Frozen soil full-field deformation self-adaptive measurement method, system and device

Abstract

This invention provides an adaptive measurement method, system, and device for full-field deformation of permafrost, relating to the field of permafrost deformation technology. The method includes: acquiring multimodal data of the permafrost environment to obtain initial data; labeling the ROIs in the sequential image data of the initial data to obtain the ROI mask and confidence map of the permafrost environment, and inputting them into a constructed reinforcement learning agent, outputting linked action data; wherein, the reward function of the reinforcement learning agent is coupled with multiple physical consistency indices calculated based on DIC; the action space of the reinforcement learning agent includes ROI mask morphology editing, multi-level ROI gating, and DIC parameter adjustment; updating the ROI mask and DIC parameters based on the linked action data to perform full-field deformation calculation, outputting the full-field displacement field and full-field strain field, and calculating the quality index; when the quality index deteriorates, optimizing and updating the reinforcement learning agent. This invention achieves adaptive measurement of full-field deformation of permafrost, improving the stability of long-term measurements.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of engineering monitoring and computer vision, and particularly relates to the field of frozen soil deformation technology, especially to an adaptive measurement method, system and equipment for full-field deformation of frozen soil. Background Technology

[0002] Digital Image Correlation (DIC), as a non-contact optical measurement method, has been widely used in engineering deformation monitoring. However, its traditional implementation faces fundamental challenges in complex dynamic environments such as permafrost. First, the reliance on manually pre-defined static Regions of Interest (ROIs) is severely mismatched with the dynamic deformation field caused by frost heave and thaw settlement in permafrost. This results in the ROI failing to effectively cover the evolving deformation zone in the later stages of measurement, introducing significant boundary errors and reducing the signal-to-noise ratio. Second, the high empirical dependence of DIC core parameters (such as subset size and step size) makes it difficult to cope with changes in image quality and deformation gradients during long-term experiments, lacking online adaptive adjustment capabilities.

[0003] To improve automation, existing solutions attempt to use image feature-based automatic ROI extraction or deep learning-based visual segmentation models. However, their optimization objectives remain limited to image appearance features (such as grayscale and texture), completely disconnected from the intrinsic physical processes of DIC calculation (such as minimizing correlation residuals and strain field smoothness). Other studies employ reinforcement learning or Bayesian optimization for parameter search, but their reward functions often unilaterally pursue a single metric (such as maximizing the overall correlation coefficient), neglecting the crucial multi-source physical consistency (such as the fit between ROI and temperature field boundaries and the smoothness of displacement time series) in multi-field coupled monitoring of permafrost. Essentially, existing technologies, whether traditional methods or automated solutions, suffer from a core bottleneck that remains unresolved: the three key aspects of ROI generation, parameter optimization, and physical quantity enhancement are fragmented, leading to insufficient measurement accuracy, robustness, and automation in complex environments. Summary of the Invention

[0004] This invention provides a method, system, and device for adaptive measurement of full-field deformation in frozen soil. The aim is to construct a closed-loop feedback collaborative adaptive system with the ultimate measurement accuracy as the unified objective, in order to solve the problem of how to improve measurement accuracy, robustness, and automation in complex environments.

[0005] In a first aspect, embodiments of the present invention provide an adaptive measurement method for full-field deformation of frozen soil, including:

[0006] Multimodal data of the permafrost environment under test are collected, and the multimodal data are preprocessed to obtain initial data; wherein, the multimodal data includes sequence image data and temperature data;

[0007] The ROIs are labeled on the sequence image data in the initial data to obtain the ROI mask and confidence map of the permafrost environment under test;

[0008] The ROI mask and the confidence map are input into the constructed reinforcement learning agent, and the linked action data is output. The reward function of the reinforcement learning agent is coupled with multiple physical consistency metrics calculated based on DIC. The action space of the reinforcement learning agent includes ROI mask morphology editing, multi-level ROI gating, and DIC parameter adjustment. The linked action data includes actions for morphology editing of candidate ROI masks and actions for adjusting DIC parameters. The DIC parameters include subset size, stride, and weight. The multi-level ROI gating is mainly divided into core region, transition region, and buffer region.

[0009] The ROI mask and DIC parameters are updated based on the linkage action data, and the full-field deformation is calculated to output the full-field displacement field and the full-field strain field.

[0010] Based on the full-field displacement field and the full-field strain field, a quality index is calculated, and when the quality index deteriorates, the reinforcement learning agent is optimized and updated.

[0011] In one possible implementation, the physical consistency metric includes DIC-related residuals, boundary mismatch, high strain gradient coverage, temperature field consistency, and temporal consistency.

[0012] In one possible implementation, the DIC-related residual is either a zero-mean normalized sum of squared differences (ZNSSD) or a zero-mean normalized cross-correlation (ZNCC).

[0013] In one possible implementation, the reward function is:

[0014]

[0015] in, The mean of ZNSSD or ZNCC residuals. For boundary mismatch degree, To achieve coverage of regions with high strain gradients, For timing consistency, To ensure consistency of the temperature field between the ROI and the thermal anomaly region, , For the measurement domain; Let t be the ROI mask. , The confidence threshold is in the range of (0–1). For confidence graph The resulting candidate contour set, , This is a region with a high strain gradient. A high strain gradient discrimination threshold is used to determine the strain gradient norm. Extracting high strain gradient regions , , For the temperature field in The mean within, For the temperature field in within the standard deviation, This is a thermal anomaly region. As weight, .

[0016] In one possible implementation, calculating the quality index based on the full-field displacement field and the full-field strain field includes:

[0017] Calculate the full-field average correlation residuals based on ZNSSD or ZNCC;

[0018] Calculate the symmetric average nearest point distance between the ROI mask contour and the candidate contour set obtained from the confidence map to obtain the boundary mismatch degree;

[0019] Calculate the normalized similarity of displacement field increments at adjacent time points to obtain temporal consistency of displacement fields.

[0020] In one possible implementation, the step of performing optimization updates on the reinforcement learning agent when the quality metric deteriorates includes:

[0021] When the full-field average correlation residual is greater than the corresponding calibration threshold, the boundary mismatch is greater than the corresponding calibration threshold, or the displacement field temporal consistency is less than the corresponding calibration threshold, the current deteriorated state is re-input into the reinforcement learning agent to restart the local search process;

[0022] Based on the new linkage action data output by the search, update the ROI mask processing and the DIC parameter linkage action data.

[0023] In one possible implementation, after obtaining the ROI mask and confidence map of the permafrost environment under test, the method further includes:

[0024] The ROI mask is morphologically refined to remove isolated patches;

[0025] The morphological refinement process includes:

[0026] Using a square structuring element, an opening operation and a closing operation are performed sequentially on the candidate ROI mask to remove isolated patches with an area smaller than the structuring element; wherein the size of the square structuring element is 3×3 pixels.

[0027] In one possible implementation, labeling the ROIs of the sequence image data in the initial data includes:

[0028] The attention mechanism U-Net, Mask R-CNN, or ViT-ROI is used to label the ROIs of the sequence image data in the initial data.

[0029] In one possible implementation, the preprocessing of the multimodal data to obtain initial data includes:

[0030] The sequence image data and the temperature data are subjected to spatiotemporal alignment processing;

[0031] The sequence of image data is denoised, and the denoised sequence of image data is then normalized for irradiance and exposure; wherein the denoising is based on nonlocal mean or bilateral filtering.

[0032] The temperature data is subjected to low-pass filtering and outlier removal.

[0033] In one possible implementation, after outputting the full-field displacement field and the full-field strain field, the following is also included:

[0034] Generate panoramic surveillance images;

[0035] The panoramic monitoring image includes one or more of the following: ROI region display image, displacement-strain field cloud map, and multi-sensor time series curve.

[0036] Secondly, embodiments of the present invention provide an adaptive measurement system for full-field deformation of permafrost, characterized in that it includes: an LED matrix light source, a high-speed camera, an infrared thermal imager, an edge computing unit, and a cloud server;

[0037] The high-speed camera and infrared thermal imager are used to collect multimodal data of the permafrost environment under test.

[0038] The edge computing unit is used to execute the methods in the first aspect or any possible implementation of the first aspect above;

[0039] The edge computing unit is a configuration chip or a device configured with the configuration chip; the configuration chip is a graphics processing unit (GPU), a field-programmable gate array (FPGA), or a system on chip (SoC).

[0040] Thirdly, embodiments of the present invention provide an adaptive measurement device for full-field deformation of frozen soil, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation of the first aspect.

[0041] In this embodiment of the invention, multimodal data is first collected and preprocessed to ensure the initial quality of the sequence images and temperature data, laying a reliable data foundation for subsequent measurements. Next, Regions of Interest (ROIs) are labeled in the sequence images to obtain ROI masks and confidence maps, enabling focus on key deformation areas and avoiding interference from irrelevant regions. These are then input into a reinforcement learning agent. The agent's reward function is coupled with multiple physical consistency indices related to DIC (Displacement-Induced Conformity), and its action space covers ROI editing and DIC parameter adjustment, enabling coordinated optimization of ROI masks and DIC parameters, breaking the traditional limitation of their separation. Subsequently, parameters are updated based on the coordinated action data, and the overall deformation is calculated, outputting accurate displacement and strain fields. Finally, through quality index evaluation, the agent is optimized when indices deteriorate, forming a closed-loop adaptive mechanism. This effectively solves the problem that traditional fixed ROIs and parameters cannot cope with boundary errors and drift accumulation caused by permafrost heave and thawing settlement, improving the stability and unmanned operation of long-term measurements. Attached Figure Description

[0042] Figure 1 This is an application scenario diagram of the adaptive measurement method for full-field deformation of frozen soil provided in an embodiment of the present invention;

[0043] Figure 2 This is a flowchart illustrating the implementation of an adaptive measurement method for full-field deformation of frozen soil according to an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram illustrating the linkage between the dimensions of multi-level ROI and DIC subsets according to an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the information flow of a reinforcement learning agent according to an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the overall framework of an adaptive measurement system for full-field deformation of frozen soil provided in an embodiment of the present invention;

[0047] Figure 6 This is a timing diagram of quality indicators and adaptive processes provided in an embodiment of the present invention;

[0048] Figure 7 This is a schematic diagram of the structure of an adaptive measurement device for full-field deformation of frozen soil provided in an embodiment of the present invention. Detailed Implementation

[0049] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0050] Figure 1 This diagram illustrates an application scenario of the adaptive measurement method for full-field deformation of frozen soil provided in this embodiment of the invention. Figure 1 As shown, it mainly consists of the object under test and its environment, the data acquisition system, the edge computing unit, and the user cloud. The data acquisition system includes an LED matrix light source, a high-speed camera, and an infrared thermal imager, used to collect multimodal data of the permafrost environment under test. The user cloud is a cloud server.

[0051] The LED matrix light source serves as the light source, enhancing the imaging effect of the high-speed camera. In practice, the high-speed camera consists of a high-resolution industrial camera, a lens, and a matching light source, used to acquire sequential digital images of the specimen surface at a set frame rate. The infrared thermal imager is a multi-physics sensing unit, integrating temperature and strain sensors, used to simultaneously monitor key physical field prior information of the frozen soil specimen and acquire temperature and strain information.

[0052] The edge computing unit is a configuration chip or a device equipped with a configuration chip; the configuration chip is a GPU, FPGA, or SoC. In this embodiment of the invention, the device can be implemented on an edge GPU, FPGA, or SoC to meet real-time analysis requirements. The cloud server is used to store raw data, processing results, and system metadata. In specific implementation, the data is stored in a compressed format in parallel to both the local cache and the cloud server, achieving comprehensive monitoring of the experimental process and dual backup archiving of the data.

[0053] The edge computing unit is used to execute the method of this invention embodiment. Under the conditions of multi-field coupling and complex optics in permafrost, it realizes dynamic refinement of ROI and adaptive linkage of DIC parameters to reduce boundary errors, improve long-term stability, and enhance unmanned operation. The following describes in detail the adaptive measurement method for full-field deformation of permafrost provided by this invention embodiment, with reference to the accompanying drawings.

[0054] Figure 2 This is a flowchart illustrating the implementation of an adaptive measurement method for full-field deformation of frozen soil according to an embodiment of the present invention. Figure 2 As shown, it includes the following steps:

[0055] S201. Collect multimodal data of the permafrost environment under test and preprocess the multimodal data to obtain initial data; the multimodal data includes sequence image data and temperature data;

[0056] The execution entity in various embodiments of the present invention can be a server, processor, microprocessor, or other device with data processing capabilities. In actual implementation, the specific implementation method of the execution entity can be selected according to actual needs. This embodiment does not impose any particular limitation on this; any device with data processing capabilities is acceptable. For ease of understanding of the overall solution, [the following is used as an example]. Figure 1 The explanation will be based on the example of a mid-edge computing unit.

[0057] Before performing full-field adaptive deformation measurement, multimodal data acquisition and preprocessing steps are first performed. Considering that permafrost deformation monitoring needs to simultaneously focus on visual morphological changes and the influence of environmental physical fields, the acquired multimodal data mainly includes sequential image data reflecting the surface texture characteristics of permafrost, as well as temperature data characterizing the thermodynamic state of permafrost.

[0058] During the acquisition process, it is necessary to ensure the spatiotemporal synchronization of the two types of data to avoid subsequent fusion deviations caused by data temporal or spatial misalignment. Preprocessing of the acquired multimodal data aims to eliminate interference factors in the original data, such as ambient light fluctuations and noise in sequence images, and transient outliers in temperature data. The preprocessed initial data must meet the basic quality requirements for subsequent ROI annotation and DIC calculation.

[0059] S202, label the ROIs of the sequence image data in the initial data to obtain the ROI mask and confidence map of the permafrost environment under test.

[0060] Since permafrost deformation is often concentrated in specific areas, such as zones of active frost heave and areas sensitive to thaw settlement, directly calculating the deformation across the entire image would introduce interference from irrelevant regions. Therefore, labeling Regions of Interest (ROIs) allows for focusing on key monitoring areas. During the labeling process, attention-based segmentation models or object detection models can be used to identify and select regions related to permafrost deformation from the image sequence. Simultaneously, ROI masks and confidence maps are generated to define the ROI range and characterize the reliability of the ROI region, respectively.

[0061] S203: Input the ROI mask and confidence map into the constructed reinforcement learning agent, and output the linked action data. The reward function of the reinforcement learning agent is coupled with multiple physical consistency metrics calculated based on DIC; the action space of the reinforcement learning agent includes ROI mask morphology editing, multi-level ROI gating, and DIC parameter adjustment; the linked action data includes actions for morphology editing of candidate ROI masks and actions for adjusting DIC parameters; the DIC parameters include subset size, stride, and weights.

[0062] The core of the reinforcement learning agent design lies in "collaborative decision-making with the goal of improving the accuracy of permafrost deformation measurement". Its reward function does not rely on a single index, but couples multiple physical consistency indices calculated based on DIC, to ensure that the agent's decision-making can take into account multiple dimensions such as DIC matching quality, ROI boundary accuracy, and the multi-field coupling characteristics of permafrost.

[0063] Figure 3 This is a schematic diagram illustrating the multi-level ROI and DIC subset size linkage provided in an embodiment of the present invention. The agent's action space covers ROI mask shape editing, multi-level ROI gating, and DIC parameter adjustment, enabling linked optimization of ROI and DIC parameters and avoiding the accuracy loss caused by the separation of the two in traditional technologies. For example, Figure 3 As shown, ROI mask morphology editing includes adjusting ROI boundaries and optimizing region integrity. Multi-level ROI gating is mainly divided into core region, transition region and buffer region.

[0064] Through agent inference, the final output includes linked action data containing ROI mask shape editing actions and DIC parameter adjustment actions. For example... Figure 3 As shown, based on the linkage action data, a high-resolution displacement field, a high signal-to-noise ratio strain field, and a resource optimization configuration map are finally obtained.

[0065] S204 updates the ROI mask and DIC parameters based on the linkage motion data, performs full-field deformation calculation, and outputs the full-field displacement field and full-field strain field.

[0066] The updated ROI mask can more accurately cover key areas of permafrost deformation, and the adjusted DIC parameters can adapt to the current image texture and deformation features. Based on this, the displacement field and strain field of the entire permafrost field are calculated through the DIC algorithm. These displacement field and strain field can reflect the displacement changes and strain distribution at various points on the permafrost surface.

[0067] S205 calculates the quality index based on the full-field displacement field and the full-field strain field, and performs optimization updates on the reinforcement learning agent when the quality index deteriorates.

[0068] The calculation of quality indicators must reflect the reliability of measurement results. For example, when the quality indicators show that the DIC matching accuracy has decreased, the ROI boundary deviates from the actual deformation zone, or the displacement field fluctuates too much, it indicates that the current ROI and DIC parameters can no longer adapt to the changes in permafrost deformation. At this time, the current deterioration state is re-input into the reinforcement learning agent, triggering the agent's local search and strategy update, generating new linkage action data to adjust the parameters, forming a closed-loop adaptive mechanism of "measurement-evaluation-optimization" to ensure measurement stability during long-term monitoring.

[0069] In this embodiment, multimodal data is first acquired and preprocessed to ensure the initial quality of the sequence images and temperature data, laying a reliable data foundation for subsequent measurements. Next, Regions of Interest (ROIs) are labeled in the sequence images to obtain ROI masks and confidence maps, enabling focus on key deformation areas and avoiding interference from irrelevant regions. These are then input into a reinforcement learning agent. The agent's reward function is coupled with multiple physical consistency indices related to DIC (Displacement-Induced Conformity), and its action space covers ROI editing and DIC parameter adjustment, enabling coordinated optimization of ROI masks and DIC parameters, breaking the traditional limitation of their separation. Subsequently, parameters are updated based on the coordinated action data, and the overall deformation is calculated, outputting accurate displacement and strain fields. Finally, through quality index evaluation, the agent is optimized when indices deteriorate, forming a closed-loop adaptive mechanism. This effectively solves the problem that traditional fixed ROIs and parameters cannot cope with boundary errors and drift accumulation caused by permafrost heave and thawing settlement, improving the stability and unmanned operation of long-term measurements.

[0070] In one possible implementation, the physical consistency metric includes DIC-related residuals, boundary mismatch, high strain gradient coverage, temperature field consistency, and temporal consistency.

[0071] Among them, the DIC-related residual directly reflects the matching accuracy of the sequence images in the DIC algorithm and is the basis for judging the accuracy of deformation calculation. If the residual is too large, it indicates that there is a deviation in image matching, which needs to be improved by adjusting the DIC parameters or the ROI range.

[0072] Boundary mismatch is used to assess the degree of fit between the ROI mask boundary and the actual deformation area boundary of the frozen soil, so as to avoid the omission of key deformation areas or the inclusion of irrelevant areas due to ROI boundary offset, and ensure the accuracy of the monitoring range.

[0073] In long-term monitoring, permafrost deformation exhibits continuity and is closely related to the temperature field. For example, frost heave intensifies as temperature decreases, while thaw settlement occurs as temperature increases. Furthermore, permafrost deformation is often accompanied by localized high strain gradients, such as crack initiation zones and areas of concentrated frost heave.

[0074] High strain gradient coverage is used to measure the coverage ratio of the ROI mask to such critical areas, ensuring that high strain gradient areas can be effectively monitored and avoiding the loss of core deformation information.

[0075] Temperature field consistency is used to associate the ROI mask with temperature anomaly areas, such as low-temperature frost heave zones and high-temperature thaw settlement zones, to ensure that the ROI can adjust with changes in the temperature field and adapt to the multi-field coupling characteristics of frozen soil.

[0076] Temporal consistency is used to evaluate the smoothness of the displacement field at adjacent time points, avoid displacement field drift caused by parameter abrupt changes, and ensure the stability of long-term measurements.

[0077] After the above five types of physical consistency indicators are coupled into the reward function of the reinforcement learning agent, the agent will simultaneously weigh matching accuracy, boundary accuracy, key region coverage, multi-field coupling adaptation, and temporal stability when generating linkage action data, rather than pursuing the optimization of a single indicator. For example, optimizing only the DIC-related residuals may cause the ROI to shrink excessively to the textured area, ignoring the temperature-related deformation area. However, by constraining the temperature field consistency indicator, it can be ensured that the ROI covers both the textured area and the temperature-related deformation area, achieving more comprehensive deformation monitoring. At the same time, the introduction of the temporal consistency indicator can avoid long-term displacement field fluctuations caused by the agent adjusting parameters for short-term matching accuracy, thus improving the overall monitoring reliability.

[0078] In this embodiment, DIC-related residuals, boundary mismatch, high strain gradient coverage, temperature field consistency, and temporal consistency are incorporated into the physical consistency index and coupled into the reward function of the reinforcement learning agent. This allows the agent to make decisions that are no longer limited to a single index, but rather comprehensively consider the DIC matching quality, ROI boundary accuracy, key deformation area coverage, multi-field coupling characteristics of permafrost, and temporal stability. This avoids measurement deviations caused by optimizing a single index, and is particularly well-suited to the multi-field coupling characteristics of the permafrost environment. It can more comprehensively ensure the rationality of ROI mask and DIC parameter adjustments, and further improve the accuracy and robustness of full-field deformation measurement.

[0079] In one possible implementation, the DIC-related residual is ZNSSD or ZNCC.

[0080] ZNSSD, by calculating the sum of normalized gray-level squared differences, can quickly reflect the degree of difference in image regions, with high computational efficiency. It is suitable for permafrost monitoring scenarios with certain real-time requirements, such as on-site dynamic monitoring. ZNCC, by calculating the normalized gray-level cross-correlation coefficient, has stronger robustness to gray-level linear transformations. It can maintain high matching accuracy even in permafrost image scenarios with blurred textures or large gray-level fluctuations, such as scenarios where frost coverage weakens the texture.

[0081] In practical implementation, an appropriate residual calculation method can be selected based on the characteristics of the permafrost monitoring scenario. Optionally, in indoor frost heave tests, where stable image quality and extremely high accuracy are required, ZNCC is selected; in field roadbed monitoring, where both real-time performance and anti-interference capabilities need to be considered, ZNSSD is selected.

[0082] In this embodiment, ZNSSD or ZNCC are two relatively accurate correlation measurement methods in the field of DIC. They can more sensitively and accurately quantify the matching degree between sequential images. Compared with other residual calculation methods, they can provide more reliable residual data support for the reward function of reinforcement learning agents, reduce the decision bias of agents caused by inaccurate residual calculation, optimize ROI masks and DIC parameters, and thus improve the accuracy of full-field displacement field and strain field calculation.

[0083] In one possible implementation, the reward function is:

[0084]

[0085] in, The mean of ZNSSD or ZNCC residuals. For boundary mismatch degree, To achieve coverage of regions with high strain gradients, For timing consistency, To ensure consistency of the temperature field between the ROI and the thermal anomaly region, , The measurement domain is the set of all pixel coordinates within the camera's field of view, after registration / cropping, that allow for DIC and temperature mapping (also understood as the "effective computation domain"). The ROI mask is taken at time t. The ROI mask is derived from the segmentation / detection model annotation and can be used for morphological refinement to remove isolated patches. ; The confidence threshold (0–1) is used to... Binarization / screening to create candidate boundaries; The higher the value, the more "conservative and reliable" the profile (but may miss boundaries). The lower the value, the more "sensitive and comprehensive" the contour becomes (but this may introduce noise); For confidence graph The resulting candidate contour set; ; A high strain gradient discrimination threshold is used to determine the strain gradient norm. Extract key deformation regions; It can be determined by calibration tests or statistical quantiles; This is a region with a high strain gradient; ; For the temperature field in The mean (or, more robustly, the median) within the range. The coefficient for judging thermal anomalies controls the sensitivity of anomaly extraction. For the temperature field in Within the standard deviation (or robust standard deviation); This is a thermal anomaly region. As weight, .

[0086] When using ZNSSD analysis on DIC-related residuals, the mean residual value is:

[0087]

[0088] The ZNSSD within a single pixel window is:

[0089]

[0090] For a window defined by a subset size s, Displacement from DIC , , , They are respectively Mean and standard deviation of the two graphs in the window; Image 1 to be analyzed; Image 2 is to be analyzed.

[0091] Boundary mismatch degree (symmetric Chamfer / ASSD) is:

[0092]

[0093] Wherein: minimum distance:

[0094]

[0095] High strain gradient coverage is:

[0096]

[0097] Temporal consistency (normalized similarity of displacement field increments) is:

[0098]

[0099] It is a normalization constant (i.e., greater than the expected maximum single-step displacement increment) so that this term falls within [0,1].

[0100] Temperature field consistency / alignment is:

[0101]

[0102] The quality degradation trigger condition (used for closed-loop rollback / reset) is:

[0103]

[0104] : Calibration threshold for experience / validation sets;

[0105] in, The threshold for the overall average correlation residual (i.e., the ZNSSD or ZNCC residual mean threshold) is set to 0.50-0.75; the recommended default value is 0.60; the appropriate value for different scenarios is 0.50-0.60 for high-precision laboratory settings and 0.65-0.75 for complex lighting / texture degradation in real-world environments.

[0106] The threshold for boundary mismatch is 3%-10%; the recommended default is 5%, and the appropriate value for different scenarios is 3%-5% in laboratory scenarios and 6%-10% in field scenarios (where the boundary is affected by frost, reflection, or shading).

[0107] The displacement / strain time series consistency threshold has a range of 0.80-0.95; the recommended default value is 0.85-0.90. Scenario-specific values: 0.90-0.95 can be set for laboratory long-term stability and low noise; 0.80-0.88 should be used when there is a large disturbance in the field to avoid frequent false triggering.

[0108] state With action :

[0109] state: ;in For the temperature field, For optical flow / displacement prediction;

[0110] Actions (ROI and DIC parameter linkage): {Mask shape editing (expansion / etching / local free deformation), multi-level ROI gating (core / transition / buffer), DIC subset size} ∈[11,61] odd number, step size ∈[3,15], weight (Core layer weight ≥ 0.6) The function is:

[0111]

[0112] Offline training is performed using simulation and real-world data, while online fine-tuning is achieved using ε-greedy or strategy distillation. Constrained RL ensures that the action space meets the requirements for instrument safety and DIC convergence.

[0113] Among them, each weight parameter It can be flexibly adjusted according to the needs of different frozen soil measurement scenarios. For example, in areas with significant frozen soil thermal anomalies, the weight of η can be appropriately increased to prioritize ensuring the consistency of the temperature field, and in areas with high strain gradients, the weight of γ can be increased to enhance the coverage of high strain gradients.

[0114] Figure 4 This is a schematic diagram of the information flow of a reinforcement learning agent according to an embodiment of the present invention, as shown below. Figure 4 As shown, state In practical implementation, either the Proximal Policy Optimization (PPO) algorithm or the Soft Actor-Critic (SAC) algorithm is trained based on simulation and actual collected data. Combined with pre-trained parameters, linkage data is generated based on state data inference. Action execution includes: ROI shape editing, multi-level ROI, subset size adjustment, step size adjustment, and weight adjustment. Measurement results and new states are obtained based on the adjusted subset size, step size, and weights. The reward calculation module is used for value network and policy updates, optimizing the decision logic of the policy network and improving the rationality of subsequent actions. Furthermore, it re-inputs the new state into the state perception layer to initiate the next round of decision-making.

[0115] In this embodiment, the reward function covers multiple dimensions of indicators such as residual, boundary, strain, time series, and temperature, ensuring that the agent's decision-making can fully meet the actual needs of the full-field deformation measurement of permafrost, making the linkage optimization of ROI mask and DIC parameters more targeted and reliable, and improving the optimization effect of the agent.

[0116] The foregoing embodiments introduced the quality degradation triggering conditions. When performing closed-loop control, it is necessary to calculate the full-field average correlated residual, boundary mismatch, and timing consistency in the quality degradation triggering conditions.

[0117] In one possible implementation, the quality indices are calculated based on the full-field displacement field and the full-field strain field, including:

[0118] Calculate the full-field average correlation residuals based on ZNSSD or ZNCC;

[0119] Calculate the symmetric average nearest point distance between the ROI mask profile and the candidate profile set obtained from the confidence map to obtain the boundary mismatch degree;

[0120] Calculate the normalized similarity of displacement field increments at adjacent time points to obtain temporal consistency of displacement fields.

[0121] The calculation of quality indicators revolves around three core dimensions of frozen soil deformation measurement: "overall matching accuracy", "ROI boundary accuracy", and "long-term stability".

[0122] The average correlation residual calculated using ZNSSD or ZNCC can intuitively reflect the overall matching quality of the DIC algorithm for the sequence of images. A low average correlation residual indicates a high degree of image texture matching and strong reliability of the displacement field calculation; a high residual suggests potential problems such as image noise, texture degradation, or insufficient parameter fit.

[0123] The boundary mismatch is obtained by calculating the symmetric average nearest point distance between the ROI mask profile and the candidate profile set. This avoids the deviation caused by distance calculation in a single direction and can accurately assess the accuracy of ROI mask positioning. If the boundary mismatch is small, it means that the ROI can accurately cover the key areas of permafrost deformation, avoiding interference from irrelevant areas or omission of core areas. If the mismatch is large, it means that the ROI needs to be adjusted through morphological editing to adapt to the dynamic changes in the deformation area caused by frost heave and thaw settlement of permafrost.

[0124] The temporal consistency of the displacement field is obtained by calculating the normalized similarity of the displacement field increments at adjacent time points. Normalization eliminates the influence of differences in displacement magnitude at different time points, allowing the index to directly reflect the smoothness of the displacement field changes over time and effectively monitor displacement drift in long-term measurements. High temporal consistency indicates that the displacement field changes conform to physical laws and there is no abnormal drift; low temporal consistency suggests possible parameter mutations, data interference, or agent decision-making bias, requiring further optimization of DIC parameters or ROI range.

[0125] In this embodiment, three quality indicators comprehensively evaluate the measurement results from three key dimensions: overall matching accuracy, ROI boundary accuracy, and temporal stability. Compared with a single quality indicator, these indicators can more comprehensively identify problems in the measurement process, providing a reliable basis for subsequent judgment on whether the reinforcement learning agent needs to be optimized, and avoiding a decrease in measurement accuracy due to the omission of problems by a single indicator.

[0126] In one possible implementation, when the quality metric deteriorates, an optimization update is performed on the reinforcement learning agent, including:

[0127] When the average correlation residual of the entire field is greater than the corresponding calibration threshold, the boundary mismatch is greater than the corresponding calibration threshold, or the temporal consistency of the displacement field is less than the corresponding calibration threshold, the current deteriorated state is re-input into the reinforcement learning agent to restart the local search process;

[0128] Based on the new linkage action data output by the search, update the ROI mask processing and DIC parameter linkage action data.

[0129] The document specifies the concrete triggering conditions for quality index degradation, including: the overall average correlation residual exceeding the corresponding calibration threshold, the boundary mismatch exceeding the corresponding calibration threshold, and the temporal consistency of the displacement field falling below the corresponding calibration threshold. When any one of these conditions is met, the reinforcement learning agent is triggered to optimize and update. In practice, the "calibration threshold" is not a fixed value but can be dynamically adjusted based on the environmental complexity and accuracy requirements of the actual application scenario (such as indoor frost heave tests or on-site roadbed monitoring), reflecting the flexibility of the solution. Optionally, in complex lighting conditions, the correlation residual threshold can be appropriately relaxed to avoid frequent optimization triggers; however, in high-precision laboratory scenarios, a strict threshold is required to ensure measurement accuracy.

[0130] Furthermore, global training is time-consuming and cannot meet the needs of real-time monitoring. In this embodiment of the invention, when a degradation condition is triggered, the agent is not globally retrained. Instead, the current degradation measurement state, including the degradation quality index data, the current ROI mask shape, the current DIC parameter configuration, and real-time temperature field data, is re-inputted into the reinforcement learning agent. Based on the coupling characteristics of the reward function, a local search is performed. The agent focuses on the key issues leading to quality degradation. For example, if the ROI boundary is offset, mask shape editing actions are searched first; if the DIC matching accuracy is low, parameter adjustment actions are searched first. This eliminates the need to traverse the entire action space, significantly improving optimization efficiency.

[0131] Finally, the reinforcement learning agent outputs new linkage action data through local search, simultaneously updating the ROI mask (e.g., performing morphological editing to correct boundary offsets, or adjusting multi-level ROI gating to cover high-strain regions) and DIC parameters. After the update is complete, the full-field deformation calculation and quality index evaluation are re-executed. If the quality index recovers to the acceptable range, monitoring continues; if it still does not meet the standard, the above "state input - local search - parameter update" process is repeated until the measurement requirements are met.

[0132] In this embodiment, the deterioration trigger conditions for quality indicators are clearly defined. Optimization is initiated when any indicator is met, accurately capturing measurement problems. The deteriorated state is input into the agent to restart the local search. This reduces computational resource consumption while ensuring the repair of measurement problems, and quickly outputs new linkage action data to update the ROI mask and DIC parameters. This allows the measurement system to quickly restore a high-precision measurement state, avoiding the accumulation of measurement errors caused by continuous quality deterioration, effectively improving the system's real-time performance and robustness, and ensuring the accuracy and stability of long-term frozen soil deformation measurements.

[0133] In practice, isolated patches and other interference in the generated ROI mask can reduce the accuracy of measurement results. Therefore, it is necessary to process isolated patches in the ROI mask.

[0134] In one possible implementation, after obtaining the ROI mask and confidence map of the permafrost environment under test, the following is also included:

[0135] Morphological refinement of the ROI mask was performed to remove isolated patches;

[0136] The morphological refinement process includes:

[0137] Using a square structuring element, perform an opening operation and a closing operation on the candidate ROI mask in sequence to remove isolated patches with an area smaller than the structuring element.

[0138] Optionally, a 3×3 pixel square structural element can be used to adapt to the texture scale of the permafrost image. In other possible implementations, a 5×5 pixel or 7×7 pixel square structural element can be used.

[0139] During the refinement process, an opening operation is first performed to remove isolated bright patches smaller than the structuring element, and then a closing operation is performed to fill in small dark area gaps, avoiding excessive erosion of the main outline of the mask.

[0140] In this embodiment, after obtaining the ROI mask, the method performs morphological refinement processing. A 3×3 pixel square structuring element is used to sequentially perform opening and closing operations. The opening operation removes isolated bright spots in the ROI mask whose area is smaller than the structuring element, while the closing operation fills in small dark spots. This effectively removes useless isolated patches, preventing them from interfering with DIC calculations and causing inaccurate local deformation measurements. It also does not damage the main outline of the ROI mask, ensuring the integrity and purity of the ROI region. This provides a more accurate ROI basis for subsequent decision-making by the reinforcement learning agent and DIC full-field deformation calculations, reduces computational interference caused by isolated patches, and further improves the accuracy of full-field deformation measurements.

[0141] In one possible implementation, the ROIs are labeled on the sequence image data in the initial data, including:

[0142] The attention mechanism U-Net, Mask R-CNN, or ViT-ROI is used to label the ROIs in the sequence image data of the initial data.

[0143] The attention mechanism U-Net can focus on texture feature regions related to permafrost deformation and reduce background interference; Mask R-CNN can simultaneously achieve object detection and instance segmentation, accurately outputting ROI masks and confidence maps; ViT-ROI captures global image correlations through the Transformer architecture, adapting to the ROI localization needs in permafrost large-scale deformation scenarios.

[0144] In this embodiment, attention mechanisms such as U-Net, Mask R-CNN, or ViT-ROI are used to label regions of interest in sequential images. Compared with traditional manual labeling or simple segmentation methods, this approach can automatically adapt to complex situations such as texture changes, lighting differences, and ice crystal reflections that may exist in permafrost images. It can accurately generate ROI masks and confidence maps, significantly reducing manual intervention and improving the automation and accuracy of ROI labeling. This lays a good foundation for subsequent reinforcement learning optimization and DIC calculation, and further promotes the unmanned and precise operation of the entire measurement process.

[0145] In one possible implementation, the multimodal data is preprocessed to obtain initial data, including:

[0146] Spatiotemporal alignment processing is performed on sequence image data and temperature data;

[0147] The sequence image data is denoised, and the irradiance and exposure are normalized after denoising; the denoising is based on nonlocal mean or bilateral filtering.

[0148] Low-pass filtering and outlier removal are performed on the temperature data.

[0149] In the specific implementation process, preprocessing of multimodal data aims to achieve standardization and spatiotemporal alignment of multi-source heterogeneous data. Non-local mean or bilateral filtering is applied to the acquired raw image sequences to suppress high-frequency noise in speckle patterns while preserving texture details. Irradiance and exposure normalization are also performed to eliminate inter-frame intensity differences caused by fluctuations in ambient light or slight changes in camera parameters, ensuring grayscale consistency of the image data. Simultaneously, low-pass filtering and outlier removal are applied to physical sensor signals such as temperature.

[0150] At the spatiotemporal alignment level, the system employs a hardware triggering and high-precision clock synchronization strategy to ensure that the time synchronization error between image acquisition and sampling of physical quantities such as temperature is no greater than 50 milliseconds, thus meeting the quasi-static measurement requirements of the permafrost deformation process. During implementation, a mapping model between the image coordinate system and the world coordinate system is established through hand-eye calibration, and point sensor data is interpolated to generate a priori physical field map, such as the temperature field, that matches the image space.

[0151] In this embodiment, the method's preprocessing of multimodal data encompasses spatiotemporal alignment, image denoising and normalization, and temperature data filtering and outlier removal. Spatiotemporal alignment ensures strict temporal and spatial matching between the sequence images and temperature data, avoiding multimodal information deviations caused by data asynchrony. Image denoising based on nonlocal means or bilateral filtering removes noise while preserving image texture details, preventing noise from affecting DIC matching accuracy. Irradiance and exposure normalization eliminate inter-frame image grayscale differences caused by ambient light fluctuations or camera parameter changes, ensuring stable image quality. Low-pass filtering and outlier removal of temperature data remove high-frequency interference and outliers from the temperature signal, ensuring reliable temperature data. These preprocessing steps collectively improve the quality of the initial data, providing high-quality data input for subsequent ROI labeling, reinforcement learning agent decision-making, and DIC full-field deformation calculation, reducing measurement errors caused by problems with the original data from the source and improving overall measurement accuracy.

[0152] The foregoing embodiments mainly elaborate on the adaptive measurement method for full-field deformation of frozen soil. Based on the foregoing embodiments, and combined with the overall block diagram of the adaptive measurement system for full-field deformation of frozen soil, the adaptive measurement process for full-field deformation of frozen soil is explained.

[0153] Figure 5 This is a schematic diagram of the overall framework of an adaptive measurement system for full-field deformation of frozen soil provided in an embodiment of the present invention. Figure 6 This is a timing diagram of quality indicators and adaptive processes provided in an embodiment of the present invention.

[0154] Combination Figure 5 and Figure 6 After the multimodal data acquisition is completed, multimodal data preprocessing and spatiotemporal alignment processing are performed, and candidate ROIs are generated through a lightweight segmentation network.

[0155] Subsequently, ROI masks and confidence maps (morphologically refined if necessary) are generated from candidate ROIs, and integrated with information such as strain gradient and temperature into the input information of the state space. This state information is fed into the reinforcement learning (RL) agent optimization decision network, and after inference, it outputs linked actions in the action space for ROI morphology adjustment and DIC parameter optimization. The action commands are applied to the DIC calculation and physics field solution stages, obtaining full-field deformation data while calculating rewards based on physical quantity constraints. (Coupling DIC residuals, boundary mismatch, and other metrics) to feed back strategy network optimization.

[0156] Deformation data is simultaneously entered into the quality assessment stage. If the indicators meet the standards, it proceeds to the output and storage stage. If the quality deteriorates, a backtracking is triggered, and the current state is re-entered into the strategy network to restart the local search, forming an adaptive optimization loop of "data-processing-decision-calculation-evaluation-feedback" to ensure the accuracy and stability of permafrost deformation measurement.

[0157] In one possible implementation, after outputting the full-field displacement field and the full-field strain field, the following is also included:

[0158] Generate panoramic surveillance images;

[0159] The panoramic monitoring images include one or more of the following: ROI area display images, displacement-strain field cloud maps, and multi-sensor time-series curves.

[0160] Among them, after outputting the full-field displacement field and full-field strain field, the ROI area, displacement-strain field cloud map and multi-sensor time series curves are displayed in real time, which makes it convenient for relevant personnel to intuitively view the frozen soil deformation information.

[0161] In this embodiment, the method generates a panoramic monitoring image after outputting the full-field displacement and strain fields. The ROI display allows operators to clearly grasp the location of key areas currently being measured. The displacement-strain field cloud map transforms abstract deformation data into intuitive visual images, facilitating the rapid identification of key deformation areas such as high strain and large displacement. The multi-sensor time-series curves can display the changing trends of parameters such as displacement, strain, and temperature over time, facilitating the analysis of long-term deformation patterns in permafrost. This visualization method significantly reduces the difficulty for operators in interpreting measurement data, enabling real-time monitoring of the measurement process and timely detection of anomalies. It also provides intuitive evidence for subsequent experimental data review and deformation mechanism analysis, significantly improving the practicality and ease of use of the entire measurement system.

[0162] 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.

[0163] In practical industrial applications, the solutions provided in these embodiments are suitable for indoor frozen soil tests (frost heave columns, model boxes) and on-site monitoring of roadbeds / tunnels / slopes, supporting long-term unattended operation and rapid comparison. The following embodiments were conducted to verify the applicability of the solutions provided in these embodiments.

[0164] In the following embodiments, the basic parameters are set as follows:

[0165] Camera and Optics: Focal length 12-25mm; Auto exposure limiting; Active light source 5-10kLux;

[0166] Speckle: Black and white contrast > 0.6, average dot diameter 3-5px, fill rate 45-60%;

[0167] Segmentation / Candidate Network: Lightweight U-Net (parameters < 3M), input 512 2 The training set contains samples with different lighting conditions and frost / ice crystal reflections.

[0168] DIC parameter range: subset ∈[11,61] (odd number), step size ∈[3,15]; Relevance threshold 0.8; Maximum iterations 50;

[0169] RL Training: State channels 6-8; Action dimensions 6-10; Reward weights (example) =1, =0.5, =0.8, =0.4, =0.3;

[0170] Synchronization and Alignment: NTP / Hardware Trigger; Temperature Sampling 1-5Hz;

[0171] Quality control: When the boundary mismatch is >5% or A value greater than 0.6 triggers a local rollback.

[0172] Example 1

[0173] In the indoor frost heave column test, the specimen was an artificial frost heave column with a diameter of 200 mm and a height of 400 mm.

[0174] The data acquisition configuration included: an industrial camera with a resolution of 2448×2048 and a frame rate of 1fps; a speckle application method with an average dot diameter of approximately 4 pixels and a fill rate of 50%; and three temperature sensors positioned at the top, bottom, and middle of the specimen. Initially, the Region of Interest (ROI) automatically covered the entire cross-section of the specimen. As frost heave deformation progressed, a reinforcement learning strategy gradually shrank the ROI to the active frost heave zone, while dynamically adjusting the DIC subset size from 41 to 21. Experimental results showed that in this embodiment, the coverage of the ROI and the high-strain zone remained above 90%, the displacement field drift rate decreased by approximately 30% compared to the fixed-parameter scheme, and the measurement accuracy was significantly improved.

[0175] Example 2

[0176] In the on-site monitoring of frozen soil subgrade, a monitoring area with a width of about 3m was set up on the cross section of the subgrade in a high-altitude and cold region.

[0177] The data acquisition setup employed a dual-camera stereo system with a baseline length of 1.2m and a frame rate of 0.2fps. Infrared lights with an illuminance of 8kLux were used for supplemental lighting to ensure stability during nighttime monitoring. Temperature and strain sensors were embedded at depths of 0.5m and 1.0m in the roadbed. The initial ROI was generated using a Mask R-CNN network, covering the road shoulder area. During reinforcement learning optimization, a multi-level ROI gating mechanism was introduced, with a "buffer layer" used to mitigate vehicle load disturbances. The dynamic range of the DIC parameters was set to: subset size 25-55, step size 7-15. Results showed that this embodiment maintained a pixel correlation convergence rate of over 85% even in extremely cold environments, and the boundary mismatch was reduced by approximately 28% compared to traditional fixed-parameter methods.

[0178] Example 3

[0179] In monitoring the lining of permafrost tunnels, a wide-angle camera with a focal length of 12mm was used for data acquisition, covering a monitoring area of ​​6m. A pre-fabricated speckle was used, with an average dot diameter of approximately 3px. Data processing was performed on an edge GPU (NVIDIA Jetson AGX Orin) for real-time analysis. In this scenario, prior information on an "abnormal humidity field" was added to the reinforcement learning state. The ROI adopted a hierarchical structure, with the core layer focusing on the crack initiation zone and the transition layer extending to the thermal and humidity anomaly region. The reward function further incorporated a humidity field consistency term. Monitoring results show that the system can operate stably in tunnels under dripping and frosting conditions, automatically tracking the development of lining cracks, with crack width measurement errors controlled within ±0.15mm.

[0180] Example 4

[0181] In the freeze-thaw cycle model box experiment, the MAB (Modal-Assisted Boolean) algorithm was used instead of reinforcement learning to optimize the ROI and DIC parameters. The action set still included ROI morphology editing and subset parameter adjustment. The experimental subjects were soil samples from the model box subjected to periodic freeze-thaw cycles. The MAB algorithm showed faster convergence speed in short-term experiments and is suitable for situations with relatively small ROI ranges and relatively stable environments. Experimental results showed that the method reduced the average residual by about 25% and shortened the experimental preparation time by about 15% over three freeze-thaw cycles compared to the fixed-parameter method, verifying the feasibility of the alternative optimization strategy.

[0182] Based on the foregoing description of the embodiments, the solutions provided by the embodiments of the present invention have the following beneficial effects.

[0183] 1. It has achieved a fundamental shift from static, experience-based judgment to dynamic, intelligent decision-making:

[0184] Existing technologies primarily rely on operators' prior knowledge for parameter presets and manual adjustments, representing a static, passive, offline optimization mode. This invention, by introducing a reinforcement learning agent, constructs a real-time perception-decision-feedback closed-loop system. This system can autonomously optimize parameters online based on dynamic changes in the measurement state, achieving a fundamental methodological leap and solving the core problem of performance degradation under complex and changing conditions caused by fixed parameters in traditional methods.

[0185] 2. Effectively resolved the inherent technical contradictions in the configuration of measurement parameters:

[0186] To address the long-standing technical contradictions in DIC, such as the difficulty in reconciling measurement accuracy and computational efficiency, and spatial resolution and noise resistance, this invention achieves spatial adaptive allocation of parameters through a reinforcement learning agent. It can implement differentiated optimization strategies for different features in different regions of the image (such as uniformly deformed areas and high-gradient feature areas), while simultaneously ensuring the stability of the overall measurement and the precision of key areas—something traditional methods cannot achieve.

[0187] 3. Full automation and intelligentization of the entire process have been achieved:

[0188] Existing technologies only automate the calculation process, while the crucial parameter optimization stage still heavily relies on manual intervention. This invention fully automates the parameter tuning process, which requires expert experience, truly achieving intelligent optimization throughout the entire process from image acquisition to result output. This not only reduces the technical requirements for operators but also significantly improves the consistency and repeatability of the measurement process.

[0189] 4. Possesses excellent technical versatility and system compatibility:

[0190] Unlike traditional methods where parameter sets optimized for specific scenarios are difficult to transfer, this invention constructs a universal technical framework. As shown in the embodiments, the system can be implemented using different reinforcement learning algorithms (such as PPO, SAC, etc.) and can be seamlessly integrated with various hardware platforms (such as industrial PLCs and MES systems), demonstrating strong technical inclusiveness and adaptability, and providing a unified solution for different application scenarios.

[0191] 5. Achieve performance improvements simultaneously across multiple technical dimensions:

[0192] Through the aforementioned technological innovations, this invention achieves significant improvements in multiple dimensions, including measurement accuracy, robustness, automation, and applicability. Particularly in complex conditions such as strain concentration regions and discontinuous deformation fields, the reliability and stability of its measurement results represent a qualitative leap compared to traditional methods, providing a completely new technical means for precision measurement in related fields.

[0193] like Figure 7As shown, this embodiment of the invention also provides an adaptive measurement device 7 for full-field deformation of frozen soil, including a memory 71 and a processor 70. The memory 71 stores a computer program 72. When the processor 70 executes the computer program 72, it implements the steps of the method in the above-described method embodiments, including but not limited to: multimodal data acquisition and synchronization, candidate ROI generation, physical consistency index calculation, ROI-DIC linkage optimization, DIC calculation, quality assessment and degradation trigger rollback / reset, and result output and storage. Exemplarily, the adaptive measurement device 7 for full-field deformation of frozen soil can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., and is not limited thereto.

[0194] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0195] 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. An adaptive measurement method for full-field deformation of frozen soil, characterized in that, include: Multimodal data of the permafrost environment under test are collected, and the multimodal data are preprocessed to obtain initial data; wherein, the multimodal data includes sequence image data and temperature data; The regions of interest (ROIs) are labeled on the sequential image data in the initial data to obtain the ROI mask and confidence map of the permafrost environment under test; The ROI mask and the confidence map are input into the constructed reinforcement learning agent, and the linked action data is output. The reward function of the reinforcement learning agent is coupled with multiple physical consistency metrics calculated based on the Digital Image Correlation (DIC) method. The action space of the reinforcement learning agent includes ROI mask morphology editing, multi-level ROI gating, and DIC parameter adjustment. The linked action data includes actions for morphology editing of candidate ROI masks and actions for adjusting DIC parameters. The DIC parameters include subset size, step size, and weight. The multiple physical consistency metrics include DIC-related residuals, boundary mismatch, high strain gradient coverage, temperature field consistency, and temporal consistency. The reward function is a weighted sum of these five metrics. The multi-level ROI gating divides the ROI into core regions, transition regions, and buffer regions to achieve different DIC calculation weights for different regions. The ROI mask and DIC parameters are updated based on the linkage action data, and the full-field deformation is calculated to output the full-field displacement field and the full-field strain field. Based on the full-field displacement field and the full-field strain field, a quality index is calculated, and when the quality index deteriorates, the reinforcement learning agent is optimized and updated. Wherein, the DIC-related residual is zero-mean normalized least squares deviation (ZNSSD) or zero-mean normalized cross-correlation (ZNCC); The reward function is: in, The mean of ZNSSD or ZNCC residuals. For boundary mismatch degree, To achieve coverage of regions with high strain gradients, For timing consistency, To ensure consistency of the temperature field between the ROI and the thermal anomaly region, , For the measurement domain; Let t be the ROI mask. , The confidence threshold is in the range of (0–1). For confidence graph The resulting candidate contour set, , This is a region with a high strain gradient. A high strain gradient discrimination threshold is used to determine the strain gradient norm. Extracting high strain gradient regions , Let be the optical flow displacement at time t. The optical flow displacement at time t-1; , For the temperature field in The mean within, For the temperature field in within the standard deviation, This is a thermal anomaly region. coordinates Temperature value at point, The coefficient for distinguishing thermal anomalies. As weight, ; Among them, improvements were made in areas with significant permafrost thermal anomalies. η The corresponding weight values; increased in the high strain gradient region. γ The corresponding weight value.

2. The adaptive measurement method for full-field deformation of frozen soil according to claim 1, characterized in that, The calculation of quality indices based on the full-field displacement field and the full-field strain field includes: Calculate the full-field average correlation residuals based on ZNSSD or ZNCC; Calculate the symmetric average nearest point distance between the ROI mask contour and the candidate contour set obtained from the confidence map to obtain the boundary mismatch degree; Calculate the normalized similarity of displacement field increments at adjacent time points to obtain temporal consistency of displacement fields.

3. The adaptive measurement method for full-field deformation of frozen soil according to claim 2, characterized in that, When the quality index deteriorates, the process of performing optimization updates on the reinforcement learning agent includes: When the full-field average correlation residual is greater than the corresponding calibration threshold, the boundary mismatch is greater than the corresponding calibration threshold, or the displacement field temporal consistency is less than the corresponding calibration threshold, the current deteriorated state is re-input into the reinforcement learning agent to restart the local search process; Based on the new linkage action data output by the search, update the ROI mask processing and the DIC parameter linkage action data.

4. The adaptive measurement method for full-field deformation of frozen soil according to claim 1, characterized in that, The preprocessing of the multimodal data to obtain initial data includes: The sequence image data and the temperature data are subjected to spatiotemporal alignment processing; The sequence of image data is denoised, and the denoised sequence of image data is then normalized for irradiance and exposure; wherein the denoising is based on nonlocal mean or bilateral filtering. The temperature data is subjected to low-pass filtering and outlier removal.

5. The adaptive measurement method for full-field deformation of frozen soil according to claim 1, characterized in that, Following the output of the full-field displacement field and the full-field strain field, the following is also included: Generate panoramic surveillance images; The panoramic monitoring image includes one or more of the following: ROI region display image, displacement-strain field cloud map, and multi-sensor time series curve.

6. An adaptive measurement system for full-field deformation of frozen soil, characterized in that, include: LED matrix light source, high-speed camera, infrared thermal imager, edge computing unit and cloud server; The high-speed camera and infrared thermal imager are used to collect multimodal data of the permafrost environment under test. The edge computing unit is used to perform the method as described in any one of claims 1 to 5; The edge computing unit is a setting chip or a device configured with the setting chip; the setting chip is a GPU, FPGA or SoC.

7. An adaptive measurement device for full-field deformation of frozen soil, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 5.

Images