Method for predicting force chain distribution of different graded particle sets under different strains

By establishing a mapping relationship between particle sets and force chain distribution through Pix2Pix generative adversarial network, the problems of parameter dependence and high computational complexity in existing technologies are solved, and efficient and accurate force chain particle identification and material property prediction are achieved, which has important engineering application value.

CN118981932APending Publication Date: 2024-11-19TAIYUAN UNIVERSITY OF TECHNOLOGY
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411063249.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

Existing technologies suffer from strong parameter dependence, high computational complexity, and high time cost when identifying force chain particles in particle assemblies, making it difficult to efficiently and accurately identify force chain particles under different strain states in large-scale particle assemblies.

Method used

Pix2Pix adversarial generative network is used to establish a mapping relationship between the numerical image of particle set and the force chain distribution image through the game between the generator and the discriminator. This generates a predicted image that is highly similar to the actual force chain image under different strain states, thus achieving accurate prediction of the mechanical properties of particulate materials.

Benefits of technology

It improves the accuracy and reliability of force chain particle identification, reduces data collection and labeling costs, simplifies the process, improves work efficiency, and provides a scientific basis for material design and optimization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981932A_ABST
    Figure CN118981932A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of computational mechanics, and particularly relates to a method for predicting force chain distribution of different graded particle sets under different strains. The method comprises the following steps: constructing particle set numerical samples under different grading distribution conditions by adopting a discrete element method based on a preset grading formula; mechanical properties of all the samples are tested through a biaxial compression numerical simulation test, and particle sizes and position information under different strain conditions are obtained; screening the obtained particle information, extracting force chain information, importing the data into an integrated development environment PyCharm, respectively drawing a particle set image and a force chain distribution image by using Matplotlib, and establishing a corresponding data set; and based on a Pix2Pix adversarial generative network, capturing a mapping relation between a particle set image and a force chain image with different grading characteristics. The method can accurately capture the complex mapping relation between the particle set image and the force chain image, and has high accuracy and reliability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of computational mechanics, and in particular relates to a method for predicting force chain distribution of different graded particle sets under different strains. Background Art

[0002] Granular materials are widely present in various natural environments and engineering fields, and are closely related to people's daily lives. At present, there are still a large number of problems related to granular materials that need to be solved in many fields such as geological disasters, geotechnical engineering, chemical engineering, environmental science, and transportation. As one of the most basic and important physical indicators of granular materials, particle grading greatly affects the physical and mechanical properties of granular materials. Even if the granular materials have the same composition, their strength, deformation and permeability characteristics will show great differences if the grading is different.

[0003] As the main stress carrier in the particle collection, force chain particles form a stress chain that runs through the entire material through mutual contact when the particle material is subjected to external loads, thereby realizing the transmission and distribution of stress. Compared with ordinary particles, the change in particle size of force chain particles can better reflect the change in the overall grading characteristics of the particle collection, and has a significant dominant effect on the reaction and behavior of the particle material under stress. Therefore, accurately identifying force chain particles in a particle collection is of great significance for a deep understanding and improvement of the physical and mechanical properties of granular materials.

[0004] At present, the methods for identifying force chain particles in particle assemblies mainly include threshold method, cluster analysis method and network analysis method. Although these methods can realize the identification of force chain particles to a certain extent and help researchers understand the mechanical behavior of granular materials, they show some limitations in practical applications. First, these methods have a large subjective dependence on parameter selection and setting. Different parameter selections may lead to significant differences in identification results. In practical production, it is often very difficult to obtain accurate parameter settings, which makes it difficult to ensure the reliability of identification results. Secondly, these methods usually need to extract features such as contact force and contact angle between particles, which increases the computational complexity, especially when dealing with large-scale particle assemblies. The efficiency is low. In addition, to identify force chain particles in a specific state, the particle assembly must first be loaded into the strain state through simulation. This process is time-consuming and complicated, further increasing the overall time cost and operation difficulty. Therefore, it is necessary to invent a new method to solve the limitations of existing methods in parameter dependence, computational complexity and time cost, so as to more efficiently and accurately identify force chain particles in different strain states. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings in the prior art and provide a method for predicting the force chain distribution of particle sets with different gradations under different strains. The method is based on the Pix2Pix generative adversarial network (GAN). Through the mutual game between the generator and the discriminator, complex features are automatically extracted from the particle set data with different gradations, and a mapping relationship between the numerical image of the particle set and the force chain distribution image is established. A predicted image that is highly similar to the actual force chain image under different strain states is generated, thereby achieving accurate prediction of the mechanical properties of granular materials, thereby providing an effective basis for material design and optimization in engineering applications.

[0006] In order to solve the problems in the prior art, the present invention discloses a method for predicting the force chain distribution of different graded particle sets under different strains, comprising the following steps:

[0007] S1: Select a gradation equation, and based on the selected gradation equation, combine the values ​​of each parameter within the parameter value range to obtain gradation curve data D1 corresponding to different gradation distribution conditions;

[0008] S2: using the discrete element method, generating a particle set numerical sample D2 corresponding to the gradation curve data in step S1, and assigning values ​​to the microscopic physical parameters of the sample;

[0009] S3: Through biaxial compression numerical simulation test, the mechanical properties of the numerical sample D2 are tested to obtain the particle size and position information data D3 of each particle in the particle set under different strain conditions;

[0010] S4: Screening the extracted particle size and position information data D3 to extract force chain particle information data D4;

[0011] S5: The particle size and position information data D3 and the force chain particle information data D4 are respectively imported into the integrated development environment PyCharm, and secondary modeling is performed using the Python language. The "Matplotlib" library in Python is called to draw the particle set image and the force chain image to construct the data set D5;

[0012] S6: Randomly divide the constructed data set D5 into a training set D51 and a validation set D52, and use the training set D51 data to train the Pix2Pix adversarial generative network model and optimize its parameters;

[0013] S7: When the quality of the images generated by the model reaches the expected standard, stop training the model;

[0014] S8: Pix2Pix is ​​used to generate adversarial networks to predict the force chain distribution image corresponding to the validation set D52 particle set image. The generated results are verified through quantitative and qualitative evaluations to verify the effectiveness and reliability of the Pix2Pix model in generating force chain images.

[0015] Preferably, the microscopic physical parameters of the particle set numerical sample D2 in step S2 include: particle set size, maximum particle size, minimum particle size intercept, particle density, initial porosity, damping coefficient, effective modulus, inter-particle friction coefficient, particle-wall friction coefficient, normal stiffness of particle-wall contact, and particle stiffness ratio (normal stiffness / tangential stiffness).

[0016] Preferably, the simulation conditions that need to be set in the biaxial compression numerical simulation test in step S3 include: the confining pressure value σ in the isotropic compression stage c , axial loading velocity v in the biaxial compression stage, and target axial strain value.

[0017] Preferably, the force chain particles screened in step S4 need to meet three necessary conditions: Condition 1: The force chain particles need to be high stress particles. It is usually stipulated that particles whose absolute value of the major principal stress (compression is positive and tension is negative) is greater than the average value are high stress particles, that is, they meet formula (1).

[0018]

[0019] Among them, the calculation formulas of particle principal stress and principal direction are as follows:

[0020]

[0021] Where N is the number of particles in the particle set, σ ij is the stress in different directions of the particle, and θ is the direction of σ1.

[0022] Condition 2: The angle between the center line of the contacting particles and the direction of the major principal stress is less than α, that is, equation (4) is satisfied:

[0023]

[0024] Where l is the connecting particle P i With P j The branch vector, is the major principal stress of adjacent particles.

[0025] Determine whether any two particles P i , P j Whether it is in the same force chain, the premise is: i As the reference particle, P j is the effective adjacent particle that satisfies the formula; at the same time, P jAs the reference particle, P i It is also a valid adjacent particle that satisfies the formula.

[0026] Condition 3: Each force chain must consist of 3 or more particles.

[0027] Preferably, when calling the "Matplotlib" library in Python to draw the particle set image and the force chain image in step S5, the particle size value needs to be normalized by the maximum value normalization method, and the relative size of the particles is accurately mapped with a value between 0 and 1, and a different color is assigned to each particle according to the normalized particle size value through a custom color mapping. The formula for the maximum value normalization is shown in formula (5):

[0028]

[0029] where x scale is the particle size value after normalization, x is the particle size value before normalization, and x min With x max Select the minimum and maximum particle size values ​​of the entire data set.

[0030] Preferably, when calling the "Matplotlib" library in Python to draw the particle set image and the force chain image in step S5, after assigning different colors to each particle, the image pixels need to be standardized to map the image pixel range to a standard interval. The standardization formula is shown in formula (6):

[0031]

[0032] Among them I norm represents the standardized pixel value, I(i,j,c) represents the pixel value of channel c at position (i,j) of the original image, mean[c] is the mean of channel c, and std[c] is the standard deviation of channel c.

[0033] Preferably, the Pix2Pix adversarial generative network constructed in step S6 is mainly composed of two parts: a generator G and a discriminator D. During the training process, the discriminator D continuously optimizes its discrimination ability by calculating the binary cross entropy loss function of the input image x, the real image y and the generated image G(x,z) respectively. The calculation formula of the binary cross entropy loss function BCE(p,y) is shown in formula (7):

[0034] BCE(p,y)=-(ylog(p)+(1-y)log(1-p)) (7),

[0035] When calculating the loss function L between the input image x and the real image y realWhen p = D(x, y), that is, the output probability of the discriminator D for the real image pair (x, y), y = 1; when calculating the loss function L of the input image x and the generated image G(x, z) fake When , p = D(x, G(x, z)), that is, the output probability of the discriminator D to generate the image pair (x, G(x, z)), y = 0.

[0036] The total loss function L of the discriminator D D is the mean of the two loss functions, as shown in formula (8):

[0037]

[0038] Where L D Denotes the total loss function of the discriminator D.

[0039] The generator G jointly optimizes the generative adversarial loss in the adversarial loss and the L1 loss in the pixel-level loss to generate images that are both realistic and highly similar to the real images at the pixel level. The calculation formula of its loss function can be expressed as shown in formula (9):

[0040] L G =L GAN +λL L1 (9)

[0041] Where L G is the total loss function of the generator G, L GAN represents the adversarial loss of the generator, L Ll represents L1 loss, and λ represents the weight parameter.

[0042] And L GAN The calculation formula can be expressed as formula (10):

[0043] L GAN =BCE(D(x,G(x,z)),1) (10),

[0044] L Ll The calculation formula can be expressed as formula (11):

[0045]

[0046] Where H represents the height of the image, W represents the width of the image, and y i,j Represents the pixel value of the real image at position (i, j), G(x,z) i,j Represents the pixel value of the generated image at position (i, j).

[0047] Preferably, the termination condition of the model training in step S7 is that the training stops when the loss curves of the discriminator D and the generator G tend to converge and the output probabilities of the discriminator D for the real image y and the generated image G(x,z) are close to 0.5.

[0048] The beneficial effects of the present invention are as follows: (1) The Pix2Pix adversarial generative network can accurately capture the complex mapping relationship between the particle set image and the force chain image through adversarial training of the generator and the discriminator, and has high accuracy and reliability. (2) Pix2Pix can also show good performance on a limited data set, reducing the cost of data collection and annotation, so that the method can still be efficiently applied when data resources are limited. (3) The particle material samples constructed using the discrete element method cover a wide range of gradation distribution conditions, ensuring the applicability and versatility of the method in different scenarios. (4) The data processing and image generation process are completed in the integrated development environment PyCharm, which is highly automated, reduces the errors and workload of manual operations, and significantly improves work efficiency. (5) There is no need to load the particle set into a specific state through simulation, and accurate identification of the force chain particles in the state can be achieved, thereby simplifying the process and saving time and resources. (6) In the material design and development stage, by predicting the force chain image, a gradation scheme with excellent performance can be quickly screened out, providing a scientific basis for the design and optimization of particle materials in engineering practice, and having important engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is the algorithm flow chart;

[0050] Figure 2 To generate numerical specimens of granular materials using discrete element method;

[0051] Figure 3 It is the particle collection image and force chain distribution image under the hyperbolic gradation curve;

[0052] Figure 4 It is the particle collection image and force chain distribution image under the inverse S-shaped gradation curve;

[0053] Figure 5 Particle collection image and force chain distribution image under the linear gradation curve;

[0054] Figure 6 Comparison of the predicted results and actual results of the validation set data under various morphological grading curves. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and implementation examples. It should be understood that the specific implementation examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0056] The method of the present invention is applied to the prediction of stress chain distribution images of particle assemblies under different continuous gradation distribution conditions under different strain states. The algorithm flow is as follows: Figure 1 As shown, the specific prediction process is as follows:

[0057] S1: Select Zhu Jungao's continuous grading equation as shown in formula (12), which is widely applicable to various continuously graded granular materials:

[0058]

[0059] The parameter value range is limited to 0<m<12, -4<b<1 to reflect the continuous gradation of general granular materials. The values ​​of m and b are combined within a reasonable value range to obtain a total of 100 sets of continuous gradation curve data D1.

[0060] S2: The discrete element method is used to generate numerical samples D2 of granular materials with different gradations, such as Figure 2 The size of sample D2 is 200mm×100mm, and the particle shape is a disc; the ratio of the maximum particle size to the sample width is set to be approximately 1:12, that is, the maximum particle size is approximately 4mm, and the minimum particle size intercept is 0.5mm. The equal volume replacement method is used to replace the particles with a particle size smaller than this value in the grading with particles with a particle size of 0.5mm.

[0061] Other microscopic parameters required for discrete element calculation are: particle density is 2600kg / m 3 , initial porosity is 0.15, damping coefficient is 0.7, effective modulus is 1×10 8 Pa, the inter-particle friction coefficient is 0.5, the particle-wall friction coefficient is 0, and the normal stiffness of the particle-wall contact is 1×10 9 N / m, particle stiffness ratio (normal stiffness / tangential stiffness) is 4 / 3.

[0062] S3: Through biaxial compression numerical simulation test, the mechanical properties of each sample are tested to obtain the particle size and position information D3 of each particle in the particle set under different strain conditions. The simulation conditions that need to be set for the numerical simulation test are: Set the confining pressure value σ in the isotropic compression stage c The pressure is 600 kPa, the loading speed in the biaxial compression stage is determined to be v = 0.15 m / s, and the compression is terminated when the axial strain reaches 15%. Starting from the axial strain value of 0, the particle size and position information are recorded every time the strain increases by 1%.

[0063] S4: Screen the extracted particle information D3 to extract the force chain particle information D4. It is stipulated that the screened force chain particles must meet three necessary conditions: Condition 1: The force chain particles must be high stress particles. It is usually stipulated that particles with a major principal stress (compression is positive, tension is negative) greater than the average value in absolute value are high stress particles, as shown in formula (1),

[0064]

[0065] The calculation formulas for the principal stress and principal direction of particles are shown in equation (2) and equation (3) respectively:

[0066]

[0067] Where N is the number of particles in the particle set, σ ij is the stress in different directions of the particle, and θ is the direction of σ1.

[0068] Condition 2: The angle between the center line of the contacting particles and the direction of the major principal stress is less than α (45°), satisfying

[0069] As shown in formula 4:

[0070]

[0071] Where l is the connected particle P i With P j The branch vector, is the major principal stress of adjacent particles.

[0072] Determine whether any two particles P i , P j Whether it is in the same force chain, the premise is: i As the reference particle, P j is the effective adjacent particle that satisfies the formula; at the same time, P j As the reference particle, P i It is also a valid adjacent particle that satisfies the formula.

[0073] Condition 3: Each force chain consists of 3 or more particles.

[0074] S5: The particle size and position information data D3 and the force chain particle data D4 are respectively imported into the integrated development environment PyCharm and the Python language is used for secondary modeling, and the "Matplotlib" library in Python is called to draw the particle collection image and the force chain image. The image data display range is set to 240mm×200mm, the output image size is set to 254mm×254mm, the resolution is set to 600dpi, and the image format is converted to RGB format.

[0075] When drawing an image, the maximum value normalization method is used to normalize the particle size value, accurately mapping the relative size of the particle to a value between 0 and 1, and assigning a different color to each particle through a custom color mapping according to the normalized particle size value. The maximum value normalization formula is shown in formula (5):

[0076]

[0077] where x scale is the particle size value after normalization; x is the particle size value before normalization; x min With x max Select the minimum and maximum particle size values ​​of the entire data set.

[0078] In S5, after assigning different colors to each particle, the image pixels need to be standardized to map the image pixel range to a standard interval. The standardization formula is shown in formula (6):

[0079]

[0080] Among them I norm represents the normalized pixel value, I(i,j,c) represents the pixel value of channel c at position (i,j) in the original image, mean[c] is the mean of channel c, and std[c] is the standard deviation of channel c. In this example, both mean[c] and std[c] are 0.5.

[0081] S6: In view of the fact that people pay more attention to the distribution of force chain particles under peak stress state in engineering practice, this case uses the image of particle collection without biaxial compression as input and the force chain distribution image under peak stress state as output. Usually, the stress reaches its peak value when the axial strain value reaches 2%.

[0082] The 100 constructed data sets are randomly divided, of which 90%, or 90 data sets, are selected as training set D51 for model training and parameter optimization, and the validation set D52 accounts for 10%, or 10 data sets, for evaluating the model prediction ability. Some of the constructed data sets are as follows Figure 3-Figure 5 shown.

[0083] In S6, the Pix2Pix adversarial generative network model is trained and parameter optimized using the training set D51 data. The Pix2Pix adversarial generative network constructed in this example mainly consists of two parts: the generator G and the discriminator D. During the training process, the discriminator D continuously optimizes its discrimination ability by calculating the binary cross entropy loss function of the input image x, the real image y and the generated image G(x,z) respectively. The calculation formula of the binary cross entropy loss function is (7):

[0084] BCE(p,y)=-(ylog(p)+(1-y)log(1-p)) (7),

[0085] When calculating the loss function L between the input image x and the real image y real When p = D(x, y), that is, the output probability of the discriminator D for the real image pair (x, y), y = 1; when calculating the loss function L of the input image x and the generated image G(x, z) fake When , p = D(x, G(x, z)), that is, the output probability of the discriminator D to generate the image pair (x, G(x, z)), y = 0.

[0086] The total loss function of the discriminator D is the mean of the two loss functions, as shown in formula (8):

[0087]

[0088] Where L D Represents the total loss function of the discriminator D;.

[0089] The generator G jointly optimizes the generative adversarial loss in the adversarial loss and the L1 loss in the pixel-level loss to generate images that are both realistic and highly similar to the real images at the pixel level. The calculation formula of its loss function can be expressed as formula (9):

[0090] L G =L GAN +λL L1 (9)

[0091] Where L G is the total loss function of the generator G, L GAN represents the adversarial loss of the generator, L Ll represents L1 loss, λ represents the weight parameter, and the weight parameter is 100 in this example.

[0092] L in formula (9) GAN The calculation formula can be expressed as shown in formula (10):

[0093] L GAN =BCE(D(x,G(x,z)),1) (10),

[0094] L in formula (9) Ll The calculation formula can be expressed as shown in formula (11):

[0095]

[0096] Where H represents the height of the image, W represents the width of the image, and y i,j Represents the pixel value of the real image at position (i, j), G(x,z)i,j Indicates the pixel value of the generated image at position (i, j). In order to reduce the value range of the input data to reduce the learning difficulty of the neural network and to maintain the clarity of the image, the length and width of the image are adjusted to 512×512 pixels, and the number of color channels is set to 3.

[0097] In S6, in order to optimize the training efficiency, the Adam optimizer is introduced into the neural network to optimize the generator G and the discriminator D. The specific parameters are set as follows: the initial learning rate is set to 2×10 -4 , the exponential decay rate of the first-order moment estimate is 0.5, and the exponential decay rate of the second-order moment estimate is 0.999. The total training rounds (epochs) are set to 200 rounds, and the batch size (batch size) is set to 4.

[0098] S7: When the loss curves of the discriminator D and the generator G tend to converge and the output probabilities of the discriminator D for the real image y and the generated image G(x,z) are close to 0.5, stop training the model.

[0099] S8: Use Pix2Pix adversarial generative network to predict the force chain distribution image corresponding to the validation set D52 particle set image when the axial strain value reaches 2%. The prediction results of the validation set D52 data are compared with the actual results. Figure 6 shown.

[0100] The results show that the Pix2Pix model has high accuracy and reliability in generating force chain images.

[0101] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting force chain distribution of different graded particle assemblies under different strains, characterized by: The steps include: S1: Select a gradation equation, and based on the selected gradation equation, combine the values ​​of each parameter within the parameter value range to obtain gradation curve data D1 corresponding to different gradation distribution conditions; S2: using the discrete element method, generating a particle set numerical sample D2 corresponding to the gradation curve data in step S1, and assigning values ​​to the microscopic physical parameters of the sample; S3: Through biaxial compression numerical simulation tests, the mechanical properties of each numerical sample D2 are tested to obtain the particle size and position information data D3 of each particle in the particle set under different strain conditions; S4: Screening the extracted particle size and position information data D3 to extract force chain particle information data D4; S5: Import the particle size and position information data D3 and the force chain particle information data D4 into the integrated development environment PyCharm, use Python language to perform secondary modeling, and call the "Matplotlib" library in Python to draw the particle collection image and the force chain image to construct the data set D5; S6: Randomly divide the constructed data set D5 into a training set D51 and a validation set D52, and use the training set D51 data to train the Pix2Pix adversarial generative network model and optimize its parameters; S7: When the quality of the images generated by the model reaches the expected standard, stop training the model; S8: Pix2Pix is ​​used to generate adversarial networks to predict the force chain distribution image corresponding to the validation set D52 particle set image. The generated results are verified through quantitative and qualitative evaluations to verify the effectiveness and reliability of the Pix2Pix model in generating force chain images.

2. A method for predicting force chain distribution under different strains of different graded particle assemblies according to claim 1, characterized in that: The microscopic physical parameters of the particle set numerical sample D2 in step S2 include: particle set size, maximum particle size, minimum particle size intercept, particle density, initial porosity, damping coefficient, effective modulus, inter-particle friction coefficient, particle-wall friction coefficient, normal stiffness of particle-wall contact, and particle stiffness ratio.

3. A method for predicting force chain distribution under different strains of different graded particle assemblies according to claim 1, characterized in that: The simulation conditions that need to be set for the biaxial compression numerical simulation test in step S3 include: the confining pressure value σ in the isotropic compression stage c , axial loading velocity v in the biaxial compression stage, and target axial strain value.

4. A method for predicting force chain distribution under different strains of different graded particle assemblies according to claim 2 or 3, characterized in that: The force chain particles screened in step S4 must meet three necessary conditions: Condition 1: The force chain particles must be high stress particles. It is usually stipulated that particles whose absolute value of the major principal stress (compression is positive and tension is negative) is greater than the average value are high stress particles, that is, they satisfy formula (1). Among them, the calculation formulas of particle principal stress and principal direction are as follows: Where N is the number of particles in the particle set, σ ij is the stress in different directions of the particle, θ is the direction of σ1; Condition 2: The angle between the center line of the contacting particles and the direction of the major principal stress is less than α, that is, equation (4) is satisfied: Where l is the connecting particle P i With P j The branch vector, is the major principal stress of adjacent particles; Determine whether any two particles P i , P j Whether it is in the same force chain, the premise is: i As the reference particle, P j is the effective adjacent particle that satisfies the formula; at the same time, P j As the reference particle, P i It is also the valid adjacent particles that satisfy the formula; Condition 3: Each force chain must consist of 3 or more particles.

5. A method for predicting force chain distribution under different strains of different graded particle assemblies according to claim 1, characterized in that: When calling the "Matplotlib" library in Python to draw the particle set image and the force chain image in step S5, the maximum value normalization method is used to normalize the particle size value, accurately map the relative size of the particle to a value between 0 and 1, and assign a different color to each particle based on the normalized particle size value through a custom color mapping. The maximum value normalization formula is shown in formula (5): where x scale is the particle size value after normalization, x is the particle size value before normalization, and x min With x max Select the minimum and maximum particle size values ​​of the entire data set.

6. A method for predicting force chain distribution under different strains of different graded particle assemblies according to claim 5, characterized in that: When calling the "Matplotlib" library in Python to draw the particle set image and the force chain image in step S5, after assigning different colors to each particle, the image pixels need to be standardized to map the image pixel range to a standard interval. The standardization formula is shown in formula (6): Among them I norm represents the standardized pixel value, I(i,j,c) represents the pixel value of channel c at position (i,j) of the original image, mean[c] is the mean of channel c, and std[c] is the standard deviation of channel c.

7. A method for predicting force chain distribution of different graded particle assemblies under different strains according to claim 6, characterized in that: The Pix2Pix adversarial generative network constructed in step S6 mainly consists of two parts: the generator G and the discriminator D. During the training process, the discriminator D continuously optimizes its discrimination ability by calculating the binary cross entropy loss function of the input image x, the real image y and the generated image G(x,z) respectively. The calculation formula of the binary cross entropy loss function BCE(p,y) is shown in formula (7): BCE(p,y)=-(ylog(p)+(1-y)log(1-p)) (7), When calculating the loss function L between the input image x and the real image y real When p = D(x, y), that is, the output probability of the discriminator D for the real image pair (x, y), y = 1; when calculating the loss function L between the input image x and the generated image G(x, z) fake hour, p = D(x, G(x, z)), that is, the output probability of the discriminator D to generate the image pair (x, G(x, z)), at this time y = 0; The total loss function L of the discriminator D D is the mean of the two loss functions, as shown in formula (8): Where L D Represents the total loss function of the discriminator D; The generator G jointly optimizes the generative adversarial loss in the adversarial loss and the L1 loss in the pixel-level loss to generate an image that is both realistic and highly similar to the real image at the pixel level. The calculation formula of its loss function can be expressed as formula (9): THE G =L GAN +λL L1 (9), Where L G is the total loss function of the generator G, L GAN represents the adversarial loss of the generator, L Ll represents L1 loss, λ represents weight parameter; And L GAN The calculation formula can be expressed as formula (10): L GAN =BCE(D(x,G(x,z)),1) (10), L Ll The calculation formula can be expressed as formula (11): Where H represents the height of the image, W represents the width of the image, and y i,j Represents the pixel value of the real image at position (i, j), G(x,z) i,j Represents the pixel value of the generated image at position (i, j).

8. A method for predicting force chain distribution under different strains of different graded particle assemblies according to claim 1, characterized in that: The termination condition of the model training in step S7 is that the training stops when the loss curves of the discriminator D and the generator G tend to converge and the output probabilities of the discriminator D for the real image y and the generated image G(x,z) are close to 0.5.

Citation Information

Cited By

  • Wet granular medium mechanical property prediction method based on GAN and CNN-BPNN

    CN120012586A