Material rolling force prediction method and system, medium, terminal

Abstract

This invention provides a method, system, medium, and terminal for predicting rolling force of materials, relating to the field of rolling technology. This invention employs a Transformative Neural Network (TCN) to capture the long-term dependencies between features in rolling force prediction, utilizes a Physical Information Neural Network (PIN) to construct a physical information equation, and combines physical information with the TCN through the PSN, achieving a fusion of data-driven and physical-driven rolling force prediction. This improves the model's adaptability and prediction accuracy to non-stationary conditions and dynamic disturbances. Simultaneously, this invention innovatively proposes a data partitioning and integrated modeling strategy that combines feature subset partitioning and feature prediction result fusion. By establishing a feature interaction compensation mechanism, it achieves systematic modeling of the rolling force generation mechanism, effectively characterizing the hierarchical features of its intrinsic mechanism, laying the foundation for improving the model's prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of rolling technology, specifically to a method and system for predicting the rolling force of materials, a medium, and a terminal. Background Technology

[0002] The hot rolling process of strip steel involves complex material physical deformation and thermodynamic heat transfer processes, and has significant physical and mechanical properties. Accurately predicting the temporal changes of finishing rolling force is particularly important for improving strip steel performance, optimizing the production process, and enhancing product quality.

[0003] Traditional PI-TCN models are mathematically modeled based on rolling force theory, but these theoretical assumptions may not cover all complex behaviors, thus limiting the model's adaptability. Currently, with the development of data storage and artificial intelligence technologies, deep learning-based data-driven prediction methods have become mainstream due to their powerful feature learning and end-to-end prediction capabilities. For example, physical guidance constructs a higher-dimensional feature space, separating the static and dynamic errors of rolling force, enabling data-driven models to better adapt to complex process conditions and improving the physical consistency of predictions.

[0004] However, existing research only uses traditional physical models and feature engineering independently, which results in insufficient adaptability. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a method, system, medium, and terminal for predicting material rolling force, which solves the technical problem of insufficient adaptive capability in existing technologies.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a method for predicting the rolling force of a material, comprising:

[0010] Acquire multi-source feature data of the hot rolling and finishing process of the target material strip.

[0011] Based on contact arc length and roll wear, the data of M features in the multi-source feature data are divided into several data subsets;

[0012] The PI-TCN model is pre-trained to process several subsets of data for each feature, and several subset prediction results are obtained. The prediction results of several subsets are added together to obtain M feature prediction results. The M feature prediction results are then fused according to a dynamic ratio to obtain the rolling force prediction result of the target material.

[0013] The pre-trained PI-TCN model refers to embedding physical information through a physical information neural network during the training process of a temporal convolutional network.

[0014] Preferably, the multi-source feature data includes process parameters, physical characteristic data, and equipment status parameters.

[0015] Preferably, the data of M features in the multi-source feature data based on contact arc length and roll wear is divided into several data subsets, including:

[0016] The data of M features in the multi-source feature data are divided into layers according to the interval of contact arc length, and the disturbance level is divided according to the amount of roll wear to obtain several data subsets after each feature is divided.

[0017] Preferably, the step of fusing the prediction results of the M features according to a dynamic ratio to obtain the rolling force prediction result of the target material includes:

[0018]

[0019] in,

[0020]

[0021]

[0022]

[0023] In the formula, For the first Predicted rolling force results for a single strip steel finishing mill; Indicates the first The characteristics of strip steel finishing rolling m The predicted value; This represents the total number of strip steel. ; The characteristics are divided into contact arc length and roll wear. The total error in the prediction of rolling force; This represents the total error in the amplified rolling force prediction. The characteristics are represented by the contact arc length and roll wear. Adaptive soft weighting for rolling force prediction; Indicates error The number of non-zero digits after the decimal point; This indicates that the PI-TCN model is performing well in terms of features during training. The root mean square error on; This indicates that the PI-TCN model is performing well in terms of features during training. The mean absolute error over time; This indicates that the PI-TCN model is performing well in terms of features during training. The mean absolute percentage error.

[0024] Preferably, the loss function during the training process of the pre-trained PI-TCN model includes:

[0025]

[0026] In the formula, This represents the loss function of the PI-TCN model. This represents the data-driven loss function. This represents the hyperparameter used to control the weights of the physical monotonicity constraint loss subfunction. Represents the physical singleness constraint loss subfunction;

[0027] Considering the dimensionality issues of data loss and physical loss, a dynamic normalization principle is adopted when calculating the loss function, and exponential smoothing is used to track the magnitude changes of each loss term. and :

[0028]

[0029] In the formula, As a smoothing factor, Control the length of historical memory; adopt an adaptive weight allocation normalized total loss function:

[0030]

[0031]

[0032]

[0033] in, It represents a very small positive number.

[0034] Preferably, the calculation method of the physical singleness constraint loss sub-function includes:

[0035]

[0036]

[0037] In the formula, This represents the predicted rolling force after embedding the physical equation. This represents the rolling force predicted by the TCN neural network. Indicates the first The first in the block strip steel Each input feature , These represent the rolling temperature and strain rate, respectively. R Represents the gas constant. To activate energy, Indicates the strain rate sensitivity index. The number of strip steel blocks, This represents a symbolic factor used to indicate the direction of physical relationships. This represents the derivatives of the PI-TCN model with respect to the property parameters and stress parameters. The function only penalizes violations of physical laws, and the square term... Used to strengthen the penalty for outliers.

[0038] Preferably, the temporal convolutional network includes dilated causal convolution, in which, for a filter... For one-dimensional sequences Dilated convolution operation of element s F It can be represented as follows:

[0039]

[0040] in, As the expansion factor, Indicates expansion rate Convolutional filters , Number the filter. For filter size, For convolution operators, Explain the direction of the past, and .

[0041] In a second aspect, the present invention provides a material rolling force prediction system, characterized in that the material rolling force prediction system is used to perform the material rolling force prediction method as described above.

[0042] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for predicting the rolling force of a material, wherein the computer program causes a computer to execute the method for predicting the rolling force of a material as described above.

[0043] Fourthly, the present invention provides an electronic device, comprising:

[0044] One or more processors;

[0045] Memory; and

[0046] One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the prediction of material rolling forces as described above.

[0047] (III) Beneficial Effects

[0048] This invention provides a method, system, medium, and terminal for predicting material rolling force. Compared with existing technologies, it has the following advantages:

[0049] This invention employs a Transformative Neural Network (TCN) to capture long-term dependencies between features in rolling force prediction. It utilizes a Physical Information Neural Network (PIN) to construct physical information equations, combining physical information with the TCN to achieve a fusion of data-driven and physical-driven rolling force prediction. This enhances the model's adaptability to non-stationary conditions and dynamic disturbances, as well as its prediction accuracy. Furthermore, this invention innovatively proposes a data partitioning and integrated modeling strategy that integrates feature subset partitioning and feature prediction results. By establishing a feature interaction compensation mechanism, it achieves systematic modeling of the rolling force generation mechanism, effectively characterizing the hierarchical features of its intrinsic mechanism and laying the foundation for improving the model's prediction accuracy. Attached Figure Description

[0050] 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a block diagram of a method for predicting the rolling force of a material according to an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the influence of roll wear and contact arc length on rolling force.

[0053] Figure 3 A schematic diagram of dilated causal convolution;

[0054] Figure 4 This is a schematic diagram of a residual block in a temporal convolutional network.

[0055] Figure 5 This is a schematic diagram of the physical information neural network in the PI-TCN model.

[0056] Figure 6 To verify the evaluation index results of various methods in predicting rolling force in pass F2;

[0057] Figure 7 To verify the evaluation index results of various methods in predicting rolling force in pass F3. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] This application provides a method, system, medium, and terminal for predicting material rolling force, which solves the technical problem of poor generalization ability of existing technologies, and achieves accurate prediction of finishing rolling force with high generalization.

[0060] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0061] The accuracy of rolling force prediction is influenced by multiple factors, including dynamic fluctuations in material deformation resistance, the coupling effect of process parameters, and the time-varying drift characteristics of equipment conditions, exhibiting highly nonlinear and uncertain characteristics. Among these, material deformation resistance, influenced by the synergistic effects of temperature, strain rate, and chemical composition (such as carbon and niobium content), is prone to abrupt changes near the phase transformation critical point, causing the rolling force response to deviate from the linear assumptions of traditional models. Furthermore, switching rolling specifications (e.g., narrow strip to wide strip) or adjusting cooling strategies instantaneously alters the thermal boundary conditions between the workpiece and rolls, making rolling force fluctuations random and intermittent. In addition, the time-varying drift of equipment conditions is not only reflected in the hysteresis of the hydraulic system and fluctuations in electrical control accuracy, but also in the dynamic changes in roll diameter caused by frictional heat accumulation and metal-to-metal contact wear during long-term rolling, which also significantly affect rolling force prediction. In summary, rolling force is influenced by the coupling effect of multiple sources, including material properties, process parameters, and equipment conditions, exhibiting highly nonlinear and strong uncertain characteristics.

[0062] Existing rolling prediction models and their related parameters suffer from insufficient accuracy and poor adaptability, failing to meet the demands of modern markets for multi-steel-grade, small-batch, and highly flexible production. These production characteristics necessitate more intelligent and dynamic prediction models to adapt to rapid switching and efficient production of different materials and specifications.

[0063] To address the aforementioned issues, the present invention provides a method and system for predicting material rolling force, a feature modeling framework that integrates prior physical structure and temporal dependencies in the construction of the medium and terminal, and introduces feature partitioning, fusion, and temporal modeling methods to improve the model's adaptability to non-stationary operating conditions and dynamic disturbances, as well as its prediction accuracy.

[0064] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0065] This invention provides a method for predicting the rolling force of a material, such as... Figure 1 As shown, it includes:

[0066] S1. Obtain multi-source feature data of the hot rolling and finishing process of the target material strip;

[0067] S2. Based on the contact arc length and roll wear, the data of M features in the multi-source feature data are divided into several data subsets;

[0068] S3. The PI-TCN model is pre-trained to process several subsets of data for each feature to obtain several subset prediction results. The prediction results of several subsets are added together to obtain M feature prediction results. The M feature prediction results are then fused according to a dynamic ratio to obtain the rolling force prediction result of the target material.

[0069] The pre-trained PI-TCN model refers to embedding physical information through a physical information neural network during the training process of a temporal convolutional network.

[0070] In one embodiment, S1, multi-source characteristic data of the hot rolling and finishing process of the target material strip are acquired, wherein the multi-source characteristic data includes process parameters, physical property data, and equipment status parameters. The specific implementation process is as follows:

[0071] In this embodiment of the invention, the target material is any hot-rolled steel. Process parameters include roughing mill exit temperature, finishing mill inlet speed, inlet temperatures of F1-F7, inlet surface temperatures of F1-F7, linear speeds of F1-F7, deformation rates of F1-F7, cooling water flow rates after F3-F6, and stand intervals of F1-F7. Here, F1-F7 refer to the rolling sequence numbers of the seven four-high mills in the finishing mill unit. The finishing mill unit consists of these seven mills (F1-F7). The intermediate billet is sequentially centered by the side guide plates between mills and then continuously rolled into finished strip steel by mills F1-F7. The rolling process of each mill is one pass; F1 is the first pass, F2 is the second pass, and so on, until F7 completes the final pass.

[0072] Physical property data includes material parameters and dimensional parameters. Material parameters refer to the types and proportions of components in the target material, while dimensional parameters refer to the width and thickness of the intermediate billet, the inlet thickness of F1 to F7, and the reduction of F1 to F7.

[0073] The equipment status parameters include: F1~F7 working roll radius, F1~F7 contact arc length, F1~F7 wear value, F1~F7 thermal expansion, F1~F7 zero point offset, F1~F7 stiffness, and F1~F7 bending roll.

[0074] In one embodiment, S2, based on the contact arc length and roll wear, the data of M features in the multi-source feature data are divided into several data subsets. The specific implementation process is as follows:

[0075] Before making the division, we first analyze the factors that affect the accuracy of rolling force prediction:

[0076] The accuracy of rolling force prediction is influenced by multiple factors, including the dynamic fluctuations in material deformation resistance, the coupling effect of process parameters, and the time-varying drift characteristics of equipment conditions, exhibiting highly nonlinear and uncertain characteristics. Among these, material deformation resistance, influenced by the synergistic effects of temperature, strain rate, and chemical composition (such as carbon and niobium content), is prone to abrupt changes near the phase transformation critical point, causing the rolling force response to deviate from the linear assumptions of traditional models. Furthermore, switching rolling specifications (e.g., narrow strip to wide strip) or adjusting cooling strategies instantaneously alters the thermal boundary conditions between the workpiece and rolls, making rolling force fluctuations random and intermittent. Existing rolling prediction models and their related parameters, due to insufficient accuracy and poor adaptability, cannot meet the demands of modern markets for multi-steel grade, small-batch, and highly flexible production. These production characteristics necessitate more intelligent and dynamic prediction models to adapt to rapid switching and efficient production of different materials and specifications.

[0077] Furthermore, the time-varying drift of equipment status is not only reflected in the lag in hydraulic system response and fluctuations in electrical control accuracy, but also in the dynamic changes in roll diameter caused by frictional heat accumulation and metal-to-metal contact wear during long-term rolling, which become an important factor affecting rolling force prediction. The changes in roll diameter during the rolling process and their impact on the rolling process are as follows: Figure 2 As shown.

[0078] The traditional formula for calculating rolling force is:

[0079] (1)

[0080] in,

[0081] (2)

[0082] (3)

[0083] In the formula, Rack number, ; Indicates the first Block steel strip; For the first Block strip steel on the frame The rolling force on; This is the pressure correction factor; For strip steel on the frame The hardness; For the first The reduction ratio of the frame; For the first The unit hardness of the frame; For the first Contact area between the mill rolls and the strip; No. The strip steel in the first Rack inlet thickness; For the first The strip steel in the first The thickness of the rack exit, i.e., the thickness at the first... Rack entrance thickness.

[0084] Contact area in equation (2) This can be expressed as the contact arc length and the width of the rolled piece. The product of:

[0085] (4)

[0086] When the Block steel frame The initial radius of the rolls on the rolls is Roll wear is The effective radius of the worn roll is

[0087] (5)

[0088] Due to the amount of reduction Then the expression for the contact arc length is:

[0089] (6)

[0090] Referring to the Bland-Ford model (a classic rolling force calculation model in the cold rolling field), the pressure correction coefficient is known. It is positively correlated with the contact arc length, and its expression is:

[0091] (7)

[0092] in, The pressure coefficient, For the average thickness, substitute... The expression is

[0093] (8)

[0094] Substituting equations (4) and (8) into equation (1), we get the complete rolling force formula as follows:

[0095] (9)

[0096] It is evident that the rolling force is directly proportional to the contact arc length and inversely proportional to the wear of the rolls.

[0097] In summary, rolling force is influenced by multiple factors, including material properties, process parameters, and equipment status, exhibiting highly nonlinear and uncertain characteristics. Analytically, rolling force is closely related to geometric and physical features such as contact arc length and roll wear, which dynamically evolve with rolling rhythm and cooling strategies, exhibiting significant time-series characteristics. Therefore, it is crucial to construct a feature modeling framework that integrates prior physical structure and temporal dependencies, introducing feature partitioning, fusion, and temporal modeling methods to improve the model's adaptability and prediction accuracy to non-stationary conditions and dynamic disturbances.

[0098] In this embodiment of the invention, the data of M features in the multi-source feature data are divided into several data subsets based on the contact arc length and roll wear. Specifically, this includes:

[0099] Contact arc length As a key control variable in rolling force calculation, it essentially characterizes the dynamic contact characteristics and deformation coordination mechanism of the workpiece-roll system. Its physical connotation is reflected in the geometric coupling relationship of the contact area. Therefore, constructing a dataset partitioning criterion based on the multi-scale characterization characteristics of the contact arc length can achieve spatial decoupling between the rolling force generation mechanism and process parameters. This mechanism-driven data organization strategy provides a coupled modeling basis for the deformation coordination equation and data distribution characteristics for constructing a physically interpretable PI-TCN model. Equation (6) shows the calculation formula for the contact arc length. Considering the roll flattening effect, the roll will undergo local deformation in actual production. The Hitchcock modified formula (an improvement on the classic Hitchcock roll elastic flattening formula) is used to calculate the contact arc length after considering the elastic flattening of the roll, as shown in the equation:

[0100] (10)

[0101] In the formula, P This represents the total vertical pressure applied by the rolls to the workpiece during the rolling process. This represents the roll's resistance to elastic flattening. Since the actual pressure value experienced by the roll cannot be obtained in advance during the rolling force prediction process, the Jacobi iteration method is used for numerical solution, gradually converging to a final stable solution. .

[0102] Based on big data feature mining of the rolling process combined with kernel density estimation and process mechanism verification, it was found that the contact arc length exhibits a three-peak distribution during the rolling process. Through coupled analysis of process parameters and equipment status, the following interval division criteria were determined, and these criteria were used as the threshold for dividing the contact arc length. Since the threshold is not exactly the same at different rolling points, the specific values ​​will be shown in the subsequent experimental section.

[0103] (11)

[0104] (12)

[0105] (13)

[0106] That is, two thresholds are set, and the data of M features in the multi-source feature data are stratified according to the three intervals determined by the two thresholds. Then, the disturbance level is divided by the amount of roll wear, and several data subsets are obtained after each feature is divided.

[0107] Roll wear evolution, as a key time-varying disturbance source for rolling force prediction, directly determines the generalization performance of the prediction model across the entire operating cycle through its dynamic quantitative characterization. Based on the synergistic analysis of online monitoring data and the mechanistic model, and by reasonably dividing the dataset, the rolling force response law under different wear stages can be effectively modeled. The formula for calculating roll wear is as follows:

[0108] (14)

[0109] In the formula, Indicates the radius of the rolls before rolling. This indicates the radius of the rolled roll. In this embodiment of the invention, it is represented by... As an important dividing point for the wear state of the rolls, it reflects the nonlinear trend of wear behavior:

[0110] (15)

[0111] (16)

[0112] Upon testing, the feature segmentation method used in the embodiments of the present invention can pass the Kruskal-Wallis test (a non-parametric analysis of variance method) well, and the rolling force differences between the subsets after the dataset is segmented are significant.

[0113] In one embodiment, S3, the pre-trained PI-TCN model processes several subsets of data for each feature to obtain several subset prediction results. These subset prediction results are then summed to obtain M feature prediction results. Finally, the M feature prediction results are fused according to a dynamic ratio to obtain the rolling force prediction result for the target material. The specific implementation process is as follows:

[0114] The pre-trained PI-TCN model refers to embedding physical information into a physical information neural network during the training of a temporal convolutional network. The following is a detailed introduction to these two networks:

[0115] Temporal Convolutional Networks (TCNs) are convolutional neural networks used for sequence modeling tasks with causal constraints, such as time series prediction. Sequence modeling networks are non-linear mapping functions that can be obtained through supervised learning. They involve using the predicted sequence... and the actual output sequence The network is trained using a loss function between the mapping and mapping parameters. Furthermore, the mapping function must adhere to causal constraints. Depend on Determine, rather than using future input To achieve this, the first layer of TCN is a one-dimensional fully convolutional network, where each intermediate layer has the same size as the input layer, and zero padding is used to make subsequent layers the same size as the previous ones. Convolution operations are purely linear operations, as they only perform matrix addition and multiplication. However, since most data in the real world is nonlinear and complex, the proposed model uses nonlinear activation functions (such as rectified linear units) to address this issue. ReLU The TCN network introduces nonlinearity through the sigmoid activation function and the sigmoid function. Like recurrent neural networks (RNNs), TCN networks can process sequences or time series of arbitrary lengths and map them to output data of the same length.

[0116] TCN networks include dilated causal convolutions to respect causal constraints, improve the corresponding field, ensure that future information is not leaked into the past, and avoid excessive computational work. For filters... For one-dimensional sequences Dilated convolution operation of element s F It can be represented as follows:

[0117] (17)

[0118] in, As the expansion factor, Indicates expansion rate Convolutional filters , Number the filter. For filter size, For convolution operators, Explain the direction of the past, and . Figure 3 An example of dilated causal convolution is shown, where the dilation factor is . The filter size is To expand the network's receptive field, the filter size was increased. and expansion factor The larger the field of perception, the better the network can capture past situations.

[0119] After performing dilated causal convolution, the TCN network uses Activation functions are used to introduce nonlinearity, thereby improving the expressive power of the model. The formula is:

[0120] (18)

[0121] To prevent overfitting during neural network training, Dropout regularization is added to the training process of each layer. Dropout randomly selects a subset of neurons. Set their activation values ​​to 0. Typically, each neuron has a certain probability of being dropped. That is, each neuron propagates forward with probability The data is discarded, and the remaining data remains unchanged. The Dropout calculation formula is:

[0122] (19)

[0123] TCN consists of a one-dimensional fully convolutional network and residual blocks, used to train models and avoid the vanishing gradient problem. The residual blocks used in this study are as follows: Figure 4 As shown. It consists of two layers of dilated causal convolutions and a nonlinear layer, and... Used as an activation function. A Dropout layer is added after each dilated convolution to avoid overfitting. An additional 1×1 convolution is added to return to the original number of channels.

[0124] TCN is a novel convolutional architecture specifically designed for sequential modeling. While LSTM is also used for time series modeling, TCN achieves an exponentially growing receptive field through dilated convolutions, enabling it to capture long-term dependencies in feature attributes and making it more suitable for long-term rolling force prediction. In rolling force prediction, TCN was compared with several mainstream neural networks, and experiments verified that TCN outperforms other models in accuracy.

[0125] The physical information embedding model is designed as follows:

[0126] Although TCN can effectively capture the time dependence of the rolling process and achieve high-precision prediction of long-term series, the limitations of purely data-driven models are particularly prominent in industrial scenarios:

[0127] (1) Lack of physical consistency: The prediction results may violate the basic laws of rolling mechanics (such as the conservation of plastic deformation work and stress-strain relationship), leading to non-physical solutions under extreme working conditions;

[0128] (2) Generalization bottleneck: Insufficient adaptability to process parameters (new steel grades, ultra-high rolling speed) outside the training data distribution.

[0129] To address this, this invention proposes a physics-informed temporal convolutional network (PI-TCN) model that integrates Physical Information Neural Network (PINN) and TCN. PINN embeds physical laws (energy conservation, constitutive equations) into the loss function, forcing the model output to conform to known mechanisms, thus achieving a fusion of data-driven and physically constrained approaches. When measured data is insufficient, it utilizes physical equations to provide prior knowledge, reducing reliance on labeled data. By combining temperature and stress field equations, it predicts the impact of features such as contact arc length and roll wear on rolling force prediction, enhancing interpretability under complex conditions. PI-TCN embeds the rolling force balance equation as a hard constraint into the loss function, enabling the model to simultaneously satisfy the dual objectives of data fit goodness and adherence to physical laws. The architecture of the Physical Information Neural Network is as follows: Figure 5 As shown. This fusion is achieved through three key designs: (1) symbolic encoding of physical equations; (2) dynamic weighting of hybrid loss functions; and (3) prediction correction mechanism based on residual connections.

[0130] In this embodiment of the invention, a hybrid modeling strategy combining the Arrhenius equation (the core formula in chemical kinetics describing the quantitative relationship between the reaction rate constant and temperature) and data-driven correction is employed in loss-driven models to balance physical rationality and data adaptability. The expression for the Arrhenius constitutive equation is as follows:

[0131] (20)

[0132] In the formula, This represents the predicted rolling force after embedding the physical equation. This represents the rolling force predicted by the TCN neural network. Indicates the first The first in the block strip steel Each input feature , These represent the rolling temperature and strain rate, respectively. R Represents the gas constant. To activate energy, This represents the strain rate sensitivity index.

[0133] After representing the physical laws of the rolling process as differentiable operators and embedding them into the neural network training process, physical consistency constraints are constructed through symbolic differentiation.

[0134] (twenty one)

[0135] In the formula, The number of strip steel blocks, This represents a sign factor (with a value of -1 or 1) used to indicate the direction of physical relationships. This represents the derivative of the PI-TCN model with respect to the property parameters and stress parameters. The function only penalizes violations of physical laws, and the square term... This is used to strengthen the penalty for outliers, making the optimization process more inclined to provide smooth and physically consistent solutions.

[0136] Therefore, the loss function of the model is expressed as the following formula:

[0137] (twenty two)

[0138] In the formula, Represents the model loss function. This represents the data-driven loss function. This represents the hyperparameter used to control the weights of the physical monotonicity constraint loss subfunction. This represents the physical singleness constraint loss function.

[0139] Considering the dimensionality issues related to data loss and physical loss, a dynamic normalization principle is adopted when calculating the loss function, and exponential smoothing (EMA) is used to track the magnitude changes of each loss term:

[0140] (twenty three)

[0141] In the formula, As a smoothing factor, Control the length of historical memory. Adaptive weight allocation is used to normalize the total loss function.

[0142] (twenty four)

[0143] (25)

[0144] (26)

[0145] in, It represents a very small positive number.

[0146] In the specific implementation process, before predicting the rolling force of the target material using the PI-TCN model, the PI-TCN model needs to be trained and optimized. The training process of the PI-TCN model is shown in Table 1:

[0147] Table 1 Training process of PI-TCN model

[0148]

[0149] The data processing methods during the training and application of the PI-TCN model are similar, and will not be elaborated here.

[0150] The trained PI-TCN model processes several subsets of data for each feature, obtaining several subset prediction results. These subset prediction results are then summed to obtain M feature prediction results. These M feature prediction results are then dynamically fused to obtain the rolling force prediction result for the target material. In this embodiment, to further improve the prediction effect of feature fusion by efficiently fusing the rolling force prediction results based on contact arc length and roll wear, an improved soft-weight enhancement strategy is used as the final prediction output of the model. This strategy employs three evaluation metrics based on different feature prediction results during the PI-TCN model training process. RMSE m , MAEm and MAPEm These are the input data. These metrics are commonly used indicators for evaluating predictive performance; their values ​​gradually decrease as prediction accuracy improves. Furthermore, the values ​​for all three sets of data range from [0, +∞).

[0151]

[0152]

[0153]

[0154] In the formula, , This represents the total number of strip steel. Indicates the first Measured values ​​of the finishing rolling force of a strip steel. Indicates the first The characteristics of strip steel finishing rolling m The predicted value of the rolling force model. Both during model training and subsequent model use can be based on... , and These three indicators yield the final rolling force prediction result. This indicates that the PI-TCN model is performing well in terms of features during training. The root mean square error on; This indicates that the PI-TCN model is performing well in terms of features during training. The mean absolute error over time; This indicates that the PI-TCN model is performing well in terms of features during training. The mean absolute percentage error.

[0155] Based on the above performance index formula, the total error in predicting rolling force under the conditions of contact arc length and roll wear can be expressed as: :

[0156]

[0157] In order to enhance To improve the discriminative power of different modules when their prediction accuracy is similar, and thus enhance the influence of each module, an error amplification formula is used:

[0158]

[0159] Based on the magnified error Adaptive soft weighting for rolling force prediction under contact arc length and roll wear classification It can be calculated using the following formula:

[0160]

[0161] in, Indicates error The number of non-zero digits after the decimal point.

[0162] Finally, the output of the soft weight enhancement strategy is obtained through the following formula:

[0163]

[0164] Here This is the final prediction result.

[0165] The effectiveness of the embodiments of the present invention will be verified through the following verification experiments:

[0166] The experiment used actual production data from the 1780 hot rolling production line of a steel group for verification. A total of 8,546 strip steel data were collected during the verification process, and 7,745 strip steel data were retained after data preprocessing.

[0167] Because the original data sets have inconsistent dimensions, and to ensure the method provided in this embodiment of the invention has good rolling force prediction capabilities, 190 initial node features were manually selected. Equation (27) was used to normalize the mean and variance of the original features, and the data was divided into training and test sets in a 3:1 ratio. The training set contains 5808 data points, and the test set contains 1937 data points. The specific data items used for modeling are described in Table 2.

[0168] (27)

[0169] (28)

[0170] (29)

[0171] In the formula, Indicates the first strip steel in the first Monitoring information collected by each sensor Indicates the first The sample mean of each feature, Indicates the sample standard deviation. Represents the normalized i-th strip steel in the first The feature values ​​collected by each sensor.

[0172] Table 2 Modeling Data Items

[0173]

[0174] The TCN model was trained using data from the hot strip mill production line in Table 2 to obtain its optimal prediction performance. Simultaneously, the proposed feature subset partitioning and feature prediction result fusion method, as well as the prediction effect of incorporating physical information, were verified. This verification process compared the evaluation index results of the MF-TCN method (i.e., using the same method as the embodiments of this invention, but the prediction model only uses the TCN model and does not include a physical information neural network incorporating physical information), the PI-TCN method (i.e., using a common method, excluding the overlap between feature subset partitioning and prediction results, and using the PI-TCN model proposed in the embodiments of this invention during the prediction process), the PIMF-TCN method (i.e., the method proposed in the embodiments of this invention), and existing comparative models in the rolling force prediction of passes F2 and F3. Figure 6 , Figure 7 It is evident that in passage F2, the PI-TCN method outperforms the MF-TCN method across all metrics; however, in passage F3, the PI-TCN method's metrics are slightly lower than MF-TCN. Nevertheless, both methods achieve significant improvements compared to the basic TCN model. Furthermore, by simultaneously introducing feature subset partitioning and feature prediction result fusion with physical information embedding, the PIMF-TCN method achieves even better results in RMSE, MAE, MAPE, and... It performed best across all evaluation metrics. This indicates that there is no conflict between feature subset partitioning and feature prediction result fusion and physical information embedding; rather, they complement each other, further improving the model's prediction accuracy and stability.

[0175] This invention also provides a material rolling force prediction system, which is used to perform the material rolling force prediction method as described above.

[0176] This invention also provides a computer-readable storage medium storing a computer program for predicting material rolling forces, wherein the computer program causes a computer to execute the material rolling force prediction method as described above.

[0177] Embodiments of the present invention also provide an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the prediction of material rolling forces as described above.

[0178] In summary, compared with existing technologies, it has the following beneficial effects:

[0179] This invention employs a Transformer-Channel Network (TCN) to capture long-term dependencies between features in rolling force prediction. It utilizes a Physical Information Neural Network (PIN) to construct physical information equations, combining physical information with the TCN to achieve a fusion of data-driven and physical-driven rolling force prediction. This enhances the model's adaptability to non-stationary conditions and dynamic disturbances, as well as its prediction accuracy. Furthermore, this invention innovatively proposes a data partitioning and integrated modeling strategy that combines feature subset partitioning with feature prediction result fusion. By establishing a feature interaction compensation mechanism, it achieves systematic modeling of the rolling force generation mechanism, effectively characterizing the hierarchical features of its intrinsic mechanism and laying the foundation for improving the model's prediction accuracy.

[0180] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0181] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the rolling force of a material, characterized in that, include: Acquire multi-source feature data of the hot rolling and finishing process of the target material strip. Based on contact arc length and roll wear, the data of M features in the multi-source feature data are divided into several data subsets; The PI-TCN model is pre-trained to process several subsets of data for each feature, and several subset prediction results are obtained. The prediction results of several subsets are added together to obtain M feature prediction results. The M feature prediction results are then fused according to a dynamic ratio to obtain the rolling force prediction result of the target material. The pre-trained PI-TCN model refers to embedding physical information through a physical information neural network during the training process of a temporal convolutional network. The step of fusing the prediction results of M features according to a dynamic ratio to obtain the rolling force prediction result of the target material includes: in, In the formula, For the first Predicted rolling force results for a single strip steel finishing mill; Indicates the first The characteristics of strip steel finishing rolling m The predicted value; This represents the total number of strip steel. ; The characteristics are represented by the contact arc length and roll wear. The total error in the prediction of rolling force; This represents the total error in the amplified rolling force prediction. The characteristics are divided into contact arc length and roll wear. Adaptive soft weighting for rolling force prediction; Indicates error The number of non-zero digits after the decimal point; This indicates that the PI-TCN model is performing well in terms of features during training. The root mean square error on; This indicates that the PI-TCN model is performing well in terms of features during training. The mean absolute error over time; This indicates that the PI-TCN model is performing well in terms of features during training. The mean absolute percentage error.

2. The method for predicting material rolling force as described in claim 1, characterized in that, The multi-source feature data includes process parameters, physical property data, and equipment status parameters.

3. The method for predicting material rolling force as described in claim 1, characterized in that, The method of dividing the M features in the multi-source feature data into several data subsets based on contact arc length and roll wear includes: The data of M features in the multi-source feature data are divided into layers according to the division interval of the contact arc length, and the disturbance level is divided according to the amount of roll wear to obtain several data subsets after each feature is divided.

4. The method for predicting material rolling force as described in any one of claims 1 to 3, characterized in that, The loss function during the training process of the pre-trained PI-TCN model includes: In the formula, This represents the loss function of the PI-TCN model. This represents the data-driven loss function. This represents the hyperparameter used to control the weights of the physical monotonicity constraint loss subfunction. Represents the physical singleness constraint loss subfunction; Considering the dimensionality issues of data loss and physical loss, a dynamic normalization principle is adopted when calculating the loss function, and exponential smoothing is used to track the magnitude changes of each loss term. and : In the formula, As a smoothing factor, Control the length of historical memory; adopt an adaptive weight allocation normalized total loss function. : in, It represents a very small positive number.

5. The method for predicting material rolling force as described in claim 4, characterized in that, The calculation method for the physical monotonicity constraint loss sub-function includes: In the formula, This represents the predicted rolling force after embedding the physical equation. This represents the rolling force predicted by the TCN neural network. Indicates the first The first in the block strip steel Each input feature , These represent the rolling temperature and strain rate, respectively. R Represents the gas constant. To activate energy, Indicates the strain rate sensitivity index. The number of strip steel blocks, Indicates the use of the first The sign factor of the physical relationship direction of each input feature. This represents the derivatives of the PI-TCN model with respect to the property parameters and stress parameters. The function only penalizes violations of physical laws, and the square term... Used to strengthen the penalty for outliers.

6. The method for predicting the rolling force of materials as described in any one of claims 1 to 3, characterized in that, The temporal convolutional network includes dilated causal convolution, in which, for the filter... For one-dimensional sequences Dilated convolution operation of element s F It is expressed as follows: in, As the expansion factor, Indicates the presence of an expansion factor Convolutional filters , Number the filter. For filter size, For convolution operators, Explain the direction of the past, and .

7. A system for predicting the rolling force of a material, characterized in that, The material rolling force prediction system is used to perform the material rolling force prediction method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, It stores a computer program for predicting the rolling force of a material, wherein the computer program causes a computer to execute the method for predicting the rolling force of a material as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors; Memory; as well as One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the prediction of material rolling forces as described in any one of claims 1 to 6.

Images