Cement-based material displacement field identification method based on one-dimensional convolution module

Abstract

This invention discloses a method for displacement field recognition of cement-based materials based on a one-dimensional convolutional model. The method includes: S1: extracting the displacement field from a speckle image; using speckle image pairs as feature data, and using the lateral displacement field, longitudinal displacement field, and combined displacement field as label data, and normalizing the label data to obtain a sample dataset; S2: constructing a displacement field recognition model for cement-based materials based on a one-dimensional convolutional model; S3: constructing a model loss function containing the design parameters to be optimized by introducing physical loss, and training the displacement field recognition model based on the one-dimensional convolutional model for cement-based materials based on the sample dataset to obtain the optimal displacement field recognition model; and realizing the displacement field recognition of cement-based materials based on the optimal displacement field recognition model. This invention solves the problem of insufficient accuracy and efficiency in the displacement field recognition of cement-based materials by existing methods.

Description

Technical Field

[0001] This invention relates to the field of digital image correlation (DIC) technology based on deep learning, and more particularly to a method for displacement field recognition of cement-based materials based on a one-dimensional convolutional model. Background Technology

[0002] Cement-based materials are a typical class of heterogeneous brittle materials, whose minute deformations during compression are difficult to discern with the naked eye, and the deformation field spans significantly different between the early and later stages of loading. Under these conditions, existing prediction models generally suffer from shortcomings such as neglecting small deformations in the early stages of loading, excessive computational resource consumption, and excessively long training times.

[0003] Existing algorithm-based traditional Deformation-Oriented Conversion (DIC) methods are highly dependent on empirical parameters in deformation field prediction. The selection of subset size, step size, and shape function directly affects measurement accuracy, while computational efficiency is low. To address this, deep learning-based DIC methods have emerged in recent years, most of which employ the U-Net network architecture for end-to-end deformation field prediction and rely on generated datasets for training. However, the displacement patterns used in the generated datasets are often overly idealized, making it difficult to reproduce the complex non-uniform deformation patterns of cement-based materials under uniaxial compression; the trained models also cannot guarantee high temporal resolution recognition. Furthermore, commonly used loss functions are often MAE and MAPE, which, in cement-based material deformation field prediction research, tend to overlook small deformations in the early stages of loading (MAE) and converge slowly and have insufficient accuracy (MAPE). In addition, directly using high-resolution deformation field data for model training leads to high computational resource consumption and long training times, while existing lightweight designs mostly focus on reducing the number of convolutional layers, failing to balance efficiency and accuracy. These shortcomings mean that existing methods still have significant limitations in high-resolution and complex deformation field prediction. Summary of the Invention

[0004] This invention provides a method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model, in order to overcome the above-mentioned technical problems.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model, specifically including the following steps:

[0007] S1: Obtain a speckle image pair of cement-based materials, where the speckle image pair includes a reference speckle image and a deformed speckle image with the same viewpoint and resolution; and extract the displacement field based on the deformed speckle image.

[0008] Furthermore, the displacement field includes a transverse displacement field and a longitudinal displacement field;

[0009] The combined displacement field is obtained from the transverse and longitudinal displacement fields, and the speckle image pairs are used as feature data. The transverse, longitudinal, and combined displacement fields are used as label data, and the label data is normalized to obtain the sample dataset.

[0010] S2: Construct a displacement field recognition model for cement-based materials based on a one-dimensional convolutional model, which includes a two-dimensional feature reduction module, a first linear layer, a one-dimensional convolutional encoding module, a second linear layer, a third linear layer, a one-dimensional deconvolutional decoding module, and a displacement field output module.

[0011] The two-dimensional feature reduction module is used to convert speckle image pairs into grayscale tensors and then perform channel-dimensional stitching to obtain a stitched image. Based on a preset two-dimensional multi-channel feature extraction layer, the stitched image is subjected to two-dimensional reduction processing to extract the two-dimensional speckle image deformation features of the stitched image after two-dimensional reduction processing.

[0012] The first linear layer is used to transform the deformation features of the two-dimensional speckle image into one-dimensional signal sequence features;

[0013] The one-dimensional convolutional coding module is used to perform multi-scale feature extraction on the features of a one-dimensional signal sequence to obtain one-dimensional deformable coding features; and the one-dimensional deformable coding features are used to characterize the displacement correlation between the reference speckle image and the deformable speckle image pair.

[0014] The second linear layer is used to perform dimensionality reduction encoding on the one-dimensional deformed coding features, and the third linear layer is used to perform dimensionality increase on the output of the second linear layer to output multiple channel sequence features with the same signal length as the one-dimensional signal sequence features.

[0015] The one-dimensional deconvolution decoding module is used to decode the sequence features of each channel and obtain the displacement sequence features of each channel;

[0016] The displacement field output module is used to first map the displacement sequence features of each channel to a two-dimensional displacement field space with a preset resolution to obtain the mapped two-dimensional displacement field; then, based on the cubic spline interpolation method, it performs interpolation processing on the mapped two-dimensional displacement field to achieve pixel-level registration with the speckle image in the input two-dimensional feature reducer, so as to predict and output the transverse displacement field, longitudinal displacement field and the combined displacement field.

[0017] S3: By introducing physical loss, a model loss function containing the design parameters to be optimized is constructed, and the model is trained on the displacement field recognition model of cement-based materials based on one-dimensional convolutional model according to the sample dataset to obtain the optimal displacement field recognition model; and the displacement field recognition of cement-based materials is realized according to the optimal displacement field recognition model.

[0018] Furthermore, the one-dimensional convolutional coding module includes several convolutional kernels and one-dimensional convolutional layers with different convolutional strides connected in sequence. The number of channels in each one-dimensional convolutional layer is 64, 128, 256, 128, 64 and 4 respectively, forming a coding bottleneck structure. This structure is used to extract and compress the sequence features of the one-dimensional signal sequence output by the first linear layer to obtain one-dimensional deformable coding features.

[0019] Furthermore, the one-dimensional deconvolution decoding module includes a plurality of convolution kernels and deconvolution arrangers with different convolution strides connected in sequence;

[0020] Furthermore, the deconvolution scrambler is a deconvolution scrambler block structure formed by connecting a single deconvolution layer and a convolution layer through a ReLU activation function layer;

[0021] The deconvolutional layer is used to perform a deconvolution operation on its input;

[0022] The ReLU activation function layer is used to perform non-linear activation operations on the output of the deconvolution layer;

[0023] The convolutional layer is used to perform convolution operations on the output of the ReLU activation function layer;

[0024] By using several deconvolutional rectifiers with different convolutional kernels and strides, the one-dimensional deformed coding features are restored to multi-channel sequence features with the same signal length as the one-dimensional signal sequence features. The displacement sequence features of each channel are obtained by decoding the sequence features of each channel.

[0025] Furthermore, the method for obtaining the optimal displacement field identification model in S3 specifically includes the following steps:

[0026] S31: Divide the sample dataset into a training set and a test set according to a preset ratio:

[0027] S32: Train the cement-based material displacement field recognition model based on the training set to obtain the trained cement-based material displacement field recognition model.

[0028] S33: Construct a model loss function that includes the design parameters to be optimized by introducing physical loss, and confirm whether the output of the cement-based material displacement field identification model after training converges based on the test set;

[0029] If so, the cement-based material displacement field identification model trained at this time is taken as the optimal displacement field identification model; if not, the parameter weights of the cement-based material displacement field identification model trained are adaptively adjusted based on the backpropagation method, and the design parameters to be optimized in the model loss function are optimized based on the Grey Wolf Optimization Algorithm (GWO) to obtain the latest model loss function and repeat step S32.

[0030] Furthermore, the model loss function constructed in S33, which includes the design parameters to be optimized, has the following expression:

[0031] (1)

[0032] (2)

[0033] (3)

[0034] (4)

[0035] In the formula: Represents the model loss function; This represents the weight coefficient corresponding to each loss function term; The average absolute error, or physical loss, represents the predicted physical quantities, namely the transverse displacement field, the longitudinal displacement field, and the resultant displacement field. This represents the mean absolute error after normalization for each sample. N Indicates the number of pixels in the image dimension; These represent the predicted and experimental values ​​of the physical quantity corresponding to a certain pixel, respectively. Represents the mixed loss function; This indicates the optimization weight parameters, i.e., the design parameters to be optimized; sum() represents the summation function; Represents the corresponding sample The predicted physical quantity; Represents the corresponding sample The experimental values; PL This represents physical constraint terms.

[0036] Furthermore, the method in S33 for optimizing the design parameters in the model loss function based on the Grey Wolf Optimization Algorithm (GWO) includes the following steps:

[0037] S331: Randomly obtain the initial population and maximum number of iterations for the Grey Wolf optimization algorithm;

[0038] Furthermore, the position of each gray wolf in the initial group is defined as the combination of design parameters to be optimized. A feasible solution;

[0039] S322: Constructing the fitness function for the Grey Wolf optimization algorithm:

[0040] And the expression for the fitness function is:

[0041] (5)

[0042] (6)

[0043] In the formula:S Represents the fitness function; These represent the average errors generated during the training process for the transverse displacement field, longitudinal displacement field, and resultant displacement field, respectively. mean() means taking the average value;

[0044] S333: Obtain the fitness value of each gray wolf individual in the initial population according to the fitness function, and obtain the sequence list by arranging the gray wolf individuals in descending order according to the fitness value;

[0045] The first three gray wolf individuals in the sequence list were selected as Alpha wolves, Beta wolves, and Delta wolves, respectively.

[0046] S334: Confirm whether the current iteration count has reached the maximum iteration count;

[0047] If so, then the current position of the Alpha Wolf will be taken as the optimal combination of design parameters. ;

[0048] If not, the initial group is iteratively updated based on the locations of the Alpha wolves, Beta wolves, and Delta wolves to obtain a new generation of group;

[0049] Define the new generation as the initial group and repeat step S333.

[0050] Furthermore, in S1, the labeled data is normalized to obtain a sample dataset, the expression of which is:

[0051] (7)

[0052] In the formula: This represents the normalized label data, i.e., the physical quantity value. These represent the upper and lower limits of the physical quantity values ​​in the original dataset, respectively. This represents the label data to be normalized. This represents the given upper and lower bounds of normalization.

[0053] Beneficial Effects: This invention provides a method for displacement field recognition of cement-based materials based on a one-dimensional convolutional model. It acquires speckle images of cement-based materials and extracts the displacement field from these images. The lateral, longitudinal, and combined displacement fields are used as label data, while the speckle images are used as feature data to obtain a sample dataset. A model loss function containing the design parameters to be optimized is constructed by introducing physical loss. The model is then trained using the sample dataset to obtain the optimal displacement field recognition model for cement-based materials. This invention introduces physical loss into the loss function and optimizes the weights of the physical loss using the Grey Wolf optimization algorithm. This establishes connections between the channels of each physical quantity for auxiliary calibration and mutual constraint when channel deviations occur, making model training more stable and significantly improving training speed. The main body of the model uses one-dimensional convolution, which is more lightweight and computationally efficient. Simultaneously, dimensionality reduction processing is performed on image features and training data. Cross-dimensional reduction modeling (2D→1D→2D) reduces training difficulty and achieves higher recognition accuracy than training directly with high-resolution data. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of the displacement field identification method for cement-based materials based on a one-dimensional convolutional model according to the present invention;

[0056] Figure 2 This is a schematic diagram of the network structure of the cement-based material displacement field recognition model based on a one-dimensional convolutional model in this embodiment;

[0057] Figure 3 This is the core block diagram of the displacement field identification method for cement-based materials in this embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] This embodiment provides a method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model, such as... Figure 1 As shown, the specific steps include:

[0060] S1: Obtain a speckle image pair of cement-based materials, wherein the speckle image pair includes a reference speckle image and a deformed speckle image with the same viewpoint and resolution; and extract the displacement field from the deformed speckle image using the preset Stream Pix 8 software; wherein, in this embodiment, the extraction of the displacement field from the speckle image using the preset Stream Pix 8 software is an existing working technique, and will not be described in detail here.

[0061] Furthermore, the displacement field includes a transverse displacement field and a longitudinal displacement field;

[0062] The combined displacement field is obtained from the transverse and longitudinal displacement fields, and the speckle image pairs are used as feature data. The transverse, longitudinal, and combined displacement fields are used as label data, and the label data is normalized to obtain the sample dataset.

[0063] In this embodiment, because the displacement amplitude varies significantly at different loading times within the same dataset, and small deformations (especially lateral displacements) account for a high proportion in the initial loading stage, they are easily dominated by large value regions, causing small deformation regions to be ignored. Therefore, the identification values ​​and label data are first normalized according to the label value range. The normalization expression is:

[0064] (1)

[0065] In the formula: This represents the normalized label data, i.e., the physical quantity value. These represent the upper and lower limits of the physical quantity values ​​in the original dataset, respectively. This represents the label data to be normalized. Represents the given upper and lower bounds of normalization;

[0066] S2: Construct a displacement field recognition model for cement-based materials based on a one-dimensional convolutional model, such as... Figure 2 As shown, it includes a two-dimensional feature reduction module, a first linear layer, a one-dimensional convolutional coding module, a second linear layer, a third linear layer, a one-dimensional deconvolutional decoding module, and a displacement field output module.

[0067] The two-dimensional feature reduction module is used to convert speckle image pairs into grayscale tensors and then perform channel-dimensional stitching to obtain a stitched image. Based on a preset two-dimensional multi-channel feature extraction layer, the stitched image is subjected to two-dimensional reduction processing to extract the two-dimensional speckle image deformation features of the stitched image after two-dimensional reduction processing.

[0068] Specifically, the two-dimensional feature reduction module includes four convolutional kernels connected in sequence and two-dimensional convolutional layers with different convolutional strides, and the number of channels in each two-dimensional convolutional layer is 64, 128, 64 and 1 respectively.

[0069] The first linear layer is used to transform the deformation features of the two-dimensional speckle image into one-dimensional signal sequence features;

[0070] The one-dimensional convolutional coding module performs multi-scale feature extraction on the one-dimensional signal sequence features to obtain one-dimensional deformable coding features; and the one-dimensional deformable coding features are used to characterize the displacement correlation between the reference speckle image and the deformable speckle image pair.

[0071] Specifically, the one-dimensional convolutional coding module includes several convolutional kernels and one-dimensional convolutional layers with different convolutional strides connected in sequence, and the number of channels of each one-dimensional convolutional layer is 64, 128, 256, 128, 64 and 4 respectively to form a coding bottleneck structure, which is used to extract and compress the one-dimensional signal sequence features output by the first linear layer in a multi-scale sequence to obtain one-dimensional deformable coding features.

[0072] The second linear layer is used to perform dimensionality reduction encoding on the one-dimensional deformed coding features, and the third linear layer is used to perform dimensionality increase on the output of the second linear layer to output multiple channel sequence features with the same signal length as the one-dimensional signal sequence features.

[0073] The one-dimensional deconvolution decoding module is used to decode the sequence features of each channel and obtain the displacement sequence features of each channel;

[0074] Specifically, the one-dimensional deconvolution decoding module includes a number of convolution kernels and deconvolution arrangers with different strides connected in sequence; and the number of channels of the deconvolution arrangers are 64, 128, 256, 128, 64 and 3 in sequence; and the deconvolution arranger is a deconvolution arranger block structure formed by connecting a single deconvolution layer and a convolution layer through a ReLU activation function layer.

[0075] The deconvolutional layer is used to perform a deconvolution operation on its input;

[0076] The ReLU activation function layer is used to perform non-linear activation operations on the output of the deconvolution layer;

[0077] The convolutional layer is used to perform convolution operations on the output of the ReLU activation function layer;

[0078] By using several deconvolution rectifiers with different convolution kernels and strides, the one-dimensional deformed coding features are restored to multi-channel sequence features with the same signal length as the one-dimensional signal sequence features. The three-channel displacement sequence features are obtained by decoding each channel sequence feature.

[0079] The displacement field output module is used to first map the displacement sequence features of each channel to a two-dimensional displacement field space with a preset resolution to obtain the mapped two-dimensional displacement field; then, based on the cubic spline interpolation method, it performs interpolation processing on the mapped two-dimensional displacement field to achieve pixel-level registration with the speckle image in the input two-dimensional feature reducer, so as to predict and output the transverse displacement field, longitudinal displacement field and the combined displacement field.

[0080] This embodiment specifically involves the displacement field during the uniaxial compressive process of cement-based materials: (1) Lateral displacement U (2) Longitudinal displacement V And (3) resultant displacement D The recognition of displacement fields is particularly important for small deformation displacement fields with high spatiotemporal resolution. In this embodiment, a one-dimensional convolutional architecture is used as the core of the model to construct a pixel-level displacement regression network of "two-dimensional reduction mapping - one-dimensional deformation encoding extraction - two-dimensional reconstruction and resolution restoration". This is a cement-based material displacement field recognition model based on a one-dimensional convolutional model. Since the model maps two-dimensional compressed features to one-dimensional signals and then uses one-dimensional convolution to extract image deformation features from the signals and generate corresponding deformation fields, this cross-dimensional reduction modeling (2D→1D→2D) greatly reduces the requirements for computing resources, and model training can be completed subsequently. The model input consists of reference (Ref) and deformed (Def) speckle image pairs with the same viewpoint and resolution. After channel-dimensional concatenation, a two-dimensional reduction process is performed to compress and extract relevant image features. Subsequently, the reduced-order features are serialized into a one-dimensional signal and fed into a multi-scale one-dimensional convolutional coding module. The one-dimensional convolutional coding module extracts a one-dimensional deformable code, which is used to explicitly represent the displacement correlation of the reference-deformed image pair. Based on this deformable code, a one-dimensional deconvolutional decoding module reconstructs the multi-channel sequence step by step and integrates it into a two-dimensional displacement field. Then, cubic spline interpolation is used to restore the original resolution. Meanwhile, the sampling frequency of the deformed speckle image used for model training is 0.01s, which ensures that the model has high spatiotemporal resolution.

[0081] S3: By introducing physical loss, a model loss function containing the design parameters to be optimized is constructed, and the model of displacement field recognition of cement-based materials based on one-dimensional convolutional model is trained according to the sample dataset to obtain the optimal displacement field recognition model; and displacement field recognition of cement-based materials is realized according to the optimal displacement field recognition model.

[0082] Specifically, in the model training phase, this embodiment employs sample-by-sample min-max interval normalization. BL 1 and the original L The weighted mixed loss of 1 improves the training effect in small deformation regions, while the physical constraint loss is added. PLIt is used to couple multiple physical quantities and provide supervision for the missing physical quantity when some physical quantity data is missing. At the same time, it jointly optimizes multiple weight parameters of mixed data loss and physical constraint loss based on comprehensive evaluation index, so that the model can obtain the best displacement field recognition performance.

[0083] The method for obtaining the optimal displacement field identification model in this embodiment specifically includes the following steps:

[0084] S31: Divide the sample dataset into a training set and a test set according to a preset ratio:

[0085] S32: The adaptive momentum (Adam) optimizer is used to train the cement-based material displacement field recognition model based on the one-dimensional convolutional model according to the training set, and the trained cement-based material displacement field recognition model is obtained.

[0086] S33: Construct a model loss function that includes the design parameters to be optimized by introducing physical loss. Its expression is:

[0087] (2)

[0088] (3)

[0089] (4)

[0090] (5)

[0091] In the formula: Represents the model loss function; This represents the weight coefficient corresponding to each loss function term; The average absolute error, or physical loss, represents the predicted physical quantities, namely the transverse displacement field, the longitudinal displacement field, and the resultant displacement field. This represents the mean absolute error after normalizing each sample (min–max normalization mapping each sample to [0.1, 0.2]). MAE ; N Indicates the number of pixels in the image dimension; These represent the predicted and experimental values ​​of the physical quantity corresponding to a certain pixel, respectively. Represents the mixed loss function; This indicates the optimization weight parameters, i.e., the design parameters to be optimized; sum() represents the summation function; Represents the corresponding sample The predicted physical quantity; Represents the corresponding sample The experimental values; PL Represents physical constraint terms;

[0092] In this embodiment, the model employs an adaptive momentum (Adam) optimizer and is trained with a piecewise learning rate (i.e., the learning rate is gradually reduced). The amplitude variations in displacement fields between different loading stages of cement-based materials are significant, especially the lateral deformation generated during uniaxial compression. To address the amplitude differences between and within deformation fields, the following measures are taken: Based on the loss function, it was proposed Loss function; used alone in this embodiment. While the loss function can avoid the problems of insufficient gradient contribution and excessive dimensional difference when the value is close to 0 in small deformation regions, The loss function initially suffers from weak constraints, slow convergence, and insignificant reduction in training loss. Therefore, a hybrid loss function is adopted to train the model, combining both aspects to form a hybrid loss function. The project is responsible for training the entire deformation field while emphasizing small deformation regions and avoiding being dominated by large-value samples. The first step is to lead the loss reduction, increase the training depth, and make the model converge toward the common optimization goal of both.

[0093] Furthermore, due to the lateral displacement field ( U ), longitudinal displacement field ( V and resultant displacement field ( D There is still a problem of inconsistent dimensions among the three physical quantities. If the values ​​of the three physical quantities are normalized to a range […], then… S min , S max If the range of values ​​for each physical quantity is not uniform, then during multi-displacement field recognition training, the calculated physical quantity values ​​of the three channels are normalized to the specified range one by one. S min , S max Then calculate the hybrid loss function L channel by channel. n Meanwhile, in this embodiment, three physical quantities ( U , V and D The recognition accuracy decreased during training due to interference between the three-channel mixed loss; simultaneously, the recognition performance of each channel (i.e., the coefficient of determination) decreased. R 2 This results in varying degrees of reduction, slower training speed, and difficulty in combining the optimal effects of multiple physical quantities; although this can be mitigated by increasing the number of training epochs. L The value was reduced sufficiently to address the optimal combination of effects, but the simultaneous identification of each physical quantity still couldn't match the identification effect of a single physical quantity. Therefore, relying on... U , V , DThe mathematical relationship between them imposes physical constraints on the predicted values ​​of each channel, forming a physical loss. PL In this embodiment, the predicted values ​​of three physical quantities are mutually corrected to increase the connection between channels; training has proven that after adding the physical loss term, its... R 2 This represents an improvement over previous results. The introduction of this term links the recognition channels for each physical quantity, thus accelerating model convergence during training. Simultaneously, due to this constraint, all three physical quantities can be synchronously combined with good recognition performance. However, attention still needs to be paid to the parameters. α Setting the value to 3, within an appropriate range, can accelerate model convergence and improve recognition accuracy. If this parameter is too small, it will not achieve the desired constraint effect, while if it is too large, it will affect the overall loss function. L This actually reduces the model's recognition performance.

[0094] In addition, this embodiment also includes training with missing datasets:

[0095] Because traditional model training relies on purely data-driven methods, this requires that the physical quantities used in training be equal to the physical quantities used in the dataset. D Data acquisition relies on matrix synthesis, which is not difficult but time-consuming, and consumes storage resources when the dataset is large. The introduction of physical loss changes model training from traditional pure data-driven to data + physical-driven, providing a solution for training with missing datasets. During model training, the original number of channels is maintained, and... L n In the item, physical quantity D Weighting coefficients ( α 3,d Adjust to 0 to simulate the dataset. D Missing, that is:

[0096] (6)

[0097] At this point, the primary task of physical loss is constraint. D Channel convergence, theoretically U and V The loss value introduced by the channel decreases continuously during model training. D Dependent on physical constraints PL Able to U and V The training effect is transferred to itself, while the model performs well on the simulated dataset. D Training under missing conditions revealed that... D The recognition effect is only slightly lower than before, but it can still maintain high accuracy.

[0098] Confirm whether the output of the trained cement-based material displacement field recognition model has converged based on the test set.

[0099] If so, the cement-based material displacement field identification model trained at this time is taken as the optimal displacement field identification model; if not, the parameter weights of the cement-based material displacement field identification model trained are adaptively adjusted based on the backpropagation method, and the design parameters to be optimized in the model loss function are optimized based on the Grey Wolf Optimization Algorithm (GWO) to obtain the latest model loss function and repeat step S32.

[0100] This embodiment describes a method for optimizing design parameters in the model loss function based on the Grey Wolf Optimization Algorithm (GWO). The specific steps include:

[0101] S331: Randomly obtain the initial population and maximum number of iterations for the Grey Wolf optimization algorithm;

[0102] Furthermore, the position of each gray wolf in the initial group is defined as the combination of design parameters to be optimized. A feasible solution;

[0103] S322: Constructing the fitness function for the Grey Wolf optimization algorithm:

[0104] And the expression for the fitness function is:

[0105] (7)

[0106] (8)

[0107] In the formula: S Represents the fitness function; These represent the average errors generated during the training process for the transverse displacement field, longitudinal displacement field, and resultant displacement field, respectively. mean() means taking the average value;

[0108] S333: Obtain the fitness value of each gray wolf individual in the initial population according to the fitness function, and obtain the sequence list by arranging the gray wolf individuals in descending order according to the fitness value;

[0109] The first three gray wolf individuals in the sequence list were selected as Alpha wolves, Beta wolves, and Delta wolves, respectively.

[0110] S334: Confirm whether the current iteration count has reached the maximum iteration count;

[0111] If so, then the current position of the Alpha Wolf will be taken as the optimal combination of design parameters. ;

[0112] If not, the initial group is iteratively updated based on the locations of the Alpha wolves, Beta wolves, and Delta wolves to obtain a new generation of group;

[0113] Define the new generation as the initial group and repeat step S333.

[0114] In this embodiment, the model exhibits strong generalization ability; simply changing the final number of channels is sufficient to modify the single / multi-displacement recognition model. Before performing multi-physical quantity recognition, each physical quantity was trained individually. This embodiment employs... R 2 and root mean square error ( RMSE The training and prediction performance of the model are evaluated. During training, the average values ​​of each physical quantity on the training and test sets are recorded. R 2 This is used as an evaluation index for the recognition effect of each physical quantity. Based on the changing trends of both (overfitting / underfitting signals), the learning rate and loss weights are iteratively adjusted. BL 1 / L Parameters such as the ratio (1), batch size, and number of training epochs are used to continuously optimize the model's recognition performance and generalization ability. The formulas for calculating the two evaluation metrics are as follows:

[0115] (9)

[0116] (10)

[0117] This embodiment, by training on a single physical quantity, allows us to understand the training difficulty and evaluation index range of each physical quantity. This work provides a reference for training multi-physical quantity models, such as the comprehensive evaluation index when training a model with multiple physical quantities. S The weights, learning rate range, input batch size, training epochs, etc., facilitate parameter optimization later. In this embodiment, multi-physical quantity training differs from single-physical quantity training; there is a problem that the three physical quantities may not simultaneously achieve optimal recognition results. Therefore, it is necessary to quantitatively design a comprehensive model evaluation index for multi-physical quantity training scenarios. This embodiment is based on... U , V as well as D The evaluation index is set by considering the difficulty and importance of individual training tasks. S This embodiment aims to avoid a certain item R 2Excessive weighting of one parameter negatively impacts the other two. The aforementioned evaluation metrics take effect when they are greater than 0.960, 0.990, and 0.990 respectively in the validation set; otherwise, they are not calculated. Based on the comprehensive evaluation metric S, i.e., the fitness function, parameter optimization can be performed on the model. When using the Grey Wolf Optimization (GWO) algorithm for model optimization, the maximum S value obtained by the parameter combination during training is used as the evaluation of that combination. The optimal parameter combination is selected based on the magnitude of the S value to ensure the best comprehensive recognition effect during the training of the three physical quantities. During training, the model weight with the maximum current S value is saved in each round to obtain the "optimal weight" used for final recognition. This weight is used to batch process the uniaxial compressive process deformation speckle map sequence, outputting the displacement matrix of the corresponding physical quantity in .txt text document format and drawing a displacement cloud map with a unified color scale. Simultaneously, error visualization is performed during training: generating scatter plots of the relative errors of each sequence and its average, as well as an error matrix in .txt text document format, to facilitate error distribution analysis and model parameter adjustment.

[0118] like Figure 3 As shown, this embodiment is based on a pre-trained end-to-end model. The recognition process requires no additional parameter settings, avoiding the influence of parameter settings on the results. The model benefits from training with a hybrid loss mechanism, achieving good recognition performance even in small deformation regions, while also considering the complete deformation process from the initial loading stage to before failure. The recognition result is directly output at the same resolution as the original image. U , V and D The displacement matrix and cloud map support batch export by number, and the recognition speed is faster than the algorithm. The main body of the model adopts one-dimensional convolution and performs dimensionality reduction processing on image features and training data, which reduces the training difficulty and achieves higher recognition accuracy than training directly with high-resolution data, while requiring less computing resources. In multi-physical quantity prediction tasks, the lateral displacement component often exhibits slow convergence, thus slowing down the overall training. To solve this problem, this model incorporates the physical constraint relationship U2 + V2 = D2 into the training process, so that the model can be subject to the physical consistency constraint of displacement components V and D when learning the U displacement component. This method not only significantly improves the convergence efficiency, and the prediction results show that the overall training time is shortened by about half, but also enhances the physical rationality of the prediction results. Table 1 shows the prediction comparison results between the method described in this embodiment and the traditional method:

[0119] Table 1. Prediction Comparison Results

[0120]

[0121] Note: U , V as well as D These are the lateral displacement, longitudinal displacement, and resultant displacement, respectively. e x ande y These represent the transverse and longitudinal strains, respectively; STAR represents the star-shaped displacement field, with zero longitudinal displacement and the amplitude of a sinusoidal function for transverse displacement, used to test the model's recognition of unidirectional displacement; px represents the number of pixels in the analysis area calculated by the model per second; the "Reliability of Prediction Results" column reflects the evaluation index of the worst-performing physical quantity prediction results in the literature.

[0122] The beneficial effects of the method described in this embodiment are as follows:

[0123] (1) Normalizing the displacement matrix during the loading process to (0.1, 0.2) avoids the gradient near 0 being too small and thus ignored during training. At the same time, compressing the numerical range further avoids large value regions dominating the training, especially the displacement field in the later stages of loading. In addition, unnormalized data loss is added, which reduces the dominant loss. The combination of these two methods with reasonable parameters trains towards the common optimization objective, thereby improving the model's accuracy in recognizing displacements in small deformation regions even when the deformation amplitudes differ significantly.

[0124] (2) Without lacking labeled data, physical loss is introduced into the loss function, and based on reasonable weights of physical loss, the channels of each physical quantity are linked. When a channel produces a deviation, the channels are calibrated and mutually constrained, making the model training more stable and significantly improving the training speed.

[0125] (3) When there is a lack of physical quantity data, physical loss is used to assist in supervising the training of the missing label channel, and at the same time, reasonable weights are used to avoid affecting the training of other channels, thereby making up for the data loss caused by the lack of data.

[0126] (4) The main body of the model uses one-dimensional convolution, which is more lightweight and computationally efficient. In addition, the image features and training data are dimensionality reduced, and cross-dimensional reduction modeling (2D→1D→2D) reduces the training difficulty and achieves higher recognition accuracy than training directly with high-resolution data. This improves the overall recognition effect of deformation fields and reduces the computational resource requirements for model training.

[0127] The publicly available model links and references involved in this embodiment are as follows:

[0128] Wang and Zhao (2023): Wang Y, Zhao J. DIC-Net: Upgrade the performance of traditional DIC with Hermite dataset and convolution neural network[J]. Optics and Lasers in Engineering, 2023, 160: 107278.

[0129] Boukhtache et al. (2023): Boukhtache S, Abdelouahab K, Bahou A, et al. Alightweight convolutional neural network as an alternative to DIC to measure in-plane displacement fields[J]. Optics and lasers in engineering, 2023, 161:107367.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model, characterized in that, Specifically, the following steps are included: S1: Obtain a speckle image pair of cement-based materials, where the speckle image pair includes a reference speckle image and a deformed speckle image with the same viewpoint and resolution; and extract the displacement field based on the deformed speckle image. Furthermore, the displacement field includes a transverse displacement field and a longitudinal displacement field; The combined displacement field is obtained from the transverse and longitudinal displacement fields, and the speckle image pairs are used as feature data. The transverse, longitudinal, and combined displacement fields are used as label data, and the label data is normalized to obtain the sample dataset. S2: Construct a displacement field recognition model for cement-based materials based on a one-dimensional convolutional model, which includes a two-dimensional feature reduction module, a first linear layer, a one-dimensional convolutional encoding module, a second linear layer, a third linear layer, a one-dimensional deconvolutional decoding module, and a displacement field output module. The two-dimensional feature reduction module is used to convert speckle image pairs into grayscale tensors and then perform channel-dimensional stitching to obtain a stitched image. Based on a preset two-dimensional multi-channel feature extraction layer, the stitched image is subjected to two-dimensional reduction processing to extract the two-dimensional speckle image deformation features of the stitched image after two-dimensional reduction processing. The first linear layer is used to transform the deformation features of the two-dimensional speckle image into one-dimensional signal sequence features; The one-dimensional convolutional coding module is used to perform multi-scale feature extraction on the features of a one-dimensional signal sequence to obtain one-dimensional deformable coding features; and the one-dimensional deformable coding features are used to characterize the displacement correlation between the reference speckle image and the deformable speckle image pair. The second linear layer is used to perform dimensionality reduction encoding on the one-dimensional deformed coding features, and the third linear layer is used to perform dimensionality increase on the output of the second linear layer to output multiple channel sequence features with the same signal length as the one-dimensional signal sequence features. The one-dimensional deconvolution decoding module is used to decode the sequence features of each channel and obtain the displacement sequence features of each channel; The displacement field output module is used to first map the displacement sequence features of each channel to a two-dimensional displacement field space with a preset resolution to obtain the mapped two-dimensional displacement field; then, based on the cubic spline interpolation method, it performs interpolation processing on the mapped two-dimensional displacement field to achieve pixel-level registration with the speckle image in the input two-dimensional feature reducer, so as to predict and output the transverse displacement field, longitudinal displacement field and the combined displacement field. S3: By introducing physical loss, a model loss function containing the design parameters to be optimized is constructed, and the model is trained on the displacement field recognition model of cement-based materials based on one-dimensional convolutional model according to the sample dataset to obtain the optimal displacement field recognition model; and the displacement field recognition of cement-based materials is realized according to the optimal displacement field recognition model.

2. The method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model according to claim 1, characterized in that, The one-dimensional convolutional coding module includes several convolutional kernels and one-dimensional convolutional layers with different strides connected in sequence. The number of channels in each one-dimensional convolutional layer is 64, 128, 256, 128, 64 and 4 respectively, forming a coding bottleneck structure. This structure is used to extract and compress the one-dimensional signal sequence features output by the first linear layer at multiple scales to obtain one-dimensional deformable coding features.

3. The method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model according to claim 2, characterized in that, The one-dimensional deconvolution decoding module includes a number of convolution kernels and deconvolution arrangers with different convolution strides connected in sequence. Furthermore, the deconvolution scrambler is a deconvolution scrambler block structure formed by connecting a single deconvolution layer and a convolution layer through a ReLU activation function layer; The deconvolutional layer is used to perform a deconvolution operation on its input; The ReLU activation function layer is used to perform non-linear activation operations on the output of the deconvolution layer; The convolutional layer is used to perform convolution operations on the output of the ReLU activation function layer; By using several deconvolutional rectifiers with different convolutional kernels and strides, the one-dimensional deformed coding features are restored to multi-channel sequence features with the same signal length as the one-dimensional signal sequence features. The displacement sequence features of each channel are obtained by decoding the sequence features of each channel.

4. The method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model according to claim 3, characterized in that, The method for obtaining the optimal displacement field identification model in S3 includes the following steps: S31: Divide the sample dataset into a training set and a test set according to a preset ratio: S32: Train the cement-based material displacement field recognition model based on the training set to obtain the trained cement-based material displacement field recognition model. S33: Construct a model loss function that includes the design parameters to be optimized by introducing physical loss, and confirm whether the output of the cement-based material displacement field identification model after training converges based on the test set; If so, the cement-based material displacement field identification model trained at this time is taken as the optimal displacement field identification model; if not, the parameter weights of the cement-based material displacement field identification model trained are adaptively adjusted based on the backpropagation method, and the design parameters to be optimized in the model loss function are optimized based on the Grey Wolf Optimization Algorithm (GWO) to obtain the latest model loss function and repeat step S32.

5. The method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model according to claim 4, characterized in that, The model loss function constructed in S33, which includes the design parameters to be optimized, is expressed as follows: （1） （2） （3） （4） In the formula: Represents the model loss function; This represents the weight coefficient corresponding to each loss function term; The average absolute error, or physical loss, represents the predicted physical quantities, namely the transverse displacement field, the longitudinal displacement field, and the resultant displacement field. This represents the mean absolute error after normalization for each sample. N Indicates the number of pixels in the image dimension; These represent the predicted and experimental values ​​of the physical quantity corresponding to a certain pixel, respectively. Represents the mixed loss function; This indicates the optimization weight parameters, i.e., the design parameters to be optimized; sum() represents the summation function; Represents the corresponding sample The predicted physical quantity; Represents the corresponding sample The experimental values; PL This represents physical constraint terms.

6. The method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model according to claim 5, characterized in that, The method for optimizing the design parameters in the model loss function based on the Grey Wolf Optimization Algorithm (GWO) in S33 includes the following steps: S331: Randomly obtain the initial population and maximum number of iterations for the Grey Wolf optimization algorithm; Furthermore, the position of each gray wolf in the initial group is defined as the combination of design parameters to be optimized. A feasible solution; S322: Constructing the fitness function for the Grey Wolf optimization algorithm: And the expression for the fitness function is: （5） （6） In the formula: S Represents the fitness function; These represent the average errors generated during the training process for the transverse displacement field, longitudinal displacement field, and resultant displacement field, respectively. mean() means taking the average value; S333: Obtain the fitness value of each gray wolf individual in the initial population according to the fitness function, and obtain the sequence list by arranging the gray wolf individuals in descending order according to the fitness value; The first three gray wolf individuals in the sequence list were selected as Alpha wolves, Beta wolves, and Delta wolves, respectively. S334: Confirm whether the current iteration count has reached the maximum iteration count; If so, then the current position of the Alpha Wolf will be taken as the optimal combination of design parameters. ; If not, the initial group is iteratively updated based on the locations of the Alpha wolves, Beta wolves, and Delta wolves to obtain a new generation of group; Define the new generation as the initial group and repeat step S333.

7. The method for identifying the displacement field of cement-based materials based on a one-dimensional convolutional model according to claim 1, characterized in that, In S1, the label data is normalized to obtain the sample dataset, and its expression is: （7） In the formula: This represents the normalized label data, i.e., the physical quantity value. These represent the upper and lower limits of the physical quantity values ​​in the original dataset, respectively. This represents the label data to be normalized. This represents the given upper and lower bounds of normalization.

Images