monitoring models, the graph neural network model has the characteristics of better fitting, fast calculation speed and high calculation accuracy. The present invention improves the model accuracy by constructing a graph structure that is consistent with the actual mechanism model, and has a certain degree of interpretability for the data model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0039] Figure 1 Schematic diagram of neural network message passing paradigm in an embodiment of the present invention;

[0040] Figure 2 Schematic diagram of convolution operation based on traditional CNN in an embodiment of the present invention;

[0041] Figure 3 Schematic diagram of convolution operation based on graph structure in an embodiment of the present invention;

[0042] Figure 4 Schematic diagram of the LeakyReLU activation function in an embodiment of the present invention;

[0043] Figure 5 It is a schematic diagram of the cold rolling process structure in an embodiment of the present invention;

[0044] Figure 6 This is a GCN model diagram of the cold rolling process in an embodiment of the present invention;

[0045] Figure 7It is a line graph of the actual value of the IU value of the cold-rolled steel coil in the embodiment of the present invention;

[0046] Figure 8 It is a line chart of the predicted IU value of the cold-rolled steel coil in the embodiment of the present invention;

[0047] Fig. 9 This is a comparison chart between the predicted and actual IU values ​​of the cold-rolled steel coil in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0049] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0050] The present invention proposes a cold rolling industrial process monitoring method based on graph neural network, comprising:

[0051] Step 1: Data cleaning and feature selection

[0052] The steel coil data sample library is selected as the industrial data set. The specific steel coil data contained in the data sample library is shown in Table 1.

[0053] Table 1

[0054]

[0055] Industrial data sets contain redundant information such as reference variables and repeated values. Data cleaning can be used to delete redundant information and ensure training speed and modeling accuracy.

[0056] There are too many feature variables in the industrial data set. Select some more representative features and remove some low-quality features that are irrelevant to the target features to simplify the data set, ultimately reducing the training time and improving the accuracy of the model. Currently, the Pearson correlation coefficient is used as the standard to select representative features. The details are as follows:

[0057] The Pearson Correlation Coefficient is a statistic mainly used to measure the degree of correlation between different features. It can be used to calculate the correlation between input parameters in data preprocessing. The number of input parameters can be reduced by eliminating input parameters with larger Pearson parameters than other input parameters. The calculation method of the Pearson Correlation Coefficient is as follows:

[0058]

[0059] Step 2: Create a graph structure dataset

[0060] Based on the cold rolling process time-space graph and expert experience knowledge base sharing, the graph structure of the cold rolling process is established. Figure 5 The figure shows a schematic diagram of the cold rolling process structure, in which each dotted box represents each cold rolling rack separated in space, the points of different colors in each rack represent the features collected by different sensors, and the connecting lines between the points represent the transmission relationship between the two connected points.

[0061] Step 3: Generate node representation

[0062] The main goal of graph neural networks is to generate node representations for nodes. The paradigm for generating node representations in graph neural networks is also called the message passing paradigm. The message passing paradigm refers to the general operation of passing messages between nodes in a graph network. Figure 1 As shown, each update of node h1 requires h1 node and all nodes in its neighbor set to perform message aggregation operations first.

[0063] The message passing paradigm can be expressed as a function as follows:

[0064]

[0065] In the formula, h v ′——output feature of the current node v at this layer; h v ——Input features of the current node v; ——Indicates that a message is transmitted to a certain node; N v ——The neighbor set of node v; h u |u∈N v ——Traverse the neighbor set of node v; g(·)——message aggregation function.

[0066] Among them, g(·) is generally a Sum, Avg, or Max function. In the network layer of GCN, g(·) is a weighted sum based on degree, while in the graph attention neural network, it is a weighted sum based on attention. f(·) represents the activation operation of the neural network on the node features after message aggregation.

[0067] GCN is similar to convolutional neural network (CNN) and has powerful feature learning capabilities. Their essence is also similar. The convolution of a point can be regarded as the weighted sum of the neighbors of the point. Figure 2The figure shows the traditional image convolution operation. Each point represents a pixel. The pixels of the image can also be regarded as a graph structure. Adjacent pixels are connected by edges, and the convolution is obtained by weighted averaging of pixel neighbors.

[0068] For graph structures, a similar approach is used for convolution, such as Figure 3 As shown in the figure, the convolution of the red node is equivalent to taking the weighted average of its neighboring nodes.

[0069] The computational paradigm of a graph neural network layer is as follows:

[0070]

[0071] The calculation paradigm for GCN is as follows:

[0072]

[0073] In the formula, H l ——Input features of layer l; H l+1 ——output features of layer l; w l ——Linear transformation matrix; σ(·)——Non-linear activation function; ——Self-connected adjacency matrix; ——The degree matrix of the self-connected matrix.

[0074] in The definition is as follows:

[0075]

[0076] Where, I is the identity matrix; A is the adjacency matrix.

[0077] in The definition is as follows:

[0078]

[0079] Step 4: Determine the activation function for nonlinear operation

[0080] After calculating GCN, the calculation results need to be input into the activation function. The activation function enables the model that can only fit linear mapping relationships to fit nonlinear problems. The more popular activation functions are: Sigmoid, Tanh, Relu and LeakyRelu.

[0081] The traditional ReLU activation function is as follows:

[0082]

[0083] From the above formula, we can see that when x is less than 0, the node value is 0 after the activation function, and the weight is updated to 0 during back propagation, which eventually leads to neuron inactivation, which affects the accuracy of the model. The LeakyRelu activation function is an improved form of the Relu activation function, such as Figure 4 As shown in the figure, the difference between the two is that the LeakyRelu activation function has a smaller slope in the negative area, which can avoid the problem that the output of Relu is 0 when the input is negative.

[0084] LeakyReLU activation function, the formula is as follows:

[0085]

[0086] From the above formula, we can see that when x is less than 0, a smaller weight is assigned, which can effectively avoid neuron deactivation when backpropagating to update the weights and improve the model accuracy. Based on this model, the LeakyReLU function is finally selected as the activation function.

[0087] Step 5: Build a graph neural network process monitoring model

[0088] Drawing on the method of determining the threshold of statistical quantity in traditional multivariate statistical methods, the 3σ rule is used to determine the threshold of abnormal IU value for cold-rolled strip. The GCN process monitoring model is established using all parameter values ​​at the current moment, and the IU value of the plate shape two moments later is estimated, and then compared with the IU control limit. When the estimated IU value two moments later is greater than the IU control limit, it is judged that the plate shape will be abnormal two moments in the future, thereby realizing process monitoring of the actual cold rolling production process.

[0089] The GCN prediction model of the board shape deviation value is established through the graph convolutional neural network. The GCN prediction model includes an input layer, three graph convolution layers, a fully connected layer and an output layer, such as Figure 6 This model has good prediction accuracy for plate shape IU, as shown in Figure 7 , Figure 8 , Fig. 9 The input data is the parameter values ​​at the current moment, and the output data is the IU value two moments later. The output data is compared with the abnormal threshold to build a graph neural network process monitoring model.

[0090] All test sets in the sample library were subjected to process monitoring based on GCN. The results of the GCN process monitoring model for the test sets were good, with the monitoring accuracy maintained above 98.7% and the average monitoring accuracy being 99.3%.

[0091] Compared with existing technologies, the graph neural network-based process monitoring model can accurately locate anomalies in existing production processes.

[0092] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A cold rolling industrial process monitoring method based on graph neural network, characterized in that: include: Acquire industrial data, process the industrial data, and establish a row-based data set; Based on the row unit data set, a graph structure data set of the cold rolling process is established according to the spatiotemporal graph of the strip rolling process and the expert experience knowledge base; Constructing a graph neural network process monitoring model, determining a calculation method for a statistical threshold value according to data characteristics and calculating an abnormal threshold value; comparing a value obtained from the monitoring model with the abnormal threshold value, and monitoring the cold rolling industrial process through the comparison result; Before constructing the graph neural network process monitoring model, the process includes: generating node representation, determining the activation function for nonlinear operation, generating node features according to the GCN calculation paradigm, and inputting the graph convolutional neural network results into the LeakyReLU activation function to transform the model that originally fits the linear mapping into a model that can fit the nonlinear result; Constructing the graph neural network process monitoring model includes: Based on the determination method of statistical threshold in traditional multivariate statistical methods, the 3σ rule is used to determine the abnormal IU value threshold of cold-rolled strip. A GCN process prediction model is established to estimate the flatness IU value after any two moments. If the flatness IU estimated value exceeds the threshold, it is considered that an abnormality has occurred, and then the graph neural network process monitoring model is constructed; A GCN prediction model of plate shape deviation value is established through graph convolutional neural network. The GCN prediction model includes one input layer, three graph convolution layers, one fully connected layer and one output layer. The input data is the parameter value at the current moment, and the output data is the IU value two moments later. The output data is compared with the abnormal threshold to construct a graph neural network process monitoring model.

2. The cold rolling industrial process monitoring method based on graph neural network according to claim 1 is characterized in that: Processing the industrial data includes: Cleaning the industrial data to remove reference values ​​and duplicate values; Feature selection is performed on the industrial data to remove low-quality features that are irrelevant to the target.

3. The cold rolling industrial process monitoring method based on graph neural network according to claim 2 is characterized in that: The Pearson correlation coefficient is used as a standard to perform feature selection on the industrial data. The calculation method of the Pearson correlation coefficient is: Among them, X and Y are variables, and N is the sample size.

4. The cold rolling industrial process monitoring method based on graph neural network according to claim 1 is characterized in that: Establishing the graph structure dataset includes: The original data set in units of rows or columns is converted into a graph structure in units of graphs, and the graph structure data set is obtained.

5. The cold rolling industrial process monitoring method based on graph neural network according to claim 1 is characterized in that: Establishing a graph structure data set of the cold rolling process, including: Based on the spatiotemporal graph of the cold rolling process and the sharing of expert experience knowledge base, a graph structure dataset of the cold rolling process is established.

6. The cold rolling industrial process monitoring method based on graph neural network according to claim 1 is characterized in that: The method according to the GCN calculation paradigm is: Among them, H l is the input feature of the lth layer, H l+1 is the output feature of the lth layer; w l is the linear transformation matrix; σ(·) is the nonlinear activation function; is the self-connected adjacency matrix; is the degree matrix of the self-connected matrix; The LeakyReLU activation function is: Among them, x is the activation value and a is the hyperparameter.

Citation Information

Patent Citations

  • System and method for structure learning for graph neural networks

    CA3131843A1

  • Virtual automatic driving test optimization method based on improved DQN algorithm

    CN111625457A

  • Cold continuous rolling industrial process monitoring and anomaly detection method based on data driving

    CN113468479A

  • Industrial steam generation quantity prediction method and system based on Pearson correlation and neural network

    CN113705888A

  • Electric power Internet-of-things equipment anomaly detection method based on graph neural network

    CN114077811A