Artificial intelligence-driven three-dimensional digital volume-dependent deformation measurement system and method

Abstract

The invention discloses an artificial intelligence-driven three-dimensional digital volume-dependent deformation measurement system and method. The system comprises a three-dimensional speckle volume data acquisition unit, a three-dimensional coordinate normalization unit, a deep neural network prediction unit, a three-dimensional physical constraint unit and a three-dimensional field reconstruction output unit. The numerical stability of network input is ensured through three-dimensional coordinate normalization, synchronous prediction of a three-dimensional displacement field and a strain tensor field is realized by adopting an end-to-end neural network architecture, and physical rationality between the displacement field and the strain field is ensured by introducing volume reconstruction consistency constraint and three-dimensional deformation compatibility constraint. According to the invention, self-supervised learning is carried out only by relying on volume image pairs before and after deformation and physical constraints, synchronous and high-precision measurement of a three-dimensional displacement field and a strain tensor field is realized, and the problem of error amplification caused by three-dimensional numerical differentiation in a traditional method is avoided; the method is suitable for three-dimensional deformation measurement and analysis in the fields of material science, mechanical engineering, biomedical engineering and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional deformation measurement, in particular to an artificial intelligence driven three-dimensional digital volume related deformation measurement system and method, which is used in the fields of material internal deformation analysis, three-dimensional displacement and strain field measurement, etc. BACKGROUND

[0002] Three-dimensional digital volume correlation (DVC) technology is an extension of digital image correlation technology in three-dimensional space, which measures the internal displacement field and strain field of materials by analyzing three-dimensional volume images before and after deformation. Traditional DVC technology usually adopts a method based on volume sub-region matching, which divides the three-dimensional volume before and after deformation into multiple sub-volumes, and calculates the spatial displacement information of each sub-volume through a correlation algorithm.

[0003] However, the traditional DVC method has the following technical bottlenecks: first, the measurement accuracy is highly dependent on the manual selection of sub-volume size, and improper parameter setting will lead to measurement errors; second, it is difficult for the traditional method to ensure the physical compatibility between the three-dimensional displacement field and the strain tensor field; third, when dealing with large deformation or discontinuous deformation inside the material, the stability and accuracy of the matching algorithm will significantly decrease; finally, the numerical derivation process from the displacement field to the strain field will introduce error accumulation, affecting the reliability of the final measurement results. In recent years, deep learning has made breakthroughs in the field of computer vision, providing a new way to improve traditional DVC technology. However, existing intelligent three-dimensional deformation measurement methods mostly rely on a large number of pre-labeled training samples, which are difficult to obtain in actual engineering applications; at the same time, existing methods lack effective integration of three-dimensional continuous medium mechanics constraints, leading to predicted results that may violate basic physical principles.

[0004] Therefore, it is urgent to develop an intelligent three-dimensional DVC technology that can ensure both measurement accuracy and physical reasonableness, in order to improve the accuracy and engineering practicability of three-dimensional deformation measurement. SUMMARY

[0005] The purpose of the present application is to provide an artificial intelligence driven three-dimensional digital volume correlation deformation measurement method, which includes a three-dimensional speckle volume data acquisition unit, a three-dimensional coordinate normalization unit, a deep neural network prediction unit, a three-dimensional physical constraint unit, and a three-dimensional field reconstruction output unit.

[0006] The three-dimensional speckle volume data acquisition unit is used to read and preprocess the physical material reference speckle volume and deformed speckle volume data obtained by a three-dimensional imaging device.

[0007] The three-dimensional coordinate normalization unit is used to standardize the original three-dimensional voxel coordinates in the reference speckle volume and deformed speckle volume data to a unified range.

[0008] The deep neural network prediction unit trains the deep neural network model using a training sample set to obtain a three-dimensional displacement field and strain field joint prediction model; the training sample set includes multiple groups of three-dimensional voxel coordinate-three-dimensional displacement field and strain field sample pairs;

[0009] The deep neural network prediction unit processes the normalized three-dimensional voxel coordinates using the three-dimensional displacement field and strain field joint prediction model to obtain target three-dimensional displacement field and strain field prediction results;

[0010] During the training of the deep neural network model, the three-dimensional physical constraint unit is used to ensure that the three-dimensional displacement field and strain field satisfy the basic relationship of continuum mechanics;

[0011] The three-dimensional field reconstruction output unit is used to generate and display the three-dimensional distribution results of the displacement field and the strain field.

[0012] Further, the three-dimensional speckle volume data acquisition unit includes a volume data reading module, a voxel intensity normalization module, and a three-dimensional coordinate grid construction module;

[0013] The volume data reading module is used to load the three-dimensional array data of the reference speckle volume and the deformed speckle volume;

[0014] The three-dimensional coordinate grid construction module is used to generate a regularized three-dimensional voxel coordinate system;

[0015] The voxel intensity normalization module is used to standardize the voxel gray value of the three-dimensional array data of the reference speckle volume and the deformed speckle volume, and write the standardized voxel gray value into the three-dimensional voxel coordinate system.

[0016] Further, the three-dimensional coordinate normalization unit normalizes the original three-dimensional voxel coordinates (x, y, z) in the reference speckle volume and the deformed speckle volume data according to the volume dimension, and maps them to [0, 1] 3 standardized space.

[0017] Further, the three-dimensional displacement field and strain field joint prediction model adopts an end-to-end learning architecture, including a multi-layer fully connected network and an activation function layer;

[0018] The three-dimensional displacement field and strain field joint prediction model takes three-dimensional normalized coordinates as input, and takes three spatial displacement components u, v, w and six strain tensor components ε xx , ε yy , ε zz , ε xy , ε xz , ε yz as output.

[0019] Further, the three-dimensional physical constraint unit comprises a volume reconstruction consistency constraint verification module and a three-dimensional deformation compatibility constraint verification module;

[0020] The volume reconstruction consistency constraint verification module calculates image consistency loss L_image=(1 / N)Σ(I_recon(x,y,z)-I_def(x,y,z)) through voxel-level difference between deformed volume and actual deformed volume 2 , to measure accuracy of target three-dimensional displacement field and strain field prediction result;

[0021] Wherein, N is the total number of voxels in the volume, I_recon(x,y,z) is the voxel gray value at coordinate (x,y,z) of the deformed volume reconstructed based on the predicted displacement field, I_def(x,y,z) is the voxel gray value at coordinate (x,y,z) of the actual deformed volume; I_recon(x,y,z)=TrilinearInterpolation(I_ref,sample_x,sample_y,sample_z); sample_x=x-u(x,y,z), sample_y=y-v(x,y,z), sample_z=z-w(x,y,z) is the coordinate of inverse mapping; TrilinearInterpolation is a trilinear interpolation operation.

[0022] The three-dimensional deformation compatibility constraint verification module calculates physical compatibility loss L_physics=Σ(ε_pred-ε_calc) according to the relationship between three-dimensional displacement gradient and strain tensor prediction value 2 , so as to determine whether the three-dimensional displacement gradient and strain tensor prediction value meet the basic relationship between strain tensor and displacement gradient in three-dimensional continuous medium mechanics;

[0023] Wherein, ε_pred is the strain tensor prediction value directly output by the deep neural network, containing six components: ε_pred=[εxx_pred,εyy_pred,εzz_pred,εxy_pred,εxz_pred,εyz_pred]; ε_calc is the strain tensor value calculated through the relationship of continuous medium mechanics according to the displacement field predicted by the neural network; ε_calc=[εxx_calc,εyy_calc,εzz_calc,εxy_calc,εxz_calc,εyz_calc], wherein u, v, w are displacement components in x, y, z directions respectively, ε xx , ε yy , ε zz are normal strain components, and εxy , ε xz , ε yz are shear strain components.

[0024] Further, the three-dimensional field reconstruction output unit comprises a three-dimensional displacement field visualization module, a three-dimensional strain field visualization module, a volume slice display module, an equivalent strain calculation module and a stereoscopic data export module.

[0025] The three-dimensional displacement field visualization module is used to show the displacement distribution in the u, v and w directions.

[0026] The three-dimensional strain field visualization module is used to show the spatial distribution of the six strain tensor components.

[0027] The volume slice display module is used to show the field distribution at different depth levels.

[0028] The equivalent strain calculation module is used to calculate the three-dimensional equivalent strain indicator.

[0029] The three-dimensional equivalent strain indicator von_Mises = Sqrt(0.5x[(ε xx -ε yy ) 2 +(ε yy -ε zz ) 2 +(ε zz -ε xx ) 2 +6x(ε xy 2 +ε xz 2 +ε yz 2 ]) where ε xx , ε yy , ε zz are principal strain components, and ε xy , ε xz , ε yz are shear strain components.

[0030] The stereoscopic data export module is used to save the three-dimensional analysis results.

[0031] A method of an artificial intelligence driven three-dimensional digital volume related deformation measurement system, comprising the following steps:

[0032] Step 1: load the reference speckle volume and the deformed speckle volume, pre-process, and create a three-dimensional normalized coordinate grid;

[0033] Step 2: normalize the three-dimensional voxel coordinates and convert them into standardized inputs;

[0034] Step 3: Constructing the deep neural network model and setting the architecture parameters; the architecture parameters include the number of network layers and the number of neurons;

[0035] Step 4: Defining the hybrid loss function L_total = λ_image × L_image + λ_physics × L_physics containing the volume gray value reconstruction consistency constraint and the three-dimensional deformation compatibility constraint; wherein λ_image is the image consistency weight, and λ_physics is the physical constraint weight;

[0036] Step 5: Training the deep neural network model end to end, optimizing the network weight parameters until the loss function reaches the convergence standard;

[0037] Step 6: Using the trained model to predict the three-dimensional displacement field and strain tensor field, and calculating the three-dimensional equivalent strain;

[0038] Step 7: Generating the three-dimensional field distribution stereoscopic visualization result, and saving the displacement field and strain field.

[0039] Further, the three-dimensional coordinate normalization in step 2 includes: normalizing the original voxel coordinates (x, y, z) according to the volume dimension to obtain the standardized coordinate representation.

[0040] Further, the three-dimensional deformation compatibility constraint in step 4 is as follows:

[0041] L_physics = Σ (ε_pred - ε_calc) 2 (1)

[0042] Further, the three-dimensional equivalent strain in step 6 is as follows:

[0043] von_Mises = Sqrt (0.5 × [(ε xx -ε yy ) 2 +(ε yy -ε zz ) 2 +(ε zz -ε xx ) 2 +6×(ε xy 2 +

[0044] ε xz 2 +ε yz 2 )])(2)

[0045] Wherein, ε xx , ε yy , ε zz are the principal strain components, and εxy ε xz ε yz For shear strain components.

[0046] The technical effects of this invention are undeniable, and its beneficial effects are as follows:

[0047] (1) The present invention adopts a three-dimensional coordinate normalization method, which significantly improves the model's ability to express complex three-dimensional deformation patterns and measurement accuracy through the nonlinear mapping capability of neural networks compared with traditional volume sub-region matching.

[0048] (2) This invention, through an end-to-end learning strategy, can simultaneously focus on the local details and overall features of the three-dimensional deformation field, thereby improving the stability and robustness of the measurement.

[0049] (3) The present invention integrates three-dimensional physical constraints to ensure physical compatibility between the displacement field and the strain tensor field, and avoids unreasonable predictions that violate the principles of three-dimensional continuous medium mechanics.

[0050] (4) This invention does not require pre-annotated 3D training data, but relies solely on the volume image pairs before and after deformation and physical constraints for self-supervised learning, which significantly improves the engineering practicality of the method.

[0051] (5) This invention realizes the synchronous and high-precision measurement of the three-dimensional displacement field and the strain tensor field, avoiding the error amplification problem caused by three-dimensional numerical differentiation in traditional methods;

[0052] (6) This invention supports complete three-dimensional strain tensor measurement, including six independent components, providing technical support for the comprehensive characterization of complex deformation behavior inside materials. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the system architecture of the AI-driven three-dimensional digital volume correlation deformation measurement technology of the present invention;

[0054] Figure 2 This is a schematic diagram of the workflow of the three-dimensional coordinate normalization unit of the present invention;

[0055] Figure 3 This is a schematic diagram of the deep neural network prediction unit structure of the present invention;

[0056] Figure 4 This is a schematic diagram illustrating the working principle of the three-dimensional physical constraint unit of the present invention;

[0057] Figure 5 This is a schematic diagram of the three-dimensional volume reconstruction and consistency verification of the present invention; Detailed Implementation

[0058] The application will be further described below with reference to the drawings, but should not be understood as limiting the above-mentioned subject matter of the application to the following examples. Various replacements and changes can be made according to ordinary technical knowledge and conventional means in the art without departing from the above-mentioned technical idea of the application, and all of them should be included in the protection scope of the application.

[0059] Embodiment 1

[0060] Referring to Figures 1 to 5 An artificial intelligence driven three-dimensional digital volume related deformation measurement method comprises a three-dimensional speckle volume data acquisition unit, a three-dimensional coordinate normalization unit, a deep neural network prediction unit, a three-dimensional physical constraint unit and a three-dimensional field reconstruction output unit.

[0061] The three-dimensional speckle volume data acquisition unit is used to read and preprocess the physical material reference speckle volume and deformation speckle volume data acquired by a three-dimensional imaging device.

[0062] The volume data is acquired by a three-dimensional imaging device such as CT scanning, MRI or optical coherence tomography, and is derived from the physical deformation process of actual engineering materials and biological tissues, including three-dimensional volume imaging data of real physical phenomena such as internal deformation of metal material samples under mechanical load, interlaminar shear deformation of composite materials, crack propagation of concrete and rock, deformation of biological soft tissues under physiological load, and local stress concentration of engineering structural parts.

[0063] The three-dimensional coordinate normalization unit is used to standardize the original three-dimensional voxel coordinates in the reference speckle volume and deformation speckle volume data to a unified range.

[0064] The deep neural network prediction unit trains the deep neural network model using a training sample set to obtain a three-dimensional displacement field and strain field joint prediction model; the training sample set includes multiple groups of three-dimensional voxel coordinate-three-dimensional displacement field and strain field sample pairs.

[0065] The deep neural network prediction unit processes the standardized three-dimensional voxel coordinates using the three-dimensional displacement field and strain field joint prediction model to obtain target three-dimensional displacement field and strain field prediction results.

[0066] During the training process of the deep neural network model, the three-dimensional physical constraint unit is used to ensure that the three-dimensional displacement field and strain field satisfy the basic relationship of continuum mechanics.

[0067] The three-dimensional field reconstruction output unit is used to generate and display the three-dimensional displacement field and strain field stereoscopic distribution results.

[0068] Embodiment 2

[0069] An artificial intelligence driven three-dimensional digital volume related deformation measurement method, the technical content is same as embodiment 1, further, the three-dimensional speckle volume data acquisition unit includes a volume data reading module, a voxel intensity normalization module and a three-dimensional coordinate grid construction module;

[0070] The volume data reading module is used to load the three-dimensional array data of the reference speckle volume and the deformed speckle volume;

[0071] The three-dimensional coordinate grid construction module is used to generate a regularized three-dimensional voxel coordinate system;

[0072] The voxel intensity normalization module is used to standardize the voxel gray value of the three-dimensional array data of the reference speckle volume and the deformed speckle volume, and write the standardized voxel gray value into the three-dimensional voxel coordinate system.

[0073] Embodiment 3:

[0074] An artificial intelligence driven three-dimensional digital volume related deformation measurement method, the technical content is same as any one of embodiments 1-2, further, the three-dimensional coordinate normalization unit normalizes the original three-dimensional voxel coordinates (x, y, z) in the reference speckle volume and the deformed speckle volume data according to the volume dimension, and maps to [0, 1] 3 standardized space.

[0075] Embodiment 4:

[0076] An artificial intelligence driven three-dimensional digital volume related deformation measurement method, the technical content is same as any one of embodiments 1-3, further, the three-dimensional displacement field and strain field joint prediction model adopts an end-to-end learning architecture, including a multi-layer fully connected network and an activation function layer;

[0077] The three-dimensional displacement field and strain field joint prediction model takes three-dimensional normalized coordinates as input, and takes three spatial displacement components u, v, w and six strain tensor components xx , ε yy , ε zz , ε xy , ε xz , ε yz as output.

[0078] Embodiment 5:

[0079] An artificial intelligence driven three-dimensional digital volume related deformation measurement method, the technical content is same as any one of embodiments 1-4, further, the three-dimensional physical constraint unit includes a volume reconstruction consistency constraint verification module and a three-dimensional deformation compatibility constraint verification module;

[0080] The volume reconstruction consistency constraint verification module calculates image consistency loss L_image = (1 / N)∑(I_recon(x,y,z)-I_def(x,y,z)) through voxel-level difference between the deformed volume and the actual deformed volume 2 , which measures the accuracy of the predicted three-dimensional displacement field and strain field; wherein N is the total number of voxels in the volume, I_recon(x,y,z) is the voxel gray value of the deformed volume reconstructed based on the predicted displacement field at coordinates (x,y,z), and I_def(x,y,z) is the voxel gray value of the actual deformed volume at coordinates (x,y,z); the voxel gray value I_recon(x,y,z) of the reconstructed volume is obtained by inverse mapping coordinate transformation, specifically: sample_x = x-u(x,y,z), sample_y = y-v(x,y,z), sample_z = z-w(x,y,z), and then using trilinear interpolation method to obtain from the reference speckle volume I_ref: I_recon(x,y,z) = TrilinearInterpolation(I_ref,sample_x,sample_y,sample_z);

[0081] The three-dimensional deformation compatibility constraint verification module calculates physical compatibility loss L_physics = ∑(ε_pred-ε_calc) according to the relationship between the predicted values of three-dimensional displacement gradient and strain tensor 2 , so as to determine whether the predicted values of three-dimensional displacement gradient and strain tensor satisfy the basic relationship between strain tensor and displacement gradient in three-dimensional continuous medium mechanics; wherein ε_pred is the predicted value of strain tensor directly output by the deep neural network, including six components: ε_pred = [εxx_pred,εyy_pred,εzz_pred,εxy_pred,εxz_pred,εyz_pred]; ε_calc is the strain tensor value calculated from the predicted displacement field of the neural network through the relationship of continuous medium mechanics, including six components: ε_calc = [εxx_calc,εyy_calc,εzz_calc,εxy_calc,εxz_calc,εyz_calc], wherein The physical compatibility loss function is expanded as: L_physics = ∑[(εxx_pred-εxx_calc) 2 +(εyy_pred-εyy_calc) 2 +(εzz_pred-εzz_calc) 2 +(εxy_pred-εxy_calc) 2+(εxz_pred-εxz_calc) 2 +(εyz_pred-εyz_calc) 2 ];

[0082] The relationship between the three-dimensional displacement gradient and the strain tensor prediction value is as follows:

[0083]

[0084] wherein u, v, w are displacement components in x, y, z directions respectively, ε xx , ε yy , ε zz are normal strain components, and ε xy , ε xz , ε yz are shear strain components.

[0085] Embodiment 6:

[0086] An artificial intelligence driven three-dimensional digital volume related deformation measurement method, the technical content is the same as any one of embodiments 1-5, further, the three-dimensional field reconstruction output unit comprises a three-dimensional displacement field visualization module, a three-dimensional strain field visualization module, a volume slice display module, an equivalent strain calculation module and a stereoscopic data export module.

[0087] The three-dimensional displacement field visualization module is used to show the displacement distribution in u, v, w three directions.

[0088] The three-dimensional strain field visualization module is used to show the spatial distribution of six strain tensor components.

[0089] The volume slice display module is used to show the field distribution of different depth levels.

[0090] The equivalent strain calculation module is used to calculate a three-dimensional equivalent strain index.

[0091] The three-dimensional equivalent strain index von_Mises=Sqrt(0.5×[(ε xx -ε yy ) 2 +(ε yy -ε zz ) 2 +(ε zz -ε xx ) 2 +6×(ε xy 2 +ε xz 2 +ε yz 2 )]) wherein ε xx , ε yy , εzz ε is the principal strain component xy ε is the principal strain component xz ε is the principal strain component yz ε is the shear strain component

[0092] The stereoscopic data derivation module is configured to save the three-dimensional analysis result.

[0093] Embodiment 7:

[0094] A method of a three-dimensional digital volume related deformation measurement system driven by artificial intelligence according to any one of embodiments 1-6, comprising the following steps:

[0095] Step 1: loading a reference speckle volume and a deformed speckle volume, preprocessing, and creating a three-dimensional normalized coordinate grid;

[0096] Step 2: normalizing the three-dimensional voxel coordinates and converting them into standardized inputs;

[0097] Step 3: constructing a deep neural network model and setting architecture parameters; the architecture parameters include the number of network layers and the number of neurons;

[0098] Step 4: defining a hybrid loss function L_total = λ_image x L_image + λ_physics x L_physics containing volume gray value reconstruction consistency constraints and three-dimensional deformation compatibility constraints; wherein λ_image is the image consistency weight, and λ_physics is the physical constraint weight;

[0099] Step 5: training the deep neural network model end-to-end, optimizing the network weight parameters, and stopping until the loss function reaches the convergence standard;

[0100] Step 6: predicting the three-dimensional displacement field and strain tensor field using the trained model, and calculating the three-dimensional equivalent strain;

[0101] Step 7: generating a stereoscopic visualization result of the three-dimensional field distribution, and saving the displacement field and strain field.

[0102] Embodiment 8:

[0103] A method of a three-dimensional digital volume related deformation measurement system driven by artificial intelligence according to any one of embodiments 1-6, the technical content is the same as embodiment 7, further, the three-dimensional coordinate normalization in step 2 includes: normalizing the original voxel coordinates (x, y, z) according to the volume dimension to obtain the standardized coordinate representation.

[0104] Embodiment 9:

[0105] A method based on the artificial intelligence driven three-dimensional digital volume related deformation measurement system according to any one of embodiments 1-6, the technical content is the same as any one of embodiments 7-8, further, the three-dimensional deformation compatibility constraint in step 4 is as follows:

[0106] L_physics =∑(ε_pred-ε_calc) 2 (6)

[0107]

[0108] where u, v, w are displacement components in x, y, z directions respectively, ε xx , ε yy , ε zz are normal strain components, and ε xy , ε xz , ε yz are shear strain components.

[0109] Embodiment 10:

[0110] A method based on the artificial intelligence driven three-dimensional digital volume related deformation measurement system according to any one of embodiments 1-6, the technical content is the same as any one of embodiments 7-9, further, the three-dimensional equivalent strain in step 6 is as follows:

[0111] von_Mises = Sqrt(0.5×[(ε xx -ε yy ) 2 +(ε yy -ε zz ) 2 +(ε zz -ε xx ) 2 +6×(ε xy 2 +

[0112] ε xz 2 +ε yz 2 )])(12)

[0113] where ε xx , ε yy , ε zz are principal strain components, and ε xy , ε xz , ε yz are shear strain components.

[0114] Embodiment 11:

[0115] An artificial intelligence driven three-dimensional digital volume related deformation measurement system, which realizes high-precision and high-reliability three-dimensional displacement field and strain tensor field synchronous measurement by fusing deep learning, spatial three-dimensional coordinates and physical constraints.

[0116] The three-dimensional digital volume related deformation measurement system comprises a three-dimensional speckle volume data acquisition unit, a three-dimensional coordinate normalization unit, a deep neural network prediction unit, a three-dimensional physical constraint unit and a three-dimensional field reconstruction output unit; wherein:

[0117] The three-dimensional speckle volume data acquisition unit is used for reading and preprocessing reference speckle volume and deformed speckle volume data.

[0118] The three-dimensional coordinate normalization unit is used for standardizing three-dimensional voxel coordinates to a unified range, enhancing network training stability.

[0119] The deep neural network prediction unit is used for constructing an end-to-end three-dimensional displacement field and strain field joint prediction model.

[0120] The three-dimensional physical constraint unit is used for ensuring that the three-dimensional displacement field and strain field satisfy the basic relationship of continuum mechanics.

[0121] The three-dimensional field reconstruction output unit is used for generating and displaying the three-dimensional displacement field and strain field stereoscopic distribution results.

[0122] The three-dimensional speckle volume data acquisition unit comprises a volume data reading module, a voxel intensity normalization module and a three-dimensional coordinate grid construction module; the volume data reading module is used for loading three-dimensional speckle volume array data; the voxel intensity normalization module is used for standardizing voxel gray value; and the three-dimensional coordinate grid construction module is used for generating a regular three-dimensional voxel coordinate system.

[0123] The three-dimensional coordinate normalization unit normalizes the original three-dimensional coordinates (x, y, z) according to the volume dimension, and maps them to [0, 1] 3 standard space, ensuring the numerical stability and convergence of neural network input data.

[0124] The deep neural network prediction unit adopts an end-to-end learning architecture, including a multi-layer fully connected network and an activation function layer; the network accepts three-dimensional normalized coordinates as input, and synchronously outputs nine components: three spatial displacement components u, v, w and six strain tensor components xx , yy , zz , xy , xz , yz , realizing parallel prediction of displacement field and strain field.

[0125] The three-dimensional physical constraint unit includes a volume gray value reconstruction consistency constraint and a three-dimensional deformation compatibility constraint; the volume gray value reconstruction consistency constraint is measured by a voxel-level difference between a reconstructed deformed volume gray value and an actual deformed volume gray value; the three-dimensional deformation compatibility constraint is realized by a relationship between a three-dimensional displacement gradient and a strain tensor prediction value, and ensures that a basic relationship between the strain tensor and the displacement gradient in three-dimensional continuous medium mechanics is met, that is:

[0126]

[0127] wherein u, v and w are displacement components in x, y and z directions respectively, ε xx , ε yy , and ε zz are normal strain components, and ε xy , ε xz , and ε yz are shear strain components.

[0128] The three-dimensional field reconstruction output unit includes a three-dimensional displacement field visualization module, a three-dimensional strain field visualization module, a volume slice display module, an equivalent strain calculation module and a three-dimensional data export module; the three-dimensional displacement field visualization module is used to display displacement distributions in u, v and w directions; the three-dimensional strain field visualization module is used to display spatial distributions of six strain tensor components; the volume slice display module is used to display field distributions at different depth levels; the equivalent strain calculation module is used to calculate a three-dimensional equivalent strain index; and the three-dimensional data export module is used to save three-dimensional analysis results.

[0129] Embodiment 12:

[0130] An implementation method of an artificial intelligence driven three-dimensional digital volume related deformation measurement system, comprising the following steps:

[0131] Step 1: loading a reference speckle volume and a deformed speckle volume, preprocessing, and creating a three-dimensional normalized coordinate grid;

[0132] Step 2: normalizing and encoding the three-dimensional voxel coordinates;

[0133] Step 3: constructing a deep fusion neural network model, setting architecture parameters such as network layer number and neuron number;

[0134] Step 4: defining a hybrid loss function including a volume reconstruction consistency constraint and a three-dimensional deformation compatibility constraint;

[0135] Step 5: training the model end to end, optimizing network weight parameters, until the loss function reaches a convergence standard;

[0136] Step 6: predicting a three-dimensional displacement field and a strain tensor field using the trained model, and calculating a three-dimensional equivalent strain;

[0137] Step 7: Generate stereoscopic visualization results of three-dimensional field distribution, and save displacement field, strain field, and other volume data.

[0138] Example 13:

[0139] As Figure 1 shown, the present application proposes an artificial intelligence driven three-dimensional digital volume related deformation measurement system, which includes a three-dimensional speckle volume data acquisition unit, a three-dimensional coordinate normalization unit, a deep neural network prediction unit, a three-dimensional physical constraint unit, and a three-dimensional field reconstruction output unit.

[0140] The three-dimensional speckle volume data acquisition unit is responsible for reading and preprocessing three-dimensional measurement data. As Figure 1 shown, this unit loads the three-dimensional speckle volume of the reference state and the deformation state through the volume data reading module, normalizes the voxel intensity to normalize the gray value range, and then generates a regularized three-dimensional voxel coordinate grid, laying the foundation for subsequent three-dimensional analysis. In actual operation, the data acquisition unit first reads the three-dimensional speckle volume, obtains the volume size D x H x W, and then creates a three-dimensional normalized coordinate grid, normalizing the coordinate value to the range of [0, 1] 3 , ensuring the numerical stability of subsequent processing.

[0141] The three-dimensional coordinate normalization unit, as Figure 2 shown, adopts a simple and effective coordinate standardization method. This method normalizes the original three-dimensional voxel coordinates (x, y, z) according to the volume dimensions and maps them to the normalized three-dimensional space [0, 1] 3 . Through this normalization method, the neural network can obtain numerically stable input, improving training convergence and prediction accuracy.

[0142] The deep neural network prediction unit, as Figure 3 shown, adopts an end-to-end learning architecture. The network accepts three-dimensional normalized coordinates (x, y, z) as input, processes them through multiple layers of fully connected networks and nonlinear activation functions, and simultaneously outputs nine components: three spatial displacement components (u, v, w) and six strain tensor components (ε xx , ε yy , ε zz , ε xy , ε xz , ε yz . As Figure 3 shown, the network includes an input layer, multiple hidden layers, and an output layer. Each hidden layer uses a hyperbolic tangent activation function to improve nonlinear modeling capability. The output layer is directly mapped to 9 channels, realizing parallel prediction of displacement field and strain tensor field.

[0143] The three-dimensional physical constraint unit, as Figure 4 shown, adopts a simple and effective physical constraint method. This method uses the continuity of the displacement field and the strain tensor field to constrain the output of the neural network, ensuring the physical consistency of the predicted results.As shown, it includes two parts, volume reconstruction consistency constraint and three-dimensional deformation compatibility constraint. The volume reconstruction consistency constraint is realized by comparing the voxel-level mean square error between the reconstructed deformation volume and the actual deformation volume, ensuring the accuracy of the displacement field prediction. As shown on the left, Figure 4 As shown on the left, this constraint first reconstructs the deformation volume based on the predicted three-dimensional displacement field (u, v, w) through trilinear interpolation, and then calculates the difference between the reconstructed volume and the actual deformation volume as an important part of the loss function. The three-dimensional deformation compatibility constraint is realized by comparing the difference between the strain tensor calculated by the three-dimensional displacement gradient and the strain tensor directly predicted by the model, ensuring that the displacement field and the strain field satisfy the basic relationship of three-dimensional continuous medium mechanics. Figure 4

[0144] The three-dimensional field reconstruction output unit is responsible for generating and displaying three-dimensional analysis results. This unit includes a three-dimensional displacement field visualization module, a three-dimensional strain field visualization module, a volume slice display module, an equivalent strain calculation module, and a stereoscopic data export module. The three-dimensional displacement field visualization module displays the spatial distribution of the u, v, and w displacement components; the three-dimensional strain field visualization module displays the three-dimensional distribution of the six strain tensor components; the volume slice display module provides field distribution slice views at different depth levels; the equivalent strain calculation module calculates the von Mises equivalent strain based on the strain tensor; and the stereoscopic data export module supports saving the three-dimensional analysis results in volume data format.

[0145] The three-dimensional field reconstruction output unit is responsible for generating and displaying three-dimensional analysis results. This unit includes a three-dimensional displacement field visualization module, a three-dimensional strain field visualization module, a volume slice display module, an equivalent strain calculation module, and a stereoscopic data export module. The three-dimensional displacement field visualization module displays the spatial distribution of the u, v, and w displacement components; the three-dimensional strain field visualization module displays the three-dimensional distribution of the six strain tensor components; the volume slice display module provides field distribution slice views at different depth levels; the equivalent strain calculation module calculates the von Mises equivalent strain based on the strain tensor; and the stereoscopic data export module supports saving the three-dimensional analysis results in volume data format.

[0146] Example 14:

[0147] The three-dimensional digital volume-related deformation measurement technology implementation method of the present application includes the following specific steps:

[0148] Step 1: Load three-dimensional speckle volume data and perform preprocessing

[0149] First, read the three-dimensional speckle volume data of the reference state and the deformation state, check whether the sizes of the two volumes are consistent, then standardize the voxel gray value to an appropriate range, and create a normalized three-dimensional voxel coordinate grid system. As shown on the left, Figure 1 ​As shown, this step establishes the basic data structure for subsequent 3D analysis.

[0150] Step 2: Perform 3D coordinate normalization.

[0151] like Figure 2 As shown, the three-dimensional voxel coordinates (x, y, z) are normalized according to the volume dimensions D×H×W and mapped to [0, 1]. 3 Standardization space. The specific calculation formulas are: x_norm=x / (W-1), y_norm=y / (H-1), z_norm=z / (D-1), to ensure the numerical stability of the network input.

[0152] Step 3: Construct a deep neural network model

[0153] like Figure 3 As shown, a deep neural network prediction unit is constructed, and architectural parameters such as the number of network layers and neurons are set. The network adopts an end-to-end learning architecture, including an input layer, multiple hidden layers, and an output layer. The input layer accepts three-dimensional normalized coordinates as input and contains three neurons corresponding to x_norm, y_norm, and z_norm, respectively. The hidden layer adopts a four-layer fully connected network structure, with 40 neurons in each layer, and uses the hyperbolic tangent (Tanh) activation function to improve nonlinear modeling capability. The output layer contains nine neurons, simultaneously outputting three spatial displacement components u, v, and w, and six strain tensor components εxx, εyy, εzz, εxy, εxz, and εyz, realizing parallel prediction of displacement and strain fields. The network weights adopt the Xavier uniform initialization method, and the biases are initialized with zero to maintain the stability of the variance of activation values ​​and gradients during forward and backward propagation.

[0154] Step 4: Define the composite loss function for physical information constraints

[0155] like Figure 4 As shown, a comprehensive loss function is constructed by integrating volume reconstruction consistency constraints and 3D deformation compatibility constraints. The volume reconstruction consistency constraint measures the difference between the reconstructed deformed volume and the actual deformed volume, ensuring the accuracy of the displacement field prediction. The specific implementation process is as follows: First, based on the predicted 3D displacement field (u,v,w), the deformed volume is reconstructed using a trilinear interpolation algorithm, employing a reverse mapping coordinate calculation method: sample_x=xu, ​​sample_y=yv, sample_z=zw. Then, trilinear interpolation is used to obtain the voxel values ​​of the reconstructed volume from the reference volume. Finally, the mean square error between the reconstructed volume and the actual deformed volume is calculated as the image consistency loss: L_image=(1 / N)Σ(I_recon(x,y,z)-I_def(x,y,z)). 2 .

[0156] The three-dimensional deformation compatibility constraint is realized by comparing the difference between the strain tensor calculated by the three-dimensional displacement gradient and the strain tensor directly predicted by the model, ensuring that the displacement field and the strain field satisfy the basic relationship of three-dimensional continuous medium mechanics. The central difference method is used to calculate the displacement gradient: u_x = (u[i,j,k+1]-u[i,j,k-1]) / 2, u_y = (u[i,j+1,k]-u[i,j-1,k]) / 2, u_z = (u[i+1,j,k]-u[i-1,j,k]) / 2, and the gradients of v and w in each direction are calculated in the same way. Based on the displacement gradient, the strain components are calculated: ε_xx = u_x, ε_yy = v_y, ε_zz = w_z, ε_xy = 0.5×(u_y+v_x), ε_xz = 0.5×(u_z+w_x), ε_yz = 0.5×(v_z+w_y). The physical compatibility loss is: L_physics = Σ(ε_pred-ε_calc) 2 .

[0157] The comprehensive loss function is defined as: L_total = λ_image×L_image+λ_physics×L_physics, where λ_image is the image consistency weight, and λ_physics is the physical constraint weight.

[0158] Step 5: Iterative training of the model

[0159] The model is trained end-to-end, and the network weight parameters are optimized until the loss function reaches the convergence criteria. The Adam optimization algorithm is used, with an initial learning rate and adaptive learning rate adjustment capability. During training, a learning rate scheduling strategy is used to automatically reduce the learning rate when the loss function stops decreasing, improving optimization efficiency. Set the total training period, calculate and print the training progress, including the total loss, image consistency loss and physical constraint loss values. The training process uses a full voxel parallel processing method, which simultaneously inputs all D×H×W voxel coordinates into the network for batch prediction, improving training efficiency. When the loss function remains stable for multiple epochs or the learning rate drops to the minimum threshold, the training is determined to be converged.

[0160] Step 6: Use the trained model to predict the three-dimensional displacement field and strain tensor field

[0161] The trained model is used to predict displacement field and strain field for the entire three-dimensional volume, and three-dimensional equivalent strain is calculated. The trained neural network is switched to evaluation mode, and the normalized three-dimensional coordinates are input to obtain the predicted output of 9 components. The prediction results are reshaped into a volume shape: the displacement field shape is DxHxWx3, and the strain field shape is DxHxWx6. Based on the predicted displacement field, the deformed volume is reconstructed by trilinear interpolation, and the reconstruction quality is verified. The three-dimensional equivalent strain is calculated, and the von Mises strain formula is used: von_Mises=sqrt(0.5×[(εxx-εyy) 2 +(εyy-εzz) 2 +(εzz-εxx) 2 +6×(εxy 2 +εxz 2 +εyz 2 )]) to provide a comprehensive strain index for material mechanics analysis.

[0162] Step 7: Generate three-dimensional field distribution visualization results, and save displacement field, strain field, and volume data

[0163] As shown in Figure 5 , generate complete three-dimensional visualization results and save analysis data. The visualization process includes: three-dimensional displacement field visualization, which respectively shows the spatial distribution of u, v, and w displacement components, and uses pseudo-color maps and isosurface rendering technology; three-dimensional strain field visualization, which shows the three-dimensional distribution of six strain tensor components, and uses volume rendering and slice display technology; volume slice display, which provides XY, XZ, and YZ slice views in three main directions, and supports field distribution observation at any depth level; equivalent strain calculation and visualization, which calculates the von Mises equivalent strain based on the strain tensor and generates a three-dimensional cloud chart.

[0164] Data saving includes: saving the predicted displacement field and strain field data in.npz format for subsequent analysis and processing; generating CSV format numerical data files to support data exchange with other analysis software; saving visualization images in PNG and PDF formats to provide high-quality technical report charts; recording reconstruction quality indicators, including mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient evaluation indicators; generating a technical analysis report to summarize key technical parameters such as measurement accuracy, spatial resolution, and computing performance.

[0165] The artificial intelligence driven three-dimensional digital volume related deformation measurement technology of the present application realizes high-precision and high-reliability three-dimensional displacement field and strain tensor field synchronous measurement by fusing three-dimensional coordinate normalization, deep neural network modeling and three-dimensional physical information constraint. Compared with the traditional method based on volume sub-region correlation, the present application does not need to manually set parameters such as sub-volume size, and automatically adapts to different three-dimensional deformation modes; compared with the existing intelligent DVC method, the present application ensures the physical reasonableness of the measurement results through three-dimensional physical information constraint, and improves the reliability and accuracy of the measurement. The present application is suitable for three-dimensional deformation measurement and analysis in the fields of material science, mechanical engineering, civil engineering, biomedical engineering and the like, and is particularly suitable for processing complex conditions such as large deformation and internal discontinuous deformation of materials, and provides reliable technical support for material three-dimensional mechanical property evaluation, structure internal safety monitoring, biological tissue deformation analysis and the like applications.

Claims

1. An artificial intelligence driven three-dimensional digital volume related deformation measurement method, characterized in that, The three-dimensional speckle volume data acquisition unit, the three-dimensional coordinate normalization unit, the deep neural network prediction unit, the three-dimensional physical constraint unit and the three-dimensional field reconstruction output unit are included. The three-dimensional speckle volume data acquisition unit is used to read and preprocess the physical material reference speckle volume and the deformed speckle volume data acquired by the three-dimensional imaging device. The three-dimensional coordinate normalization unit is used to standardize the original three-dimensional voxel coordinates in the reference speckle volume and the deformed speckle volume data to a unified range. The deep neural network prediction unit trains the deep neural network model by using a training sample set to obtain a three-dimensional displacement field and strain field joint prediction model; the training sample set includes a plurality of three-dimensional voxel coordinate-three-dimensional displacement field and strain field sample pairs. The deep neural network prediction unit processes the standardized three-dimensional voxel coordinates by using the three-dimensional displacement field and strain field joint prediction model to obtain a target three-dimensional displacement field and strain field prediction result. In the deep neural network model training process, the three-dimensional physical constraint unit is used to ensure that the three-dimensional displacement field and the strain field satisfy the basic relationship of continuum mechanics. The three-dimensional field reconstruction output unit is used to generate and display the three-dimensional displacement field and the strain field stereoscopic distribution result.

2. The artificial intelligence driven three-dimensional digital volume related deformation measurement system of claim 1, wherein, The three-dimensional speckle volume data acquisition unit includes a volume data reading module, a voxel intensity normalization module and a three-dimensional coordinate grid construction module. The volume data reading module is used to load the three-dimensional array data of the reference speckle volume and the deformed speckle volume. The three-dimensional coordinate grid construction module is used to generate a regularized three-dimensional voxel coordinate system. The voxel intensity normalization module is used to standardize the voxel gray value of the three-dimensional array data of the reference speckle volume and the deformed speckle volume, and write the standardized voxel gray value into the three-dimensional voxel coordinate system.

3. The artificial intelligence driven three-dimensional digital volume related deformation measurement system of claim 1, wherein, The three-dimensional coordinate normalization unit normalizes the original three-dimensional voxel coordinates (x, y, z) in the reference speckle volume and the deformed speckle volume data according to the volume dimension, and maps to [0, 1] according to the volume dimension. 3 Standardized space.

4. The artificial intelligence driven three-dimensional digital volume related deformation measurement system of claim 1, wherein, The three-dimensional displacement field and strain field joint prediction model adopts an end-to-end learning architecture, including a multi-layer fully connected network and an activation function layer. The three-dimensional displacement field and strain field joint prediction model takes three-dimensional normalized coordinates as input, and takes three spatial displacement components u, v, w and six strain tensor components ε xx 、 yy 、 zz 、 xy 、 xz 、 yz as output.

5. The artificial intelligence driven three-dimensional digital volume related deformation measurement system of claim 1, wherein, The three-dimensional physical constraint unit includes a volume reconstruction consistency constraint verification module and a three-dimensional deformation compatibility constraint verification module. The volume reconstruction consistency constraint verification module calculates image consistency loss L_image=(1 / N)∑(I_recon(x,y,z)-I_def(x,y,z)) through voxel-level difference between deformed volume and actual deformed volume 2 , measure target three-dimensional displacement field and strain field prediction result accuracy; Wherein, N is the total number of voxels in the volume, I_recon(x,y,z) is the voxel gray value of the deformed volume reconstructed based on the predicted displacement field at coordinates (x,y,z), I_def(x,y,z) is the voxel gray value of the actual deformed volume at coordinates (x,y,z); I_recon(x,y,z) = TrilinearInterpolation(I_ref,sample_x,sample_y,sample_z); sample_x = x-u(x,y,z), sample_y = y-v(x,y,z), sample_z = z-w(x,y,z) are the coordinates of the inverse mapping; TrilinearInterpolation is a trilinear interpolation operation. The three-dimensional deformation compatibility constraint verification module calculates a physical compatibility loss L_physics=∑(ε_pred-ε_calc) according to a relationship between the three-dimensional displacement gradient and the strain tensor prediction value 2 , so as to determine whether the three-dimensional displacement gradient and the strain tensor prediction value satisfy a basic relationship between the strain tensor and the displacement gradient in three-dimensional continuous medium mechanics; wherein, ε_pred is the strain tensor prediction value directly output by the deep neural network, containing six components: ε_pred = [εxx_pred, εyy_pred, εzz_pred, εxy_pred, εxz_pred, εyz_pred]; ε_calc is the strain tensor value calculated from the displacement field predicted by the neural network through the continuum mechanics relationship; ε_calc = [εxx_calc, εyy_calc, εzz_calc, εxy_calc, εxz_calc, εyz_calc], wherein u, v, w are displacement components in x, y, z directions, respectively, and ε xx , ε yy , ε zz are normal strain components, and ε xy , ε xz , ε yz are shear strain components.

6. The artificial intelligence driven three-dimensional digital volume related deformation measurement system of claim 1, wherein, The three-dimensional field reconstruction output unit includes a three-dimensional displacement field visualization module, a three-dimensional strain field visualization module, a volume slice display module, an equivalent strain calculation module and a stereoscopic data export module. The three-dimensional displacement field visualization module is used to show the displacement distribution in u, v and w directions. The three-dimensional strain field visualization module is used to show the spatial distribution of six strain tensor components. The volume slice display module is used to show the field distribution at different depth levels. The equivalent strain calculation module is used to calculate the three-dimensional equivalent strain indicator. Three-dimensional equivalent strain indicator von_Mises = Sqrt(0.5 x [(ε xx -ε yy ) 2 +(ε yy -ε zz ) 2 +(ε zz -ε xx ) 2 +6 x (ε xy 2 +ε xz 2 +ε yz 2 )]) where ε xx , ε yy , ε zz are the principal strain components and ε xy , ε xz , ε yz are the shear strain components; The three-dimensional data export module is used to save the three-dimensional analysis results.

7. A method of an artificial intelligence driven three-dimensional digital volume related deformation measurement system, characterized in that, The method comprises the following steps: Step 1: load the reference speckle volume and the deformed speckle volume, pre-process, and create a three-dimensional normalized coordinate grid; Step 2: normalize the three-dimensional voxel coordinates and convert them into standardized inputs; Step 3: build a deep neural network model and set architecture parameters; the architecture parameters include the number of network layers and the number of neurons; Step 4: define a hybrid loss function L_total = λ_image × L_image + λ_physics × L_physics that contains volume gray value reconstruction consistency constraints and three-dimensional deformation compatibility constraints; wherein λ_image is the image consistency weight, and λ_physics is the physical constraint weight; Step 5: perform end-to-end training on the deep neural network model, optimize the network weight parameters, and continue until the loss function reaches the convergence standard; Step 6: use the trained model to predict the three-dimensional displacement field and strain tensor field, and calculate the three-dimensional equivalent strain; Step 7: generate three-dimensional field distribution stereoscopic visualization results, and save the displacement field and strain field.

8. The method of claim 7, wherein the method is performed by an artificial intelligence driven three-dimensional digital volume related deformation measurement system. The three-dimensional coordinate normalization in step 2 includes normalizing the original voxel coordinates (x, y, z) according to the volume dimensions to obtain the standardized coordinate representation.

9. The method of claim 7, wherein the method is performed by an artificial intelligence driven three-dimensional digital volume related deformation measurement system. The three-dimensional deformation compatibility constraint in step 4 is as follows: L_physics =∑(ε_pred - ε_calc) 2 (1) wherein ε_pred = [εxx_pred, εyy_pred, εzz_pred, εxy_pred, εxz_pred, εyz_pred]; ε_calc is the strain tensor value calculated from the neural network predicted displacement field through the continuum mechanics relationship; ε_calc = [εxx_calc, εyy_calc, εzz_calc, εxy_calc, εxz_calc, εyz_calc].

10. The method of claim 7, wherein the system is an artificial intelligence driven three-dimensional digital volume related deformation measurement system. The three-dimensional equivalent strain in step 6 is as follows: von Mises = Sqrt(0.5 x [(ε xx - ε yy ) 2 + (ε yy - ε zz ) 2 + (ε zz - ε xx ) 2 + 6 x (ε xy 2 + ε xz 2 +ε yz 2 )])(2) where ε xx , ε yy , ε zz are the principal strain components and ε xy , ε xz , ε yz are the shear strain components.

Images