A cold-rolled strip steel product quality diagnosis method based on graph structure modeling

Abstract

This invention belongs to the field of steel rolling automation technology, and particularly relates to a method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling. The method includes: acquiring production data, which includes process parameters and quality indicators; preprocessing and quality category encoding the production data to obtain encoded production data; dividing the encoded production data into training, validation, and test sets; constructing a graph neural network model based on an attention mechanism, the graph neural network model including a first graph attention network layer connected to the graph neural network, and a third graph attention layer generating modulation parameters; performing feature linear modulation on the graph neural network using the ProGAT framework based on the modulation parameters; training the graph neural network model using the training and validation sets; and inputting the test set into the trained graph neural network model to achieve quality diagnosis of cold-rolled strip steel products. This invention improves the ability to identify a few abnormal states.

Description

Technical Field

[0001] The present invention belongs to the technical field of steel rolling automation, and particularly relates to a method for diagnosing the quality of cold-rolled strip products based on graph structure modeling. Background Art

[0002] In the modern steel industry, cold-rolled sheet and strip are widely used in fields such as automobile manufacturing, high-end household appliances, and precision instruments. The product quality has high requirements for thickness accuracy and shape quality. A cold-rolled tandem rolling production line is usually composed of multiple stands connected in series, and complex process coupling relationships are formed through strip transfer between stands.

[0003] Existing cold-rolled production lines are generally equipped with online monitoring devices such as tension sensors, X-ray thickness gauges, and shape detection devices for real-time monitoring of the production process. However, actual production data shows that quality abnormalities of cold-rolled strip are not evenly distributed throughout the production process, but are concentrated in specific non-steady-state stages, and the proportion of abnormal samples in the overall data is relatively low, showing an obvious data imbalance problem. Currently, some data-driven quality diagnosis methods mainly model based on a single stand or an independent process section, without fully considering the material transfer relationship and process coupling characteristics between stands in the tandem rolling mill, resulting in insufficient diagnostic accuracy and stability under the condition of multi-stand linkage. In addition, existing methods are prone to missed detections or misjudgments when facing a small number of quality abnormal samples and uneven distributions, and it is difficult to meet the actual production requirements. Summary of the Invention

[0004] In view of the above-mentioned technical problems existing in the diagnosis of the quality of cold-rolled strip products, such as the insufficient consideration of the process coupling relationship between stands and the insufficient diagnostic accuracy caused by data imbalance, a method for diagnosing the quality of cold-rolled strip products based on graph structure modeling is provided. The present invention mainly uses the graph structure modeling method to map multiple stands of a cold-rolled tandem rolling mill into graph nodes, constructs a directed topological edge structure to represent the material transfer relationship and process coupling characteristics between stands, and uses a graph attention network to extract the spatial dependence characteristics during the strip rolling process; at the same time, global process parameter modulation is introduced to inject macroscopic process constraints into the local feature evolution process, so as to fully mine the process coupling information between stands in the cold-rolling process, effectively address the data imbalance problem, improve the accuracy and stability of quality abnormality diagnosis under the condition of multi-stand linkage, and achieve the technical effects of precise classification and early warning of the quality of cold-rolled strip products.

[0005] The technical means adopted by the present invention are as follows:

[0006] A method for diagnosing the quality of cold-rolled strip products based on graph structure modeling, comprising the following steps: Acquire production data, which includes process parameters and quality indicators. The process parameters include rolling force, speed, tension, and roll gap. The quality indicators include thickness deviation and plate shape deviation. The production data is preprocessed and quality category coded to obtain coded production data; The encoded production data is divided into a training set, a validation set, and a test set; A graph neural network model based on an attention mechanism is constructed. The graph neural network model includes a first graph attention network layer, a second graph attention network layer, and a third graph attention network layer connected in sequence. The first graph attention network layer is connected to the graph neural network, and the third graph attention layer generates modulation parameters. Based on the modulation parameters, the graph neural network is subjected to feature linear modulation using the ProGAT framework. The graph neural network model is trained using the training set and validation set to obtain a trained graph neural network model. The test set is input into a trained graph neural network model to diagnose the quality of cold-rolled strip steel products.

[0007] Furthermore, the workflow of the graph neural network model includes: Establish the node mapping relationship of the rack:

[0008] in, This represents the node mapping relationship of the rack. The feature container for the k-th stand, k=1, 2, 3, 4, 5, encapsulates the local process dynamic information of the stand, including the rolling force, bending roll force, rolling speed, inlet tension and outlet tension of the stand; Based on the aforementioned local process dynamic information, a directed topological edge structure is constructed, and the path of the directed topological edge structure is as follows:

[0009] in, For paths with directed topological edge structures, For the forward process edge, As a self-holding edge, the calculation formula for the forward process edge is:

[0010] in, For the feature container of the (k+1)th rack, the formula for calculating the self-holding edge is: ; The encoded production data is input into the first graph attention network layer. In the first graph attention network layer, the updated feature representation is calculated based on the node mapping relationship and the path of the directed topological edge structure. The updated feature representation is passed sequentially through the second graph attention network layer and the third graph attention network layer to obtain the modulation parameters; The updated feature representation is processed by a graph neural network to obtain the output of the graph neural network; Based on the modulation parameters, the graph neural network is subjected to feature linear modulation using the ProGAT framework; The output features of the modulated graph neural network are used for graph classification to obtain the output of the graph neural network model.

[0011] Furthermore, the network architectures of the first attention network layer, the second attention network layer, and the third attention network layer are identical, and the workflow of the first attention network layer includes: The encoded production data input is defined as a node feature set, and the node feature set is as follows:

[0012] in, h i For the i-th encoded production data, For node feature set; The node feature set is linearly transformed using a weight matrix; Calculate the attention coefficients of nodes in the node feature set after linear transformation; The attention coefficients are normalized using weight normalization to obtain the final attention weights; Using the final attention weight, the updated feature representation of each node is generated by weighted aggregation of neighboring nodes and processed by a nonlinear activation function.

[0013] Furthermore, the formula for calculating the attention coefficient is as follows:

[0014] in, Attention coefficient It is a non-linear activation function. h i For the i-th encoded production data, h j For the j-th encoded production data, The final attention weight is calculated using the following formula, which is the weight matrix:

[0015] in, As the final weight of attention, It is a natural exponential function. The LeakyReLU activation function is used. For nodes i The neighborhood set, For the concatenation operation, the formula for calculating the updated feature representation is:

[0016] in, Let i be the updated feature representation. It is a non-linear activation function.

[0017] Furthermore, the workflow of the ProGAT framework includes: Construct a global feature vector and a local feature vector. The local feature vector is the updated feature representation. The global feature vector includes the inlet set thickness, the outlet set thickness, the actual inlet thickness, the actual outlet thickness, and the strip width. The global feature vector is processed by a multilayer perceptron to obtain modulation parameters, which include scaling vectors and translation vectors. The global feature vector is modulated using Hadamard element-wise multiplication and addition to generate a conditional feature representation:

[0018] in, For conditional feature representation, For scaling vectors, For Hadamard element-wise multiplication, For the k-th updated feature representation, It is a translation vector; The conditional feature representation is then subjected to numerical stabilization and activation function processing to obtain the stabilized features:

[0019] in, These are the characteristics after stabilization. It is an exponential linear unit. For batch normalization; The stabilized features are then subjected to global average pooling to form a graph representation. The calculation formula for the graph representation is as follows:

[0020] in, For the atlas representation, This represents the total number of nodes in the graph atlas. These are the nodes in the atlas representation.

[0021] Furthermore, when the training set is input into the graph structure network model for training, a loss function based on class weights is used for boundary sample classification. The calculation process of the loss function includes: Based on the real-time distribution of samples in each training batch, the number of samples in each category in the current batch is counted, and the weights corresponding to each category are calculated using an exponential decay method. The formula for calculating the weights corresponding to each category is as follows:

[0022] in, The weight corresponding to category c, For hyperparameters, This represents the number of samples in category c in the current training batch; To ensure numerical stability and maintain consistent gradient scaling, the weights for each category are normalized:

[0023] in, The normalized weights corresponding to category c are: The total number of categories, The weight corresponding to the m-th category; The atlas representation is input into the linear classification head to generate a prediction probability vector. The prediction probability vector is then processed by the softmax function to obtain the prediction probability corresponding to the category. Based on the predicted probabilities corresponding to the categories, a power-law modulation factor is calculated. Combining the dynamic category balancing weights and the power-law modulation factor, the loss function is calculated as follows:

[0024] in, For loss function, For the sample size, The predicted probability for each category. To focus parameters, The normalized weights corresponding to the true class labels of sample i. Let i be the true category label for sample i.

[0025] Furthermore, the preprocessing of the production data includes: The mode imputation method is used to fill in missing values ​​in production data, replacing missing values ​​with the values ​​that appear most frequently in the feature dimensions. Outliers in each feature dimension can be identified and removed using the 3σ principle or box plot method, or truncation can be performed using upper and lower thresholds. The feature data is scaled using Z-score normalization or Min-Max normalization. The SMOTE oversampling algorithm is used to synthesize and augment minority class samples on the training set. The specific steps include: Step 1: Based on the number of samples in each category, divide the training set into a minority class sample set and a majority class sample set; Step 2: Randomly select a sample from the minority class sample set as the reference sample, and calculate the Euclidean distance between the sample and its nearest neighbor samples; Step 3: Based on the calculated distance, interpolate and generate new samples in the feature space; Step 4: Repeat steps 1 to 4 until the number of samples in each category reaches the preset balance ratio, and obtain the enhanced training set.

[0026] Furthermore, the quality coding process for the production data includes: Based on thickness deviation and shape deviation, strip steel products are divided into four quality categories, including qualified category, abnormal shape category, abnormal thickness category, and severe abnormality category. The qualified category is defined as having an absolute thickness deviation within 1.0% and a shape deviation ≤ 6I, where I is a shape unit. The abnormal shape category is defined as having an absolute thickness deviation within 1.0% and a shape deviation > 6I. The abnormal thickness category is defined as having an absolute thickness deviation > 1.0% and a shape deviation ≤ 6I. The severe abnormality category is defined as having an absolute thickness deviation > 1.0% and a shape deviation > 6I. The quality categories are coded using the hot rolling coding method, and the categorical variables are converted into numerical vector form: the code for qualified category is [1,0,0,0], the code for abnormal plate shape category is [0,1,0,0], the code for abnormal thickness category is [0,0,1,0], and the code for severe abnormality category is [0,0,0,1].

[0027] Furthermore, the division of the encoded production data into a training set, a validation set, and a test set includes: The encoded production data is divided into a development set and a test set, with 80% used as the development set and 20% as the test set. Within the development set, it is further divided into a training set and a validation set in an 8:2 ratio.

[0028] Compared with the prior art, the present invention has the following advantages: This invention introduces a graphical modeling approach that reflects the physical structure of cold rolling mills, enabling a unified expression of quality information under multi-stand coupled operating conditions and improving the stability of diagnostic results. It effectively identifies the gradual accumulation of quality degradation trends during cold rolling production, avoiding alarms only being triggered after severe anomalies occur. Addressing the characteristics of industrial data—low proportion and uneven distribution of quality anomaly samples—it improves the ability to identify a few abnormal states. The method has a clear structure, is easy to integrate with existing cold rolling automation systems, and has significant engineering application value.

[0029] Based on the above reasons, this invention can be widely promoted in fields such as steel rolling automation. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of a cold rolling production line and sensor layout for a cold-rolled strip steel product quality diagnosis method based on graph structure modeling according to the present invention. Figure 2 This is a schematic diagram of the system architecture of a cold-rolled strip steel product quality diagnosis method based on graph structure modeling according to the present invention; Figure 3 This is a schematic diagram of the global process parameter modulation method for a cold-rolled strip steel product quality diagnosis method based on graph structure modeling according to the present invention; Figure 4 This is a flowchart illustrating the product quality diagnosis process of a graph-based modeling method for diagnosing the quality of cold-rolled strip steel products according to the present invention. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] Figure 1 In this cold rolling production line, there are five stands, each equipped with sensors such as tension sensors, X-ray thickness gauges, and strip shape detection devices. Production data includes, but is not limited to, process parameters such as rolling force, speed, tension, and roll gap for each stand, as well as quality indicators such as strip thickness and strip shape.

[0035] The rolling force is calculated using the pressure data fed back by the pressure sensor of the hydraulic pressing system, along with the diameters of the rod chamber and rodless chamber of the hydraulic cylinder; the speed is the actual speed detected by the encoder on the back of the motor; the tension is the actual tension detected by the tension gauge; the roll gap is calculated based on the position value detected by the SONY magnetic ruler inside the hydraulic cylinder; the thickness deviation is detected by the thickness gauge to measure the actual thickness difference; and the plate shape deviation is detected by the plate shape roll.

[0036] Based on the operational experience of cold rolling production lines, process characteristics closely related to the quality of cold-rolled strip steel are selected as input variables, including but not limited to the inlet tension, outlet tension, rolling force, rolling speed, inlet and outlet thickness of each stand, as well as the final shape deviation and thickness deviation as quality labels.

[0037] like Figure 2 As shown, this invention provides a method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling, including the following steps: S1. Obtain production data, which includes process parameters and quality indicators. Process parameters include rolling force, speed, tension, and roll gap. Quality indicators include thickness deviation and shape deviation.

[0038] Production data generated during the cold rolling continuous rolling production line was collected. This data included real-time measurements of process parameters and corresponding cold-rolled strip quality indicators from multiple stands. The collected process parameters included rolling force, rolling speed, roll gap, inlet tension, outlet tension, inlet thickness, outlet thickness, and shape-related parameters for each stand. Feature filtering was performed on the collected production data, retaining key features related to cold-rolled strip thickness deviation and shape quality as input variables for subsequent modeling. To capture complex dynamic relationships, a dataset containing 26-dimensional process features was constructed.

[0039] S2. Preprocess and classify the production data to obtain the coded production data.

[0040] The preprocessing of production data includes: The mode imputation method is used to fill in missing values ​​in each feature dimension, replacing missing values ​​with the most frequent values ​​in that feature dimension to reduce the impact of outliers on model training. Outliers in each feature dimension are identified and removed using the 3σ principle or box plot method, or truncation is performed using upper and lower thresholds to ensure the validity of the input data. The SMOTE oversampling algorithm is used to synthesize and expand the minority class samples on the training set. Specific steps include: Step 1: Divide the training set into a minority class sample set and a majority class sample set based on the number of samples in each class. Step 2: Randomly select a sample from the minority class sample set as the reference sample and calculate the Euclidean distance between this sample and its nearest neighbors. Step 3: Generate new samples by interpolating them in the feature space based on the calculated distances. Step 4: Repeat steps 1 to 4 until the number of samples in each class reaches a preset balanced ratio, resulting in the enhanced training set.

[0041] The Min-Max normalization method is used to normalize the production data after handling missing values, mapping each feature to the 0~1 interval to obtain the preprocessed production data.

[0042] The above steps achieve initial equilibration of extremely imbalanced industrial data, improving the model's ability to learn from rare defective samples.

[0043] The quality coding process for production data includes: The first step involves classifying strip steel products into four quality categories based on thickness and shape deviations: Acceptable, Shape Abnormal, Thickness Abnormal, and Severe Abnormal. The Acceptable category is defined as a thickness deviation with an absolute value within 1.0% and a shape deviation ≤ 6I. The Shape Abnormal category is defined as a thickness deviation with an absolute value within 1.0% and a shape deviation > 6I. The Severe Abnormal category is defined as a thickness deviation with an absolute value > 1.0% and a shape deviation ≤ 6I. I represents the shape unit, calculated using the following formula:

[0044] in, I For plate units, For elongation, For the length difference, This is a reference length.

[0045] The second step is to encode the quality categories using the hot rolling coding method, converting the categorical variables into numerical vector forms: the code for the qualified category is [1,0,0,0], the code for the plate shape abnormality category is [0,1,0,0], the code for the thickness abnormality category is [0,0,1,0], and the code for the severe abnormality category is [0,0,0,1].

[0046] S3. Divide the encoded production data into training set, validation set and test set.

[0047] The encoded production data is divided into a development set and a test set, with 80% allocated to the development set and 20% to the test set for final model performance evaluation. Within the development set, it is further divided into a training set and a validation set in an 8:2 ratio, with 80% used for model training and 20% for hyperparameter tuning.

[0048] The training set is represented as ,in For the training set, For the production data in the l-th sample of the training set, for Category This represents the number of samples in the training set.

[0049] S4. Construct a graph neural network model based on an attention mechanism. The graph neural network model includes a first graph attention network layer, a second graph attention network layer, and a third graph attention network layer connected in sequence. The first graph attention network layer is connected to the graph neural network. The third graph attention layer generates modulation parameters. Based on the modulation parameters, the graph neural network is subjected to feature linear modulation using the ProGAT framework.

[0050] In the model, each stand in the cold rolling production line is abstracted as a node in a graph structure. The features of each node represent the dynamic process parameters of that stand, including but not limited to rolling force, speed, tension, and roll gap. Through this node mapping, the process state of each stand can be effectively described. The edges in the graph represent the material transfer relationships and process coupling relationships between stands. Graph Attention Network (GAT) can capture the influence between different stands by calculating the association strength between nodes, thereby effectively modeling the information transfer in the process.

[0051] The workflow of a graph neural network model includes: The first step involved constructing a mapping strategy based on the physical layout of the production line and the movement path of the strip steel, by defining a set of nodes. Model each physical rack in chronological order to establish the node mapping relationship of the racks:

[0052] in, This represents the node mapping relationship of the rack. Let K be the feature container for the k-th stand, where k = 1, 2, 3, 4, 5. This feature container encapsulates the local process dynamics information of the stand, including rolling force, bending roll force, rolling speed, inlet tension, and outlet tension. This feature vector integrates mechanical loads, kinematic parameters, and inter-stand constraints, and is typically represented as a concatenation of various sensor variables.

[0053] In the above formula, and These represent the rolling force, bending roll force, rolling speed, inlet tension, and outlet tension of the k-th stand, respectively.

[0054] The second step is to construct a directed topological edge structure based on local process dynamic information, and to set the edge set... The path is constructed as the union of forward process-propagating edges and self-holding edges, representing the causal path of process disturbance propagation. The path of the directed topological edge structure is:

[0055] in, For paths with directed topological edge structures, For the forward process edge, For self-maintaining edges. From the set The forward process edge is defined to model the downstream material transport process and the resulting coupling effects. The formula for calculating the forward process edge is:

[0056] in, This is the feature container for the (k+1)th rack. It is derived from the set... The self-holding edge is defined to characterize the time inertia and mechanical stability of each rack. The formula for calculating the self-holding edge is: ; The third step is to input the encoded production data into the first graph attention network layer. In the first graph attention network layer, the updated feature representation is calculated based on the node mapping relationship and the path of the directed topological edge structure.

[0057] Fourth step: Pass the updated feature representation through the second graph attention network layer and the third graph attention network layer in sequence to obtain the modulation parameters.

[0058] Step 5: The updated feature representation is processed through a graph neural network to obtain the output of the graph neural network.

[0059] Step 6: Perform feature linear modulation on the graph neural network using the ProGAT framework based on the modulation parameters.

[0060] Step 7: Perform graph classification on the output features of the modulated graph neural network to obtain the output of the graph neural network model.

[0061] The first, second, and third attention network layers have the same network architecture. Adding a second graph attention network layer further updates the feature information of the nodes.

[0062] The workflow of the first attention network layer includes: The first step involves analyzing the features of each stand node, including its process parameters such as rolling force, roll gap, and inlet tension. A graph attention network (GAN) is used as input to process these node features. The GAN is chosen as the backbone of the feature extraction module, its core function being to calculate the correlation strength between nodes, which represents how information propagates through the rolling mill. The encoded production data input is defined as the node feature set, which is as follows:

[0063] in, h i For the i-th encoded production data, This is the node feature set.

[0064] The second step, to enhance the expressive power of these features, involves applying a shared linear transformation parameterized by the weight matrix W to each node. This linear transformation of the node feature set is performed using the weight matrix.

[0065] The third step involves calculating the association strength (attention coefficient) between nodes using the graph attention mechanism. This strength represents the degree to which one rack influences the state of other racks. Through these calculations, the graph network can dynamically adjust the flow of information between nodes. The attention coefficient of each node in the linearly transformed feature set is calculated; this coefficient represents the importance of node j's features to node i. In the cold rolling environment, this coefficient can be interpreted as the intensity of process disturbance propagation from one rack to another. The formula for calculating the attention coefficient is:

[0066] in, Attention coefficient It is a non-linear activation function used for nodes. i and nodes j The features are processed using a weighted representation. h i For the i-th encoded production data, h j For the j-th encoded production data, This is the weight matrix.

[0067] Step 4: Features between nodes are aggregated using a weighted average. The aggregated features are then passed to the next layer to form a more accurate representation of the node state. This process considers not only the state of a single rack but also the influence of the states of adjacent racks. To make these coefficients comparable across different nodes and considering the topological structure of the graph, attention coefficients are normalized using weight normalization to obtain the final attention weight. The formula for calculating the final attention weight is:

[0068] in, As the final weight of attention, It is a natural exponential function, which can alleviate the "dead neuron" problem that ReLU may cause. The LeakyReLU activation function is used. For nodes i The neighborhood set, For the splicing operation, the nodes are... i and nodes j The transformed feature vectors are concatenated to form a new feature vector, α. T It is a transpose operation, typically a learnable parameter vector, Wh i and Wh j It is a node i and nodes j After passing through the weight matrix respectively W The transformed feature vector.

[0069] Step 5: The node feature representations processed by the non-linear activation function of the graph attention network further optimize the quality diagnosis results, ensuring high sensitivity and accuracy to abnormal states. Using the final attention weights, weighted aggregation of neighboring nodes is processed by a non-linear activation function to generate updated feature representations for each node. The calculation formula for the updated feature representations is as follows:

[0070] in, Let i be the updated feature representation. It is a non-linear activation function.

[0071] To achieve comprehensive quality assessment, the proposed ProGAT framework needs to integrate local dynamic features with global process constraint information. In this embodiment, the local feature set X... loc Defined as the attribute of each node in the graph, each h k Characterizes the high-frequency dynamic fluctuation features of the k-th rack. Global feature vector x glb This describes the macroscopic boundary conditions for the entire rolling process, including key parameters such as the set inlet thickness, set outlet thickness, actual inlet thickness, actual outlet thickness, and strip width. A global process parameter modulation unit is introduced to inject global process constraint information into local stand features. A modulation mechanism is used to weight and adjust the dynamic features between stands, such as... Figure 3 As shown, this modulation process enables consistency between local features and the global objective, thereby improving the accuracy and robustness of quality diagnosis. To address the severe class imbalance problem and difficult-to-classify boundary samples in cold-rolled data, an adaptive class weight loss function is introduced. By combining the number of effective samples in a mini-batch with the focus modulation term, the sample weights are adaptively adjusted, ensuring that the model maintains high sensitivity to rare defect patterns.

[0072] The ProGAT framework's workflow includes: The first step is to construct global and local feature vectors. The local feature vectors are the updated feature representations, while the global feature vectors include the inlet set thickness, the outlet set thickness, the actual inlet thickness, the actual outlet thickness, and the strip width.

[0073] The second step of this invention employs a global process parameter modulation unit. By performing an affine transformation on the latent graph features, global process information is injected into the stand-level feature evolution process, thereby achieving comprehensive quality assessment of the cold rolling process. The global feature vector is processed by a multilayer perceptron to obtain modulation parameters, which include scaling and translation vectors. The calculation formula for the modulation parameters is as follows:

[0074] in, For scaling vectors, It is a translation vector. It is a generator for Multi-Layer Perceptron (MLP).

[0075] Step 3, Feature Modulation and Conditionalization: The global feature vector is modulated using Hadamard element-wise multiplication and addition to generate a conditional feature representation.

[0076] in, For conditional feature representation, For Hadamard element-wise multiplication, This represents the k-th updated feature representation.

[0077] Step 4: Numerical Stabilization and Activation Function Processing: The conditional feature representation undergoes numerical stabilization and activation function processing to ensure feature stability and enhance the model's training robustness. The stabilized features are obtained as follows:

[0078] in, These are the characteristics after stabilization. It is an exponential linear unit. For batch normalization.

[0079] Step 5, Graph-level Feature Pooling: To obtain a representative embedding representation of the entire five-rack system, Global Mean Pooling (GMP) is performed on the modulated node set V. This operation aggregates local and conditionally processed rack-level information. Global mean pooling is then applied to the stabilized features to form a graph representation. The calculation formula for the graph representation is as follows:

[0080] in, For the atlas representation, This represents the total number of nodes in the graph atlas. These are the nodes in the atlas representation.

[0081] S5. Use the training set and validation set to train the graph neural network model to obtain the trained graph neural network model.

[0082] When the training set is input into the graph structure network model for training, a loss function based on adaptive class weights is used for boundary sample classification. By adaptively adjusting the weights of samples in each training batch, high sensitivity to rare defect patterns is ensured, avoiding the dominance of majority class samples in model training. This process effectively solves the inter-class imbalance problem in imbalanced datasets and improves the accurate identification of abnormal states. The calculation process of the loss function includes: Step 1: Based on the real-time distribution of samples in each training batch, count the number of samples in each category in the current batch, and calculate the weight of each category using an exponential decay method. The formula for calculating the weight of each category is as follows:

[0083] in, The weight corresponding to category c, To control the effective sample size, n is determined by performing dynamic binning statistics on the true labels in each iteration. c This allows the model to adapt to fluctuations in the composition of the current batch of samples. This represents the number of samples in category c in the current training batch.

[0084] The second step, to ensure numerical stability and maintain the consistency of gradient scale, is to normalize the weights corresponding to each category:

[0085] in, The normalized weights corresponding to category c are: The total number of categories, represents the weight corresponding to the m-th category.

[0086] The third step involves inputting the atlas representation into the linear classification head to generate a predicted probability vector s. This predicted probability vector is then processed using the softmax function to obtain the predicted probabilities corresponding to each category. .

[0087] Step 4: To highlight the importance of difficult-to-classify samples, calculate the power-law modulation factor based on the predicted probabilities corresponding to the categories. Combining dynamic class balancing weights and a power-law modulation factor, the loss function is calculated as follows:

[0088] in, For loss function, For the sample size, The predicted probability for each category. To focus parameters, The normalized weights corresponding to the true class labels of sample i. Let be the true class label for sample i. This loss function form avoids the ProGAT framework being dominated by easily classifiable samples in the majority class, and instead guides the model to focus on key outliers that have a low frequency but a decisive impact on the final quality grade of cold-rolled strip steel.

[0089] Specifically, step four includes two main steps: Dynamic class weight calculation: In each training batch, the class weights are dynamically adjusted according to the distribution of samples, so that the model can focus on minority class samples, especially those abnormal samples that account for a small proportion of the training data but are crucial for quality diagnosis.

[0090] Focus modulation mechanism: To avoid the model from over-focusing on easily classifiable majority class samples during training, the focus modulation mechanism dynamically adjusts the sample weights, enabling the model to focus on those difficult-to-classify boundary samples, thereby improving the model's sensitivity and classification accuracy.

[0091] S6. Use the test set as input to the trained graph neural network model to diagnose the quality of cold-rolled strip steel products.

[0092] Based on the above processing flow, this invention can achieve quality diagnosis of cold-rolled strip steel and adjust the production process according to the diagnosis results. When the model detects a quality anomaly, the system will output the corresponding quality diagnosis results to help operators take timely measures to optimize production parameters or adjust production processes to ensure stable product quality. The diagnosis results include a prediction of the quality level. For example, if an anomaly is found in thickness or shape, the system can prompt adjustments to the control parameters of the relevant stand to prevent the quality problem from escalating further.

[0093] Figure 4 This is a flowchart illustrating a graph learning modeling process for quality diagnosis of cold-rolled products: First, key process parameters related to product quality are selected, and the raw data undergoes missing value processing, data augmentation, and normalization, with sample labeling completed based on quality grades. Then, a graph structure based on the coupling relationship between the stand and tension is constructed, node features are represented, and weights between nodes are calculated using an attention mechanism to achieve weighted information aggregation, thereby obtaining a feature expression of the coupling characteristics of the fusion process. On this basis, a dynamic feature modulation mechanism and numerical stabilization processing (such as BN and ELU) are introduced to improve model robustness. Simultaneously, dynamic class weight calculation and an improved focus loss function are used to alleviate class imbalance problems. Finally, high-precision abnormal state identification and quality diagnosis decision support are achieved, forming a complete quality diagnosis framework integrating data preprocessing, graph attention modeling, robust optimization, and cost-sensitive learning.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling, characterized in that, Includes the following steps: Acquire production data, which includes process parameters and quality indicators. The process parameters include rolling force, speed, tension, and roll gap. The quality indicators include thickness deviation and plate shape deviation. The production data is preprocessed and quality category coded to obtain coded production data; The encoded production data is divided into a training set, a validation set, and a test set; A graph neural network model based on an attention mechanism is constructed. The graph neural network model includes a first graph attention network layer, a second graph attention network layer, and a third graph attention network layer connected in sequence. The first graph attention network layer is connected to the graph neural network, and the third graph attention layer generates modulation parameters. Based on the modulation parameters, the graph neural network is subjected to feature linear modulation using the ProGAT framework. The graph neural network model is trained using the training set and validation set to obtain a trained graph neural network model. The test set is input into a trained graph neural network model to diagnose the quality of cold-rolled strip steel products.

2. The method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling according to claim 1, characterized in that, The workflow of the graph neural network model includes: Establish the node mapping relationship of the rack: in, This represents the node mapping relationship of the rack. The feature container for the k-th stand, k=1, 2, 3, 4, 5, encapsulates the local process dynamic information of the stand, including the rolling force, bending roll force, rolling speed, inlet tension and outlet tension of the stand; Based on the aforementioned local process dynamic information, a directed topological edge structure is constructed, and the path of the directed topological edge structure is as follows: in, For paths with directed topological edge structures, For the forward process edge, As a self-holding edge, the calculation formula for the forward process edge is: in, For the feature container of the (k+1)th rack, the formula for calculating the self-holding edge is: ； The encoded production data is input into the first graph attention network layer. In the first graph attention network layer, the updated feature representation is calculated based on the node mapping relationship and the path of the directed topological edge structure. The updated feature representation is passed sequentially through the second graph attention network layer and the third graph attention network layer to obtain the modulation parameters; The updated feature representation is processed by a graph neural network to obtain the output of the graph neural network; Based on the modulation parameters, the graph neural network is subjected to feature linear modulation using the ProGAT framework; The output features of the modulated graph neural network are used for graph classification to obtain the output of the graph neural network model.

3. The method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling according to claim 2, characterized in that, The network architectures of the first attention network layer, the second attention network layer, and the third attention network layer are the same. The workflow of the first attention network layer includes: The encoded production data input is defined as a node feature set, and the node feature set is as follows: in, h i For the i-th encoded production data, For node feature set; The node feature set is linearly transformed using a weight matrix; Calculate the attention coefficients of nodes in the node feature set after linear transformation; The attention coefficients are normalized using weight normalization to obtain the final attention weights; Using the final attention weight, the updated feature representation of each node is generated by weighted aggregation of neighboring nodes and processed by a nonlinear activation function.

4. The method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling according to claim 3, characterized in that, The formula for calculating the attention coefficient is: in, Attention coefficient It is a non-linear activation function. h i For the i-th encoded production data, h j For the j-th encoded production data, The final attention weight is calculated using the following formula, which is the weight matrix: in, As the final weight of attention, It is a natural exponential function. The LeakyReLU activation function is used. For nodes i The neighborhood set, For the concatenation operation, the formula for calculating the updated feature representation is: in, Let i be the updated feature representation. It is a non-linear activation function.

5. The method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling according to claim 2, characterized in that, The workflow of the ProGAT framework includes: Construct a global feature vector and a local feature vector. The local feature vector is the updated feature representation. The global feature vector includes the inlet set thickness, the outlet set thickness, the actual inlet thickness, the actual outlet thickness, and the strip width. The global feature vector is processed by a multilayer perceptron to obtain modulation parameters, which include scaling vectors and translation vectors. The global feature vector is modulated using Hadamard element-wise multiplication and addition to generate a conditional feature representation: in, For conditional feature representation, For scaling vectors, For Hadamard element-wise multiplication, For the k-th updated feature representation, It is a translation vector; The conditional feature representation is then subjected to numerical stabilization and activation function processing to obtain the stabilized features: in, These are the characteristics after stabilization. It is an exponential linear unit. For batch normalization; The stabilized features are then subjected to global average pooling to form a graph representation. The calculation formula for the graph representation is as follows: in, For the atlas representation, This represents the total number of nodes in the graph atlas. These are the nodes in the atlas representation.

6. The method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling according to claim 1, characterized in that, When the training set is input into the graph structure network model for training, boundary sample classification is performed using a loss function based on class weights. The calculation process of the loss function includes: Based on the real-time distribution of samples in each training batch, the number of samples in each category in the current batch is counted, and the weights corresponding to each category are calculated using an exponential decay method. The formula for calculating the weights corresponding to each category is as follows: in, The weight corresponding to category c. For hyperparameters, This represents the number of samples in category c in the current training batch; To ensure numerical stability and maintain consistent gradient scaling, the weights for each category are normalized: in, The normalized weights corresponding to category c are: The total number of categories, The weight corresponding to the m-th category; The atlas representation is input into the linear classification head to generate a prediction probability vector. The prediction probability vector is then processed by the softmax function to obtain the prediction probability corresponding to the category. Based on the predicted probabilities corresponding to the categories, a power-law modulation factor is calculated. Combining the dynamic category balancing weights and the power-law modulation factor, the loss function is calculated as follows: in, For loss function, For the sample size, The predicted probability for each category. To focus parameters, The normalized weights corresponding to the true class labels of sample i. Let i be the true category label for sample i.

7. The method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling according to claim 1, characterized in that, The preprocessing of the production data includes: The mode imputation method is used to fill in missing values ​​in production data, replacing missing values ​​with the values ​​that appear most frequently in the feature dimensions. Outliers in each feature dimension can be identified and removed using the 3σ principle or box plot method, or truncation can be performed using upper and lower thresholds. The feature data is scaled using Z-score normalization or Min-Max normalization. The SMOTE oversampling algorithm is used to synthesize and augment minority class samples on the training set. The specific steps include: Step 1: Based on the number of samples in each category, divide the training set into a minority class sample set and a majority class sample set; Step 2: Randomly select a sample from the minority class sample set as the reference sample, and calculate the Euclidean distance between the sample and its nearest neighbor samples; Step 3: Based on the calculated distance, interpolate and generate new samples in the feature space; Step 4: Repeat steps 1 to 4 until the number of samples in each category reaches the preset balance ratio, and obtain the enhanced training set.

8. The method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling according to claim 1, characterized in that, The quality coding process for the production data includes: Based on thickness deviation and shape deviation, strip steel products are divided into four quality categories, including qualified category, abnormal shape category, abnormal thickness category, and severe abnormality category. The qualified category is defined as having an absolute thickness deviation within 1.0% and a shape deviation ≤ 6I, where I is a shape unit. The abnormal shape category is defined as having an absolute thickness deviation within 1.0% and a shape deviation > 6I. The abnormal thickness category is defined as having an absolute thickness deviation > 1.0% and a shape deviation ≤ 6I. The severe abnormality category is defined as having an absolute thickness deviation > 1.0% and a shape deviation > 6I. The quality categories are coded using the hot rolling coding method, and the categorical variables are converted into numerical vector form: the code for qualified category is [1,0,0,0], the code for abnormal plate shape category is [0,1,0,0], the code for abnormal thickness category is [0,0,1,0], and the code for severe abnormality category is [0,0,0,1].

9. The method for quality diagnosis of cold-rolled strip steel products based on graph structure modeling according to claim 1, characterized in that, The division of the encoded production data into training, validation, and test sets includes: The encoded production data is divided into a development set and a test set, with 80% used as the development set and 20% as the test set. Within the development set, it is further divided into a training set and a validation set in an 8:2 ratio.

Images