A deep learning-based frozen wall thickness online evaluation method

Abstract

The application discloses a kind of frozen wall thickness online evaluation methods based on deep learning, it is related to frozen method construction monitoring technical field, comprising the following steps: S1, generate theoretical intersection bottom drawing, extract theoretical extreme point set;S2, construct global information entropy matrix, construct improved UNETR model;S3, introduce exclusive-or logical constraint mechanism, construct hybrid objective function space;S4, based on Hawkes process constructs risk space-time evolution model, exports optimal risk node coordinates as risk prediction center point;S5, execute physical drilling to obtain measured temperature and formation parameters;S6, calculate the bayesian surprise degree of measured data and prior prediction, based on surprise degree constructs dynamic weighted loss function;S7, output frozen wall thickness evaluation result.The application overcomes the limitations of passive monitoring lag, poor characteristic physical interpretability and ignoring active sampling in traditional methods, and provides an efficient solution for frozen method construction safety.

Description

Technical Field

[0001] This invention relates to the field of construction monitoring technology using the freezing method, and in particular to an online assessment method for the thickness of frozen walls based on deep learning. Background Technology

[0002] With the widespread application of artificial ground freezing engineering in urban underground space development, the formation mechanism and thickness monitoring of frozen walls under complex geological conditions face severe computational and modeling challenges. Existing traditional numerical simulation methods and shallow machine learning algorithms, such as the equivalent thickness method and multiple regression analysis, while improving computational efficiency by utilizing simplified forms of the heat conduction equation, mainly rely on discrete point data from sparse temperature measurement wells and pre-set physical empirical parameters for deduction. This method, based solely on sparse sampling points, ignores the complex nonlinear dynamic characteristics implicit in the freezing temperature field (such as heat flow advection effects and abrupt changes in formation thermal parameters) and deep global spatiotemporal correlations, leading to structural distortion of spatial information when constructing a three-dimensional temperature field model, thus limiting the accuracy of freezing front prediction and thickness assessment. Furthermore, existing static monitoring and assessment methods often struggle to fully utilize the spatiotemporal evolution patterns in historical monitoring data to dynamically optimize prediction models, resulting in delayed responses when dealing with formation hydrogeological anomalies, increasing the risk of frozen wall failure or water infiltration accidents.

[0003] Therefore, how to provide a deep learning-based online evaluation method for frozen wall thickness is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a deep learning-based online assessment method for frozen wall thickness. Through a Hawkes process-driven active Bayesian surprise learning closed-loop mechanism, MH random walk attention calculation and Monte Carlo tree search expansion are performed on the acquired frozen field physical decomposition features and the mixed objective function space. A high-dimensional vector feature, including an irrotational potential field vector, a divergence-free flow field vector, and a global information entropy matrix, is constructed and input into an improved UNETR model. An MH random walk attention encoding layer is used to capture long-distance heat conduction dependencies. Combining a Hawkes process risk spatiotemporal evolution model and an XOR logic constraint mechanism, a comprehensive reward value is calculated to drive the Monte Carlo tree search, outputting the optimal risk node coordinates and performing physical drilling. Measured temperatures and formation parameters are mapped back to the frozen field state vector to form incremental training samples. The Bayesian surprise of the measured data and prior predictions is calculated, a dynamic weighted loss function is constructed, and the network weight parameters are iteratively corrected using a gradient descent algorithm. This mechanism, by establishing a feedback pathway from proactive search for spatiotemporal risks to dynamic correction of model uncertainties, effectively eliminates geological anomaly blind spots in frozen field prediction. It ensures that the corrected, improved UNETR model can dynamically adapt to nonlinear changes in stratigraphic structure, achieving the technical effect of improving the spatial assessment accuracy of frozen wall thickness while maintaining high computational performance. This invention overcomes the limitations of traditional methods, such as passive monitoring lag, poor interpretability of characteristic physics, and neglect of active sampling, providing an efficient solution for safe construction using the freezing method.

[0005] According to an embodiment of the present invention, a deep learning-based online evaluation method for frozen wall thickness specifically includes:

[0006] S1. Obtain the topological coordinates and zero-degree duration of the temperature measuring hole and the freezing hole, construct the geometric model of the freezing column and generate the theoretical intersection base map, and extract the theoretical extreme point set;

[0007] S2. Based on the theoretical intersection base map, the frozen field data is decomposed into an irrotational potential energy field vector and a divergence-free flow field vector and mapped to the grid nodes. The global information entropy matrix is ​​constructed using the field numerical distribution. Combined with the Laplace operator gradient characteristics, an improved UNETR model is constructed.

[0008] S3. Introduce the XOR logic constraint mechanism to perform logical bit operations on the global information entropy matrix and the theoretical extreme point set to identify abnormal regions, construct a logical mask screening search space, and integrate the geometric constraint of the freezing hole spacing to construct a hybrid objective function space.

[0009] S4. Construct a risk spatiotemporal evolution model based on Hawkes process in the mixed objective function space, calculate the spatiotemporal evoked effect and combine it with information gain to obtain the comprehensive reward value, drive the Monte Carlo tree search expansion and convergence, and output the optimal risk node coordinates as the risk prediction center point.

[0010] S5. Starting from the risk prediction center point, calculate the geological structure trend based on the visit frequency gradient of Monte Carlo tree backpropagation, generate borehole coordinate sequence along the trend direction, and perform physical drilling to obtain measured temperature and formation parameters.

[0011] S6. Map the measured temperature and formation parameters back to the frozen field state vector to form incremental training samples, calculate the Bayesian surprise of the measured data and prior predictions, construct a dynamic weighted loss function, optimize the network weight parameters, and correct and improve the UNETR model.

[0012] S7. Based on the modified UNETR model, spatial distribution characteristics are extracted. Geostatistical methods are used to model the spatial correlation of frozen soil thickness and make optimal unbiased estimation. Uncertainty quantification index is constructed by combining the estimation variance, and the frozen wall thickness assessment results are output.

[0013] Optionally, S1 specifically includes:

[0014] S11. Obtain the three-dimensional spatial coordinates of the temperature measuring hole and the freezing hole, as well as the geometric parameters of the borehole axis, and construct a borehole spatial distribution model.

[0015] S12. Based on the borehole spatial distribution model, the zero-degree duration is mapped to the freezing temperature field evolution curve to obtain the radial radius of the freezing column as time increases, and the geometric model of the freezing column is generated by combining the radial radius and the axis geometric parameters.

[0016] S13. Perform a Boolean union operation on the frozen column geometric model to generate the frozen intersection region, and cut the bottom surface along the vertical direction to generate the theoretical intersection base map;

[0017] S14. Based on the boundary contour of the theoretical intersection base map, extract the coordinate extreme points and generate the theoretical extreme point set.

[0018] Optionally, S2 specifically includes:

[0019] S21. Map the temperature readings of the temperature measuring holes to the theoretical intersecting base map grid, calculate the three-dimensional temperature gradient vector of each node, and combine them to generate the initial state vector field of the freezing field.

[0020] S22. Calculate divergence and curl based on vector field components. Solve the Poisson equations for scalar potential and vector potential respectively with divergence and curl as source terms. Obtain the irrotational potential energy field vector by calculating the gradient of the scalar potential and the irrotational flow field vector by calculating the curl of the vector potential.

[0021] S23. Calculate the Laplace operator magnitude of the irrotational potential energy field vector and the divergence-free flow field vector. Weight the Laplace operator magnitude with the irrotational and divergence-free magnitudes to obtain the local joint characteristic. Normalize the local joint characteristic to obtain the node energy distribution probability. Calculate the local information entropy value of each node based on the node energy distribution probability. Arrange the nodes according to their grid positions to generate the global information entropy matrix.

[0022] S24. The irrotational potential field vector, the divergence-free flow field vector, the Laplace operator gradient feature and the global information entropy matrix are concatenated to construct a high-dimensional vector feature that includes physical decomposition features and uncertainty measures, and an improved UNETR model is constructed.

[0023] Optionally, the improved UNETR model includes a 3D voxel embedding layer, an MH random walk attention encoding layer, a multi-scale feature aggregation layer, a 3D voxel decoding layer, a Bayesian uncertainty estimation head, and a frozen front segmentation layer.

[0024] The 3D voxel embedding layer is used to perform block embedding processing on the high-dimensional vector features after splicing the irrotational potential energy field vector, the divergence-free flow field vector, the Laplacian operator gradient features, and the global information entropy matrix. The three-dimensional voxel grid is divided into image blocks of fixed size. An initial voxel embedding vector sequence is generated through linear projection mapping, and three-dimensional position encoding is added to output the initial voxel embedding vector sequence containing spatial geometric structure information.

[0025] The MH random walk attention coding layer is used to receive the initial voxel embedding vector sequence, replacing the traditional multi-head self-attention mechanism. It constructs a sparse adjacency graph through random sampling to capture long-distance heat conduction dependencies and outputs a sparse global context feature map. Specifically, it includes a proposal distribution construction module, an acceptance probability calculation module, and a feature aggregation module.

[0026] The proposed distribution construction module is used to parse the three-dimensional spatial coordinates and Laplacian gradient features from the initial voxel embedding vector sequence, define candidate neighborhoods based on physical distance with the current voxel node as the center, calculate the probability density of the current node in combination with the gradient magnitude, and perform random walks from the current node to the nodes in the neighborhood.

[0027] The acceptance probability calculation module is used to calculate the probability density of the target node based on the irrotational potential energy modulus and the global information entropy, and to calculate the ratio of the probability density of the target node to the probability density of the current node. The acceptance rate is calculated in combination with the Metropolis-Hastins criterion. If the acceptance rate is greater than the generated random threshold, the corresponding wandering node is accepted as a relevant neighbor; otherwise, it is rejected.

[0028] The feature aggregation module is used to collect all accepted nodes to form the most relevant node set after walking through a preset number of steps, and to perform attention weight calculation and feature update based on the current node to query the nodes in the most relevant node set, and output a sparse global context feature map.

[0029] The multi-scale feature aggregation layer is used to receive intermediate layer features from different stages of the MH random walk attention coding layer, and to perform skip connections and fusion of local detail features and global walk features under different walk radii to reconstruct an enhanced feature map that preserves spatial details.

[0030] The 3D voxel decoding layer is used to receive the enhanced feature map and perform progressive upsampling operations to restore the spatial resolution of the feature map; combined with the intermediate layer features of different stages of the MH random walk attention coding layer, the features are refined step by step through convolution and upsampling operations to generate a high-resolution predicted feature map with the same size as the input.

[0031] The Bayesian uncertainty estimation head is used to receive high-resolution predicted feature maps, generate the mean parameter of the frozen temperature field via a linear mapping of a parallel first convolutional branch, and generate the log-variance parameter of the frozen temperature field via a linear mapping of a second convolutional branch.

[0032] The freezing front segmentation layer is used to receive high-resolution predicted feature maps, fuse the mean parameter and the logarithmic variance parameter, perform semantic segmentation on the fused features, and output the probability value of each grid node belonging to the permafrost region. Based on the probability value and the result of the mean parameter after being verified by the freezing temperature threshold, the thickness distribution boundary of the frozen wall is jointly extracted, and the final frozen field state vector is generated by combining the segmented feature mapping.

[0033] Optionally, S3 specifically includes:

[0034] S31. Read the global information entropy matrix and the theoretical extreme point set, map the theoretical extreme point set to the grid coordinates of the global information entropy matrix, calculate the binary state difference at the mapped coordinate position through logical XOR operation, identify abnormal regions based on the difference results and generate an initial difference mask.

[0035] S32. Based on the initial difference mask, lock the mesh search space, obtain the spatial three-dimensional coordinates of the freezing holes, and calculate the Euclidean distance between each pair of freezing holes to generate the freezing hole spacing matrix.

[0036] S33. Set the distance weight coefficient according to the freezing hole spacing matrix, map the distance weight coefficient to the grid search space, and construct a hybrid objective function space that integrates spatial geometric constraints.

[0037] Optionally, S4 specifically includes:

[0038] S41. Read the grid nodes in the mixed objective function space as candidate nodes and historical anomaly information, calculate the time interval and Euclidean distance between candidate nodes and historical anomalies, calculate the exponent using the time decay parameter and spatial decay parameter, perform exponentiation on the natural constant to obtain the excitation intensity value, and sum them up to obtain the spatiotemporal influence value.

[0039] S42. Extract the initial entropy value of the candidate node from the global information entropy matrix, calculate the average entropy value of the neighborhood within the preset neighborhood range, take the absolute value of the difference between the initial entropy value and the average entropy value of the neighborhood as the information gain, multiply the spatiotemporal influence value and the information gain by the corresponding weight coefficients respectively and add them together to obtain the comprehensive reward value.

[0040] S43. Construct a Monte Carlo tree structure, use the confidence interval upper bound algorithm to select child nodes for expansion, sample the expanded nodes and calculate the comprehensive reward value, backpropagate to update the node parameters, and after iterating to a preset number of times, output the node with the highest average reward value as the risk prediction center point.

[0041] Optionally, S5 specifically includes:

[0042] S51. Using the risk prediction center point as the reference point, extract the number of visits to each child node in the neighborhood of the current node in the Monte Carlo tree, and construct the number of visits distribution matrix.

[0043] S52. Calculate the directional derivatives of the visit frequency distribution matrix in the X, Y and Z axes of the three-dimensional spatial coordinate system, and combine the directional derivatives to obtain the visit frequency gradient vector.

[0044] S53. Normalize the gradient vector of the number of visits, and set the direction of the normalized vector to the direction of the geological structure trend.

[0045] S54. Using the risk prediction center point as the starting point of the sequence, and following the geological structure trend direction, perform ray interpolation in combination with the preset borehole spacing to generate a borehole coordinate sequence.

[0046] S55. Perform physical drilling operations based on the borehole coordinate sequence, and collect measured temperature data and formation parameters during the drilling process.

[0047] Optionally, S6 specifically includes:

[0048] S61. Obtain the coordinates, measured temperature, and formation parameters of the physical drilling borehole. Convert the coordinates into grid indices of the frozen field state vector. Assign the measured temperature and formation parameters to the vector features of the corresponding grid indices. Then, concatenate the assigned vectors with the data in the historical sample database in chronological order to construct an incremental training sample set.

[0049] S62. Input the incremental training sample set into the improved UNETR model, calculate the prediction mean and prediction variance of each node in the output layer through forward propagation, construct a Gaussian distribution using the prediction mean and prediction variance as the prior prediction distribution, calculate the probability density value of the measured data in the prior prediction distribution, take the natural logarithm of the probability density value and take the opposite number to obtain the Bayesian surprise value.

[0050] S63. Construct the basic mean square error loss function, obtain the preset maximum surprise threshold, divide the Bayesian surprise value by the maximum surprise threshold, add the calculation result to the constant 1 to obtain the dynamic weighting coefficient, multiply the dynamic weighting coefficient by the basic mean square error loss function to construct the dynamic weighting loss function.

[0051] S64. Calculate the gradient of the dynamic weighted loss function with respect to the network weight parameters, set the learning rate and momentum parameters, and use the gradient descent algorithm to iteratively update the network weight parameters until the preset convergence condition is met, and output the corrected improved UNETR model.

[0052] Optionally, S7 specifically includes:

[0053] S71. Based on the modified and improved UNETR model, extract the spatial three-dimensional coordinates and temperature field prediction values ​​of the frozen soil unit to generate a spatial discrete point set of frozen soil.

[0054] S72. Using the spatial discrete point set of permafrost as a regional variable, geostatistical methods are used to calculate the semivariogram, and a variogram model is fitted to analyze the variability and continuity of permafrost thickness in the spatial direction, and a spatial correlation model is constructed.

[0055] S73. Based on the spatial correlation model, the Kriging interpolation algorithm is used to make the optimal unbiased estimate of the frozen soil thickness at non-observation locations. The thickness estimate and corresponding estimation variance of the grid nodes are calculated to generate full-field thickness distribution data containing deterministic prediction and uncertainty measurement.

[0056] S74. Convert the estimated variance into a spatial confidence index, identify areas with thinness, and quantitatively assess the sealing and stability of the frozen wall, outputting the assessment results.

[0057] The beneficial effects of this invention are:

[0058] (1) This invention achieves precise decoupling of deep physical dynamics of the frozen temperature field and efficient capture of long-distance thermal conduction dependence by constructing an improved UNETR model and an MH random walk attention mechanism. The frozen field data is decomposed into an irrotational potential energy field vector, a divergence-free flow field vector, and Laplace operator gradient features and mapped to grid nodes. A high-dimensional vector feature input is constructed by combining the global information entropy matrix to improve the UNETR model. The MH random walk attention encoding layer performs sparse adjacency graph construction and feature aggregation based on the Metropolis-Hastins criterion through a proposal distribution construction module and an acceptance probability calculation module, replacing the traditional multi-head self-attention mechanism. This mechanism transforms the continuous physical field into a sparse graph attention space containing physical prior constraints, effectively utilizing the Laplace gradient magnitude and irrotational potential energy magnitude to guide the search path, ensuring that the model captures long-distance dependence while significantly reducing computational complexity, and significantly improving the representation accuracy and inference efficiency of the frozen field state vector.

[0059] (2) This invention establishes a boundary extraction and active detection closed-loop correction system that integrates probability-driven and physical verification by employing a dual constraint mechanism of the frozen front segmentation layer and a Bayesian surprise-weighted feedback mechanism. The frozen front segmentation layer fuses the mean parameter and log-variance parameter for feature extraction and performs semantic segmentation. Based on the results of the probability value and the mean parameter after verification by the freezing temperature threshold, the frozen wall thickness distribution boundary is jointly extracted. Physical drilling is performed by combining the optimal risk node coordinates output by the Hawkes process risk spatiotemporal evolution model and Monte Carlo tree search. A dynamic weighted loss function is constructed using Bayesian surprise to optimize the network weight parameters and correct and improve the UNETR model. This system accurately locks the frozen front through joint verification of probability distribution and temperature threshold, and achieves targeted learning of abnormal samples through Bayesian surprise weighting, ensuring that the output frozen wall thickness assessment results and uncertainty quantification indicators have extremely high adaptability and accuracy to complex geological changes. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is an overall flowchart of a deep learning-based online evaluation method for frozen wall thickness proposed in this invention.

[0062] Figure 2 This is a flowchart illustrating the working principle of the improved UNETR model, a deep learning-based online evaluation method for frozen wall thickness proposed in this invention. Detailed Implementation

[0063] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1 and Figure 2 A deep learning-based online method for evaluating the thickness of frozen walls, specifically including:

[0065] S1. Obtain the topological coordinates and zero-degree duration of the temperature measuring hole and the freezing hole, construct the geometric model of the freezing column and generate the theoretical intersection base map, and extract the theoretical extreme point set;

[0066] S2. Based on the theoretical intersection base map, the frozen field data is decomposed into an irrotational potential energy field vector and a divergence-free flow field vector and mapped to the grid nodes. The global information entropy matrix is ​​constructed using the field numerical distribution. Combined with the Laplace operator gradient characteristics, an improved UNETR model is constructed.

[0067] S3. Introduce the XOR logic constraint mechanism to perform logical bit operations on the global information entropy matrix and the theoretical extreme point set to identify abnormal regions, construct a logical mask screening search space, and integrate the geometric constraint of the freezing hole spacing to construct a hybrid objective function space.

[0068] S4. Construct a risk spatiotemporal evolution model based on Hawkes process in the mixed objective function space, calculate the spatiotemporal evoked effect and combine it with information gain to obtain the comprehensive reward value, drive the Monte Carlo tree search expansion and convergence, and output the optimal risk node coordinates as the risk prediction center point.

[0069] S5. Starting from the risk prediction center point, calculate the geological structure trend based on the visit frequency gradient of Monte Carlo tree backpropagation, generate borehole coordinate sequence along the trend direction, and perform physical drilling to obtain measured temperature and formation parameters.

[0070] S6. Map the measured temperature and formation parameters back to the frozen field state vector to form incremental training samples, calculate the Bayesian surprise of the measured data and prior predictions, construct a dynamic weighted loss function, optimize the network weight parameters, and correct and improve the UNETR model.

[0071] S7. Based on the modified UNETR model, spatial distribution characteristics are extracted. Geostatistical methods are used to model the spatial correlation of frozen soil thickness and make optimal unbiased estimation. Uncertainty quantification index is constructed by combining the estimation variance, and the frozen wall thickness assessment results are output.

[0072] In this embodiment, S1 specifically includes:

[0073] S11. Obtain the inclination data and construction acceptance records of the borehole layout, read the three-dimensional geodetic coordinates of the borehole opening and the borehole inclination and azimuth parameters in the records, use trigonometric functions to calculate the extension vector of the borehole axis in three-dimensional space, combine the borehole depth parameters to generate a series of discrete three-dimensional coordinate points on the axis, connect all discrete points in the order of borehole number to generate line segments, and construct a borehole spatial distribution model that includes spatial position and geometric shape.

[0074] S12. Read the distance from the temperature sensor hole to the nearest freezing hole as L, count the number of days the temperature sensor reading at the temperature sensor hole is below 0 degrees Celsius as d, read the number of days in the current active freezing period as D, and use the formula:

[0075]

[0076] Calculate the development radius of the frozen cylinder, where R represents the development radius of the frozen soil, and generate a geometric model of the frozen cylinder by combining the development radius of the frozen cylinder with the axial geometric parameters.

[0077] S13. Traverse all spatial grid cells in the geometric model of frozen columns, draw the intersection area of ​​frozen columns on the freezing hole layout model according to the development radius of frozen columns, perform horizontal sectioning of the intersection area along the vertical elevation direction and extract the cross-sectional profile, and generate the theoretical intersection base map according to the filling area inside the cross-sectional profile.

[0078] S14. Read the pixel matrix data of the theoretical intersection base map, traverse each row and each column of the matrix, and count the column coordinates of each row of pixels where the gray value jumps from 0 to 255 as the left boundary point, and count the column coordinates of each row of pixels where the gray value jumps from 255 to 0 as the right boundary point. Extract the four points with the smallest and largest x-coordinate values ​​and the smallest and largest y-coordinate values ​​among the left and right boundary points, convert the planar coordinates of the four points into three-dimensional spatial coordinates, and generate the theoretical extreme point set.

[0079] In this embodiment, S2 specifically includes:

[0080] S21. Read the spatial three-dimensional coordinates of the temperature measuring hole and the corresponding temperature sensor readings. Fill the temperature readings into the grid nodes corresponding to the theoretical intersection base map. Traverse each three-dimensional coordinate point of the grid node, extract the node temperature value at a distance of 0.5 meters in the positive X-axis direction and the node temperature value at a distance of 0.5 meters in the negative X-axis direction of the current node. Subtract the temperature value in the negative X-axis direction from the temperature value in the positive X-axis direction and divide the result by 1 to calculate the temperature partial derivative in the X-axis direction. Similarly, extract the temperature values ​​at a distance of 0.5 meters in the positive and negative Y-axis directions and the temperature values ​​at a distance of 0.5 meters in the positive and negative Z-axis directions, calculate the difference, divide by the distance, and calculate the temperature partial derivatives in the Y-axis and Z-axis directions. Combine the partial derivative arrays in the X-axis, Y-axis, and Z-axis directions into a three-dimensional vector to construct the initial state vector field of the freezing field.

[0081] S22. Read the partial derivatives of the X-axis component, Y-axis component, and Z-axis component in the initial state vector field of the frozen field. Summate these three partial derivatives in the same direction to obtain the divergence value. Read the partial derivatives of the Y-axis component in the Z-axis and the Z-axis component in the Y-axis, and subtract them to obtain the first component of the curl vector. Similarly, calculate the other two components of the curl vector using the differences between the cross partial derivatives of the Z-axis and X-axis, and X-axis and Y-axis. Assign the divergence values ​​to the right-hand side of the Poisson equation. The source term is set to 0 for the scalar potential function at the grid boundary as the first type of boundary condition. The coordinate difference between the current grid node and its six adjacent nodes (up, down, left, right, front, and back) is calculated. A system of linear algebraic equations containing the potential function values ​​of the six adjacent nodes is constructed. The system of equations is solved iteratively using the preconditional conjugate gradient method. The iteration stops when the difference between the potential function values ​​of two adjacent iterations is less than 0.001. The scalar potential function values ​​of each grid node are output. The partial derivatives of the scalar potential function values ​​in the X, Y, and Z axes are calculated respectively. The partial derivatives in the three directions are combined to obtain the irrotational potential energy field vector.

[0082] S23. Calculate the second-order partial derivatives of the irrotational potential energy field vector and the divergent flow field vector in the X, Y, and Z axes. Add the second-order partial derivatives in the three directions to obtain the Laplace operator value. Extract the absolute value of the Laplace operator value as the Laplace operator gradient feature. Calculate the magnitude of the irrotational potential energy field vector as the irrotational magnitude value. Calculate the magnitude of the divergent flow field vector as the divergent magnitude value. Multiply the Laplace operator gradient feature by 0.5, the irrotational magnitude value by 0.3, and the divergent magnitude value by 0.2, and add the three products to obtain the local joint feature of each grid node. Calculate the sum of the local joint feature of all grid nodes. Divide the local joint feature of a single grid node by the sum to obtain the node energy distribution probability. Take the logarithm of the node energy distribution probability to the base 10 and take the opposite of the result to calculate the local information entropy value of each node. Arrange the grids according to their row, column, and layer positions in three-dimensional space to generate a global information entropy matrix.

[0083] S24. Extract the three component values ​​of the irrotational potential energy field vector along the X, Y, and Z axes; extract the three component values ​​of the divergent flow field vector along the X, Y, and Z axes; extract the values ​​of the Laplace operator gradient feature and the global information entropy matrix; concatenate the three component values ​​of the irrotational potential energy field vector, the three component values ​​of the divergent flow field vector, the Laplace operator gradient feature value, and the global information entropy matrix value in sequence to construct a high-dimensional vector feature containing physical decomposition features and uncertainty measures; use the high-dimensional vector feature as input data to construct an improved UNETR model.

[0084] The physical field decomposition feature extraction method proposed in this step is similar to the traditional temperature field processing method in that both are based on discrete grids to characterize the physical field and use numerical difference to calculate the spatial rate of change.

[0085] The difference lies in that this invention breaks through the limitations of traditional methods that extract features based solely on temperature scalars or simple gradients. It adds a Helmholtz vector decomposition step, constructs a Poisson equation with divergence and curl as source terms, and generates an irrotational potential energy field vector and a divergence-free flow field vector through iterative solution. Furthermore, it combines Laplace gradient features and global information entropy to construct high-dimensional joint features.

[0086] The beneficial effect of the improvement is that by decoupling the temperature field into physical potential energy and flow field through vector decomposition and Poisson solution, the accurate mapping from the observed value to the deep physical driving force is realized, which significantly enhances the model's physical perception of the nonlinear law of heat flow evolution and effectively improves the identification accuracy and robustness in complex geological environments.

[0087] In this embodiment, the improved UNETR model includes a 3D voxel embedding layer, an MH random walk attention encoding layer, a multi-scale feature aggregation layer, a 3D voxel decoding layer, a Bayesian uncertainty estimation head, and a frozen front segmentation layer.

[0088] The 3D voxel embedding layer is used to read the irrotational potential energy field vector, the divergence-free flow field vector, the Laplacian operator gradient features, and the high-dimensional vector features after the global information entropy matrix is ​​concatenated. The 3D voxel grid is divided into 4x4x4 image blocks along the length, width, and height dimensions. The feature vectors in each image block are flattened into one-dimensional vectors of length 256. The one-dimensional vectors are linearly mapped to the initial voxel embedding vectors using a weight matrix of dimension 256x128. The row index, column index, and layer index of the voxel nodes in the 3D grid are read. The index values ​​are encoded using sine and cosine functions to generate a 3D position encoding vector. The 3D position encoding vector is added to the initial voxel embedding vector to output the initial voxel embedding vector sequence containing spatial geometric structure information.

[0089] The MH random walk attention coding layer is used to receive the initial voxel embedding vector sequence, replacing the traditional multi-head self-attention mechanism. It constructs a sparse adjacency graph through random sampling to capture long-distance heat conduction dependencies and outputs a sparse global context feature map. Specifically, it includes a proposal distribution construction module, an acceptance probability calculation module, and a feature aggregation module.

[0090] The proposal distribution construction module is used to read the initial voxel embedding vector sequence, extract the three-dimensional spatial coordinates and Laplacian gradient features of the current voxel node, delineate a spherical search area with a radius of 2.5 meters centered on the current voxel node, mark other voxel nodes in the area as candidate neighbor nodes, calculate the probability density of the current node by dividing the value of the Laplacian gradient feature by the sum of the gradients of the candidate neighbor nodes, generate a proposal distribution with probability density as weight, launch random walk exploration paths to the nodes in the candidate neighbor based on the proposal distribution, and output the coordinates of the target node.

[0091] The acceptance probability calculation module is used to receive target nodes in the random walk trial path and execute the Metropolis-Hastins criterion. Specifically, it reads the irrotational potential modulus and global information entropy of the target node, multiplies the irrotational potential modulus by 0.6 and adds the global information entropy by 0.4 to calculate the target feature value of the target node, divides the target feature value by the sum of the feature values ​​of all nodes in the search area to calculate the probability density of the target node, calculates the ratio of the probability density of the target node to the probability density of the current node to obtain the acceptance ratio, generates a uniformly distributed random number between 0 and 1, compares the acceptance ratio with the random number, and if the acceptance ratio is greater than the random number, the target node is determined to be accepted as a relevant neighbor; otherwise, it is rejected, and the set of relevant neighbor nodes is output.

[0092] The feature aggregation module is used to collect all accepted relevant neighbor nodes to form the most relevant node set after a walk of a preset number of 8 steps. It reads the voxel embedding vector of each neighbor node in the most relevant node set, calculates the dot product of the voxel embedding vector of the current node and the voxel embedding vector of each neighbor node, performs an exponential operation on the dot product to obtain the attention score, uses the attention score to perform a weighted sum of the voxel embedding vectors of the neighbor nodes, updates the feature vector of the current node, traverses all voxel nodes to complete the feature update, and outputs a sparse global context feature map.

[0093] The multi-scale feature aggregation layer is used to receive intermediate layer features from different stages of the MH random walk attention coding layer. The intermediate layer features from different stages are upsampled to a uniform spatial resolution through bilinear interpolation. The feature vectors after uniform resolution are concatenated end to end in the channel dimension. The concatenated features are fused by convolution using a 1x1 convolution kernel. The fused features are added element-wise to the original input features to reconstruct an enhanced feature map that preserves spatial details.

[0094] The 3D voxel decoding layer is used to read the enhanced feature map. It uses a transposed convolution kernel to perform an upsampling operation with a stride of 2, which increases the spatial resolution of the feature map by 2 times. It combines the intermediate layer features with the same resolution in the MH random walk attention coding layer, and jumps the intermediate layer features to the upsampled feature map. It uses a 3x3 convolution kernel to refine the fused features. The upsampling, jump connection and convolution refinement steps are repeated until the input size is restored, generating a high-resolution predicted feature map with the same size as the input.

[0095] The Bayesian uncertainty estimation head is used to read the high-resolution predicted feature map, input it into the first convolutional branch, and process it with a 3x3 convolutional layer with a stride of 1 and a linear activation function to generate the mean parameter of the freezing temperature field. The high-resolution predicted feature map is then input into the second convolutional branch, and processed with a 3x3 convolutional layer with a stride of 1 and a linear activation function to generate the log-variance parameter of the freezing temperature field. The mean parameter and the log-variance parameter are then output.

[0096] The frozen front segmentation layer is used to read high-resolution predicted feature maps, mean parameters, and log-variance parameters. The mean and log-variance parameters are concatenated along the channel dimension to obtain fused features. The fused features are then convolved using a 1x1 convolution kernel with a stride of 1. The convolution result is processed using the Sigmoid activation function, and the probability value of each grid node belonging to the frozen soil region is output. Nodes with a probability value greater than 0.5 are marked as frozen soil regions. The spatial coordinates of the edge nodes of the frozen soil region are extracted, and the thickness distribution boundary of the frozen wall is jointly extracted. The final frozen field state vector is generated by combining the segmented feature maps.

[0097] The voxel feature encoding process based on MH random walk proposed in this step is similar to the traditional multi-head self-attention mechanism in that it is based on the sequence modeling theory of the Transformer architecture. That is, by mapping the input data into a high-dimensional embedding vector sequence, the global correlation weight between features is calculated using the attention mechanism, and the feature representation of the node is updated through weighted aggregation operation.

[0098] The difference lies in that this invention breaks away from the limitations of traditional methods that rely solely on dense attention calculations based on fully connected graphs or ignore the heterogeneity of the physical field space. Instead of directly calculating the similarity between all nodes in traditional models, this invention adds a Metropolis-Hastins random walk attention mechanism, constructing a proposal distribution based on Laplacian gradient features for local neighborhood sampling, rather than a global traversal search. In the neighbor node selection step, the acceptance probability is calculated using the Metropolis-Hastins criterion combined with irrotational potential and information entropy, rather than a fixed Top-K selection. Finally, in the feature aggregation step, sparse attention calculations and feature updates are performed based on the most relevant node set after the random walk convergence, rather than dense matrix operations.

[0099] The beneficial effects of the improvements are that this invention, through a physically guided random walk and probabilistic acceptance mechanism, can integrate the thermal conductivity physical properties and geometric priors of the frozen wall into the graph structure construction process. This breaks through the limitations of traditional methods in three-dimensional voxel processing, which suffers from high computational complexity and lack of physical constraints, and achieves efficient mapping from global computation in Euclidean space to physically guided sparse sampling. This design significantly enhances the model's ability to capture long-distance thermal conductivity dependence and can more accurately characterize the nonlinear evolution characteristics of the frozen front. The computation method based on sparse adjacency graphs effectively reduces memory usage and computational load, enhancing the real-time performance and robustness of the system in large-scale three-dimensional frozen field analysis.

[0100] In this embodiment, S3 specifically includes:

[0101] S31. Read the global information entropy matrix and the theoretical extreme point set, extract the three-dimensional spatial coordinates of each extreme point in the theoretical extreme point set, convert the three-dimensional spatial coordinates into the row index and column index corresponding to the global information entropy matrix, read the entropy value matrix in the 3x3 neighborhood with the row and column index as the center, mark the nodes in the neighborhood with an entropy value greater than the entropy value threshold of 0.5 as 1, and mark the nodes with an entropy value less than or equal to the entropy value threshold of 0.5 as 0, generate a binarized actual state mask, and generate a theoretical expected mask based on the theoretical extreme points. Perform a logical XOR operation on the values ​​of the actual state mask and the theoretical expected mask at the corresponding positions. When the operation result is 1, the position is determined to be an outlier. Count the position coordinates of all outliers to generate an initial difference mask.

[0102] S32. Read the initial difference mask, extract the coordinates of the outlier points with a value of 1 in the mask, calculate the maximum and minimum values ​​of the outlier point coordinates in the X, Y and Z axes, construct a cuboid mesh search space with the maximum value plus 1 meter as the upper limit and the minimum value minus 1 meter as the lower limit, obtain the spatial three-dimensional coordinates of the freezing holes, calculate the Euclidean distance between every two freezing holes, fill the Euclidean distance values ​​into the corresponding row and column positions of the matrix to generate the freezing hole spacing matrix;

[0103] S33. Read the spacing matrix of frozen holes, extract the distance values ​​on the off-diagonal lines of the matrix, divide 1 by the distance value to obtain the distance weight coefficient, multiply the distance weight coefficient by the coordinate values ​​of the grid nodes in the search space, and superimpose the calculation results onto the grid nodes corresponding to the global information entropy matrix to construct a hybrid objective function space that integrates spatial geometric constraints.

[0104] In this embodiment, S4 specifically includes:

[0105] S41. Read the grid nodes in the mixed objective function space as candidate nodes, read the occurrence timestamps and spatial three-dimensional coordinates of historical anomalies, extract the current system time, calculate the time interval by subtracting the occurrence timestamps of historical anomalies from the current system time, extract the spatial three-dimensional coordinates of candidate nodes, calculate the straight-line distance between the spatial three-dimensional coordinates of historical anomalies and the spatial three-dimensional coordinates of candidate nodes to obtain the Euclidean distance, set the time decay parameter to 24 hours and the spatial decay parameter to 5 meters, divide the time interval by 24 hours to obtain the time decay term, divide the Euclidean distance by 5 meters to obtain the spatial decay term, add the time decay term and the spatial decay term and take the opposite of the result as the exponent, calculate the exponent of the natural constant to obtain the excitation intensity value, sum the excitation intensity values ​​of all historical anomalies to the same candidate node to obtain the spatiotemporal influence value;

[0106] S42. Extract the information entropy values ​​of candidate nodes from the global information entropy matrix as the initial entropy value. Delineate a spherical neighborhood with a radius of 1 meter centered on the candidate node. Calculate the information entropy values ​​of all grid nodes within the neighborhood. Add the information entropy values ​​of all grid nodes within the neighborhood and divide by the total number of nodes to obtain the average entropy value of the neighborhood. Calculate the absolute value of the difference between the initial entropy value and the average entropy value of the neighborhood as the information gain. Set the spatiotemporal weight coefficient to 0.7 and the information gain weight coefficient to 0.3. Multiply 0.7 by the spatiotemporal influence value to obtain the spatiotemporal weighted term. Multiply 0.3 by the information gain to obtain the gain weighted term. Add the spatiotemporal weighted term and the gain weighted term to obtain the comprehensive reward value.

[0107] S43. Construct a Monte Carlo tree structure containing a root node and child nodes. Use the grid nodes in the mixed objective function space as the candidate set of child nodes. Calculate the average reward value of the child nodes at the root node and divide it by the square root of the number of visits. Add the product of the constant 1.4 and the natural logarithm of the ratio of the total number of visits to obtain the selection score. Select the child node with the highest score for expansion. Perform random state sampling on the expanded leaf nodes and extract the comprehensive reward value corresponding to the sampled node as the simulated reward value. Propagate the simulated reward value back to the root node along the search path. Add the simulated reward value to the total reward value of the parent node and increment the number of visits of the parent node by 1. Repeat the selection, expansion, simulation and back propagation steps until the number of visits reaches 1000. Select the node with the highest average reward value as the optimal risk node and output the spatial three-dimensional coordinates of the optimal risk node as the risk prediction center point.

[0108] The spatiotemporal evolution risk prediction process proposed in this step is similar to the traditional Hawkes process point prediction method in that both are based on the spatiotemporal point process modeling theory, that is, using the exponential decay function to simulate the spatiotemporal propagation effect of historical events, and using cumulative summation to measure the historical impact.

[0109] The difference lies in that this invention breaks away from the limitations of traditional methods that rely solely on a single spatiotemporal feature. It adds a step to construct a spatiotemporal influence value, combining historical timestamps and spatial coordinates to calculate the excitation intensity. In the reward value calculation step, the neighborhood difference of the global information entropy is integrated as information gain and weighted together with the spatiotemporal influence value. Finally, in the risk localization step, Monte Carlo tree search and a confidence interval upper bound algorithm are used to iteratively optimize and output the optimal risk node.

[0110] The beneficial effects of the improvements are that by weighting with spatiotemporal decay and entropy gain, the evolution characteristics of the frozen wall and uncertainty anomalies are integrated into the risk assessment, realizing accurate deduction from static monitoring to dynamic risk field; Monte Carlo tree search significantly enhances the ability to capture nonlinear spatiotemporal risk patterns, effectively improving the positioning accuracy and reliability of safety early warning in deep well freezing projects.

[0111] In this embodiment, S5 specifically includes:

[0112] S51. Read the three-dimensional spatial coordinates of the risk prediction center point, extract the neighboring child nodes of the current node within a radius of 5 meters in the Monte Carlo tree, count the number of times each neighboring child node is visited during the search process, use the spatial three-dimensional coordinates of the neighboring child nodes as matrix indexes, fill the corresponding number of visits into the matrix values, and construct the number of visits distribution matrix.

[0113] S52. Read the access frequency distribution matrix. Subtract the value corresponding to the current X-axis index plus 1 from the value corresponding to the current X-axis index minus 1, and divide the difference by 2 to obtain the X-axis derivative. Subtract the value corresponding to the current Y-axis index plus 1 from the value corresponding to the current Y-axis index minus 1, and divide the difference by 2 to obtain the Y-axis derivative. Subtract the value corresponding to the current Z-axis index plus 1 from the value corresponding to the current Z-axis index minus 1, and divide the difference by 2 to obtain the Z-axis derivative. Combine the X-axis derivative, Y-axis derivative, and Z-axis derivative into a vector to obtain the access frequency gradient vector.

[0114] S53. Read the gradient vector of the number of visits, calculate the sum of squares of the X-axis, Y-axis and Z-axis components in the vector, take the square root of the sum of squares to obtain the vector magnitude, divide the X-axis component by the vector magnitude to obtain the X-axis unit component, divide the Y-axis component by the vector magnitude to obtain the Y-axis unit component, divide the Z-axis component by the vector magnitude to obtain the Z-axis unit component, combine the X-axis unit component, Y-axis unit component and Z-axis unit component into a unitized vector, and set the direction of the unitized vector to the geological structure trend direction;

[0115] S54. Read the three-dimensional spatial coordinates of the risk prediction center point as the starting point of the sequence. Set the borehole spacing to 3 meters along the geological structure trend direction. Add the product of 3 meters and the unit vector of the geological structure trend direction to the three-dimensional spatial coordinates of the starting point of the sequence to obtain the coordinates of the first borehole. Add the product of 3 meters and the unit vector of the geological structure trend direction to the coordinates of the first borehole to obtain the coordinates of the second borehole. Repeat the coordinate addition step to generate 10 borehole coordinates and construct the borehole coordinate sequence.

[0116] S55. Read the borehole coordinate sequence, control the drilling equipment to move sequentially to each coordinate position in the borehole coordinate sequence, perform physical drilling operations, collect temperature data every 0.5 meters of depth during the drilling process using a temperature sensor, record the formation lithology parameters, and store the collected measured temperature data and formation parameters.

[0117] In this embodiment, S6 specifically includes:

[0118] S61. Read the coordinates, measured temperature, and formation parameters of the physical drilling borehole. Based on the size and origin coordinates of the frozen field state vector grid, convert the physical spatial coordinates of the borehole into the grid index position. Fill the measured temperature values ​​into the temperature feature channel corresponding to the grid index, and fill the formation parameters into the lithological feature channel corresponding to the grid index. Read the data vector in the historical sample library, and concatenate the vector after filling in the measured data to the end of the historical sample library data vector. Arrange them in chronological order to construct an incremental training sample set.

[0119] S62. Input the incremental training sample set into the improved UNETR model for forward propagation calculation, obtain the predicted mean and predicted variance of the output layer nodes, use the predicted variance value as the variance parameter of the Gaussian distribution, use the predicted mean value as the mean parameter of the Gaussian distribution, construct a Gaussian distribution centered on the predicted mean value as the prior prediction distribution, extract the measured temperature data of the borehole location in the incremental training sample set, calculate the probability density value of the measured temperature data in the prior prediction distribution, perform natural logarithmic operation on the probability density value, and take the opposite of the natural logarithmic operation result to obtain the Bayesian surprise value.

[0120] S63. Calculate the square of the difference between the model's predicted mean and the measured temperature data. Sum the squared results and divide by the sample size to construct the basic mean square error loss function. Set the maximum surprise threshold to 5. Divide the Bayesian surprise value by 5 to obtain the quotient. Add the quotient to the constant 1 to obtain the dynamic weighting coefficient. Multiply the dynamic weighting coefficient by the basic mean square error loss function value to construct the dynamic weighting loss function.

[0121] S64. Calculate the partial derivative of the dynamic weighted loss function with respect to the weight parameters in the improved UNETR model to obtain the gradient value. Set the learning rate to 0.001 and the momentum parameter to 0.9. Calculate the new weight parameters by subtracting the product of the learning rate and the gradient value from the weight parameters and adding the product of the momentum parameter and the previous gradient update amount. Repeat the update steps until the value of the dynamic weighted loss function is less than 0.0001, and output the corrected improved UNETR model.

[0122] The incremental dynamic update learning process proposed in this step is similar to the traditional deep learning static training method in that it is based on the backpropagation theory of deep neural networks. That is, the error between the predicted output and the true label is calculated through forward propagation, the model bias is quantified by the loss function, and the gradient descent algorithm is used to optimize the network weight parameters to minimize the prediction error.

[0123] The difference lies in that this invention breaks away from the limitations of traditional methods that rely solely on fixed-distribution datasets for offline training or ignore model prediction uncertainties. Building upon the traditional model's direct calculation of loss using mean squared error, this invention adds a physical drilling data fusion step. This maps measured borehole temperatures and formation parameters to grid features and constructs an incremental sample set, rather than solely using historical simulation data. In the loss calculation step, a Gaussian prior distribution is constructed using the predicted mean and variance, and the Bayesian surprise factor is calculated. This surprise factor is then transformed into a dynamically weighted coefficient and fused with the basic loss function, rather than relying solely on mean squared error calculation. Finally, in the parameter update step, the UNETR model weights are iteratively corrected based on gradient descent with momentum, jointly outputting a corrected model adapted to the new geological conditions, rather than using the initial model with frozen parameters.

[0124] The beneficial effects of this improvement are that, through incremental updates driven by physical drilling data and dynamic weighting based on Bayesian surprise, the time-varying characteristics of frozen field evolution and the uncertainty of model cognition can be integrated into the loss function calculation. This breaks through the limitation of traditional methods where rigid model parameters lead to prediction failure when geological conditions change abruptly, and achieves accurate approximation from offline static modeling to online dynamic evolution. This design significantly enhances the model's real-time perception of the thermodynamic response of complex formations in deep well freezing projects, and can more accurately correct systematic biases in temperature field prediction. The surprise-based adaptive loss adjustment mechanism effectively improves the model's sensitivity to new data and training efficiency, and enhances the system's robustness and prediction accuracy in the face of geological uncertainties during long-term freezing.

[0125] In this embodiment, S7 specifically includes:

[0126] S71. Read the three-dimensional mesh data output by the corrected and improved UNETR model, extract the spatial three-dimensional coordinates and temperature field prediction values ​​of each mesh node, set the freezing temperature threshold to 0 degrees Celsius, mark the mesh nodes with temperature field prediction values ​​less than 0 degrees Celsius as frozen soil units, count the spatial positions of all mesh nodes marked as frozen soil units, combine the spatial position coordinates into a set, generate a set of spatial discrete points of frozen soil, and at the same time extract the coordinate information of the frozen holes in the freezing hole layout diagram, calculate the straight-line distance between each pair of frozen holes as the hole spacing.

[0127] S72. Read the spatial discrete point set of frozen soil, extract the spatial distance difference between discrete points and the numerical difference of frozen soil thickness, calculate the product of the square of the spatial distance difference and the square of the thickness difference, divide the product by 2 to obtain the semivariogram value, set the distance grouping tolerance to 0.5 meters, group the discrete point pairs according to spatial distance and calculate the average semivariogram value of each group, use the spherical model to fit the average semivariogram value and the distance to obtain the spatial correlation model;

[0128] S73. Using the spatial correlation model, with the non-observed grid node as the center, search for observed frozen soil units within a 10-meter radius as neighborhood samples. Calculate the covariance between the neighborhood samples and the central node based on the spatial correlation model. Construct a Kriging equation system using the covariance values, solve the equation system to obtain the weight coefficients of the neighborhood samples, multiply the weight coefficients by the frozen soil thickness values ​​of the neighborhood samples, and sum them to obtain the thickness estimate of the grid node. Calculate the estimated variance using the weight coefficients and covariance values. Read the corresponding brine temperature, borehole spacing, excavation face temperature, and double-row additional temperature at that location, and substitute them into the formula:

[0129] ;

[0130] Where E is the thickness estimate, l is the hole spacing, For the temperature of the brine, Additional temperature for double rows, To determine the temperature of the excavation face, the theoretical average temperature is calculated, and full-field thickness distribution data including thickness estimates, estimated variance, and theoretical average temperature is generated.

[0131] S74. Read the full-field thickness distribution data, calculate the estimated variance, and subtract the estimated variance from 1 to obtain the spatial confidence index. Set the design safety thickness threshold to 4 meters. Identify areas with thickness estimates less than 4 meters and spatial confidence index less than 0.8 as thin-thickness areas. Calculate the absolute value of the difference between the theoretical average temperature and the predicted temperature field value as the temperature deviation. Set the deviation threshold to 1 degree Celsius. Issue a thickness anomaly warning when the temperature deviation in the thin-thickness area is greater than 1 degree Celsius. Statistically calculate the thickness estimates of all nodes in the thin-thickness area, sum the thickness estimates and divide by the total number of nodes to obtain the average thickness value. Extract the minimum value among the thickness estimates as the minimum thickness value. When the average thickness value is greater than 4 meters and the minimum thickness value is greater than 3.5 meters, it is determined that the frozen wall meets the sealing and stability requirements, and the frozen wall thickness evaluation result is output.

[0132] Example 1: To verify the feasibility of this invention in the construction monitoring of underground frozen engineering projects, the method of this invention was applied to the intelligent monitoring system of the frozen engineering project of the shield tunnel connecting passage of a certain city's subway construction group company (hereinafter referred to as "Company M"). In traditional frozen wall thickness assessment systems, linear interpolation or simplified heat conduction formula inversion based on sparse temperature measurement well data is usually used. These methods not only make it difficult to accurately analyze the nonlinear evolution of the freezing front under complex hydrogeological conditions, but also cannot quantify the spatial uncertainty of the prediction results, which can easily lead to misjudgment of the frozen wall junction state or missed detection of weak areas. To solve the above problems, Company M decided to adopt the online frozen wall thickness assessment method based on deep learning proposed in this invention.

[0133] During implementation, Company M first used a high-precision total station to acquire the spatial three-dimensional coordinates of the temperature measuring holes and freezing holes, as well as the geometric parameters of the borehole axis, and constructed a spatial distribution model of the boreholes. Simultaneously, the zero-degree duration was mapped to the freezing temperature field evolution curve, and a freezing column geometric model was generated by combining the radial radius and axis geometric parameters. A theoretical intersection base map was generated through Boolean union operations, and the theoretical extreme point set was extracted. Subsequently, the temperature readings of the temperature measuring holes were acquired, mapped to grid nodes, and the temperature gradient vector was calculated. The freezing field data was then decomposed into an irrotational potential energy field vector and a divergent flow field vector by solving the Poisson equation. A global information entropy matrix was constructed by combining the Laplace operator gradient features, serving as a high-dimensional input feature for improving the UNETR model.

[0134] Company M improved the MH random walk attention encoding layer in the UNETR model, replacing the traditional multi-head self-attention mechanism. Utilizing the Metropolis-Hastins criterion, it performed sparse neighborhood sampling guided by physical distance and gradient magnitude, effectively capturing long-distance heat conduction dependence. Simultaneously, it introduced an XOR logic constraint mechanism, performing logical bit operations between the global information entropy matrix and the theoretical extreme point set to construct a hybrid objective function space incorporating geometric constraints on the spacing between frozen boreholes. In the core evaluation phase, this invention constructed a risk spatiotemporal evolution model based on the Hawkes process, calculated spatiotemporal excitation effects, and combined information gain to drive Monte Carlo tree search, outputting the optimal risk node coordinates. The system calculated the geological structure trend based on the risk prediction center point, generated a borehole coordinate sequence, and performed physical drilling to obtain measured temperature and formation parameters. Finally, it calculated the Bayesian surprise of the measured data and prior predictions, constructed a dynamic weighted loss function to correct the model weights, and used geostatistical methods combined with variance estimation to output the final frozen wall thickness assessment results and uncertainty quantification indicators.

[0135] During implementation, the technical team at Company M discovered that, compared to traditional manual interpolation and numerical simulation methods, the method of this invention significantly improves the accuracy and adaptability of frozen wall thickness assessment. Traditional methods cannot effectively utilize deep physical field characteristics and exhibit lag in response to stratigraphic anomalies. In contrast, the method of this invention, through physical field decomposition, proactive risk search, and Bayesian surprise factor closed-loop correction, effectively achieves refined extraction of frozen front boundaries and proactive quantitative early warning of risks.

[0136] To further verify the actual performance of the method of the present invention, Company M conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1:

[0137] Table 1. Performance Comparison of Frozen Wall Thickness Evaluation for Connecting Passages in Company M Shield Tunnel Section

[0138] index Traditional methods Method of the present invention Increase Average error in predicted frozen wall thickness (mm) 210 35 -83.3% Freezing front positioning deviation (mm) 180 25 -86.1% Abnormal area identification false negative rate (%) 15.2 1.5 -90.1% Time taken for a single evaluation data iteration (minutes) 45 12 -73.3% Lead time for risk and hazard prediction (days) 2 7 +250.0% The number of new samples required for model correction (in terms of samples) 50 8 -84.0% Formation parameter abrupt change adaptation time (hours) 24 4 -83.3% Confidence level of uncertainty in the assessment results (%) 65 92 +41.5% Accident rate in engineering projects (%) 3.5 0 -100.0%

[0139] As shown in Table 1, the performance of the frozen wall thickness assessment system was comprehensively improved after applying the method of this invention. The average error in frozen wall thickness prediction was reduced from 210 mm to 35 mm using traditional methods, and the positioning deviation of the frozen front was reduced from 180 mm to 25 mm, significantly improving the accuracy of state perception and providing a reliable basis for construction decisions. The missed detection rate of abnormal areas was reduced from 15.2% to 1.5%, effectively avoiding the risk of water seepage. The lead time for risk and hazard prediction was significantly extended from 2 days to 7 days, significantly enhancing the system's timeliness. In addition, the number of new samples required for model correction was reduced from 50 to 8, and the adaptation time for abrupt changes in formation parameters was shortened from 24 hours to 4 hours, significantly improving the efficiency of online assessment. The confidence level of the assessment results uncertainty was also significantly improved, from 65% to 92%, the number of invalid boreholes was reduced by 45.0%, and the incidence of engineering safety accidents was reduced to 0, achieving zero-accident operation.

[0140] Through the method of this invention, Company M has successfully achieved refined perception and proactive quantitative early warning of risks in the formation process of frozen walls, effectively avoiding safety hazards in the construction of the freezing method, ensuring the smooth connection of the subway connecting passage, significantly improving the intelligence and digitalization level of frozen engineering monitoring, significantly reducing the data analysis burden of technical personnel, enhancing the stability and robustness of the monitoring system, and providing strong technical support for the construction of smart underground engineering.

[0141] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based online method for evaluating the thickness of frozen walls, characterized in that, Includes the following steps: S1. Obtain the topological coordinates and zero-degree duration of the temperature measuring hole and the freezing hole, construct the geometric model of the freezing column and generate the theoretical intersection base map, and extract the theoretical extreme point set; S2. Based on the theoretical intersection base map, the frozen field data is decomposed into an irrotational potential energy field vector and a divergence-free flow field vector and mapped to the grid nodes. The global information entropy matrix is ​​constructed using the field numerical distribution. Combined with the Laplace operator gradient characteristics, an improved UNETR model is constructed. S3. Introduce the XOR logic constraint mechanism to perform logical bit operations on the global information entropy matrix and the theoretical extreme point set to identify abnormal regions, construct a logical mask screening search space, and integrate the geometric constraint of the freezing hole spacing to construct a hybrid objective function space. S4. Construct a risk spatiotemporal evolution model based on Hawkes process in the mixed objective function space, calculate the spatiotemporal evoked effect and combine it with information gain to obtain the comprehensive reward value, drive the Monte Carlo tree search expansion and convergence, and output the optimal risk node coordinates as the risk prediction center point. S5. Starting from the risk prediction center point, calculate the geological structure trend based on the visit frequency gradient of Monte Carlo tree backpropagation, generate borehole coordinate sequence along the trend direction, and perform physical drilling to obtain measured temperature and formation parameters. S6. Map the measured temperature and formation parameters back to the frozen field state vector to form incremental training samples, calculate the Bayesian surprise of the measured data and prior predictions, construct a dynamic weighted loss function, optimize the network weight parameters, and correct and improve the UNETR model. S7. Based on the modified UNETR model, spatial distribution characteristics are extracted. Geostatistical methods are used to model the spatial correlation of frozen soil thickness and make optimal unbiased estimation. Uncertainty quantification index is constructed by combining the estimation variance, and the frozen wall thickness assessment results are output.

2. The online evaluation method for frozen wall thickness based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Obtain the three-dimensional spatial coordinates of the temperature measuring hole and the freezing hole, as well as the geometric parameters of the borehole axis, and construct a borehole spatial distribution model. S12. Based on the borehole spatial distribution model, the zero-degree duration is mapped to the freezing temperature field evolution curve to obtain the radial radius of the freezing column as time increases, and the geometric model of the freezing column is generated by combining the radial radius and the axis geometric parameters. S13. Perform a Boolean union operation on the frozen column geometric model to generate the frozen intersection region, and cut the bottom surface along the vertical direction to generate the theoretical intersection base map; S14. Based on the boundary contour of the theoretical intersection base map, extract the coordinate extreme points and generate the theoretical extreme point set.

3. The online evaluation method for frozen wall thickness based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Map the temperature readings of the temperature measuring holes to the theoretical intersecting base map grid, calculate the three-dimensional temperature gradient vector of each node, and combine them to generate the initial state vector field of the freezing field. S22. Calculate divergence and curl based on vector field components. Solve the Poisson equations for scalar potential and vector potential respectively with divergence and curl as source terms. Obtain the irrotational potential energy field vector by calculating the gradient of the scalar potential and the irrotational flow field vector by calculating the curl of the vector potential. S23. Calculate the Laplace operator magnitude of the irrotational potential energy field vector and the divergence-free flow field vector. Weight the Laplace operator magnitude with the irrotational and divergence-free magnitudes to obtain the local joint characteristic. Normalize the local joint characteristic to obtain the node energy distribution probability. Calculate the local information entropy value of each node based on the node energy distribution probability. Arrange the nodes according to their grid positions to generate the global information entropy matrix. S24. The irrotational potential field vector, the divergence-free flow field vector, the Laplace operator gradient feature and the global information entropy matrix are concatenated to construct a high-dimensional vector feature that includes physical decomposition features and uncertainty measures, and an improved UNETR model is constructed.

4. The online evaluation method for frozen wall thickness based on deep learning according to claim 1, characterized in that, The improved UNETR model includes a 3D voxel embedding layer, an MH random walk attention encoding layer, a multi-scale feature aggregation layer, a 3D voxel decoding layer, a Bayesian uncertainty estimation head, and a frozen front segmentation layer. The 3D voxel embedding layer is used to perform block embedding processing on the high-dimensional vector features after splicing the irrotational potential energy field vector, the divergence-free flow field vector, the Laplacian operator gradient features, and the global information entropy matrix. The three-dimensional voxel grid is divided into image blocks of fixed size. An initial voxel embedding vector sequence is generated through linear projection mapping, and three-dimensional position encoding is added to output the initial voxel embedding vector sequence containing spatial geometric structure information. The MH random walk attention coding layer is used to receive the initial voxel embedding vector sequence, replacing the traditional multi-head self-attention mechanism. It constructs a sparse adjacency graph through random sampling to capture long-distance heat conduction dependencies and outputs a sparse global context feature map. Specifically, it includes a proposal distribution construction module, an acceptance probability calculation module, and a feature aggregation module. The proposed distribution construction module is used to parse the three-dimensional spatial coordinates and Laplacian gradient features from the initial voxel embedding vector sequence, define candidate neighborhoods based on physical distance with the current voxel node as the center, calculate the probability density of the current node in combination with the gradient magnitude, and perform random walks from the current node to the nodes in the neighborhood. The acceptance probability calculation module is used to calculate the probability density of the target node based on the irrotational potential energy modulus and the global information entropy, and to calculate the ratio of the probability density of the target node to the probability density of the current node. The acceptance rate is calculated in combination with the Metropolis-Hastins criterion. If the acceptance rate is greater than the generated random threshold, the corresponding wandering node is accepted as a relevant neighbor; otherwise, it is rejected. The feature aggregation module is used to collect all accepted nodes to form the most relevant node set after walking through a preset number of steps, and to perform attention weight calculation and feature update based on the current node to query the nodes in the most relevant node set, and output a sparse global context feature map. The multi-scale feature aggregation layer is used to receive intermediate layer features from different stages of the MH random walk attention coding layer, and to perform skip connections and fusion of local detail features and global walk features under different walk radii to reconstruct an enhanced feature map that preserves spatial details. The 3D voxel decoding layer is used to receive the enhanced feature map and perform progressive upsampling operations to restore the spatial resolution of the feature map; combined with the intermediate layer features of different stages of the MH random walk attention coding layer, the features are refined step by step through convolution and upsampling operations to generate a high-resolution predicted feature map with the same size as the input. The Bayesian uncertainty estimation head is used to receive high-resolution predicted feature maps, generate the mean parameter of the frozen temperature field via a linear mapping of a parallel first convolutional branch, and generate the log-variance parameter of the frozen temperature field via a linear mapping of a second convolutional branch. The freezing front segmentation layer is used to receive high-resolution predicted feature maps, fuse the mean parameter and the logarithmic variance parameter, perform semantic segmentation on the fused features, and output the probability value of each grid node belonging to the permafrost region. Based on the probability value and the result of the mean parameter after being verified by the freezing temperature threshold, the thickness distribution boundary of the frozen wall is jointly extracted, and the final frozen field state vector is generated by combining the segmented feature mapping.

5. The online evaluation method for frozen wall thickness based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Read the global information entropy matrix and the theoretical extreme point set, map the theoretical extreme point set to the grid coordinates of the global information entropy matrix, calculate the binary state difference at the mapped coordinate position through logical XOR operation, identify abnormal regions based on the difference results and generate an initial difference mask. S32. Based on the initial difference mask, lock the mesh search space, obtain the spatial three-dimensional coordinates of the freezing holes, and calculate the Euclidean distance between each pair of freezing holes to generate the freezing hole spacing matrix. S33. Set the distance weight coefficient according to the freezing hole spacing matrix, map the distance weight coefficient to the grid search space, and construct a hybrid objective function space that integrates spatial geometric constraints.

6. The online evaluation method for frozen wall thickness based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Read the grid nodes in the mixed objective function space as candidate nodes and historical anomaly information, calculate the time interval and Euclidean distance between candidate nodes and historical anomalies, calculate the exponent using the time decay parameter and spatial decay parameter, perform exponentiation on the natural constant to obtain the excitation intensity value, and sum them up to obtain the spatiotemporal influence value. S42. Extract the initial entropy value of the candidate node from the global information entropy matrix, calculate the average entropy value of the neighborhood within the preset neighborhood range, take the absolute value of the difference between the initial entropy value and the average entropy value of the neighborhood as the information gain, multiply the spatiotemporal influence value and the information gain by the corresponding weight coefficients respectively and add them together to obtain the comprehensive reward value. S43. Construct a Monte Carlo tree structure, use the confidence interval upper bound algorithm to select child nodes for expansion, sample the expanded nodes and calculate the comprehensive reward value, backpropagate to update the node parameters, and after iterating to a preset number of times, output the node with the highest average reward value as the risk prediction center point.

7. The online evaluation method for frozen wall thickness based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Using the risk prediction center point as the reference point, extract the number of visits to each child node in the neighborhood of the current node in the Monte Carlo tree, and construct the number of visits distribution matrix. S52. Calculate the directional derivatives of the visit frequency distribution matrix in the X, Y and Z axes of the three-dimensional spatial coordinate system, and combine the directional derivatives to obtain the visit frequency gradient vector. S53. Normalize the gradient vector of the number of visits, and set the direction of the normalized vector to the direction of the geological structure trend. S54. Using the risk prediction center point as the starting point of the sequence, and following the geological structure trend direction, perform ray interpolation in combination with the preset borehole spacing to generate a borehole coordinate sequence. S55. Perform physical drilling operations based on the borehole coordinate sequence, and collect measured temperature data and formation parameters during the drilling process.

8. The online evaluation method for frozen wall thickness based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Obtain the coordinates, measured temperature, and formation parameters of the physical drilling borehole. Convert the coordinates into grid indices of the frozen field state vector. Assign the measured temperature and formation parameters to the vector features of the corresponding grid indices. Then, concatenate the assigned vectors with the data in the historical sample database in chronological order to construct an incremental training sample set. S62. Input the incremental training sample set into the improved UNETR model, calculate the prediction mean and prediction variance of each node in the output layer through forward propagation, construct a Gaussian distribution using the prediction mean and prediction variance as the prior prediction distribution, calculate the probability density value of the measured data in the prior prediction distribution, take the natural logarithm of the probability density value and take the opposite number to obtain the Bayesian surprise value. S63. Construct the basic mean square error loss function, obtain the preset maximum surprise threshold, divide the Bayesian surprise value by the maximum surprise threshold, add the calculation result to the constant 1 to obtain the dynamic weighting coefficient, multiply the dynamic weighting coefficient by the basic mean square error loss function to construct the dynamic weighting loss function. S64. Calculate the gradient of the dynamic weighted loss function with respect to the network weight parameters, set the learning rate and momentum parameters, and use the gradient descent algorithm to iteratively update the network weight parameters until the preset convergence condition is met, and output the corrected improved UNETR model.

9. The online evaluation method for frozen wall thickness based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Based on the modified and improved UNETR model, extract the spatial three-dimensional coordinates and temperature field prediction values ​​of the frozen soil unit to generate a spatial discrete point set of frozen soil. S72. Using the spatial discrete point set of permafrost as a regional variable, geostatistical methods are used to calculate the semivariogram, and a variogram model is fitted to analyze the variability and continuity of permafrost thickness in the spatial direction, and a spatial correlation model is constructed. S73. Based on the spatial correlation model, the Kriging interpolation algorithm is used to make the optimal unbiased estimate of the frozen soil thickness at non-observation locations. The thickness estimate and corresponding estimation variance of the grid nodes are calculated to generate full-field thickness distribution data containing deterministic prediction and uncertainty measurement. S74. Convert the estimated variance into a spatial confidence index, identify areas with thinness, and quantitatively assess the sealing and stability of the frozen wall, outputting the assessment results.

Images