CN-121564518-B - Cement-based material displacement field identification method based on one-dimensional convolution module

CN121564518BCN 121564518 BCN121564518 BCN 121564518BCN-121564518-B

Abstract

The invention discloses a cement-based material displacement field identification method based on a one-dimensional convolution module, which comprises the steps of S1, extracting a displacement field according to a speckle image, taking a speckle image pair as characteristic data, taking a transverse displacement field, a longitudinal displacement field and a proper displacement field as tag data, carrying out normalization processing on the tag data to obtain a sample data set, S2, constructing a cement-based material displacement field identification model based on the one-dimensional convolution module, S3, constructing a model loss function containing design parameters to be optimized by introducing physical loss, carrying out model training on the cement-based material displacement field identification model based on the one-dimensional convolution module according to the sample data set to obtain an optimal displacement field identification model, and carrying out cement-based material displacement field identification according to the optimal displacement field identification model. The method solves the problems of insufficient recognition precision and efficiency of the cement-based material displacement field in the existing method.

Inventors

CHEN XUEGUANG
Mu Hanping
ZHANG YU
LIU SHIWEI
ZHOU YINGXIANG
HU ZHIQIANG
LIU YI
NIU WENJUAN
Han Shoudu
LUO WEIBANG
Niu Wanji

Assignees

大连理工大学
新疆水利水电勘测设计研究院有限责任公司

Dates

Publication Date: 20260505
Application Date: 20260121

Claims (7)

1. The cement-based material displacement field identification method based on the one-dimensional convolution module is characterized by comprising the following steps of: S1, acquiring a speckle image pair of a cement-based material, wherein the speckle image pair comprises a reference speckle image and a deformation speckle image which are at the same visual angle and the same resolution; and the displacement field comprises a transverse displacement field and a longitudinal displacement field; Acquiring a closed displacement field according to the transverse displacement field and the longitudinal displacement field, and taking the speckle image pair as characteristic data; s2, constructing a cement-based material displacement field identification model based on a one-dimensional convolution module, wherein the cement-based material displacement field identification model comprises a two-dimensional characteristic order reduction module, a first linear layer, a one-dimensional convolution coding module, a second linear layer, a third linear layer, a one-dimensional deconvolution decoding module and a displacement field output module; The two-dimensional feature order reduction module is used for converting the speckle image pair into a gray tensor, then carrying out channel dimension stitching to obtain a stitched image, and carrying out two-dimensional order reduction processing on the stitched image based on a preset two-dimensional multi-channel feature extraction layer so as to extract the two-dimensional speckle image deformation feature of the stitched image after the two-dimensional order reduction processing; the first linear layer is used for converting the deformation characteristic of the two-dimensional speckle image into a one-dimensional signal sequence characteristic; The one-dimensional convolution coding module is used for carrying out multi-scale feature extraction on the one-dimensional signal sequence features to obtain one-dimensional deformation coding features, and the one-dimensional deformation coding features are used for representing displacement association relations between reference speckle images and deformation speckle image pairs; The second linear layer is used for performing dimension reduction coding operation on the one-dimensional deformation coding feature, and the third linear layer is used for performing dimension increasing operation on the output of the second linear layer so as to output a plurality of channel sequence features with the same signal length as the one-dimensional signal sequence features; The one-dimensional deconvolution decoding module is used for decoding the sequence characteristics of each channel and acquiring the displacement sequence characteristics of each channel; The displacement field output module is used for mapping the displacement sequence characteristics of each channel to a two-dimensional displacement field space with preset resolution to obtain a mapped two-dimensional displacement field, and then performing interpolation processing on the mapped two-dimensional displacement field based on a cubic spline interpolation method to realize pixel-level registration with a speckle image input into the two-dimensional characteristic reducer so as to predict and output a transverse displacement field, a longitudinal displacement field and an appropriate displacement field; And S3, constructing a model loss function containing design parameters to be optimized by introducing physical loss, carrying out model training on a cement-based material displacement field identification model based on a one-dimensional convolution model according to a sample data set to obtain an optimal displacement field identification model, and realizing the displacement field identification of the cement-based material according to the optimal displacement field identification model.
2. The cement-based material displacement field identification method based on the one-dimensional convolution module according to claim 1, wherein the one-dimensional convolution coding module comprises a plurality of one-dimensional convolution layers with different convolution kernels and convolution step sizes which are sequentially connected, the channel number of each one-dimensional convolution layer is 64, 128, 256, 128, 64 and 4 sequentially to form a coding bottleneck structure, and the method is used for carrying out multi-scale sequence feature extraction and compression on one-dimensional signal sequence features output by a first linear layer to obtain one-dimensional deformation coding features.
3. The cement-based material displacement field identification method based on a one-dimensional convolution module according to claim 2, wherein the one-dimensional deconvolution decoding module comprises a plurality of deconvolution collators with different convolution kernels and convolution step sizes which are connected in sequence; the deconvolution collator is a deconvolution collator block structure formed by connecting a single deconvolution layer with a convolution layer through RELU activation function layers; The deconvolution layer is used for executing deconvolution operation on the input of the deconvolution layer; the RELU activation function layer is used for executing nonlinear activation operation on the output of the deconvolution layer; The convolution layer is used for performing convolution operation on the output of the RELU activation function layer; and recovering the one-dimensional deformation coding characteristic to a multi-channel sequence characteristic with the same signal length as the one-dimensional signal sequence characteristic through a plurality of deconvolution collators with different convolution kernels and convolution step sizes, and decoding each channel sequence characteristic to obtain each channel displacement sequence characteristic.
4. The method for identifying the displacement field of the cement-based material based on the one-dimensional convolution module according to claim 3, wherein the method for obtaining the optimal displacement field identification model in the step S3 comprises the following steps: s31, dividing the sample data set into a training set and a testing set according to a preset proportion: s32, carrying out model training on the cement-based material displacement field identification model based on the one-dimensional convolution module according to the training set to obtain a trained cement-based material displacement field identification model; s33, constructing a model loss function containing design parameters to be optimized by introducing physical loss, and confirming whether the output of the trained cement-based material displacement field identification model is converged or not according to a test set; if not, the parameter weight of the trained cement-based material displacement field recognition model is self-adaptively adjusted based on a back propagation method, and meanwhile, the design parameters to be optimized in the model loss function are optimized based on a gray wolf optimization algorithm GWO, so that the latest model loss function is obtained, and the step S32 is repeatedly executed.
5. The method for recognizing displacement fields of cement-based materials based on one-dimensional convolution model as set forth in claim 4, wherein the model loss function constructed in S33 includes design parameters to be optimized, and its expression is as follows (1) (2) (3) (4) Wherein: representing a model loss function; representing the weight coefficient corresponding to each loss function item; representing the predicted physical quantity, namely the transverse displacement field, the longitudinal displacement field and the mean absolute error of the resultant displacement field, namely the physical loss; N represents the number of pixels in the dimension of the image; respectively representing a predicted value and a test value of a physical quantity corresponding to a certain pixel; representing a mixing loss function; representing optimization weight parameters, namely design parameters to be optimized; Representing corresponding samples Is a predicted physical quantity of (a); Representing corresponding samples PL represents the physical constraint term.
6. The method for identifying the displacement field of the cement-based material based on the one-dimensional convolution model according to claim 5, wherein the method for optimizing the design parameters to be optimized in the model loss function based on the gray wolf optimization algorithm GWO in the step S33 specifically comprises the following steps: s331, randomly acquiring an initial population and the maximum iteration number of a gray wolf optimization algorithm; and the position of each gray wolf in the initial group is defined as the combination of design parameters to be optimized Is a feasible solution of (a); s322, constructing an adaptability function of a gray wolf optimization algorithm: And the fitness function has the expression of (5) (6) Wherein S represents an fitness function; Respectively represent the average error generated by the transverse displacement field, the longitudinal displacement field and the resultant displacement field in the training process Mean () represents taking the average; s333, acquiring fitness values of all the wolf individuals in the initial population according to the fitness function, and acquiring a sequence table by descending the sequence of the wolf individuals according to the fitness values; The first three gray wolves in the sequence table are selected as Alpha wolves, beta wolves and Delta wolves in sequence; s334, confirming whether the current iteration number reaches the maximum iteration number or not; if yes, the position of Alpha wolf is taken as the optimal design parameter combination ; If not, carrying out iterative updating on the initial group according to the positions of the Alpha wolves, the Beta wolves and the Delta wolves to obtain a new generation group; the new generation population is defined as the initial population, and step S333 is repeatedly performed.
7. The cement-based material displacement field identification method based on a one-dimensional convolution module as claimed in claim 1, wherein the normalization processing is performed on the tag data in S1 to obtain a sample data set, and the expression is as follows (7) Wherein: Representing the physical quantity value of the label data after normalization processing; Respectively representing an upper limit and a lower limit of the physical quantity value in the original data set; representing tag data to be normalized; Representing a given upper and lower normalization limit.

Description

Cement-based material displacement field identification method based on one-dimensional convolution module Technical Field The invention relates to the technical field of Digital Image Correlation (DIC) based on deep learning, in particular to a cement-based material displacement field identification method based on a one-dimensional convolution module. Background Cement-based materials are a typical class of heterogeneous brittle materials that are difficult to distinguish by the naked eye during compression with small deformations and with significant differences in deformation field span between the early and late stages of loading. Under the condition, the existing prediction model has the defects of neglecting small deformation in the early stage of loading, overlarge consumption of computing resources, overlong training time and the like. The traditional DIC method based on the algorithm is highly dependent on experience parameters in deformation field prediction, the selection of subset size, step length and shape function directly influences measurement accuracy, and meanwhile, the calculation efficiency is low. For this reason, deep learning based DIC approaches have emerged in recent years, most of which employ U-Net network architecture to directly conduct end-to-end predictions of deformation fields and rely on the generation of datasets for training. However, the displacement mode adopted by the constructed generated data set is often too ideal, so that complex non-uniform deformation rules in the uniaxial compression process of the cement-based material are difficult to reproduce, and the trained model cannot ensure the recognition effect with high time resolution. Meanwhile, the common loss functions are MAE, MAPE and the like, in the prediction research of the deformation field of the cement-based material, the former is easy to ignore small deformation in the earlier stage of loading, and the latter is slow to converge and has insufficient precision. In addition, the direct adoption of high-resolution deformation field data for model training can bring the problems of large calculation resource consumption and long training time consumption, and most of the existing lightweight designs only stay in reducing the number of convolution layers, so that efficiency and accuracy are difficult to be considered. These deficiencies make the existing methods still have significant limitations in high resolution and complex deformation field predictions. Disclosure of Invention The invention provides a cement-based material displacement field identification method based on a one-dimensional convolution module, which aims to overcome the technical problems. In order to achieve the above object, the technical scheme of the present invention is as follows: A cement-based material displacement field identification method based on a one-dimensional convolution module specifically comprises the following steps: S1, acquiring a speckle image pair of a cement-based material, wherein the speckle image pair comprises a reference speckle image and a deformation speckle image which are at the same visual angle and the same resolution; and the displacement field comprises a transverse displacement field and a longitudinal displacement field; Acquiring a closed displacement field according to the transverse displacement field and the longitudinal displacement field, and taking the speckle image pair as characteristic data; s2, constructing a cement-based material displacement field identification model based on a one-dimensional convolution module, wherein the cement-based material displacement field identification model comprises a two-dimensional characteristic order reduction module, a first linear layer, a one-dimensional convolution coding module, a second linear layer, a third linear layer, a one-dimensional deconvolution decoding module and a displacement field output module; The two-dimensional feature order reduction module is used for converting the speckle image pair into a gray tensor, then carrying out channel dimension stitching to obtain a stitched image, and carrying out two-dimensional order reduction processing on the stitched image based on a preset two-dimensional multi-channel feature extraction layer so as to extract the two-dimensional speckle image deformation feature of the stitched image after the two-dimensional order reduction processing; the first linear layer is used for converting the deformation characteristic of the two-dimensional speckle image into a one-dimensional signal sequence characteristic; The one-dimensional convolution coding module is used for carrying out multi-scale feature extraction on the one-dimensional signal sequence features to obtain one-dimensional deformation coding features, and the one-dimensional deformation coding features are used for representing displacement association relations between reference speckle images and deformation speckle image pairs; The