Frozen soil deformation measurement method, device and system based on deep neural network

Abstract

The invention provides a frozen soil deformation measurement method, device and system based on a deep neural network, and relates to the technical field of frozen soil deformation measurement. According to the method, multi-modal fusion analysis is carried out on the frozen soil physical characteristics, the test working conditions and the imaging conditions through the condition generation network, the optimal grading parameters highly matched with the sample can be rapidly generated, the problems of poor subjectivity and repeatability of artificial experience are solved, and the test preparation period is greatly shortened. And meanwhile, the system takes imaging quality, physical implementability and manufacturability as a joint optimization target, the physical constraints such as sedimentation resistance and segregation resistance are met on the premise of ensuring the DIC measurement feature point density and texture quality by the grading parameters, and the calculation precision of displacement field calculation is remarkably improved. According to the invention, by fusing the physical mechanism and the imaging quality, the problem of low precision and efficiency of current frozen soil deformation measurement is effectively solved, automatic and intelligent generation of the grading parameters of the tracer particles is realized, and the measurement precision and efficiency are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of frozen soil deformation measurement, and particularly relates to a frozen soil deformation measurement method, device and system based on a deep neural network. BACKGROUND

[0002] Frozen soil is widely distributed in cold regions, and its frost heaving deformation phenomenon has an important influence on the stability of highways, railways, building foundations and underground engineering in cold regions. Traditional frozen soil deformation measurement methods mainly rely on displacement meters, LVDT (Linear Variable Differential Transformer) sensors and other sensors for one-dimensional total deformation measurement. However, these methods cannot analyze the local deformation field characteristics of frozen soil, can only measure the overall axial displacement, and are difficult to reveal the local strain distribution and subtle deformation law of frozen soil in the freezing and thawing process.

[0003] In recent years, digital image correlation (DIC) technology and particle image velocimetry (PIV) technology have been gradually introduced into the field of frozen soil deformation measurement due to their non-contact, full-field and high-precision advantages. However, in tests on silt, silty clay and other fine-grained frozen soil, the application effect of these technologies is severely restricted.

[0004] Existing methods highly rely on manual experience and trial-and-error to select and optimize tracer particles, which leads to low efficiency, long cycle and high cost in the design process; at the same time, due to the lack of objective standards, subjective bias of different operators directly affects the reliability of the scheme, resulting in poor experimental repeatability. In addition, the fixed tracer particle ratio is difficult to adapt to frozen soil samples with different physical properties (such as particle composition and water content), showing poor adaptability and severely restricting the accuracy and efficiency of frozen soil deformation measurement. SUMMARY

[0005] The present application provides a frozen soil deformation measurement method, device and system based on a deep neural network, which solves the problem of low accuracy and efficiency of current frozen soil deformation measurement, and improves the accuracy and efficiency of frozen soil deformation measurement.

[0006] In a first aspect, the present application provides a frozen soil deformation measurement method based on a deep neural network, which comprises: acquiring physical property parameters, test condition parameters and imaging system parameters of a frozen soil sample; based on the physical property parameters, test condition parameters and imaging system parameters of the frozen soil sample, constructing a conditional input vector; based on the conditional input vector and a preset condition generation network, taking an imaging quality target, a physical implementability target and a manufacturability target as optimization targets, performing multi-objective optimization solving to determine initial grading parameters of tracer particles; the initial grading parameters include particle size distribution, black and white particle ratio and particle dosage; based on the initial grading parameters, performing imaging quality evaluation and parameter optimization to determine optimized grading parameters of the tracer particles; based on the optimized grading parameters, preparing a frozen soil sample to be measured for frozen soil deformation measurement.

[0007] Secondly, embodiments of the present invention provide a frozen soil deformation measurement device based on a deep neural network. The frozen soil deformation measurement device includes: a communication module and a processing module. The communication module is used to acquire physical property parameters, test condition parameters, and imaging system parameters of a frozen soil sample. The processing module is used to construct a conditional input vector based on the physical property parameters, test condition parameters, and imaging system parameters of the frozen soil sample. Based on the conditional input vector and a preset conditional generation network, multi-objective optimization is performed with imaging quality objectives, physical feasibility objectives, and manufacturability objectives as optimization objectives to determine the initial gradation parameters of the tracer particles. The initial gradation parameters include particle size distribution, black-and-white particle ratio, and particle content. Based on the initial gradation parameters, imaging quality assessment and parameter tuning are performed to determine the optimized gradation parameters of the tracer particles. Based on the optimized gradation parameters, a frozen soil sample to be tested is prepared, and frozen soil deformation measurement is performed.

[0008] Thirdly, embodiments of the present invention provide a frozen soil deformation measurement system based on a deep neural network. The frozen soil deformation measurement system includes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0009] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0010] This invention provides a method, device, and system for measuring frozen soil deformation based on deep neural networks. By using a conditional generation network to perform multimodal fusion analysis of frozen soil physical properties, experimental conditions, and imaging conditions, this invention can rapidly generate optimal gradation parameters highly adapted to the sample, overcoming the subjectivity and poor repeatability of manual experience and significantly shortening the experimental preparation cycle. Simultaneously, the system uses imaging quality, physical feasibility, and manufacturability as joint optimization objectives to ensure that the generated parameters, while guaranteeing the density and texture quality of DIC measurement feature points, meet physical constraints such as anti-settlement and anti-segregation, thereby significantly improving the accuracy of displacement field calculations. By integrating physical mechanisms and imaging quality, this invention effectively solves the current problems of low accuracy and efficiency in frozen soil deformation measurement, realizing the automated and intelligent generation of tracer particle gradation parameters, and significantly improving measurement accuracy and efficiency. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating a method for measuring frozen soil deformation based on a deep neural network, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall system architecture and process for measuring frozen soil deformation provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of input and output data provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a conditional generation network and an objective function provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a physical constraint embedding mechanism provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of an imaging-measurement closed-loop optimization and online calibration process provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of DIC measurement of a frozen soil deformation field provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a frozen soil deformation measurement device based on a deep neural network provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0014] In the description of this invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" and "more than one" refer to two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0015] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0016] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include other steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0018] As described in the background section, the gradation design process in frozen soil deformation measurement is inefficient and heavily reliant on subjective factors. Manual trial and error is not only time-consuming but also difficult to adapt to different types of frozen soil (such as silt and silty clay) or variable external environmental conditions (such as temperature gradient changes and freeze-thaw cycles). This inefficiency and uncertainty severely restrict the repeatability of experiments and the feasibility of large-scale implementation.

[0019] More importantly, existing methods often overlook the compatibility between the physical properties of frozen soil samples (such as moisture content, density, and native particle composition) and the tracer particles. This lack of "hard matching" easily leads to uneven particle distribution within the soil, inducing aggregation, sedimentation, or segregation between particles. During frost heave, these phenomena further exacerbate the degradation of surface texture, directly causing problems in DIC / PIV analysis such as increased average displacement error (usually exceeding 15%), decreased autocorrelation index (often below 0.6), and insufficient density of traceable feature points (mostly below 0.05 per pixel²), significantly affecting the accuracy and reliability of deformation field reconstruction.

[0020] Furthermore, while adding large-particle materials such as black and white quartz sand is a common practice to improve image contrast, its gradation design, including particle size distribution, dosage, and black-and-white ratio, still lacks systematic physical guidance and constraints. Key process factors such as the settlement tendency caused by density differences between particles and soil, and the weak anti-segregation ability due to material inhomogeneity during frost heave, have not received sufficient attention. The result is often that while feature recognition is improved to some extent, deformation compatibility and measurement consistency are sacrificed, ultimately introducing a significant systematic error.

[0021] In summary, there is an urgent need in the field of permafrost DIC / PIV technology to develop an intelligent generation method for tracer particle gradation parameters. This invention deeply integrates the physical properties of the permafrost medium with the technical requirements of deformation observation, and can achieve a synergistic improvement in accuracy, efficiency, and robustness while ensuring imaging quality (e.g., through MIG index and feature point density assessment) and physical feasibility (e.g., resistance to subsidence and segregation). MIG, n_f, r_peak, and Cov are used as the main imaging quality indicators, and are co-optimized with the two types of physical feasibility constraints of resistance to subsidence and segregation.

[0022] This invention combines deep neural networks to reduce the time and cost of adjusting tracer particle gradation in experiments. By training a deep neural network model, optimal tracer particle gradation parameters are automatically generated, reducing reliance on manual experience, improving design efficiency, and lowering experimental adjustment time and costs. It also improves the accuracy of refined measurements of frozen soil deformation. Based on the initial characteristics of frozen soil samples (such as water content and particle size distribution), optimized tracer particle gradation parameters are generated using a deep learning model, ensuring the uniform distribution and stability of tracer particles in the frozen soil sample, thereby improving the accuracy of DIC technology in frozen soil deformation measurements. This invention enhances the refined measurement capabilities of frozen soil experiments, achieving intelligent and precise tracer particle gradation design, and overcoming the shortcomings of existing technologies.

[0023] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a method for measuring frozen soil deformation based on a deep neural network. The method includes steps S101-S105.

[0024] S101. Obtain the physical property parameters, test condition parameters, and imaging system parameters of the frozen soil sample.

[0025] S102. Based on the physical property parameters, test condition parameters, and imaging system parameters of the frozen soil samples, construct the conditional input vector.

[0026] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1023.

[0027] S1021. Standardize the physical property parameters, test condition parameters, and imaging system parameters to eliminate dimensions and obtain standardized parameters for each category.

[0028] In some embodiments, physical property parameters include the moisture content, dry density, natural particle size distribution curve, and cemented salt content of the frozen soil sample. Test condition parameters include temperature gradient and freezing rate; imaging system parameters include camera parameters, light source conditions, and specimen size.

[0029] S1022. Based on standardized parameters, perform one-hot encoding to obtain numerical parameters for each category.

[0030] S1023. Based on the numerical parameters of each category, perform splicing and fusion to obtain the conditional input vector.

[0031] For example, a conditional input vector can be represented as follows.

[0032] ; Among them, the physical property parameters include moisture content w and initial dry density ρ d Natural particle size distribution curve P(d) and cementation / salt index ξ; experimental operating parameters include temperature gradient. The parameters are T, the expected freezing rate f, and the imaging system parameters, including camera parameters, light source conditions, and specimen size. Continuous variables are standardized, and categorical variables are one-hot encoded to form a standardized input vector.

[0033] S103. Based on the conditional input vector and the preset conditional generation network, multi-objective optimization is performed with imaging quality target, physical feasibility target and manufacturability target as optimization targets to determine the initial gradation parameters of the tracer particles.

[0034] In some embodiments, the initial gradation parameters include particle size distribution, black and white particle ratio, and particle doping amount.

[0035] As one possible implementation, such as Figure 3 As shown, in this embodiment of the invention, the conditional input vector can be input into the generator of the preset conditional generation network for forward computation, and the initial gradation parameters of the tracer particles can be output. The preset conditional generation network is trained based on the conditional input vectors and gradation parameters corresponding to multiple permafrost samples. During the training process, the comprehensive objective function composed of imaging quality target, physical feasibility target and manufacturability target is used as the optimization criterion.

[0036] In some embodiments, a pre-defined conditional generative network, i.e. a pre-defined generative model, includes, but is not limited to, a conditional generative adversarial network, a conditional diffusion model, a variational autoencoder, or a conditional Transformer. The generative model can be any implementation of cGAN, conditional diffusion, VAE, or conditional Transformer.

[0037] In some embodiments, the synthesis objective function is defined as: ; in, J imaging At least a weighted average of MIG, n_f, r_peak, and Cov, J physics It must contain at least anti-settlement and anti-segregation constraints; J cost It should include at least material costs and preparation time penalties.

[0038] In some embodiments, a comprehensive objective function J is constructed by fusing the dual objectives of imaging quality and physical feasibility.

[0039] For example, the imaging quality sub-target J imaging Based on digital image correlation (DIC) image quality evaluation metrics, this is a weighted combination of mean gray-level gradient (MIG), subset autocorrelation peak (r_peak), traceable feature point density nf, and texture coverage Cov. MIG = (1 / |Ω|)·∑_{(x,y)∈Ω}| I(x,y)|, where I is an 8-bit grayscale value, Ω is the measurement area; nf: the density of traceable feature points per unit pixel area (number / px) 2 ); r_peak: peak value of standard zero-displacement autocorrelation; Cov: texture coverage ratio.

[0040] For example, the physical implementability sub-objective J physics Introducing anti-settlement and anti-segregation constraints, and constructing a penalty term based on factors such as the density difference Δρ between tracer particles and soil, Stokes number, settlement time scale, static angle of repose, and ice migration risk.

[0041] For example, manufacturable sub-target J cost Factors such as material costs, preparation time, and process complexity should be considered.

[0042] The overall objective function J is expressed as: ; Where α, β, and γ are the weight coefficients of each sub-objective, α, β, and γ [0,1], α+β+γ=1. J imaging Contains MIG, n_f, r_peak, Cov; J physicsIncludes anti-settlement / anti-segregation constraints; J cost Includes cost / time penalties.

[0043] In some embodiments, such as Figure 4 As shown, the preset conditional generative network adopts the conditional generative adversarial network (cGAN) architecture, the core of which is to learn the complex mapping from the conditional input x to the optimal parameter y through adversarial training.

[0044] ; Among them, the particle size distribution curve P(d i Black and white particle ratio k bw Tracer particle doping φ, target surface coverage C ov And the set of process parameters u. Simultaneously, the generator can output a synthesized texture image T. syn It is used for offline evaluation of image quality.

[0045] For example, the network structure of a pre-defined conditional generation network is as follows: The generator G receives a conditional vector x and random noise z, and outputs a gradation parameter vector y and a synthesized texture image T through a series of fully connected or transposed convolutional layers. syn Discriminator D then receives the actual data pairs (x, y). real Or generate data pairs (x, y) fake They tried to distinguish between genuine and fake.

[0046] For example, a loss function for a pre-defined conditional generation network: adversarial loss (such as W). asserstein The loss function ensures that the generated distribution approximates the real data distribution. Simultaneously, the comprehensive objective function J is introduced into the generator training as an important reinforcement learning reward signal or auxiliary loss term, directly guiding the model to converge towards the optimization objective.

[0047] An example of the training process and hyperparameters for a pre-defined conditional generator network: Data partitioning: Using the hold-out method, the historical dataset is partitioned into training, validation, and test sets in a 7:2:1 ratio. Optimizer: Both the generator G and discriminator D use the Adam optimizer. Hyperparameter settings: The learning rate is set to 1e-4, the batch size is set to 32 or 64 depending on memory, the number of training iterations (epochs) is 1000, and an early stopping strategy is used to prevent overfitting (training terminates when the validation set loss no longer decreases for 20 consecutive epochs). Figure 5 As shown, constraint embedding: physical prior constraints are implemented by adding penalty terms (such as L2 regularization, penalty functions for physical inequality constraints) to the training loss. Through the above training, the generator G learns to output parameters that conform to both the data distribution and the requirements of multi-objective optimization under given conditions.

[0048] In some embodiments, such as Figure 5 As shown, this invention can also embed physical priors and manufacturability constraints. By designing a parameter projection layer and introducing a penalty term in the loss function, physical priors and manufacturability constraints are embedded: particle size availability constraint: limiting P(di) to a commercially available particle size range; blending feasibility constraint: ensuring that the doping amount φ and coverage C are consistent. ov It meets the accessibility requirements of actual coating or blending processes; anti-settling constraints: the particle settling distance is estimated based on density difference Δρ, particle size d, viscosity μ and freezing rate f, and is limited to not exceeding a set threshold; anti-segregation constraints: the variance of the volume fraction of black and white particles in the layer and between layers is controlled to not exceed the allowable range.

[0049] S104. Based on the initial gradation parameters, perform imaging quality assessment and parameter optimization to determine the optimal gradation parameters for the tracer particles.

[0050] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1045.

[0051] S1041. Based on the initial gradation parameters, a synthetic texture image is generated using a texture synthesis algorithm.

[0052] S1042. Perform DIC quality assessment on the synthesized texture image and calculate the imaging quality index.

[0053] In some embodiments, imaging quality metrics include average gray-level gradient (MIG), traceable feature point density, texture coverage, and subset autocorrelation peak.

[0054] S1043. Based on the physical feasibility objective, the initial gradation parameters are used as the starting point for optimization. The adjustable range of each parameter is determined, and the parameter optimization space is constructed.

[0055] In some embodiments, the parameter optimization space includes the particle size distribution range, the doping process range, and the black-white ratio range.

[0056] S1044. Taking the initial gradation parameters as the starting point for optimization, within the parameter optimization space, with the goal of improving the imaging quality index, Bayesian optimization algorithm is used for iterative optimization.

[0057] S1045. When the iterative optimization process reaches the convergence condition, the optimal gradation parameters are output as the optimized gradation parameters.

[0058] In some embodiments, convergence conditions include the imaging quality index reaching a preset threshold, the objective function improvement being less than a set tolerance, or the maximum number of iterations being reached.

[0059] For example, MIG≥M0, n_f≥n0 / px 2The threshold is reached when any combination of r_peak ≥ r0, or when the relative improvement of the comprehensive objective function is < δ, or when the number of iterations reaches N_max.

[0060] For example, such as Figure 6 As shown, this invention provides a schematic diagram of an imaging-measurement closed-loop optimization and online calibration process. This invention can acquire images of prepared local samples to obtain initial images of their surface texture. The acquired images are then preprocessed, including enhancement and denoising, to improve image quality and prepare for subsequent analysis. This invention can use Digital Image Correlation (DIC) or Particle Image Velocimetry (PIV) techniques to analyze the preprocessed image sequence and calculate the full-field deformation field (such as displacement and strain field) of the frozen soil sample. This invention automatically quantifies and evaluates the image quality used for calculation. Key indicators include: MIG (Mean Gray Gradient): measures image contrast; n_f (Number of Traceable Feature Points): measures the number of texture features available for DIC / PIV matching; and SSIM (Structural Similarity): assesses the integrity of the image structure. Based on the quality evaluation results, this invention can initiate an optimization module. A multi-objective optimization algorithm is used, with the evaluation indicators as the optimization target, to automatically adjust and generate new and better tracer particle gradation parameters.

[0061] Embodiments of the present invention can utilize synthetic image T syn Alternatively, DIC / PIV quality assessment can be performed using small sample real-world images, calculating indicators such as MIG, nf, spurious displacement rate, and correlation coefficient distribution. Based on the assessment results, Bayesian optimization or reinforcement learning algorithms are used to iteratively update the hyperparameters and loss weights α, β, and γ of the generator G, making the generation strategy adaptable to specific soil sample characteristics and imaging device conditions, thus achieving closed-loop optimization.

[0062] S105. Based on the optimized gradation parameters, prepare frozen soil samples to be tested and perform frozen soil deformation measurement.

[0063] As one possible implementation, step S105 can be specifically implemented as steps S1051-S1054.

[0064] S1051. Prepare a tracer particle mixture according to the particle size distribution, black and white particle ratio and particle dosage in the optimized gradation parameters.

[0065] S1052. Based on the set of process parameters in the optimized gradation parameters, the tracer particle mixture is uniformly mixed with the frozen soil sample, and the mixing time, stirring speed and vibration intensity are controlled to prepare a frozen soil sample with a surface texture that meets the set conditions.

[0066] In some embodiments, the setting conditions include the surface coverage of the tracer particles being greater than a coverage threshold, and / or the surface contrast being greater than a contrast threshold.

[0067] S1053. Place the prepared frozen soil sample to be tested in a freeze-thaw test chamber and conduct a freeze-thaw cycle test under the set temperature gradient and freezing rate conditions.

[0068] S1054. Digital image correlation (DIC) technology is used to continuously acquire sequential images of the sample surface during the freeze-thaw process through an imaging system.

[0069] S1055. Perform DIC analysis on the sequence images to calculate the deformation measurement results of the frozen soil sample.

[0070] In some embodiments, the deformation measurement results include full-field deformation field and displacement field data.

[0071] For example, such as Figure 7 As shown, this invention achieves refined acquisition of the deformation field of frozen soil through a complete digital image correlation (DIC) measurement process. The invention utilizes frozen soil samples prepared with optimized gradation parameters and embedded tracer particles. The samples are placed in a controlled freezing test chamber, and under conditions simulating real freeze-thaw environments and potentially applying axial loads, a DIC image acquisition system consisting of a high-resolution camera and a constant light source continuously acquires image sequences during the stress-deformation process. After acquiring digital images before and after deformation, the core DIC analysis stage begins: first, a calculation sub-region is selected on the reference image, and a calculation grid covering the entire field is planned; then, each sub-region is searched and matched with subpixel precision in the target image; finally, the displacement and strain of each sub-region are calculated, thus obtaining the displacement and strain data for the entire field. These data are further transformed into intuitive visualization results, including a full-field displacement cloud map showing the distribution of displacement magnitude, a vector map revealing the magnitude and direction of local deformation, and a full-field strain cloud map reflecting the distribution of each strain component. Through this systematic processing flow, the present invention can output quantitative, full-field analysis results of frozen soil deformation field, fundamentally overcoming the limitation of traditional sensors that can only perform one-dimensional point measurements, and realizing the accurate capture and characterization of local deformation details in the process of frozen soil frost heave and thaw settlement.

[0072] For example, embodiments of the present invention can perform tracer particle preparation and loading processes based on the generated parameter vector y, including according to P(di) and k bw The particles are mixed and incorporated into the soil sample at a dosage of φ. The sample preparation is completed by controlling process parameters (such as mixing time, stirring speed, layer thickness, and compaction strength) to finally obtain a testable specimen that meets the requirements of surface coverage and contrast.

[0073] This invention provides a method for measuring frozen soil deformation based on deep neural networks. By using a conditional generation network to perform multimodal fusion analysis of frozen soil physical properties, experimental conditions, and imaging conditions, it can rapidly generate optimal gradation parameters highly adapted to the sample, overcoming the subjectivity and poor repeatability of manual experience and significantly shortening the experimental preparation cycle. Simultaneously, the system uses imaging quality, physical feasibility, and manufacturability as joint optimization objectives, ensuring that the generated parameters meet physical constraints such as anti-settlement and anti-segregation while guaranteeing the density and texture quality of DIC measurement feature points, thereby significantly improving the accuracy of displacement field calculations. This invention effectively solves the current problems of low accuracy and efficiency in frozen soil deformation measurement by integrating physical mechanisms and imaging quality, realizing the automated and intelligent generation of tracer particle gradation parameters, and significantly improving measurement accuracy and efficiency.

[0074] Optionally, the frozen soil deformation measurement method based on deep neural networks provided in this embodiment of the invention further includes steps S201-S205 before step S103.

[0075] S201. Obtain data from multiple permafrost samples from historical periods.

[0076] In some embodiments, the permafrost sample data includes physical property parameters, test condition parameters, imaging system parameters, and tracer particle gradation parameters.

[0077] S202. Based on the physical property parameters, test condition parameters, and imaging system parameters of multiple frozen soil sample data, determine the conditional input vectors for multiple frozen soil samples.

[0078] S203. Using the conditional input vector of each frozen soil sample as the conditional input and the gradation parameters of each frozen soil sample as the real data, generate a real sample set.

[0079] S204. Construct a conditional generation network, which includes a generator and a discriminator. S205. Using the conditional input of each sample in the real sample set as the input of the generator and discriminator, and using the gradation parameters of each sample as the real sample, the conditional generation network is adversarially trained to obtain the preset conditional generation network.

[0080] For example, step S205 can be specifically implemented as steps A1-A10.

[0081] A1. Randomly sample a batch of samples from the real sample set to form the current batch of samples.

[0082] A2. Generate a random noise vector, and input the random noise vector and the conditional input vector of the current batch of samples into the generator to obtain the generation gradation parameters.

[0083] A3. With the generator parameters fixed, calculate the discriminator loss function based on the generated gradation parameters and the actual gradation parameters of the current batch of samples, and update the discriminator parameters using the gradient descent algorithm with the goal of minimizing the discriminator loss function.

[0084] A4. With the parameters of the discriminator fixed, the random noise vector and the conditional input vector are input into the generator again to obtain new generation gradation parameters.

[0085] A5. Calculate the discriminator output expectation of the new generation gradation parameters, and calculate the corresponding imaging quality target value, physical feasibility target value and manufacturability target value respectively.

[0086] A6. Based on the discriminator output expectation, imaging quality target value, physical feasibility target value, and manufacturability target value, the generator's comprehensive objective function is calculated by weighted summation.

[0087] A7. With the goal of minimizing the generator's comprehensive objective function, update the generator's parameters using the backpropagation algorithm.

[0088] A8. After each generator parameter update, apply physical constraints to its output gradation parameters, constrain the particle size distribution within a preset range through parameter projection, and ensure that the particle doping and surface coverage meet the process requirements through numerical trimming.

[0089] A9. Repeat the above iterative process. When the difference in the change of the comprehensive objective function value on the validation set over multiple consecutive iterations is less than the preset threshold, or when the number of iterations reaches the preset maximum number of iterations, terminate the training process.

[0090] A10. Use the finally trained generator as a preset condition to generate the network.

[0091] In this way, the present invention can pre-construct a condition generation network, comprehensively considering multiple objective factors such as imaging quality objectives, physical feasibility objectives, and manufacturability objectives, thereby improving the accuracy and efficiency of permafrost deformation measurement.

[0092] Optionally, the frozen soil deformation measurement method based on deep neural networks provided in this embodiment of the invention further includes steps S301-S304 after step S105.

[0093] S301. During the freeze-thaw test, the imaging quality index of the acquired sequence images is calculated in real time.

[0094] In some embodiments, imaging quality metrics include average gray-scale gradient (MIG) and traceable feature point density.

[0095] S302. When the imaging quality index is lower than the preset threshold, the online correction mechanism is activated to fine-tune and optimize the tracer particle parameters based on the current imaging quality index, and obtain the corrected gradation parameters.

[0096] In some embodiments, fine-tuning optimization employs a Bayesian optimization algorithm to perform a fast search within a small neighborhood of the initial optimization gradation parameters.

[0097] The small neighborhood is defined as the area within the L2 norm radius ε of the parameter vector y (ε∈[0.03,0.08], in terms of normalized parameters), and a single online correction iteration does not exceed N (N∈[3,6]). When ΔJ imaging Stop early when <δ or MIG, nf, r_peak reach the threshold.

[0098] S303. Based on the corrected gradation parameters, the tracer particle mixture is re-formulated and a new test sample is prepared to continue the frozen soil deformation measurement.

[0099] S304. Establish a measurement quality archive, recording the gradation parameters, imaging quality indicators, and deformation measurement results of each experiment, in order to optimize subsequent conditional generation network training.

[0100] For example, quality record records include sample ID, w, ρ d 、P(d),ξ、 T, f, imaging parameters, y_init / y_opt / y_corr, MIG, n_f, r_peak, Cov, error (MAE / RMSE), etc., to support subsequent retraining and adaptive threshold updates.

[0101] For example, in the embodiments of the present invention, calibration images can be acquired at the beginning of the test. If the imaging quality index (such as MIG or nf) does not reach the preset threshold M0 or n0, online fine-tuning is initiated: the parameter vector y is quickly adjusted based on small step gradient update or Bayesian optimization steps to achieve adaptive correction within a single test and ensure that the measurement start state meets the accuracy requirements.

[0102] Through the above process, this invention achieves intelligent, reliable, and adaptive generation of parameters from the physical properties and test conditions of frozen soil to the tracer particle gradation parameters. It effectively solves the problems of relying on human experience, ignoring physical constraints, and lacking system optimization in the existing technology, and significantly improves the accuracy, efficiency, and reliability of frozen soil deformation measurement.

[0103] This invention constructs a closed-loop intelligent optimization system that integrates physical mechanisms, manufacturability constraints, and imaging quality-driven approaches, rather than a single algorithm or model. Its core innovation lies in the following three levels of "fusion": 1. A precise mapping mechanism between multimodal conditional inputs and multi-objective outputs. For the first time, this mechanism maps the physical properties of permafrost (such as w, ρ) to... d , P(d), ξ), test conditions (such as T, f) and the imaging system parameters together constitute the conditional vector x, which serves as the input to the generative model. Through a conditional generative network (cGAN), a model is established to transform this complex multimodal conditional space into the optimal gradation parameter vector y (containing P(di), k)... bw The end-to-end nonlinear mapping relationship of φ, Cov, u) solves the pain point of traditional methods that rely on manual trial and error and are difficult to handle high-dimensional nonlinear problems.

[0104] 2. Deep Embedding Strategy of Physical Constraints and Manufacturability Priors. This approach creatively transforms domain knowledge (physical laws and process limitations) from "posterior checking" to "prior guidance," deeply embedding it into the optimization loop. Specifically, at the loss function level, it constructs a fusion imaging quality (J... imaging Physical Realizability (J) physics ) and economy (J cost The comprehensive objective function J is used as the core reward signal for training the generator, guiding the model to generate high-performance and feasible solutions from the root. At the network structure / post-processing level: through parameter projection layers and activation function design (such as Sigmoid, ClippedReLU), hard constraints are placed on the output parameters (such as particle size distribution and doping levels) to ensure they fall within a preset feasible domain (such as the purchasable particle size range and process-achievable doping levels), guaranteeing 100% feasibility of the generated results. Anti-settling: s(Δρ,d,μ,f)≤κd50, κ∈[0.2,0.5]; Anti-segregation: intra-layer / inter-layer variance of black and white volume fractions ≤ σ * ,σ * ∈[0.01,0.04]; Coverage: Cov≥Cov min (Recommended 0.5–0.7); Dosage: ∈[ min , max ].

[0105] 3. An adaptive optimization architecture with a dual-loop "offline-online" approach. A complete adaptive optimization process was designed, forming two collaborative closed loops: Offline training loop: Based on historical datasets, the generator G is trained to acquire basic knowledge and reasoning capabilities. Online application-optimization loop: This is the essence of the system. In online applications, the system not only generates parameters but also fine-tunes the generation strategy online (e.g., adjusting loss weights α, β, γ) or directly fine-tunes parameters y (online correction mechanism) based on the evaluation results (DIC quality index) of synthetic image previews or real-world calibration images. This enables the system to possess lifelong learning and on-site adaptive capabilities, allowing it to self-adjust for specific samples and imaging environments, continuously improving the reliability and robustness of the output.

[0106] Compared with existing technologies, this invention brings significant technological advancements and application value by deeply integrating a deep generative model with physical mechanisms and imaging quality drivers. Firstly, this invention fundamentally revolutionizes the inefficient traditional trial-and-error model that relies on manual experience, achieving intelligent and automated mapping from permafrost sample characteristics to optimal gradation parameters. The system can comprehensively consider multi-dimensional inputs such as moisture content, density, original gradation, temperature field, and imaging conditions, and instantly outputs the optimal parameter scheme that meets multiple physical constraints and imaging quality requirements through a trained generative network. This greatly shortens the experimental preparation cycle and reduces the unreliability and trial-and-error costs caused by human factors.

[0107] The closed-loop architecture of "offline training-online evaluation-adaptive optimization" constructed in this invention endows the method with strong versatility and robustness. The system can not only learn universal laws from historical data in the offline stage, but also fine-tune the generation strategy or parameters through real-time image quality evaluation feedback during online application. This allows it to flexibly adapt to the specific characteristics of different regions and types of permafrost samples, as well as various complex field imaging environments. This self-optimization capability ensures that the method can continuously provide a high-precision and high-reliability measurement foundation in various scenarios, from indoor model experiments to in-situ field monitoring, providing solid technical support for the refined study of permafrost mechanical properties and the stability evaluation of related engineering projects.

[0108] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0109] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0110] Figure 8A schematic diagram of a deep neural network-based permafrost deformation measurement device 400 is shown in an embodiment of the present invention. The permafrost deformation measurement device 400 includes a communication module 401 and a processing module 402. The processing module includes a generative model inference module, a parameter projection layer, a quality assessment module, and an online optimization module; wherein the quality assessment module calculates MIG, n_f, r_peak, and Cov, and the online optimization module performs BO / neighborhood search and stops according to ε, N, and δ.

[0111] The communication module 401 is used to acquire the physical property parameters, test condition parameters and imaging system parameters of the frozen soil sample.

[0112] The processing module 402 is used to construct a conditional input vector based on the physical property parameters, test condition parameters, and imaging system parameters of the frozen soil sample; based on the conditional input vector and a preset conditional generation network, it performs multi-objective optimization with imaging quality, physical feasibility, and manufacturability as optimization objectives to determine the initial gradation parameters of the tracer particles; the initial gradation parameters include particle size distribution, black and white particle ratio, and particle content; based on the initial gradation parameters, it performs imaging quality assessment and parameter tuning to determine the optimized gradation parameters of the tracer particles; based on the optimized gradation parameters, it prepares the frozen soil sample to be tested and performs frozen soil deformation measurement.

[0113] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 500 includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501. When the processor 501 executes the computer program 503, it implements the steps in the above-described method embodiments. Alternatively, when the processor 501 executes the computer program 503, it implements the functions of each module / unit in the above-described device embodiments.

[0114] For example, the computer program 503 may be divided into one or more modules / units, which are stored in the memory 502 and executed by the processor 501 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 503 in the electronic device 500.

[0115] The processor 501 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0116] The memory 502 can be an internal storage unit of the electronic device 500, such as a hard disk or memory of the electronic device 500. The memory 502 can also be an external storage device of the electronic device 500, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 500. Furthermore, the memory 502 can include both internal and external storage units of the electronic device 500. The memory 502 is used to store the computer program and other programs and data required by the terminal. The memory 502 can also be used to temporarily store data that has been output or will be output.

[0117] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method of measuring deformation of frozen ground based on a deep neural network, characterized by, The method comprises the following steps: acquiring physical property parameters, test condition parameters and imaging system parameters of a frozen soil sample; constructing a condition input vector based on the physical property parameters, test condition parameters and imaging system parameters of the frozen soil sample; based on the condition input vector and a preset condition generation network, taking an imaging quality target, a physical implementability target and a manufacturability target as optimization targets, performing multi-objective optimization solving to determine initial grading parameters of tracer particles; the initial grading parameters include particle size distribution, black and white particle ratio and particle content; based on the initial grading parameters, performing imaging quality evaluation and parameter tuning to determine optimized grading parameters of the tracer particles; based on the optimized grading parameters, preparing a frozen soil sample to be tested to perform frozen soil deformation measurement.

2. The deep neural network-based frozen soil deformation measurement method according to claim 1, characterized in that, The method of constructing a condition input vector based on the physical property parameters, test condition parameters and imaging system parameters of the frozen soil sample comprises the following steps: standardizing the physical property parameters, test condition parameters and imaging system parameters to eliminate dimensions and obtain standardized parameters of various categories; the physical property parameters include water content, dry density, natural particle size distribution curve and cementation salt content of the frozen soil sample; the test condition parameters include temperature gradient and freezing rate; the imaging system parameters include camera parameters, light source conditions and specimen size; based on the standardized parameters, performing one-hot encoding to obtain numerical parameters of various categories; based on the numerical parameters of various categories, performing splicing and fusion to obtain the condition input vector. 3.The deep neural network-based frozen soil deformation measurement method according to claim 1, characterized in that, The method of determining initial grading parameters of tracer particles based on the condition input vector and a preset condition generation network, taking an imaging quality target, a physical implementability target and a manufacturability target as optimization targets, and performing multi-objective optimization solving comprises the following steps: inputting the condition input vector into a generator of the preset condition generation network to perform forward calculation and output initial grading parameters of the tracer particles; the preset condition generation network is obtained by training a plurality of condition input vectors and grading parameters of frozen soil samples, and a comprehensive objective function composed of the imaging quality target, the physical implementability target and the manufacturability target is used as an optimization criterion during the training process. 4.The deep neural network-based frozen soil deformation measurement method according to claim 1, wherein, Before the method of determining initial grading parameters of tracer particles based on the condition input vector and a preset condition generation network, taking an imaging quality target, a physical implementability target and a manufacturability target as optimization targets, and performing multi-objective optimization solving, the method further comprises the following steps: acquiring a plurality of frozen soil sample data in a historical period, the frozen soil sample data including physical property parameters, test condition parameters, imaging system parameters and grading parameters of tracer particles; based on the physical property parameters, test condition parameters and imaging system parameters of the plurality of frozen soil sample data, determining condition input vectors of the plurality of frozen soil samples; generating a real sample set by taking the condition input vector of each frozen soil sample as a condition input and taking the grading parameters of each frozen soil sample as real data; constructing a condition generation network, the condition generation network comprising a generator and a discriminator; The conditional input of each sample in the real sample set is input into the generator and the discriminator, and the grading parameter of each sample is taken as a real sample to adversarially train the conditional generation network, so as to obtain the preset conditional generation network. 5.The deep neural network-based frozen soil deformation measurement method according to claim 4, characterized in that, The conditional input of each sample in the real sample set is input into the generator and the discriminator, and the grading parameter of each sample is taken as a real sample to adversarially train the conditional generation network, so as to obtain the preset conditional generation network, including: Randomly sampling a batch of samples from the real sample set as a current batch of samples; Generating a random noise vector, and inputting the random noise vector and the conditional input vector of the current batch of samples into the generator to obtain generated grading parameters; Fixing the parameters of the generator, calculating the discriminator loss function based on the generated grading parameters and the real grading parameters of the current batch of samples, and updating the parameters of the discriminator through a gradient descent algorithm to minimize the discriminator loss function; Fixing the parameters of the discriminator, inputting the random noise vector and the conditional input vector into the generator again to obtain new generated grading parameters; Calculating the discriminator output expectation of the new generated grading parameters, and calculating the imaging quality target value, the physical implementability target value and the manufacturability target value corresponding thereto, respectively; Calculating the comprehensive target function of the generator by weighted summation based on the discriminator output expectation, the imaging quality target value, the physical implementability target value and the manufacturability target value; Updating the parameters of the generator through a back propagation algorithm to minimize the comprehensive target function of the generator; After each update of the generator parameters, applying physical constraints to the grading parameters output by the generator, constraining the particle size distribution within a preset range through parameter projection, and ensuring that the particle content and surface coverage meet the process requirements through numerical clipping; Repeating the above iteration process, and terminating the training process when the change difference of the comprehensive target function value on the validation set is less than a preset threshold value for a plurality of consecutive iterations, or the number of iterations reaches a preset maximum number of iterations; Taking the generator finally trained as the preset conditional generation network. 6.The deep neural network-based frozen soil deformation measurement method according to claim 1, wherein, The imaging quality evaluation and parameter optimization are performed based on the initial grading parameters to determine the optimized grading parameters of the tracer particles, including: Based on the initial grading parameters, a synthetic texture image is generated through a texture synthesis algorithm; Performing DIC quality evaluation on the synthetic texture image to calculate imaging quality indicators, including average gray gradient MIG, trackable feature point density, texture coverage rate and subset autocorrelation peak value; Based on the physical implementability target, taking the initial grading parameters as an optimization starting point, determining the adjustable range of each parameter, and constructing a parameter optimization space; the parameter optimization space includes a particle size distribution range, a content process interval and a black and white ratio range; Taking the initial grading parameters as an optimization starting point, and using a Bayesian optimization algorithm to iteratively optimize in the parameter optimization space to improve the imaging quality indicators. When the iterative optimization process reaches a convergence condition, output the optimal grading parameters as the optimization grading parameters; the convergence condition includes that an imaging quality index reaches a preset threshold, a target function improvement amplitude is less than a set tolerance, or a maximum iteration number is reached. 7.The deep neural network-based frozen soil deformation measurement method according to claim 1, wherein, Based on the optimization grading parameters, a test sample of frozen soil is prepared, and a deformation measurement of the frozen soil is performed, including: According to the particle size distribution, the black and white particle ratio, and the particle content in the optimization grading parameters, a tracer particle mixture is prepared; According to a set of process parameters in the optimization grading parameters, the tracer particle mixture is uniformly mixed with a frozen soil sample, and the mixing time, stirring speed, and tamping strength are controlled to prepare a test sample of frozen soil whose surface texture meets a set condition; the set condition includes that the surface coverage of the tracer particles is greater than a coverage threshold, and / or the surface contrast is greater than a contrast threshold; The prepared test sample of frozen soil is placed in a freeze-thaw test chamber, and a freeze-thaw cycle test is performed under a set temperature gradient and freezing rate; A digital image correlation (DIC) technique is used to continuously collect sequence images of the sample surface during the freeze-thaw process through an imaging system; DIC analysis is performed on the sequence images to calculate deformation measurement results of the frozen soil sample, including full-field deformation field and displacement field data. 8.The deep neural network-based frozen soil deformation measurement method according to claim 7, characterized in that, After the test sample of frozen soil is prepared based on the optimization grading parameters and the deformation measurement of the frozen soil is performed, the method further includes: During the freeze-thaw test, an imaging quality index of the collected sequence images is calculated in real time; the imaging quality index includes an average gray level gradient (MIG) and a trackable feature point density; When the imaging quality index is lower than a preset threshold, an online correction mechanism is started, the tracer particle parameters are fine-tuned and optimized based on the current imaging quality index, and corrected grading parameters are obtained; the fine-tuning and optimization uses a Bayesian optimization algorithm to perform a rapid search within a small neighborhood of the initial optimization grading parameters; Based on the corrected grading parameters, a new tracer particle mixture is prepared, a new test sample is prepared, and the deformation measurement of the frozen soil is continued; A measurement quality archive is established to record the grading parameters, imaging quality index, and deformation measurement results of each test, so as to optimize the subsequent condition generation network training.

9. A frozen ground deformation measuring device based on a deep neural network, characterized by, The method includes: A communication module is configured to obtain physical characteristic parameters of a frozen soil sample, test working condition parameters, and imaging system parameters; A processing module is configured to construct a condition input vector based on the physical characteristic parameters of the frozen soil sample, the test working condition parameters, and the imaging system parameters; Based on the condition input vector and a preset condition generation network, an imaging quality target, a physical implementability target, and a manufacturability target are used as optimization targets to perform multi-objective optimization solving to determine initial grading parameters of the tracer particles; the initial grading parameters include a particle size distribution, a black and white particle ratio, and a particle content; based on the initial grading parameters, imaging quality evaluation and parameter tuning are performed to determine optimization grading parameters of the tracer particles; and based on the optimization grading parameters, a test sample of frozen soil is prepared, and a deformation measurement of the frozen soil is performed.

10. A deep neural network-based frozen soil deformation measurement system, characterized in that, The frozen soil deformation measurement system comprises an electronic device, the electronic device comprises a memory and a processor, the memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to execute the method as claimed in any one of claims 1 to 8.

Images