Small sample yarn tension prediction method based on physical information neural network
By constructing a small-sample yarn tension prediction method based on physical information neural networks, the problems of high experimental costs and insufficient accuracy in small-sample yarn tension prediction are solved, achieving efficient and accurate yarn tension prediction and process parameter optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-10
AI Technical Summary
Existing yarn tension prediction technologies face problems such as high experimental costs, low computational efficiency, and insufficient prediction accuracy for small samples.
A small-sample yarn tension prediction method based on physical information neural network is constructed. By establishing the mechanical equilibrium equation of yarn on the transmission path, a theoretical model of yarn tension is formed. The model is trained using a multilayer perceptron neural network, and physical constraints are introduced to form a physical information neural network prediction model. The parameters are optimized by minimizing the joint loss function.
It achieves high accuracy and stability in yarn tension prediction under small sample conditions, reduces experimental and computational costs, and improves the efficiency of yarn tension control and process parameter optimization.
Smart Images
Figure CN121835398A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of yarn tension control technology, and in particular to a small-sample yarn tension prediction method based on a physical information neural network. Background Technology
[0002] Braided composite materials are widely used in aerospace, automotive manufacturing, and shipbuilding due to their excellent mechanical properties and structural designability. In the production of braided composite materials, yarn tension is a key process parameter affecting braiding efficiency, fabric structural stability, and mechanical properties. Depending on the operating conditions, the yarn, during its transfer from the yarn supply mechanism to the braiding zone, needs to pass through multiple yarn guide components sequentially. Inevitably, friction occurs between the yarn and the surfaces of these components, causing continuous changes in yarn tension along the transmission path. Therefore, accurately describing and predicting the tension evolution along the transmission path is of significant engineering importance for optimizing tension control strategies, improving braiding efficiency, and enhancing fabric quality. Currently, research methods for yarn tension mainly include theoretical modeling, numerical simulation, and experimental testing. Traditional theoretical models are mostly based on the classical winch equation, typically assuming a constant coefficient of friction between the yarn and the contact interface. This makes it difficult to accurately reflect the nonlinear frictional behavior caused by changes in contact pressure, differences in material properties, and geometric factors under actual operating conditions. Although some improved models incorporate factors such as bending stiffness, Poisson's ratio, and elongation to enhance their applicability, the determination of these parameters still heavily relies on extensive experimental calibration, limited by factors such as long research cycles and high costs. Numerical simulation methods can reveal the mechanism of yarn tension variation to some extent, but they typically involve large computational loads and demanding hardware resources, making it difficult to meet the practical needs of rapid prediction and optimization of process parameters.
[0003] In recent years, data-driven models based on machine learning have been increasingly applied to yarn tension prediction. While this reduces the reliance on complex physical modeling, such methods often require large-scale sample data and lack inherent constraints on physical mechanisms. Under small sample conditions, they are prone to insufficient generalization ability and decreased prediction accuracy. Specifically: On the one hand, traditional yarn tension prediction methods typically rely on numerous repetitive physical experiments or high-fidelity numerical simulations, followed by data analysis to obtain the tension variation characteristics under different combinations of process parameters. In practical applications, this approach often faces problems such as long experimental cycles and high equipment and labor costs. Furthermore, complex numerical simulations require repeated solutions to nonlinear frictional contact and multi-condition coupling problems, necessitating high-configuration computer resources and significant computation time.
[0004] On the other hand, traditional data-driven prediction models, which use minimizing sample error as their sole optimization objective, are prone to misinterpreting random noise and local features as inherent patterns when the model has high degrees of freedom but a limited number of samples. This leads to a failure to learn the general patterns in the data, resulting in an overly broad solution space. Consequently, while the model performs well on the training set, its generalization performance on unknown samples significantly declines. This problem is particularly prominent in small-sample engineering applications, the root cause of which lies in the model's lack of inherent constraints on the real physical evolution mechanisms.
[0005] In summary, existing yarn tension prediction technologies still face problems such as high experimental costs, low computational efficiency, and insufficient prediction accuracy with small sample sizes. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a small-sample yarn tension prediction method based on a physical information neural network, which solves the technical problems of high experimental costs, low computational efficiency, and insufficient prediction accuracy in small samples in existing yarn tension prediction technologies.
[0007] The technical solution adopted in this invention is as follows: This invention provides a small-sample yarn tension prediction method based on a physical information neural network, comprising: Based on the force balance relationship of the yarn as it passes around the cylindrical guide roller along the transmission path, the mechanical balance equation of yarn tension is constructed, resulting in a theoretical model of yarn tension:
[0008] in, The output tension of the yarn after it passes over the cylindrical guide roller; The initial tension of the yarn before it passes over the cylindrical guide roller; The contact wrap angle between the yarn and the cylindrical guide roller; The radius of the cylindrical guide roller; and All parameters are power-law parameters determined by the yarn and cylindrical guide roller materials; Within the preset process parameters, the output tension of the test yarn under different initial tension, contact wrap angle and cylindrical guide roller radius conditions is used to form a small sample experimental dataset; From the small sample experimental dataset, the output tension corresponding to all contact wrap angles and cylindrical guide roller radii at a certain initial tension value is selected as the fitting dataset. The power law parameters are obtained by fitting a nonlinear curve using the fitting dataset. and The specific value; The initial tension, contact wrap angle, cylindrical guide roller radius, and power-law parameters in the small sample experimental data set are used as examples. and The specific values are inputs, and the corresponding output tension is output. A multilayer perceptron neural network is trained, and the yarn tension theoretical model is introduced into the neural network training process as a physical constraint to form a physical information neural network prediction model. During training, the neural network parameters are iteratively updated using the backpropagation algorithm by minimizing the joint loss function, which includes a data loss term and a physical constraint loss term. The data loss term is used to measure the deviation between the predicted value of the prediction model and the experimental measurement value. The physical constraint loss term is used to constrain the output of the prediction model by comparing the deviation between the predicted value of the prediction model and the calculated value of the yarn tension theoretical model, so that it conforms to physical laws. The trained prediction model is used to quickly predict the output tension of the yarn after it passes over the cylindrical guide roller.
[0009] The preferred technical solution is: The calculation of the joint loss function includes: , In the above formula, For the joint loss function, , These are the data loss term and the physical constraint loss term, respectively. , These are the corresponding weights; in: ,
[0010] In the above two equations, , These represent the output tension predicted by the neural network and the output tension tested in experiments, respectively. This represents the total number of samples in the training dataset. For the first Index of each sample.
[0011] in, , Both are 0.5.
[0012] The construction of the small sample experimental dataset includes: Set A initial tension value, B contact wrap angle value, and C cylindrical guide roller radius value to form A×B×C parameter combinations; An experimental platform was built, and a high-precision yarn tension meter was used to measure the output tension of the yarn after it passed over the guide roller under various parameter combinations. The experimental data were organized to form a small sample experimental dataset containing the initial yarn tension, contact wrap angle, guide roller radius, and corresponding output tension.
[0013] A, B, and C are all no greater than 5.
[0014] The initial tension value is set within the range of 1.5N to 2.5N; the contact wrap angle value is set within the range of 90° to 150°; and the cylindrical guide roller radius value is set within the range of 0.005m to 0.015m.
[0015] The multilayer perceptron neural network comprises one input layer, three hidden layers, and one output layer.
[0016] The number of neurons in the three hidden layers decreases in the order of 60, 50, and 40, and the activation function is uniformly the hyperbolic tangent function.
[0017] The yarn tension theoretical model adopts a nonlinear friction power law relationship. Describe the friction between the yarn and the surface of the cylindrical guide roller, where, The friction force experienced by the yarn The normal load on the yarn.
[0018] Based on the mechanical equilibrium equation of the yarn tension, a micro-element force analysis is performed on the contact arc segment between the yarn and the cylindrical guide roller, and the contact wrap angle is used as the basis for this analysis. By integrating the yarn tension as the integrand, an analytical expression for the yarn tension along the transmission path is obtained, thus yielding the theoretical model of the yarn tension.
[0019] The technical solution of the present invention can achieve at least some of the following beneficial effects: The predictive model of this invention relies on only a small amount of experimental sample data, yet it can still accurately characterize the influence of initial tension, contact wrap angle, and guide roller geometry on yarn tension evolution. It can accurately predict the tension change of yarn before and after it passes over the surface of the cylindrical guide roller along the transmission path. In terms of prediction accuracy and generalization performance, it outperforms traditional neural network models, providing a reliable technical means for yarn tension control and textile process parameter optimization. Specifically, this is demonstrated as follows: On the one hand, the prediction model of this invention incorporates physical constraints on yarn tension evolution during model construction. The mechanical equilibrium and friction transmission laws followed by the yarn before and after passing around the cylindrical guide roller during transmission are embedded into the model's training framework as part of the loss function. This ensures that the model follows fundamental physical laws such as force balance and friction laws during both parameter learning and output prediction stages. Each parameter in the model and the final prediction result can be traced back to specific physical meanings, and their changing trends highly match the theoretical model of yarn tension evolution, thus making the model interpretable. Furthermore, the introduction of physical constraints guides the model to prioritize learning mapping relationships consistent with physical laws, thereby reducing the model's sensitivity to random noise and local sample features. Especially when the number of samples is limited, physical constraints prevent the model from relying solely on empirical fitting of a limited number of samples, instead simultaneously satisfying the physical control equations and force balance relationships. Therefore, physical constraints effectively avoid the overfitting problem that traditional pure data-driven models are prone to under small sample conditions, significantly improving the model's prediction accuracy. Because the evolution of yarn tension always follows the established physical laws, even under unknown working conditions or parameter combinations not covered by the training data, the model's predictions are still constrained by the physical mechanism and will not deviate from reasonable physical change laws. This results in more stable predictive performance, allowing the model to still exhibit good predictive accuracy and physical consistency under new samples not used in training—that is, the model's generalization ability. The model's predictions are highly consistent with experimental test results, demonstrating strong predictive stability and reliability, and possessing good engineering application value.
[0020] On the other hand, the prediction model constructed in this invention, through joint learning of a small amount of experimental data and the physical laws governing yarn tension evolution, maps the computational process, which originally required repeated experiments and numerical solutions, into a rapid model inference process. For any given combination of process parameters, the corresponding yarn tension prediction result can be obtained simply by using the surrogate model for rapid calculation. This invention can serve as an equivalent alternative to large-scale repeated experimental testing of yarn tension and complex finite element simulation, significantly reducing experimental resource consumption and computational costs, shortening the time cycle of yarn tension research, and providing an efficient and low-cost approach for yarn tension characteristic analysis and comparative studies of different process schemes.
[0021] Furthermore, this invention can be flexibly applied to yarns of different material types, such as carbon fiber and glass fiber, providing reliable technical support for the design of yarn tension control systems, the optimization of textile process parameters, and the improvement of the forming quality of braided composite materials. This invention is suitable for the initial tension control of high-volume yarns, has a wide range of applications, and is highly versatile.
[0022] Other features and advantages of the invention will be set forth in the following description or may be learned by practicing the invention. Attached Figure Description
[0023] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.
[0024] Figure 2 This is a schematic diagram of the force analysis of the yarn tension theoretical model in an embodiment of the present invention.
[0025] Figure 3 This is a graph showing the result of fitting the power-law parameters to the dataset using a parameter fitting dataset according to an embodiment of the present invention.
[0026] Figure 4 This is a diagram of the physical information neural network prediction model architecture according to an embodiment of the present invention.
[0027] Figure 5 This is a graph showing the loss function results during the training process of the physical information neural network prediction model in an embodiment of the present invention.
[0028] Figure 6 This is a schematic diagram illustrating the application of the yarn tension prediction proxy model in an embodiment of the present invention. Detailed Implementation
[0029] The specific embodiments of the present invention are described below with reference to the accompanying drawings.
[0030] See Figure 1 This embodiment provides a small-sample yarn tension prediction method based on a physical information neural network, including: S1. Based on the force balance relationship when the yarn passes around the cylindrical guide roller on the transmission path, the mechanical balance equation of yarn tension is constructed, and the theoretical model of yarn tension is obtained:
[0031] in, The output tension of the yarn after it passes over the cylindrical guide roller; The initial tension of the yarn before it passes over the cylindrical guide roller; The contact wrap angle is determined by the arc between the yarn and the cylindrical guide roller; The radius of the cylindrical guide roller; and All parameters are power-law parameters determined by the yarn and cylindrical guide roller materials.
[0032] Specifically, the construction of the yarn tension theoretical model, taking carbon fiber yarn as an example, adopts a nonlinear friction power law relationship. Describe the friction between the yarn and the surface of the cylindrical guide roller. The friction force experienced by the yarn The normal load on the yarn. and That is, the power-law parameters determined by the yarn and cylindrical guide roller materials mentioned above; for the same yarn and cylindrical guide roller materials, the power-law parameters are fixed constants; establishing Figure 2 The diagram shown illustrates the force analysis of the theoretical model for yarn tension. Since the yarn is a continuous geometric body, a micro-element yarn segment is selected on the arc segment where the yarn contacts the guide roller. By performing a force analysis and based on the force balance relationship when the yarn passes over the cylindrical guide roller, the mechanical balance equation of the yarn tension can be obtained:
[0033] in, This refers to the normal reaction force generated between the yarn and the guide roller. The tangential friction generated between the yarn and the guide roller is effective for micro-yarn segments. The contact wrap angle of the yarn increased from 0 to Assuming ,but Therefore, we can obtain ,because Small enough to be considered , Furthermore, we can conclude that:
[0034]
[0035] Ignore second-order infinitesimals and assume The normal support force per unit length of yarn. Let be the frictional force per unit length of yarn, then
[0036]
[0037]
[0038] Therefore, we can conclude that:
[0039]
[0040] Solve the above equilibrium equations simultaneously, and combine them with the arc length formula. By simplification, we can obtain:
[0041] Further substituting the reaction force density of the yarn on the circular cross-section: And by integration, the theoretical model of yarn tension can be obtained:
[0042] The yarn types applicable to this embodiment include various textile yarns such as carbon fiber and glass fiber.
[0043] S2. Within the preset process parameter range, test the output tension of the yarn under different initial tension, contact wrap angle and cylindrical guide roller radius conditions to form a small sample experimental dataset.
[0044] Specifically, the construction of a small-sample experimental dataset includes: Set A initial tension value, B contact wrap angle value, and C cylindrical guide roller radius value to form A×B×C parameter combinations; An experimental platform was built, and a high-precision yarn tension meter was used to measure the output tension of the yarn after it passed over the guide roller under various parameter combinations. The experimental data were organized to form a small sample experimental dataset containing the initial yarn tension, contact wrap angle, guide roller radius, and corresponding output tension.
[0045] Among them, A, B, and C are preferably no greater than 5.
[0046] To cover typical weaving conditions, this embodiment presets the initial tension to 1.5N, 2.0N, and 2.5N within a reasonable range of process parameters. The contact wrap angle between the yarn and the cylindrical guide roller is set to 92.3°, 103.5°, 116.1°, 125.35°, and 134.8°, respectively. The radius of the cylindrical guide roller is set to 0.005m, 0.0075m, 0.01m, 0.0125m, and 0.015m, respectively. A total of 75 sets of experimental data are generated, forming a small sample experimental dataset.
[0047] S3. Select all contact wrap angles and output tensions corresponding to a certain initial tension value and the radius of the cylindrical guide roller from the small sample experimental dataset as the fitting dataset, and obtain the power law parameters by fitting a nonlinear curve using the fitting dataset. and The specific value.
[0048] Since the power-law parameters are mainly determined by the yarn material and the surface characteristics of the guide roller, and can be considered constants under the same material conditions, this embodiment uses experimental data under a fixed initial tension (e.g., 2.5N) as the fitting dataset for parameter fitting. Specifically, based on the curve fitter in MATLAB software, the contact angle under the corresponding working condition in the parameter fitting dataset is used. The radius of the cylindrical guide roller Initial tension of the yarn before it passes over the cylindrical guide roller and the output tension of the yarn after it passes over the cylindrical guide roller Power-law parameters were obtained by curve fitting using the yarn tension theoretical model. and For specific values and fitting results, please refer to [link / reference]. Figure 3 As shown.
[0049] S4. Using the small sample experimental dataset and the power law parameters... and The specific values are used as the training set, with the initial tension of the yarn. Contact corner Cylindrical guide roller radius The power-law parameter and The specific value is input, with the corresponding output tension. For output, a multilayer perceptron neural network is trained, and the yarn tension theoretical model is introduced into the neural network training process as a physical constraint to form a physical information neural network prediction model. During training, model parameters are optimized and the prediction model is constructed by minimizing the joint loss function, which includes a data loss term and a physical constraint loss term. The data loss term is used to measure the deviation between the predicted value of the prediction model and the experimental measurement value. The physical constraint loss term is used to constrain the output of the prediction model by comparing the deviation between the predicted value of the prediction model and the calculated value of the yarn tension theoretical model, so that it conforms to physical laws.
[0050] The architecture of the physical information neural network prediction model can be found in [link to relevant documentation]. Figure 4 .
[0051] Specifically, the calculation of the joint loss function includes: , In the above formula, For the joint loss function, , These are the data loss term and the physical constraint loss term, respectively. , These are the corresponding weights; in: ,
[0052] In the above two equations, , These represent the output tension predicted by the neural network and the output tension tested in experiments, respectively. This represents the total number of samples in the training dataset. For the first Index of each sample.
[0053] As a preferred approach, to balance data-driven loss and physical constraint loss, this embodiment will , All values are set to 0.5, thus balancing prediction accuracy and physical consistency under small sample conditions.
[0054] As a preferred embodiment, the multilayer perceptron neural network comprises one input layer, three hidden layers, and one output layer. The number of neurons in the three hidden layers decreases in the order of 60, 50, and 40, and the activation function is uniformly a hyperbolic tangent function to enhance the network's ability to fit nonlinear relationships.
[0055] As a preferred approach, in terms of model training parameter settings, this embodiment sets the maximum number of training rounds to 500, and updates parameters using mini-batch sampling in each training iteration, with a batch size of 1 to enhance the model's learning stability and generalization ability under small sample conditions; the learning rate is set to 5×10. -5 To avoid loss function oscillations or divergence during training, a smaller learning step size is used. Model parameter updates employ an optimization algorithm based on adaptive adjustment of first and second moments, with momentum decay coefficients set to 0.9 and 0.999 respectively, thereby improving training stability while ensuring convergence speed. Regarding dataset partitioning, this embodiment randomly shuffles the original experimental dataset and divides it proportionally into training, validation, and test sets. The training set accounts for 80% of the total data and is used for learning the parameters of the physical information neural network. Within the training set, 10% of the samples are further divided as a validation set to monitor the model's convergence during training and prevent overfitting. The remaining 20% of the samples are used as a test set to independently evaluate the model's predictive performance after training.
[0056] Specifically, before model training begins, the network input and output data are standardized to eliminate the influence of different dimensions on the training process; to ensure the repeatability of training results, the random number generation process is fixed and initialized before training begins; in the physical constraint loss calculation and prediction result output stages, the physical dimensions are restored through destandardization; and the backpropagation algorithm is used to iteratively update the neural network parameters.
[0057] Specifically, by minimizing the joint loss function mentioned above, the backpropagation algorithm is used to iteratively update the neural network parameters until the model converges and the prediction model is obtained. See the joint loss function curve during training. Figure 5 As shown in the figure, the total loss function shows a continuous decreasing trend and eventually stabilizes with the increase of iterations during the training process. Both the data loss term and the physical constraint loss term can achieve synchronous convergence without obvious oscillation or divergence. This indicates that the constructed physical information neural network has good training stability and convergence performance under small sample conditions.
[0058] S5. The trained physical information neural network prediction model is used to quickly predict the output tension of the yarn after it passes over the cylindrical guide roller.
[0059] Specifically, the trained prediction model is used as a proxy model for yarn tension prediction. In practical applications, only the initial yarn tension, contact wrap angle, radius of the cylindrical guide roller, and corresponding power law parameters need to be input to quickly predict the output tension of the yarn after it passes around the guide roller, thereby obtaining the tension change before and after passing around the roller.
[0060] To verify the accuracy of the prediction model in this embodiment, a prediction dataset was constructed to validate the application of the prediction model. Experimental tests were conducted on the yarn output tension under the process parameter conditions included in the prediction dataset for comparison with the prediction results. Specifically, this included: A yarn tension prediction dataset was constructed using the 75 small sample experimental datasets obtained in step S2 and the determined power law parameter values. In the yarn tension prediction dataset, the initial tension is set to a fixed value of 1.0N; The prediction model was used to predict and train the yarn output tension under different contact wrap angles and different combinations of cylindrical guide roller radii under an initial tension of 1.0N, and the results were compared with experimental results.
[0061] The determination coefficient of the model trained on the prediction dataset ( The mean square error (MSE) was 0.997, the root mean square error (RMSE) was 0.000363, the mean square error (RMSE) was 0.019503, the mean absolute error (MAE) was 0.014641, and the mean absolute percentage error (MAPE) was 1.263044%.
[0062] The comparison between the predicted values of the prediction model and the experimental results can be found in [link to relevant documentation]. Figure 6 It can be seen that the maximum deviation between the predicted yarn output tension and the corresponding experimental results is only 0.007N, and the prediction accuracy of the model is as high as 99.5%.
[0063] To verify the effectiveness of physical constraints, a comparative example is set up—the weights of the physical constraint loss term in the joint loss function are adjusted. Set to 0. At this point, the predictive agent model degenerates into a traditional multilayer perceptron neural network. Then, it is trained using the same 75 sets of data. The results show that the determination coefficient of the trained model (…) The mean squared error (MSE) was 0.970, the root mean squared error (RMSE) was 0.002676, the mean squared error (RMSE) was 0.051730, the mean absolute error (MAE) was 0.040873, and the mean absolute percentage error (MAPE) was 3.321517%. The prediction results of the multilayer perceptron neural network were compared with the corresponding experimental results; the maximum deviation between the two was 0.066N, and the model's prediction accuracy was 90.1%.
[0064] Comparative verification shows that the prediction model constructed in this embodiment does not require repeated extensive experimental testing or complex numerical simulations. It can achieve high-precision prediction of yarn tension under small sample conditions, significantly reducing experimental and computational costs. Simultaneously, by introducing explicit physical constraints into the model, it effectively avoids the overfitting problem that traditional pure data-driven models are prone to under small sample conditions, improving the stability and engineering applicability of the prediction results. This invention can accurately predict the output tension of yarn after winding around the roller under different combinations of process parameters in a very short time, and can be used for yarn tension control system design, process parameter optimization, and stability assessment of the weaving process.
[0065] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting yarn tension in small samples based on a physical information neural network, characterized in that, include: Based on the force balance relationship of the yarn as it passes around the cylindrical guide roller along the transmission path, the mechanical balance equation of yarn tension is constructed, resulting in a theoretical model of yarn tension: ， in, The output tension of the yarn after it passes over the cylindrical guide roller; The initial tension of the yarn before it passes over the cylindrical guide roller; The contact wrap angle between the yarn and the cylindrical guide roller; The radius of the cylindrical guide roller; and All parameters are power-law parameters determined by the yarn and cylindrical guide roller materials; Within the preset process parameters, the output tension of the test yarn under different initial tension, contact wrap angle and cylindrical guide roller radius conditions is used to form a small sample experimental dataset; From the small sample experimental dataset, the output tension corresponding to all contact wrap angles and cylindrical guide roller radii at a certain initial tension value is selected as the fitting dataset. The power law parameters are obtained by fitting a nonlinear curve using the fitting dataset. and The specific value; The initial tension, contact wrap angle, cylindrical guide roller radius, and power-law parameters in the small sample experimental data set are used as examples. and The specific values are inputs, and the corresponding output tension is output. A multilayer perceptron neural network is trained, and the yarn tension theoretical model is introduced into the neural network training process as a physical constraint to form a physical information neural network prediction model. During training, the neural network parameters are iteratively updated using the backpropagation algorithm by minimizing the joint loss function, which includes a data loss term and a physical constraint loss term. The data loss term is used to measure the deviation between the predicted value of the prediction model and the experimental measurement value. The physical constraint loss term is used to constrain the output of the prediction model by comparing the deviation between the predicted value of the prediction model and the calculated value of the yarn tension theoretical model, so that it conforms to physical laws. The trained prediction model is used to quickly predict the output tension of the yarn after it passes over the cylindrical guide roller.
2. The method according to claim 1, characterized in that, The calculation of the joint loss function includes: ， In the above formula, For the joint loss function, , These are the data loss term and the physical constraint loss term, respectively. , These are the corresponding weights; in: ， ， In the above two equations, , These represent the output tension predicted by the neural network and the output tension tested in experiments, respectively. This represents the total number of samples in the training dataset. For the first Index of each sample.
3. The method according to claim 2, characterized in that, , Both are 0.
5.
4. The method according to claim 1, characterized in that, The construction of the small sample experimental dataset includes: Set A initial tension value, B contact wrap angle value, and C cylindrical guide roller radius value to form A×B×C parameter combinations; An experimental platform was built, and a high-precision yarn tension meter was used to measure the output tension of the yarn after it passed over the guide roller under various parameter combinations. The experimental data were organized to form a small sample experimental dataset containing the initial yarn tension, contact wrap angle, guide roller radius, and corresponding output tension.
5. The method according to claim 4, characterized in that, A, B, and C are all no greater than 5.
6. The method according to claim 4, characterized in that, The initial tension value is set within the range of 1.5N to 2.5N; the contact wrap angle value is set within the range of 90° to 150°; and the cylindrical guide roller radius value is set within the range of 0.005m to 0.015m.
7. The method according to claim 1, characterized in that, The multilayer perceptron neural network comprises one input layer, three hidden layers, and one output layer.
8. The method according to claim 7, characterized in that, The number of neurons in the three hidden layers decreases in the order of 60, 50, and 40, and the activation function is uniformly the hyperbolic tangent function.
9. The method according to claim 1, characterized in that, The yarn tension theoretical model adopts a nonlinear friction power law relationship. Describe the friction between the yarn and the surface of the cylindrical guide roller, where, The friction force experienced by the yarn The normal load on the yarn.
10. The method according to claim 1, characterized in that, Based on the mechanical equilibrium equation of the yarn tension, a micro-element force analysis is performed on the contact arc segment between the yarn and the cylindrical guide roller, and the contact wrap angle is used as the basis for this analysis. By integrating the yarn tension as the integrand, an analytical expression for the yarn tension along the transmission path is obtained, thus yielding the theoretical model of the yarn tension.