Rapid prediction method for stress field of bending protection device based on data driving

Abstract

The invention provides a bending protection device stress field rapid prediction method based on data driving, and belongs to the crossing field of engineering structure health monitoring and artificial intelligence. The method comprises the following steps: firstly, performing parametric modeling on a bending protection device to obtain a training data set; secondly, building a deep learning neural network, training the neural network by using the training data set, completing the rapid prediction of the overall stress field through the bending inclination angle of the specific position of the bending protection device, and obtaining a stress field prediction result; and finally, deploying the trained neural network in an actual engineering environment, realizing rapid prediction of a stress field of the bending protection device, drawing of a stress nephogram and automatic marking of a dangerous area, and realizing full-time-domain and full-space-domain real-time monitoring of the bending protection device. According to the method, the deep learning method is used for rapid prediction of the stress field of the bending protection device, the efficiency is high, the accuracy is high, the robustness is good, the real-time monitoring requirement can be met, and the engineering applicability is high.

Description

Technical Field

[0001] The present invention belongs to the cross - field of engineering structure health monitoring and artificial intelligence, and relates to a method for rapidly predicting the stress field of a bending protection device based on data - driven, specifically to a general framework for rapidly predicting the stress field of a bending protection device by integrating finite - element simulation and deep learning. It is particularly applicable to the real - time stress monitoring of various bending protection structures such as oil and gas pipeline benders, subsea umbilical cable bending limiters, and dynamic riser protection devices, and realizes the full - process technology from data generation, model training to real - time monitoring and data post - processing. Background Art

[0002] The bending protection device prevents structural failure problems such as buckling fracture and fatigue damage of the pipeline caused by excessive bending by restricting the bending curvature of the pipeline, dispersing local concentrated stress, and adapting to controllable deformation, ensuring the safe operation and life extension of the marine flexible pipeline system under complex loads. Its structural safety directly affects the service reliability of major equipment such as oil and gas transmission pipelines and offshore platforms. Therefore, rapidly obtaining the stress field of the bending protection device is an important technical means for structural safety analysis.

[0003] At present, scholars at home and abroad have conducted extensive research on the health monitoring of marine flexible pipelines [Ma Guoming, Qin Weiqi, Wang Sihan, et al. Research Review on Distributed Optical Fiber Monitoring Technology for Submarine Cable Status [J]. Proceedings of the CSEE, 2025, 45(01): 370-388. DOI: 10.13334 / j.0258-8013.pcsee.240013], the application design of bending protection devices [Zhu Pengcheng, Yang Zhixun, Yan Jun, et al. New Design Research on Marine Flexible Cable Bending Limiters Based on Topology Optimization Method [J]. Journal of Ship Mechanics, 2023, 27(03): 415-426], and mechanical analysis [Ding Lesheng, Chen Jinlong, Chen Xiao, et al. Shear-Bending Stiffness Analysis of Bending Limiters Based on Semi-Numerical Method [J]. Ocean Engineering, 2022, 40(06): 152-159. DOI: 10.16483 / j.issn.1005-9865.2022.06.016]. However, the health monitoring of bending protection devices is almost blank, which is due to three core defects in existing monitoring technologies: the traditional finite element method [Dong Wulei, Yang Huayong, Guo Zhaoyang, et al. Finite Element Analysis and Experimental Verification of Two Submarine Cable Bending Limiters Based on Material Nonlinearity [J]. Journal of Ocean Technology, 2019, 38(06): 89-94], [Sun Kai, Yue Qianjin, Yan Jun, et al. Finite Element Analysis of Marine Flexible Riser Bending Preventer Based on Material Nonlinearity [J]. Computer Aided Engineering, 2014, 23(06): 66-69. DOI: 10.13340 / j.cae.2014.06.014] requires several hours for a single calculation, which cannot meet the requirements of real-time monitoring; strain gauge measurement can only obtain discrete point data, resulting in stress monitoring blind spots; the existing data-driven methods have insufficient adaptability to complex working conditions [Yu Qianxiang, Li Qing, Li Linlin, et al. Review of Data-Driven Fault Diagnosis Methods for Large-Scale Industrial Production Processes [J]. Chinese Journal of Engineering, 2025, 47(04): 780-793. DOI: 10.13374 / j.issn2095-9389.2024.05.24.002]. When the sensor layout changes or some sensors fail, re-modeling is required.

[0004] In summary, the existing methods are difficult to achieve real-time monitoring of bending protection devices in the full time domain and full space domain. There is an urgent need for a method for rapid prediction of the stress field of bending protection devices. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention proposes a general framework for rapid prediction of the stress field of bending protection devices. Specifically, it provides a data-driven method for rapid prediction of the stress field of bending protection devices, which can realize real-time monitoring of the stress of bending protection devices, reduce the number of measuring points arranged, and save time costs.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A data-driven rapid prediction method for the stress field of a bending protection device. The rapid prediction method for the stress field of the bending protection device first performs parametric modeling on the bending protection device to obtain a training data set; secondly, builds a deep learning neural network and trains the neural network using the training data set; finally, deploys the trained neural network into the actual engineering environment to achieve rapid prediction of the stress field of the bending protection device, drawing of stress nephograms, and automatic annotation of dangerous areas. Specifically, it includes the following steps:

[0008] S1. Parametric modeling module: Establish a finite element model of the bending protection device to obtain a training data set. Specifically:

[0009] Simplify the solid model of the bending protection device to establish a corresponding initial Abaqus finite element analysis model, that is, construct the finite element model of the bending protection device, perform mesh division on it, and set the boundary conditions, material parameters, and loads of the finite element model according to the actual engineering, type of effective load, and geometric constraint requirements.

[0010] Use the Latin hypercube sampling method to obtain a large number of working condition samples to avoid generating invalid samples. Through parametric modeling, write a Python program to calculate the full-field stress of the bending protection device and the bending inclination angle at specific positions under different working conditions, and construct a training data set. The stress is the Mises stress, and the specific position is the position where the bending inclination angle sensor is embedded in the bending protection device in actual engineering. The bending inclination angle is obtained by using the Lagrange interpolation method for adjacent nodes at specific positions in the deformed finite element model.

[0011] The training data set includes two parts. One is a vector composed of the bending inclination angles at specific positions of the bending protection device, and the other is a three-dimensional matrix composed of the full-field stress of the bending protection device;

[0012] S2. Deep learning module: Build a deep learning neural network and train it to achieve rapid prediction of the overall stress field through the bending inclination angle at specific positions of the bending protection device. The deep learning neural network is a deconvolution neural network CNN, and its network architecture includes an input layer, a preprocessing layer, a custom loss function, a feature extraction layer, and a postprocessing layer. Specifically as follows:

[0013] S21. Input layer: Transform the bending inclination angle vector in the S1 training data set into a two-dimensional matrix suitable for deconvolution operations through a linear layer and shape reconstruction as the input data for training the neural network;

[0014] S22, Pre-processing Layer: Using the 3D discrete cosine transform method, the 3D stress matrix in the S1 training dataset is converted to the frequency domain. Low-frequency coefficients with values significantly greater than zero are extracted through a linear 3D zigzag scan and stored in a 2D matrix. The index mapping between the matrix before and after the zigzag scan is stored in a table. This reduces the dimensionality of the output data and improves noise immunity. A 2D matrix consisting of the low-frequency coefficients of the full-field stress of the bending protection device is obtained, which serves as the target data for training the neural network.

[0015] The three-dimensional discrete cosine transform formula is as follows:

[0016]

[0017]

[0018] Among them, A is the three-dimensional matrix composed of the full-field stress of the bending protection device, A m,n,p is an element, m represents element A m,n,p The first dimension index in matrix A, n represents the element A m,n,p The second dimension index in matrix A, p represents the element A m,n,p The index of the third dimension in matrix A; M is the size of the first dimension of matrix A, N is the size of the second dimension of matrix A, and P is the size of the third dimension of matrix A; It is a three-dimensional matrix composed of the frequency domain coefficients corresponding to the matrix A. is an element, u represents an element In the matrix The first dimension index in, v represents the element In the matrix The second dimension index in , w represents the element In the matrix The third dimension index in E u ,E v ,E w is the weight factor,

[0019] S23. Custom loss function: Construct a weight matrix through the maximum and minimum normalization method, design a weighted loss function, and focus on the area with large loss values. The formula is as follows:

[0020] δ=|T′-T|

[0021]

[0022] Where T is the frequency domain coefficient matrix in S22 The two-dimensional matrix obtained after performing three-dimensional zigzag scanning, where Q is the size of the first dimension of matrix T, and R is the size of the second dimension of matrix T; T′ is the prediction result obtained after performing a deconvolution operation on the input data in S21, which will be specifically described in S24; |.| represents taking the absolute value; δ represents the absolute error between matrix T′ and matrix T, and δ q,r is an element thereof, ω represents the weight matrix obtained by performing min-max normalization on matrix δ, and ω q,r is an element thereof, q represents the element δ q,r and ω q,r in the first dimension index in matrices δ and ω, r represents the element δ q,r and ω q,r in the second dimension index in matrices δ and ω; η1 and η2 are control parameters. By decreasing η1 and increasing η2, the attention degree of the loss function to the high-loss value region can be improved; Loss represents the loss value between T′ and T calculated by the custom loss function.

[0023] S24. Feature extraction layer: Transform the input data matrix in S21 into the shape of the target data matrix in S22 through a certain number of deconvolution layers to obtain the low-frequency coefficient prediction result. At the same time, add activation functions, normalization, and regularization to introduce non-linearity and prevent overfitting. Calculate the loss value between the low-frequency coefficient prediction result and the target data in S22 using the custom loss function in S23 and perform backpropagation. Update the hyperparameters in the deconvolution layer through the Adam optimizer. Repeat the deconvolution and loss value backpropagation operations until the loss value converges to complete the training process;

[0024] S25. Post-processing layer: Perform three-dimensional inverse zigzag scanning on the low-frequency coefficient prediction result of the neural network completed in S24, and fill it in reverse order into a three-dimensional empty matrix with the same size as the stress three-dimensional matrix in S1 according to the order in the index mapping relationship list in S22, fill the remaining positions with zeros, and then perform three-dimensional inverse discrete cosine transform to obtain the stress field prediction result as the output data of the neural network;

[0025] The formula for three-dimensional inverse discrete cosine transform is as follows:

[0026]

[0027] where A is a three-dimensional matrix composed of the full-field stress of the bending protection device, and A m,n,p is an element thereof; is a three-dimensional matrix composed of the frequency domain coefficients corresponding to matrix A, is an element thereof.

[0028] S3. Scenario Deployment Module: Embed an inclination sensor at the specific position described in S1 of the bending protection device entity, input the bending inclination measured by the inclination sensor at a small number of actual measurement points into the neural network trained in S2, and obtain the prediction result of the stress field of the bending protection device. At the same time, import the finite element model and the stress field prediction result in S1 into Paraview, draw a real-time stress nephogram, automatically highlight the dangerous area where the stress is greater than 90% of the maximum stress value, and automatically alarm when the maximum stress value exceeds the allowable stress of the device, achieving real-time monitoring of the bending protection device in the full time domain and full space domain.

[0029] The beneficial effects of the present invention are as follows:

[0030] (1) The present invention applies the deep learning method to the rapid prediction of the stress field of the bending protection device. Compared with the traditional finite element method, this method only uses data-driven and does not require a large number of iterations for numerical solution of the mechanical equations, greatly reducing the calculation time and cost, which can reach 40 milliseconds at the fastest, meeting the requirements of real-time monitoring.

[0031] (2) The present invention uses three-dimensional discrete cosine transform for data preprocessing, only retains the low-frequency coefficients, filters out the high-frequency noise, enhances the anti-noise ability of the deep learning model, reduces the dimension of the training data, so that the training parameters are reduced by 99.5% compared with the case where this method is not used, achieving lightweight design. At the same time, it supports any sensor layout and can still maintain an accuracy of more than 85% when 20% of the sensors fail, with good robustness and strong engineering applicability.

[0032] (3) The present invention uses a custom loss function with weights, making the training of the neural network pay more attention to the areas with larger loss values, which can effectively reduce the maximum loss value and still maintain an accuracy of more than 90% in the high-gradient area of the structural stress, with high accuracy. Description of the Drawings

[0033] Figure 1 is the design flow of an embodiment of the present invention;

[0034] Figure 2 is the finite element model of an embodiment of the present invention;

[0035] Figure 3 is the material stress-strain curve of an embodiment of the present invention;

[0036] Figure 4 is the assembled finite element model of an embodiment of the present invention

[0037] Figure 5 is the schematic diagram of the neural network model of an embodiment of the present invention;

[0038] Figure 6Schematic diagram of three-dimensional zigzag scanning according to an embodiment of the present invention;

[0039] Figure 7 Variation of the loss function value during the neural network training according to an embodiment of the present invention;

[0040] Figure 8 Contour map of the predicted result of the structural stress field according to an embodiment of the present invention;

[0041] Figure 9 Automatic annotation of the stress danger area according to an embodiment of the present invention. Detailed implementation manners

[0042] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0043] The present invention proposes a method for quickly predicting the stress field of a bending protection device based on data driving. In this embodiment, it is used for a marine flexible pipe cable bending limiter (hereinafter referred to as the bending limiter). The specific design process is as Figure 1 shown and specifically includes the following steps:

[0044] S1. Parametric modeling module: Establish a finite element model of the bending limiter and obtain a training data set. Specifically:

[0045] Simplify the solid model of the bending limiter, establish a corresponding initial Abaqus finite element analysis model, that is, construct a finite element model of the bending limiter, and use C3D8R elements to perform mesh division on it. As Figure 2 shown, the number of structural units of one section of the bending limiter is 19080. Set the boundary conditions, material parameters, and loads of the finite element model according to the actual engineering, effective load type, and geometric constraint requirements. Specifically, use polyurethane material, and its stress-strain curve is as Figure 3 shown. Assemble five sections of the bending limiter according to the installation requirements and set them to the locked position, as Figure 4 shown. One end is fixed and the other end is subjected to bending moment and shear force for the constraints and loads.

[0046] Use the Latin hypercube sampling method to obtain 1000 working condition samples to avoid generating invalid samples. Through parametric modeling, write a Python program to calculate the full-field stress of the bending limiter and the bending inclination angle at a specific position under different working conditions, and construct a training data set. The stress is the Mises stress, and the specific position is the position where an inclination angle sensor is embedded in the center of each section of the bending limiter. The bending inclination angle is obtained by using the Lagrange interpolation method for adjacent nodes at a specific position in the deformed finite element model.

[0047] The training dataset includes two parts. One is a vector composed of the bending inclination angles at specific positions of the bending limiter, with a size of 1×5. The other is a three-dimensional matrix composed of the full-field stress of the bending limiter, with a size of 49×49×49;

[0048] S2. Deep learning module: Build a deep learning neural network and train it to achieve rapid prediction of the overall stress field through the bending inclination angles at specific positions of the bending limiter. The deep learning neural network is a deconvolution neural network CNN, as Figure 5 shown. Its network architecture includes an input layer, a preprocessing layer, a custom loss function, a feature extraction layer, and a postprocessing layer, specifically as follows:

[0049] S21. Input layer: Transform the bending inclination angle vector in the S1 training dataset into a two-dimensional matrix suitable for deconvolution operations through a linear layer and shape reconstruction, with a size of 3×3, as the input data for training the neural network;

[0050] S22. Preprocessing layer: Use the three-dimensional discrete cosine transform method to transform the stress three-dimensional matrix in the S1 training dataset into the frequency domain, and extract the low-frequency coefficients with numerically significant values greater than zero into a two-dimensional matrix with a size of 25×25 through the linear three-dimensional zigzag scan as shown in Figure 6 . At the same time, store the index mapping relationship of the matrix before and after the zigzag scan in a list. Thereby reducing the dimension of the output data and improving the anti-noise ability, and obtaining a two-dimensional matrix composed of the low-frequency coefficients of the full-field stress of the bending limiter as the target data for training the neural network.

[0051] The three-dimensional discrete cosine transform formula is as follows:

[0052]

[0053] where A is a three-dimensional matrix composed of the full-field stress of the bending limiter, and A m,n,p is an element therein; is a three-dimensional matrix composed of the frequency domain coefficients corresponding to matrix A, is an element therein; M, N, and P are all 49;

[0054] S23. Custom loss function: Construct a weight matrix through the maximum-minimum normalization method, and design a weighted loss function that focuses on the regions with larger loss values. The formula is as follows:

[0055] δ = |T′ - T|

[0056]

[0057] where T is the frequency domain coefficient matrix in S22 The two-dimensional matrix obtained after performing three-dimensional zigzag scanning; T′ is the prediction result obtained after deconvolution operation on the input data in S21; Q and R are both 25; δ represents the absolute error between matrix T′ and matrix T, and δ q,r is an element thereof; ω represents the weight matrix obtained after performing min-max normalization on matrix δ, and ω q,r is an element thereof; η1 and η2 are control parameters, η1 is 1, and η2 is 10; Loss represents the loss value between T′ and T calculated by a custom loss function.

[0058] S24. Feature extraction layer: The input data matrix in S21 is transformed into the shape of the target data matrix in S22 through four deconvolution layers, that is, the input data size is changed from 3×3 to 25×25 to obtain the low-frequency coefficient prediction result. At the same time, an activation function, normalization, and regularization are added to introduce non-linearity and prevent overfitting. The custom loss function in S23 is used to calculate the loss value between the low-frequency coefficient prediction result and the target data in S22 and backpropagate, and the hyperparameters in the deconvolution layer are updated through the Adam optimizer. The deconvolution and loss value backpropagation operations are repeated until the loss value converges to complete the training process. The change of the loss function value during the training process is as Figure 7 shown;

[0059] S25. Post-processing layer: The low-frequency coefficient prediction result of the neural network completed in training in S24 is subjected to three-dimensional inverse zigzag scanning, filled back into the three-dimensional matrix in reverse order according to the order in the index mapping relationship list in S22, and the remaining positions are filled with zeros, and then three-dimensional inverse discrete cosine transform is performed to obtain the stress field prediction result as the output data of the neural network;

[0060] The formula for three-dimensional inverse discrete cosine transform is as follows:

[0061]

[0062] where A is a three-dimensional matrix composed of the full-field stress of the bender limiter, and A m,n,p is an element thereof; is a three-dimensional matrix composed of the frequency domain coefficients corresponding to matrix A, is an element thereof.

[0063] S3. Scenario deployment module: An inclination sensor is embedded at the specific position described in S1 in the bender limiter entity, and the bending inclination measured by the inclination sensor at a small number of real measurement points is input into the neural network completed in training in S2, and the stress field prediction result of the bender limiter is obtained. The prediction time is only 40 milliseconds at the fastest, and its overall highest accuracy can reach more than 98%. The accuracy can still be maintained above 90% in the high stress gradient area. At the same time, the finite element model in S1 and the stress field prediction result are imported into Paraview to draw a real-time stress nephogram, such asFigure 8 as shown; automatically highlight the dangerous area where the stress is greater than 90% of the maximum stress, such as Figure 9 as shown; automatically alarm when the maximum stress exceeds the allowable stress of the device, achieving real-time monitoring of the bending limiter in the full time domain and full space domain. Randomly reduce the input value of one inclination sensor, and the overall prediction result still maintains an accuracy of more than 85%.

[0064] The above embodiments only express the implementation manners of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A rapid prediction method for the stress field of a bending protection device based on data driving, characterized in that, The rapid prediction method for the stress field of the bending protection device includes the following steps: S1. Parametrically model the bending protection device to obtain a training data set; The training data set includes two parts. The first part is a vector composed of the bending inclination angles at specific positions of the bending protection device, and the second part is a three-dimensional matrix composed of the full-field stress of the bending protection device; S2. Build a deep learning neural network and use the training data set to train the neural network to achieve rapid prediction of the overall stress field through the bending inclination angles at specific positions of the bending protection device, and obtain the stress field prediction result; The deep learning neural network is a deconvolution neural network CNN, and its network architecture includes an input layer, a preprocessing layer, a custom loss function, a feature extraction layer, and a postprocessing layer; S3. Deploy the trained neural network to the actual engineering environment to achieve rapid prediction of the stress field of the bending protection device, drawing of stress nephograms, and automatic marking of dangerous areas, and realize real-time monitoring of the bending protection device in the full time domain and full space domain.

2. The rapid prediction method of the stress field of a bending protection device based on data driving according to claim 1, characterized in that, Specifically, for S1: Simplify the solid model of the bending protection device, establish a corresponding initial Abaqus finite element analysis model, that is, construct the finite element model of the bending protection device, perform mesh division on it, and set the boundary conditions, material parameters, and loads of the finite element model according to the actual engineering, type of payload, and geometric constraint requirements; Use the Latin hypercube sampling method to obtain working condition samples. Through parametric modeling, calculate the full-field stress and the bending inclination angles at specific positions of the bending protection device under different working conditions, and construct a training data set; among them, the stress is the Mises stress, the specific position is the position where the inclination angle sensor is embedded in the bending protection device in actual engineering, and the bending inclination angle is obtained by using the Lagrange interpolation method for adjacent nodes at the specific position in the deformed finite element model.

3. A rapid prediction method for the stress field of a bending protection device based on data driving according to claim 1, characterized in that Specifically, for S2: S21. Input layer: Transform the bending inclination angle vector in the training data set of S1 into a two-dimensional matrix suitable for deconvolution operation through a linear layer and shape reconstruction, as the input data for training the neural network; S22. Preprocessing layer: Use the three-dimensional discrete cosine transform method to transform the stress three-dimensional matrix in the training data set of S1 into the frequency domain, and extract the low-frequency coefficients with numerically significant values greater than zero through a linear three-dimensional zigzag scan and store them in a two-dimensional matrix. At the same time, store the index mapping relationship of the matrices before and after the zigzag scan in a list to reduce the dimension of the output data and improve the anti-noise ability, and obtain a two-dimensional matrix composed of the low-frequency coefficients of the full-field stress of the bending protection device as the target data for training the neural network; S23. Custom loss function: Construct a weight matrix through the maximum-minimum normalization method and design a weighted loss function; S24. Feature extraction layer: The input data matrix in S21 is transformed into the shape of the target data matrix in S22 through a transposed convolutional layer to obtain the prediction result of the low-frequency coefficients. Meanwhile, an activation function, normalization, and regularization are added to introduce non-linearity and prevent overfitting. The custom loss function in S23 is used to calculate the loss value between the prediction result of the low-frequency coefficients and the target data in S22 and backpropagate it. The hyperparameters in the transposed convolutional layer are updated through the Adam optimizer. The operations of transposed convolution and loss value backpropagation are repeated until the loss value converges, completing the training process. S25. Post-processing layer: The prediction result of the low-frequency coefficients of the neural network trained in S24 is subjected to a three-dimensional inverse zigzag scan and filled backward into a three-dimensional empty matrix with the same size as the stress three-dimensional matrix in S1 according to the order in the index mapping relationship list in S22. The remaining positions are filled with zeros, and then a three-dimensional inverse discrete cosine transform is performed to obtain the stress field prediction result, which is used as the output data of the neural network.

4. A method for rapidly predicting the stress field of a bending protection device based on data driving according to claim 3, characterized in that Specifically in S22, the three-dimensional discrete cosine transform formula used for the three-dimensional discrete cosine transform method is as follows: Among them, A is a three-dimensional matrix composed of the full-field stress of the bending protection device, and A m,n,p is an element, where m represents the element A m,n,p in the first dimension index of matrix A, n represents the element A m,n,p in the second dimension index of matrix A, p represents the element A m,n,p in the third dimension index of matrix A; M is the size of the first dimension of matrix A, N is the size of the second dimension of matrix A, and P is the size of the third dimension of matrix A; is a three-dimensional matrix composed of the frequency domain coefficients corresponding to matrix A, is an element, where u represents the element in the first dimension index of matrix and v represents the element in the second dimension index of matrix and w represents the element in the third dimension index of matrix ; E u , E v , E w are weight factors.

5. A method for quickly predicting the stress field of a bending protection device based on data driving according to claim 4, characterized in that, The described E u , E v , E w Specifically:

6. A method for rapidly predicting the stress field of a bending protection device based on data driving according to claim 4, characterized in that The formula adopted in S23 is as follows: δ = |T′ - T| Where T is the frequency domain coefficient matrix in S22 The two-dimensional matrix obtained after the three-dimensional zigzag scan is: Q is the size of the first dimension of the matrix T, R is the size of the second dimension of the matrix T; T′ is the prediction result obtained after the deconvolution operation of the input data in S21; |.| means taking the absolute value; δ represents the absolute error between the matrix T′ and the matrix T, δ q,r is an element, ω represents the weight matrix obtained by normalizing the matrix δ to the maximum and minimum, ω q,r is an element, q represents the element δ q,r and ω q,r The first dimension index in matrices δ and ω, r represents the element δ q,r and ω q,r The second dimension index in matrices δ and ω; η1 and η2 are control parameters; Loss represents the loss value between T′ and T calculated by the custom loss function.

7. A method for rapidly predicting the stress field of a bending protection device based on data driving according to claim 3, characterized in that, In S25, the formula for the three-dimensional inverse discrete cosine transform is as follows: Among them, A is a three-dimensional matrix composed of the full-field stress of the bending protection device, and A m,n,p is an element thereof; is a three-dimensional matrix composed of the frequency domain coefficients corresponding to the matrix A, is an element thereof.

8. A method for rapidly predicting the stress field of a bending protection device based on data driving according to claim 1, characterized in that Specifically, S3 is: An inclination sensor is embedded at the specific position in S1 of the bending protection device entity. The bending inclination measured by the inclination sensor at a small number of actual measurement points is input into the neural network trained in S2 to obtain the stress field prediction result of the bending protection device. Meanwhile, according to the finite element model in S1 and the stress field prediction result in S2, a real-time stress nephogram is drawn, the dangerous area where the stress is greater than the stress threshold is automatically highlighted, and an alarm is automatically triggered when the maximum stress exceeds the allowable stress of the device, achieving real-time monitoring of the bending protection device in the full time domain and full space domain.

9. A method for rapidly predicting the stress field of a bending protection device based on data driving according to claim 8, characterized in that, The stress threshold is 90% of the maximum stress.

Images