Method and device for rapidly evaluating load state of multi-phase composite structure image

Abstract

The present invention relates to a method and device for quickly evaluating the load state of a multi-phase composite structure image. The method includes the following steps: obtaining a cross-sectional morphology image of the multi-phase composite structure, establishing an image restoration finite element analysis model, and calculating the load distribution under different boundary conditions and geometric features; constructing a generative adversarial network constrained by an elastic mechanics model and training it based on the data obtained from the finite element analysis model; and evaluating the load state of the multi-phase composite structure to be detected based on the trained generative adversarial network. Compared with the prior art, the present invention has the advantages of good prediction accuracy and high efficiency.

Description

Technical Field

[0001] The present invention relates to the field of load state evaluation, and particularly to a method and device for rapidly evaluating the load state of a multi-phase composite structure image. Background Art

[0002] Multi-phase composite structures widely exist in engineering applications. Generally, the differences in the properties of each phase will cause the structural failure when the boundary conditions change violently. For example, the thermal barrier coatings widely used in high-temperature components of aerospace generally are prepared by the air plasma spraying (APS) method, resulting in a complex and variable pore structure in its ceramic top layer (TC). While the internal pore structure of the TC improves the thermal insulation performance, it is easily infiltrated by the melted CaO–MgO–Al2O3–SiO2 (CMAS) during high-temperature service, forming a TC–CMAS two-phase structure. The thermal expansion coefficient of CMAS is quite different from that of the ceramic matrix. During the cooling process, the solidified CMAS can induce high thermal mismatch stress in the TC layer, causing the coating to fail prematurely. Therefore, a large number of scholars have carried out numerical simulations on the thermal cycling load of TC–CMAS and made effective progress. However, these methods face a major challenge, that is, when the boundary conditions or microstructural characteristics change, it is necessary to re-model and carry out calculations, consuming a large amount of computing resources. And the optimization of composite structure characteristics is an important technical way to improve the performance of multi-phase materials. The low efficiency of traditional numerical calculation methods has become an obstacle restricting the rapid optimization design of composite structures. With the development of artificial intelligence technology, data-driven neural network models have a high computing speed. Therefore, developing a load rapid evaluation method based on a neural network architecture is of great significance for the process and composite structure optimization of multi-phase materials.

[0003] Currently, the evaluation techniques for load states such as stress and strain of multi-phase structures are mainly divided into two types:

[0004] The first type is the experimental test based on composite structure specimens. Generally, the digital image correlation (DIC) method is used to measure the strain field of the specimen under simulated boundary conditions. However, limited by the equipment resolution, it is difficult to capture the detailed changes in the strain distribution of the composite microstructure, and the sample preparation and test cycle are long, and the test cost is high. Therefore, the experimental test method cannot meet the requirements of multi-phase composite structure optimization and on-line monitoring of load states.

[0005] The second method relies on numerical simulation of relevant finite element software. This method conducts numerical simulation analysis by means of computer simulation software to obtain the thermal strain state of the multiphase composite structure. However, the numerical simulation of complex structures places extremely high demands on computer hardware; the composite structure is generally complex and variable, and each time a new structure is encountered in numerical calculation, a geometric model needs to be rebuilt, involving a large amount of repetitive work and consuming a great deal of time and computing resources, making it difficult to support the requirements of composite structure optimization analysis; in addition, some work simplifies the composite structure into a regular shape and establishes a simplified geometric model for simulating structural parameters such as the inclusion rate for numerical simulation. Although this calculation method reduces the consumption of computing resources, the credibility of the results is relatively low.

[0006] As Figure 1 shown, taking the traditional method for evaluating stress and strain states based on numerical simulation as an example, first, it is necessary to photograph or simplify the structural morphology from the multiphase composite structure sample. After semantic segmentation of the structure, a numerical simulation model is established in combination with service parameters such as boundary conditions and material properties. Among them, the numerical simulation process includes the establishment of an image restoration geometric model, material property setting, mesh generation, boundary condition setting, submission of calculations, and post-processing of results, generally consuming a great deal of time and computing resources. For tasks that require solving multiple sets of data to obtain a certain-scale data set, all the above processes need to be repeated for the new composite structure to be measured, especially the time-consuming numerical simulation process. Therefore, it is difficult for the traditional numerical simulation method to obtain a large amount of data in a short period to support work such as optimizing the strength and process parameters of the composite structure. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and provide a method, device, and medium for rapidly evaluating the load state of images of multiphase composite structures.

[0008] The purpose of the present invention can be achieved through the following technical solutions:

[0009] According to the first aspect of the present invention, there is provided a method for rapidly evaluating the load state of images of multiphase composite structures, including the following steps:

[0010] Obtain a cross-sectional morphology image of the multiphase composite structure, establish an image restoration finite element analysis model, and calculate the load distribution under different boundary conditions and geometric features;

[0011] Construct a generative adversarial network constrained by an elastic mechanics model and train it based on the data obtained from the finite element analysis model;

[0012] Evaluate the load state of the multiphase composite structure to be detected based on the trained generative adversarial network.

[0013] Preferably, the method specifically includes the following steps:

[0014] S1: Obtain the cross-sectional morphology image of the multiphase composite structure and construct an image library;

[0015] S2: Perform semantic segmentation on the images in the image library according to different phases to obtain a semantic segmentation structure diagram;

[0016] S3: Perform boundary recognition and anti-aliasing optimization on the semantically segmented images;

[0017] S4: Establish a finite element model for image restoration based on the optimized structural boundary coordinates, and calculate the load distribution of the structure under predetermined working condition parameters;

[0018] S5: Establish an elastic mechanics constrained generative adversarial network model for fitting the relationship between the structural image and the load distribution;

[0019] S6: Use the semantic segmentation structure diagram obtained in S2 as the model input and the load distribution obtained in S4 as the model output to construct a training sample set;

[0020] S7: Train the generative adversarial network model based on the training sample set;

[0021] S8: Obtain the cross-sectional morphology image of the multiphase composite structure to be detected and perform preprocessing based on S2, and input it into the trained generative adversarial network model to obtain the load state.

[0022] Further preferably, the step S2 specifically includes:

[0023] S21: Process the cross-sectional morphology image with Gaussian filtering;

[0024] S22: Adjust the image contrast to enhance the gray level difference of each phase;

[0025] S23: Mark different phase components with different colors.

[0026] Further preferably, the step S3 specifically includes:

[0027] S31: Identify the region of each phase component in the semantic segmentation structure diagram in step S2 based on the connected domain analysis method;

[0028] S32: Perform anti-aliasing optimization on the identified structural boundary by using the coordinate averaging method.

[0029] Further preferably, the step S4 specifically includes:

[0030] Perform finite element analysis on the optimized structural boundary of step S3 to establish a geometric model of a two-dimensional multiphase structure; set corresponding material properties for each phase component and matrix material, etc.; divide the geometric model using quadrilateral meshes with the same mesh size for different groups of structures; set boundary conditions according to the predetermined working conditions and submit the calculation of the load distribution of multiple groups of structures under nonlinear conditions.

[0031] Further preferably, step S5 specifically includes:

[0032] Based on the pix2pix architecture, establish an initial generative adversarial network for pixel-to-pixel conversion: establish a multi-layer U-shaped convolutional neural network as the generator G of the generative adversarial network, and establish a downsampling convolutional neural network as the discriminator D of the generative adversarial network.

[0033] Further preferably, the load distribution is the stress distribution and strain field of the multiphase composite structure. The stress and strain of the multiphase composite structure are respectively used as outputs to train the generative adversarial network. Step S5 specifically includes:

[0034] The loss function of the initial generative adversarial network is denoted as:

[0035]

[0036] where training G maximizes the logD(r, G(r)) loss, and training D minimizes logD(r, s), E e represents the mathematical expectation, λ is the coefficient of the L1 norm loss L 1 of, r is the input of the network, s is the true reference of the output, and L 1 (G) is denoted as:

[0037] L 1 (G) = E r,s [||s - G(r)|| 1

[0038] When the prediction target is the stress distribution, s is the finite element stress field σ t ; when the prediction target is the strain distribution, s is the finite element strain field ε t ,

[0039] Introduce the generalized Hooke's law considering thermal expansion, denoted as:

[0040]

[0041] where E' = E / (1 - ν 2 ), ν' = ν / (1 - ν), α' = α(1 + ν), G = E / 2(1 + ν), E is the elastic modulus, ε 11 is the elastic strain component in the x-axis direction, ε​22 is the strain component in the y-axis direction, ε 12 is the shear strain in the x-y plane, ν represents the Poisson's ratio, α is the coefficient of thermal expansion, σ represents the stress components in each axial direction, σ 11 is the stress component along the x-axis direction, σ 22 is the stress component in the y-axis direction, σ 12 is the shear stress in the x-y plane, T r represents room temperature, T 0 represents the initial cooling temperature;

[0042] Express each stress component in terms of strain components, denoted as:

[0043]

[0044] Introduce the deformation compatibility conditions of the continuity assumption in elasticity, denoted as:

[0045]

[0046] Let the multiphase structure distribution after semantic segmentation in step S2 be x c , and the corresponding strain distribution calculated based on the finite element model in step S4 is set as ε t . Combine the three strain components into a three-dimensional array, denoted as:

[0047]

[0048] where # represents the switching layer, indicating that ε t is composed of ε t,11 , ε t,22 , stacked in order to form a three-dimensional array;

[0049] The strain array generated by the generator is denoted as:

[0050] ε m = [ε m,11 #ε m,22 #ε m,12

[0051] ε t and ε m can both be equivalent to the multi-channel image A. Let the convolutional kernel be a multi-dimensional array C. The two-dimensional convolutional calculation of three channels can be expressed as:

[0052]

[0053] where ζ and ξ are the receptive field positions, k is the number of channels, * represents convolution,

[0054] ​The strain compatibility error of the strain image is calculated by the difference method, and any pixel point P ζ,ξ The second-order partial derivatives in each direction are denoted as:

[0055]

[0056] The deformation compatibility condition of the continuity assumption in elasticity can be expressed as:

[0057]

[0058] Introduce a custom fixed convolution kernel C p , denoted as:

[0059]

[0060] Use the convolution operation to conveniently calculate the strain compatibility error L of the strain field C , denoted as:

[0061]

[0062] Among them, f is the pixel side length dimension correction coefficient,

[0063] When the prediction target is the strain distribution, set the stress load distribution calculated by finite element as the true reference, and calculate the stress-strain relationship error L based on the stress components H,12 and L H,3 , denoted as:

[0064]

[0065] Among them, L H,12 and L H,3 are used to constrain the axial strain and shear strain components respectively, and L C , L H,12 and L H,3 are simultaneously used to constrain the weight optimization of the generator,

[0066] The discriminator D loss L of the true strain load distribution S,D , denoted as:

[0067] L S,D = logD(r c , ε t , L C,t )

[0068] Among them, r c is the structural image of semantic segmentation, ε t is the true strain distribution corresponding to the structure, and L C,t is the strain compatibility error of the true strain distribution,

[0069] The strain distribution generated by the generator G, and its discriminator loss L SG,D , denoted as:

[0070] L SG,D = log(1 - D(r c , G(r c )), L C,m )

[0071] where L C,m is the strain compatibility error of the synthetic strain distribution;

[0072] The discriminator loss L D can be denoted as:

[0073]

[0074] where b is the sample batch size,

[0075] When the prediction target is the stress distribution, the strain distribution calculated by finite element is set as the true reference, and the stress-strain relationship error L H,12 and L H,3 , denoted as:

[0076]

[0077] where L H,12 and L H,3 are respectively used to constrain the axial stress and shear stress components, and L C , L H,12 and L H,3 are simultaneously used to constrain the weight optimization of the generator,

[0078] The discriminator D loss L of the true strain load distribution S,D , denoted as:

[0079] L S,D = log D(r c , σ t , L C,t )

[0080] The stress distribution generated by the generator G, and its discriminator loss L SG,D , denoted as:

[0081] L SG,D = log(1 - D(r c , G(r c )), L C,m )

[0082] where L C,mis the strain compatibility error of the synthesized strain distribution. After calculating the generated stress distribution as the strain distribution according to the generalized Hooke's law considering thermal expansion, convolution calculation should be performed based on the strain compatibility error; discriminator loss L D can be denoted as:

[0083]

[0084] Generator loss L G includes the L1 norm L 1 , and the L C of the elasticity mechanics constraint, L S,12 and L S,3 , denoted as:

[0085]

[0086] Therefore, the total loss function of the generative adversarial network with elasticity mechanics constraints is denoted as:

[0087]

[0088] wherein, is the coefficient matrix of the elasticity mechanics loss L E , and L E is denoted as:

[0089] L E = [L C,m , L S,12 , L S,3 T .

[0090] Further preferably, the step S6 specifically includes:

[0091] Cut the semantic segmentation structure diagram and its corresponding load distribution into multiple groups of local pictures that meet the input picture size conditions of the generative adversarial network established in step S5, and establish a training sample set; establish a corresponding material property and boundary condition matrix according to the structure picture for calculating the elasticity mechanics loss during the training process; select about 70% of the samples for network training, and the remaining samples for verification.

[0092] Further preferably, the step S8 specifically includes:

[0093] Preprocess the high-resolution multi-phase composite structure picture to be measured according to step S2, and adjust the image scale to the same size as the training set image; cut the preprocessed structure into multiple groups of local pictures that meet the input picture size conditions of the generative adversarial network established in step S5, and use the network model trained in step S7 to predict their corresponding strain distributions in turn; merge the predicted multiple groups of local strain distributions, and remove the overlapping areas to obtain the complete strain distribution. ​

[0094] According to a second aspect of the present invention, there is provided an electronic device including a memory and a processor, wherein a computer program is stored on the memory, and when the processor executes the program, the rapid load state evaluation method for a multi-phase composite structure image as described above is implemented.

[0095] According to a third aspect of the present invention, there is provided a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the rapid load state evaluation method for a multi-phase composite structure image as described above is implemented.

[0096] Compared with the prior art, the present invention has the following advantages:

[0097] 1. The present invention establishes a load evaluation method for a multi-phase composite structure such as stress and strain based on a generative adversarial network constrained by elastic mechanics equations. The trained network fitting model has a much higher evaluation speed for load distribution than the numerical simulation method, and the prediction accuracy meets the engineering requirements, which can provide a rapid data analysis method for the strength evaluation of the multi-phase composite structure and shorten the design cycle.

[0098] 2. The present invention combines a physics-constrained generative adversarial network and an image segmentation method, and can establish a training data set based on a small-scale numerical calculation sample data, reduce the data dependence of the neural network, enhance the generalization ability of the network model, and facilitate engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 It is a flowchart for the evaluation of the strain state of a multi-phase composite structure based on numerical simulation.

[0100] Figure 2 It is an overall architecture diagram of the rapid load state evaluation method for the multi-phase composite structure image of the present invention.

[0101] Figure 3 It is a flowchart for the rapid evaluation of the thermal strain state of a thermal barrier coating based on a microstructure image of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0102] The present invention will be described in detail below with reference to the drawings and specific embodiments. Note that the following description of the embodiments is only illustrative in essence, and the present invention is not intended to limit the objects to which it is applicable or its uses, and the present invention is not limited to the following embodiments.

[0103] Embodiment

[0104] The present invention aims to provide a method for rapidly evaluating the load state of a multi-phase composite structure image, which is used for evaluating the stress or strain load state of a complex multi-phase composite structure. The evaluation method includes: obtaining a cross-sectional morphology image of the multi-phase composite structure, establishing an image restoration finite element analysis model, and calculating the load distribution under different boundary conditions and geometric features; constructing a generative adversarial network constrained by an elastic mechanics model and training it based on the data obtained from the finite element analysis model; and evaluating the load state of the multi-phase composite structure to be detected based on the trained generative adversarial network. The present invention is achieved through the following technical steps:

[0105] Step 1: Take a cross-sectional morphology image of the multi-phase composite structure: Prepare a sample of the multi-phase material to be tested or take a sample, use an electron microscope to take a SEM image of its microstructure, determine the scale size of the image, and construct an image library;

[0106] Step 2: Perform semantic segmentation on the structure image obtained in Step 1 according to different phases: Use Gaussian filtering to process the SEM image taken in Step 1 to reduce its noise level; appropriately adjust its contrast to enhance the gray-scale difference of each phase; mark each phase component with different colors to obtain a semantic segmentation structure diagram.

[0107] Step 3: Identify the boundaries of the composite structure segmented in Step 2 and optimize the coordinates using the coordinate averaging method for anti-aliasing: Based on the connected component analysis method, identify the regions of each phase component in the image after semantic segmentation in Step 2, and denote each phase region as:

[0108] Q area ={(Q 1 ,Q 2 ,L,Q i )|1≤i≤Num(Q area )}

[0109] where Q area is the set of each phase region, Q i is one of the phase regions, i represents the region number, and Num(Q area ) represents the number of component regions;

[0110] The boundary coordinates of the region are denoted as B C :

[0111]

[0112] j represents the number of boundary pixel points, B i is the region boundary, b p,ij is a point in the boundary B i , and (x ij ,y ij ) is the coordinate of b p,ij ;

[0113] The anti-aliasing optimization of the identified composite structure boundary is carried out by using the coordinate averaging method (CA), and the expression of the CA algorithm is denoted as:

[0114]

[0115] P x and P y respectively represent the optimized x and y coordinates, and n is the number of recursion times; among them, there are b ij to b i(j+n) A total of n + 1 points participate in the CA optimization.

[0116] Step 4: Establish a finite element model for image restoration based on the optimized composite structure boundary coordinates in Step 3. Under the predetermined working condition parameters, calculate the stress and strain distributions of the structure: Import the structure boundary optimized in Step 3 into the finite element analysis software to establish a geometric model of the two-dimensional multiphase structure; Set the corresponding material properties for each phase component and matrix material, etc.; Divide the geometric model with quadrilateral meshes, and the mesh sizes of different group structures are the same; Set the boundary conditions according to the predetermined working conditions, and submit the calculation of the load distribution of multiple group structures under nonlinear conditions. In this embodiment, the load distribution is the stress distribution and strain distribution of the multiphase composite structure.

[0117] Step 5: Establish an elastic mechanics constrained generative adversarial network model for fitting the relationship between the structure image and the strain distribution: Establish an initial generative adversarial network for pixel-to-pixel conversion based on the pix2pix architecture: Establish a multi-layer U-shaped convolutional neural network as the generator G of the generative adversarial network, and establish a downsampling convolutional neural network as the discriminator D of the generative adversarial network; The loss function of the initial generative adversarial network is denoted as:

[0118]

[0119] Among them, train G to maximize the logD(r, G(r)) loss, and train D to minimize logD(r, s), where E e represents the mathematical expectation, λ is the coefficient of the L1 norm loss L 1 , r is the input of the network, s is the true reference of the output, and L 1 (G) is denoted as:

[0120] L 1 (G) = E r,s [||s - G(r)|| 1 (2)

[0121] When the prediction target is the stress distribution, y is the finite element stress distribution σ t ; when the prediction target is the strain field, y is the finite element strain field ε t .

[0122] The stress and strain in the multiphase composite structure should satisfy the generalized Hooke's law considering thermal expansion; in addition, the structure remains continuous during the cooling process, so its strain components satisfy the strain compatibility condition.

[0123] Therefore, the generalized Hooke's law considering thermal expansion is introduced and denoted as:

[0124]

[0125] where E' = E / (1 - ν 2 ), ν' = ν / (1 - ν), α' = α(1 + ν), G = E / 2(1 + ν), E is the elastic modulus, ε 11 is the elastic strain component in the x-axis direction, ε 22 is the strain component in the y-axis direction, ε 12 is the shear strain in the x - y plane, ν represents the Poisson's ratio, α is the coefficient of thermal expansion, σ represents the stress components in each axial direction, σ 11 is the stress component along the x-axis direction, σ 22 is the stress component in the y-axis direction, σ 12 is the shear stress in the x - y plane, T r represents the room temperature, T 0 represents the initial cooling temperature;

[0126] The stress components are expressed in terms of the strain components and denoted as:

[0127]

[0128] The deformation compatibility condition introducing the continuity assumption of elasticity mechanics is denoted as:

[0129]

[0130] Suppose the distribution of the multiphase structure after semantic segmentation in step 2 is x c , and the corresponding strain distribution calculated based on the finite element model in step 4 is set as ε t . The three strain components are combined into a three-dimensional array and denoted as:

[0131]

[0132] where # represents the switching layer, indicating that ε t is a three-dimensional array stacked in sequence by ε t,11 , ε t,22 , ;

[0133] The strain array generated by the generator is denoted as:

[0134] ε m= [ε m,11 #ε m,22 #ε m,12 (7)

[0135] ε t and ε m can both be equivalently represented as a multi-channel image A. Let the convolutional kernel be a multi-dimensional array C. The two-dimensional convolutional calculation for three channels can be expressed as:

[0136]

[0137] where ζ and ξ are the receptive field positions, k is the number of channels, and * represents convolution.

[0138] The differential method is used to calculate the strain compatibility error of the strain image. For any pixel point P ζ,ξ The second-order partial derivatives in each direction are denoted as:

[0139]

[0140] Equation (5) can be expressed as:

[0141]

[0142] Introduce a custom fixed convolutional kernel C p , denoted as:

[0143]

[0144] Use the convenient convolutional operation to calculate the strain compatibility error L of the strain field C , denoted as:

[0145]

[0146] where f is the pixel side length size correction coefficient.

[0147] When the prediction target is the strain field, set the FEM stress distribution as the true reference, and calculate the stress-strain relationship error L H,12 and L H,3 , denoted as:

[0148]

[0149] where L H,12 and L H,3 are respectively used to constrain the axial strain and shear strain components.

[0150] L C , L H,12 and L H,3 are simultaneously used to constrain the weight optimization of the generator. The discriminator D loss L of the true strain load distribution S,D , denoted as:

[0151] L S,D = log D(r c ,σ t ,L C,t ) (14)

[0152] where r c is the structural image of semantic segmentation, ε t is the true strain distribution corresponding to the structure, L C,t is the strain compatibility error of the true strain distribution,

[0153] The strain distribution generated by the generator G, and its discriminator loss L SG,D , is denoted as:

[0154] L SG,D = log(1 - D(r c ,G(r c ))),L C,m ) (15)

[0155] where L C,m is the strain compatibility error of the synthetic strain distribution; the discriminator loss L D can be denoted as:

[0156]

[0157] where b is the sample batch size.

[0158] When the prediction target is the stress distribution, the FEM strain distribution is set as the true reference, and the stress-strain relationship error L H,12 and L H,3 are calculated based on Equation (3), denoted as:

[0159]

[0160] where L H,12 and L H,3 are respectively used to constrain the axial stress and shear stress components.

[0161] L C , L H,12 and L H,3 are simultaneously used to constrain the weight optimization of the generator. The discriminator D loss L S,D of the true strain load distribution is denoted as:

[0162] L S,D = log D(r c ,σ t ,L C,t ) (18)

[0163] The stress distribution generated by the generator G and its discriminator loss L SG,D , denoted as:

[0164] L SG,D = log(1 - D(r c , G(r c )), L C,m ) (19)

[0165] where L C,m is the strain compatibility error of the synthetic strain distribution. After calculating the generated stress distribution as the strain distribution according to Equation (3), convolution calculation should be performed based on Equation (12); the discriminator loss L D can be denoted as:

[0166]

[0167] The generator loss L G includes the L1 norm L 1 , and the L C of the elasticity mechanics constraint, L S,12 and L S,3 , denoted as:

[0168]

[0169] Therefore, the total loss function of the generative adversarial network with elasticity mechanics constraints is denoted as:

[0170]

[0171] where is the coefficient matrix of the elasticity mechanics loss L E , and L E is denoted as:

[0172] L E = [L C,m , L S,12 , L S,3 T (23)

[0173] ​Step 6 uses the semantic segmentation structure diagram obtained in Step 2 as the model input and the load distribution obtained in Step 4 as the model output to construct a training sample set. The semantic segmentation structure diagram obtained in Step 2 and the load distribution diagrams such as stress and strain calculated in Step 4 are segmented into multiple local views, and a training data set is established based on the segmented structure diagrams and the load diagrams at the corresponding positions: the high-resolution composite structure pictures and their corresponding stress and strain pictures are cut into multiple groups of local pictures that meet the input picture size conditions of the generative adversarial network established in Step 5 to establish a training sample set; a corresponding material property and boundary condition matrix is established according to the structure pictures for calculating the elasticity loss during the training process; about 70% of the samples are selected for network training, and the remaining samples are used for verification.

[0174] Step 7 trains the generative adversarial network established in Step 6 based on the training set established in Step 5. The load distribution is the stress distribution and strain field of the multiphase composite structure, and the stress and strain of the multiphase composite structure are used as outputs respectively to train the generative adversarial network.

[0175] Step 8 repeats Step 2 for preprocessing the composite structure to be measured, and uses the network model trained in Step 7 to quickly predict the strain field of the structure to be measured: the high-resolution multiphase composite structure picture to be measured is preprocessed according to Steps 1 and 2, and the image scale is adjusted to the same size as the training set image; the preprocessed structure is cut into multiple groups of local pictures that meet the input picture size conditions of the generative adversarial network established in Step 5, and the network model trained in Step 7 is used to predict their corresponding strain distributions in turn; the predicted multiple groups of local strain distributions are merged, and the overlapping areas are removed to obtain the complete strain distribution.

[0176] In this example, a typical CaO–MgO–Al2O3–SiO2 (CMAS) infiltrated thermal barrier coating (TC–CMAS) complex two-phase microstructure is taken as the object for strength evaluation. During the cooling stage at the end of the gas turbine operation, the solidified CMAS causes a relatively complex strain state in the coating due to the different thermal expansion coefficients from the ceramic matrix. At the same time, the coating is mostly prepared by air plasma spraying method, resulting in a complex and variable microstructure. A large number of thermal load analyses need to be carried out for the variable structure. In this embodiment, a method for quickly evaluating the load state of a multiphase composite structure image is specifically provided, without the need to model for each new structure morphology every time. The overall architecture is shown in Figure 2 . The specific implementation of this method is mainly divided into two stages: the training stage and the thermal strain evaluation:

[0177] Training stage:

[0178] As Figure 3As shown, first, the cross-sectional microstructure image of the thermal barrier coating sample is taken by an electron microscope. The scales of different images are adjusted to the same value, and the scale size is recorded. The SEM images taken are processed by Gaussian filtering to reduce the noise level; the contrast is appropriately enhanced to make the gray-scale difference between CMAS inclusions and the ceramic matrix; then, based on the local threshold method, the image is binarized, and different colors are used to mark CMAS and the ceramic matrix to complete the semantic segmentation of the structure image. The boundaries of the segmented microstructure are recognized, and the CA method is used to remove the sawtooth effect at the boundaries to obtain the structure boundaries for establishing the finite element model for image restoration.

[0179] The temperature boundary conditions for the coating cooling are obtained, and the basic mechanical property parameters of CMAS and the ceramic matrix are obtained. Based on the structure boundaries, a geometric model for image restoration is established and meshed. After setting the boundary conditions and material parameters, the strain distributions of multiple groups of coatings are calculated. Based on the distributions of CMAS and the ceramic matrix, the corresponding material property and boundary condition matrices are established, and the training dataset of the generative adversarial network is established by combining the microstructure image and the strain load distribution. A generative adversarial network model for predicting the strain distribution of the coating is established, and the network model is constrained by the generalized Hooke's law and the strain compatibility condition. The calculation of finite element data is often time-consuming and has a large amount of repetitive work. Therefore, the input image resolution of the generative adversarial network is generally set relatively low, which is 256×256 in this example. In this way, the high-resolution image can be segmented into multiple low-resolution local images, and the training dataset is established based on small-scale high-definition images.

[0180] The generative adversarial network model is trained based on the established training dataset. Among them, the generalized Hooke's law considering thermal expansion and the strain compatibility condition constrain the gradient update of the generator during the training process. The strain compatibility error, strain distribution, and structure image of the generated strain field are input into the discriminator together, and the discriminator outputs the discrimination error. The network is trained until the error meets the threshold or reaches the predetermined number of steps.

[0181] Prediction stage:

[0182] The microstructure image of the coating sample to be measured is taken by an electron microscope. Semantic segmentation is performed using the same processing method as the training set images, and the image scale is adjusted to the same as the training set. According to the segmented structure image, the corresponding coating material property and boundary condition matrices are established, which are used as input data together with the structure image. In the prediction stage, there is no need to establish a model for each sample again. The trained generative adversarial network is directly used as the structure-strain fitting model. The structure image and the parameter matrix are input into the fitting model, and the strain distribution of the coating can be quickly predicted. When the number of electron microscope images of the sample to be measured is large, the above process of establishing the image and the parameter matrix is repeated until the data volume requirement of the database is met.

[0183] It should be noted that the training phase in the embodiments includes a large amount of work such as data collection, data analysis, and data extraction. Good computing hardware and parallel software design can provide more convenient means for data processing and analysis, and their work content is the prerequisite for the preparation of the thermal state evaluation in the online phase. The work in the online phase is not sensitive to the computing hardware requirements and can be quickly implemented relying on ordinary single-core computers.

[0184] The electronic device of the present invention includes a central processing unit (CPU), which can execute various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or computer program instructions loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0185] Multiple components in the device are connected to the I / O interface, including: an input unit, such as a keyboard, a mouse, etc.; an output unit, such as various types of displays, speakers, etc.; a storage unit, such as a magnetic disk, an optical disc, etc.; and a communication unit, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0186] The processing unit executes the various methods and processes described above, such as steps 1 to 8. For example, in some embodiments, steps 1 to 8 can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device via the ROM and / or the communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more of the steps 1 to 8 described above can be executed. Alternatively, in other embodiments, the CPU can be configured to execute steps 1 to 8 in any other suitable manner (for example, by means of firmware).

[0187] The functions described above herein can be at least partially executed by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on a chip (SOCs), complex programmable logic devices (CPLDs), and so on.

[0188] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to the processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, such that when the program codes are executed by the processor or controller, the functions / operations specified in the flowchart and / or block diagram are implemented. The program codes can be executed entirely on the machine, partially on the machine, executed partially on the machine as an independent software package and partially on a remote machine, or executed entirely on a remote machine or server.

[0189] In the context of the present invention, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0190] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for rapid assessment of the load state of a multiphase composite structure image, characterized in that, it includes the following steps: Obtain the cross-sectional morphology image of the multiphase composite structure, establish an image restoration finite element analysis model, and calculate the load distribution under different boundary conditions and geometric features; Construct a generative adversarial network constrained by the elastic mechanics model and train it based on the data obtained from the finite element analysis model; Evaluate the load state of the multiphase composite structure to be detected based on the trained generative adversarial network, The method specifically includes the following steps: S1: Obtain the cross-sectional morphology image of the multiphase composite structure and construct an image library; S2: Perform semantic segmentation on the images in the image library according to different phases to obtain a semantic segmentation structure diagram; S3: Perform boundary recognition and anti-aliasing optimization on the semantically segmented images; S4: Establish an image restoration finite element model based on the optimized structural boundary coordinates, and calculate the load distribution of the structure under predetermined working condition parameters; S5: Establish an elastic mechanics constraint generative adversarial network model for fitting the relationship between the structural image and the load distribution; S6: Construct a training sample set with the semantic segmentation structure diagram obtained in S2 as the model input and the load distribution obtained in S4 as the model output; S7: Train the generative adversarial network model based on the training sample set; S8: Obtain the cross-sectional morphology image of the multiphase composite structure to be detected and preprocess it based on S2, and send it into the trained generative adversarial network model to obtain the load state, The specific steps of step S5 include: Establish an initial generative adversarial network for pixel-to-pixel conversion based on the pix2pix architecture: establish a multi-layer U-shaped convolutional neural network as the generator G of the generative adversarial network, and establish a downsampling convolutional neural network as the discriminator D of the generative adversarial network, The load distribution is the stress distribution and strain field of the multiphase composite structure. The stress distribution and strain field of the multiphase composite structure are used as outputs respectively to train the generative adversarial network. The specific steps of step S5 include: The loss function of the initial generative adversarial network is denoted as: Among them, train G to maximize the logD(r, G(r)) loss, and train D to minimize logD(r, s), where E e represents the mathematical expectation, λ is the coefficient of the L1 norm loss L 1 , r is the input of the network, s is the true reference of the output, and L 1 (G) is denoted as: L 1 (G) = E r,s [||s - G(r)|| 1 ​ When the prediction target is the stress distribution, s is the finite element stress field σ t ; when the prediction target is the strain distribution, s is the finite element strain field ε t , Introduce the generalized Hooke's law considering thermal expansion, denoted as: where E' = E / (1 - ν 2 ), ν' = ν / (1 - ν), α' = α(1 + ν), G = E / 2(1 + ν), E is the elastic modulus, ε 11 is the elastic strain component in the x-axis direction, ε 22 is the strain component in the y-axis direction, ε 12 is the shear strain in the x-y plane, ν represents the Poisson's ratio, α is the coefficient of thermal expansion, σ represents the stress components in each axial direction, σ 11 is the stress component along the x-axis direction, σ 22 is the stress component in the y-axis direction, σ 12 is the shear stress in the x-y plane, T r represents the room temperature, T 0 represents the initial cooling temperature; Express each stress component in terms of strain components, denoted as: Introduce the deformation compatibility condition of the continuity assumption of elastic mechanics, denoted as: Let the multiphase structure distribution after semantic segmentation in step S2 be x c , and the corresponding strain distribution calculated based on the finite element model in step S4 is set as ε t , combine the three strain components into a three-dimensional array, denoted as: Among them, # represents the switching layer, indicating ε t is composed of ε t,11 , ε t,22 , a three-dimensional array stacked in sequence; The strain array generated by the generator Denoted as: ε m = [ε m,11 #ε m,22 #ε m,12 ​ ε t Both ε and m can be equivalently regarded as a multi-channel image A. Let the convolutional kernel be a multi-dimensional array C. The two-dimensional convolutional calculation for three channels can be expressed as: where ζ and ξ are the receptive field positions, k is the number of channels, * represents convolution, The strain coordination error of the strain image is calculated by the difference method, and any pixel point P ζ,ξ The second-order partial derivatives in each direction are denoted as: The deformation compatibility condition of the continuity assumption of elastic mechanics can be expressed as: Introduce a custom fixed convolution kernel C p , denoted as: Conveniently calculate the strain compatibility error L of the strain field by using convolution operation C , denoted as: where f is the pixel side length size correction coefficient, When the prediction target is the strain distribution, the stress load distribution calculated by finite element is set as the true reference, and the stress-strain relationship error L is calculated based on the stress components H,12 and L H,3 , denoted as: Among them, L H,12 and L H,3 are respectively used to constrain the axial strain and shear strain components. L C L H,12 and L H,3 are simultaneously used to constrain the weight optimization of the generator. Discriminator D loss L of true strain load distribution S,D , denoted as: L S,D = logD(r c , ε t , L C,t ) where r c is the structural image of semantic segmentation, ε t is the true strain distribution corresponding to the structure, L C,t is the strain compatibility error of the true strain distribution The strain distribution generated by the generator G, with its discriminator loss L SG,D , denoted as: L SG,D = log(1 - D(r c , G(r c )), L C,m ) where L C,m is the strain compatibility error of the composite strain distribution; Discriminator loss L D Can be denoted as: where b is the sample batch size, When the prediction target is the stress distribution, the strain distribution calculated by finite element is set as the true reference, and the stress-strain relationship error L is calculated based on the generalized Hooke's law considering thermal expansion H,12 and L H,3 , denoted as: Among them, L H,12 and L H,3 are respectively used to constrain the axial stress and shear stress components. L C , L H,12 and L H,3 are simultaneously used to constrain the weight optimization of the generator. Discriminator D loss \(L\) of the true strain load distribution S,D , denoted as: L S,D = logD(r c ,σ t ,L C,t ) The stress distribution generated by the generator G, and its discriminator loss L SG,D , denoted as: L SG,D = log(1 - D(r c , G(r c )), L C,m ) Among them, L C,m is the strain coordination error of the synthetic strain distribution. After calculating the generated stress distribution as the strain distribution according to the generalized Hooke's law considering thermal expansion, the convolution calculation is then performed based on the strain coordination error; the discriminator loss L D can be denoted as: Generator loss L G including the L1 norm L 1 , and the L of elasticity constraints C , L S,12 , and L S,3 , denoted as: Therefore, the total loss function of the generative adversarial network constrained by elastic mechanics is denoted as: Among them, is the coefficient matrix of the elastodynamic loss L E and L E is denoted as: L E = [L C,m , L S,12 , L S,3 T .​ 2. According to the method for rapid assessment of the load state of a multiphase composite structure image as claimed in claim 1, characterized in that, the specific steps of step S2 include: S21: Process the cross-sectional morphology image with Gaussian filtering; S22: Adjust the image contrast to enhance the gray level difference of each phase; S23: Mark different phase components with different colors.

3. According to the method for rapid assessment of the load state of a multiphase composite structure image as claimed in claim 1, characterized in that, the specific steps of step S3 include: S31: Identify the regions of each phase component in the semantic segmentation structure diagram in step S2 based on the connected domain analysis method; S32: Optimize the anti-aliasing of the recognized structural boundary by using the coordinate averaging method.

4. A method for quickly evaluating the load state of a multiphase composite structure image according to claim 1, wherein, the specific steps of step S4 include: Perform finite element analysis on the structural boundary optimized in step S3 to establish a geometric model of the two-dimensional multiphase structure; set corresponding material properties for each phase component and matrix material, etc.; divide the geometric model using quadrilateral meshes, and the mesh sizes of different group structures are the same; set boundary conditions according to the predetermined working conditions, and submit the calculation of the load distribution of multiple group structures under non-linear conditions.

5. A method for quickly evaluating the load state of a multiphase composite structure image according to claim 1, wherein, the specific steps of step S6 include: Cut the semantic segmentation structure diagram and its corresponding load distribution into multiple groups of local pictures that meet the input picture size conditions of the generative adversarial network established in step S5 to establish a training sample set; establish a corresponding material property and boundary condition matrix according to the structural picture for calculating the elastic mechanics loss during the training process; select 70% of the samples for network training, and the remaining samples for verification.

6. A method for quickly evaluating the load state of a multiphase composite structure image according to claim 1, wherein, the specific steps of step S8 include: Preprocess the high-resolution multiphase composite structure picture to be measured according to step S2, and adjust the image scale to the same size as the training set image; cut the preprocessed structure into multiple groups of local pictures that meet the input picture size conditions of the generative adversarial network established in step S5, and use the network model trained in step S7 to predict their corresponding strain distributions in turn; merge the predicted multiple groups of local strain distributions, and remove the overlapping areas to obtain the complete strain distribution.

7. An electronic device, including a memory and a processor, and a computer program is stored on the memory, wherein, when the processor executes the program, it implements a method for quickly evaluating the load state of a multiphase composite structure image according to any one of claims 1 to 6.

Images