Health management system for real-time safety state of foundation pit supporting structure

Abstract

The invention relates to the technical field of image analysis, in particular to a health management system for the real-time safety state of a foundation pit supporting structure, and the system comprises a three-dimensional scene point cloud construction module which synchronously collects left and right camera images and preset camera parameters, calculates the parallax value of each pixel, and transmits the parallax value to a display module; and according to the parallax value of each pixel, the pixel coordinate and the internal and external parameters of the camera. According to the method, pixel-level parallax extraction and camera parameters are combined, the three-dimensional space coordinates of the underground diaphragm wall are calculated, spatial reconstruction of high-density point locations on the surface of the structure is achieved, a reference form capable of being tracked quantitatively is constructed, and corresponding point iterative registration and normal offset analysis are introduced, so that the accuracy of the method is improved. A real deformation track of a wall body can be extracted from tiny deformation, a deviation value set which is highly coupled with a structure state is constructed, a pixel point extraction strategy aiming at anchor rod wall construction characteristics is adopted at a node level, and real space coordinates of constructed nodes are locked.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image analysis, and particularly relates to a health management system for real-time safety state of a foundation pit supporting structure. BACKGROUND

[0002] A foundation pit is a soil pit excavated according to a foundation design position, a foundation elevation and a foundation plane size. In the process of excavating a foundation pit, if the soil of a foundation pit wall is unstable and is affected by underground water, or is limited by a construction site or adjacent buildings and cannot be excavated by using a slope, a supporting structure needs to be used for the foundation pit. Foundation pit supporting is a supporting, reinforcing and protecting measure used for a foundation pit side wall and a surrounding environment to ensure the safety of underground structure construction and the surrounding environment.

[0003] The prior art relies on artificial control of measuring points or traditional structural parameter monitoring means, and usually judges the stability of a foundation pit supporting structure by setting component sensors such as stress gauges and displacement gauges. This method needs to be pre-constructed in the operation process, is limited by the sparseness of the number and position layout of the control points, is prone to have a detection blind area, and is difficult to comprehensively cover the entire structure surface. In a complex construction environment, for example, in the presence of adjacent building disturbance or underground water infiltration, point data is easily disturbed by accidental errors, resulting in distortion of the overall safety evaluation result. Therefore, improvement is needed. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art and propose a health management system for real-time safety state of a foundation pit supporting structure.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: a health management system for real-time safety state of a foundation pit supporting structure comprises:

[0006] A three-dimensional scene point cloud construction module, the left and right camera images synchronously collected and preset camera parameters are used to calculate the parallax value of each pixel, and the three-dimensional space coordinates of each pixel point on the surface of the underground continuous wall are solved according to the parallax value, pixel coordinates and camera internal and external parameters of each pixel, and a reference wall body space point set is established;

[0007] A wall body deformation deviation calculation module, a real-time wall body space point set is obtained through real-time images, the distance between corresponding points in the real-time wall body space point set and the reference wall body space point set is minimized through iterative operation, the pose registration transformation parameter is solved, the real-time wall body space point set is corrected by using the pose registration transformation parameter, the normal distance of each point in the corrected point set to the surface formed by the reference wall body space point set is calculated, and a surrounding structure surface normal distance set is generated;

[0008] The key node three-dimensional positioning module detects the pixel points of the intersection of the cross beam and the column of the lattice anchor rod wall and the anchor rod head position in the real-time image, solves the three-dimensional coordinates, obtains the three-dimensional coordinates of the structural feature points, traverses each point in the three-dimensional coordinates of the structural feature points, finds the distance value of the nearest neighbor point, and generates a topology node set with a deformation variable;

[0009] The topology network deformation energy calculation module calculates the length of the connection between the nodes and the included angle at the node according to the preset connection relationship, obtains the network geometric parameters, and calculates the structural distortion energy quantization value based on the network geometric parameters and the topology node set with a deformation variable.

[0010] Preferably, the acquisition step of the reference wall space point set is:

[0011] The left and right camera images synchronously collected and the preset camera parameters are used to perform absolute difference accumulation on each pixel block in the epipolar direction and aggregate in multiple directions, record the minimum matching cost index of each pixel point, and determine the uniqueness validity by a threshold value, to obtain discrete minimum disparity indexes;

[0012] According to the discrete minimum disparity index, the sub-pixel disparity value of the pixel point is calculated;

[0013] According to the sub-pixel disparity value, the three-dimensional space coordinates of each pixel point on the underground continuous wall surface are solved by combining the pixel coordinates and the focal length, the principal point coordinates, the baseline length, the rotation matrix and the displacement vector of the camera internal and external parameters, and all three-dimensional space coordinates are aggregated to generate the reference wall space point set.

[0014] Preferably, the acquisition step of the pose registration transformation parameter is:

[0015] The two-dimensional coordinates of all pixel points on the wall surface are extracted from the real-time image, the two-dimensional coordinates of each pixel point are converted into three-dimensional space coordinates by calling the camera internal and external parameters, and the real-time wall space point set is formed by point-by-point accumulation;

[0016] According to the real-time wall space point set, the corresponding points in the reference wall space point set are matched point by point, the Euclidean distance between the two groups of points is calculated round by round in an iterative manner and the minimum distance is taken, the distance sum of all point pairs is taken as the iterative target value, the rotation matrix and the translation vector are updated until the iterative target value converges, and the pose registration transformation parameter is obtained.

[0017] Preferably, the acquisition step of the enclosure surface normal distance set is:

[0018] The pose registration transformation parameter is used to perform overall coordinate system correction on the three-dimensional coordinates of the real-time wall space point set, calculate the normal distance from the corrected coordinates to the surface of the reference wall space point set point by point, and record to generate the enclosure surface normal distance set.

[0019] Preferably, the step of obtaining the three-dimensional coordinates of the structural feature points is:

[0020] In the real-time image, the pixel points of the intersection of the horizontal beams and the vertical columns of the lattice anchor wall and the anchor head position are located frame by frame, the pixel row and column coordinates are extracted, the camera internal and external parameters are read, the pixel row and column coordinates are projected to the imaging plane after distortion correction and principal point translation correction, the spatial intersection position is calculated combined with the base line and the attitude, and the three-dimensional coordinates of the structural feature points are obtained.

[0021] Preferably, the step of obtaining the topology node set with deformation variables is:

[0022] According to the three-dimensional coordinates of the structural feature points, a spatial mapping is established with the three-dimensional coordinates of the corrected points in the surface normal distance set of the enclosure as the retrieval object, a first matching point is retrieved from small to large in the Euclidean distance for each three-dimensional coordinate of the structural feature points, the normal distance recorded in the matching point is read, and a corresponding record of the three-dimensional coordinates of the structural feature points and the normal distance is generated.

[0023] According to the corresponding record of the three-dimensional coordinates of the structural feature points and the normal distance, a topology node entry is created for each three-dimensional coordinate of the structural feature points, and the read normal distance is written as a deformation variable, all topology node entries are summarized and the one-to-one correspondence between the coordinates and the deformation variables is maintained, and a topology node set with deformation variables is generated.

[0024] Preferably, the step of obtaining the network geometric parameters is:

[0025] The topology node set with deformation variables and the initial topology node set are traversed one by one for each line of the preset connection relationship, the current three-dimensional coordinates and the initial three-dimensional coordinates at both ends of the line are extracted, the current line length and the initial line length are calculated according to the Euclidean distance, and the current included angle and the initial included angle are calculated by the inner product of the adjacent line vectors, and the network geometric parameters are obtained.

[0026] Preferably, the step of obtaining the structural distortion energy quantization value is:

[0027] According to the network geometric parameters, the length change rate of each line is calculated, the difference between the current line length and the initial line length is divided by the initial line length to obtain the relative length change rate of each line, and the relative angle change rate of each node is calculated, the difference between the current included angle and the initial included angle is divided by the initial included angle to obtain the dimensionless angle change rate of each node.

[0028] The relative length change rate and the relative angle change rate are used to calculate the structural distortion energy quantization value.

[0029] Compared with the prior art, the application has the advantages and positive effects that:

[0030] In the application, by combining pixel-level disparity extraction with camera parameters, the three-dimensional space coordinates of the underground continuous wall are calculated, the spatial reconstruction of the high-density point position of the structure surface is realized, the quantifiable tracking reference form is constructed, the corresponding point iterative registration and normal offset analysis are introduced, the real deformation trajectory of the wall is extracted from the micro deformation, the deviation amount set highly coupled with the structure state is constructed, at the node level, the pixel point extraction strategy for the construction characteristics of the anchor rod wall is adopted to lock the real space coordinates of the construction node, and the deformation amount of each key node is connected based on the normal distance, so that the deformation data of the structure feature points have physical meaning, the topological geometric parameters such as the length of the connecting line and the included angle are constructed, on the basis of introducing the normalization processing of the node scale and the angle change rate, the local deformation and the overall structure response are uniformly modeled in the energy scale, so that the quantitative evaluation of the structure distortion risk is completed, and thus the chain from image extraction, spatial analysis, node identification to distortion quantization is constructed, the fuzzy disconnection between deformation identification and risk judgment is eliminated, and the real-time of structure safety state perception is enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 The system flowchart of the application. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical scheme and advantages of the application more clear, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0033] Please refer to Figure 1 The application provides a technical scheme: a health management system for real-time safety state of a foundation support structure includes:

[0034] A three-dimensional scene point cloud construction module, the left and right camera images synchronously collected and the preset camera parameters are used to calculate the disparity value of each pixel, and the three-dimensional space coordinates of each pixel point on the surface of the underground continuous wall are solved according to the disparity value, pixel coordinates and camera internal and external parameters of each pixel, and the reference wall space point set is established;

[0035] A wall deformation deviation calculation module, the real-time wall space point set is obtained through real-time image acquisition, the distance between the corresponding points in the real-time wall space point set and the reference wall space point set is minimized through iterative operation, the pose registration transformation parameter is solved, the real-time wall space point set is corrected by using the pose registration transformation parameter, the normal distance of each point in the corrected point set to the surface formed by the reference wall space point set is calculated, and the surface normal distance set of the enclosing structure is generated;

[0036] The key node three-dimensional positioning module detects the pixel points of the intersection of the cross beam and the column of the lattice anchor rod wall and the anchor rod head position in the real-time image, solves the three-dimensional coordinates, obtains the three-dimensional coordinates of the structural feature points, traverses each point in the three-dimensional coordinates of the structural feature points, finds the distance value of the nearest neighbor point, and generates a topology node set with a deformation variable;

[0037] The topology network deformation energy calculation module calculates the length of the connection line between nodes and the included angle at the node according to the initial topology node set and the preset connection relationship, obtains the network geometric parameters, and calculates the structural distortion energy quantization value according to the network geometric parameters and the topology node set with the deformation variable.

[0038] The acquisition step of the reference wall space point set is:

[0039] The left and right camera images and the preset camera parameters are synchronously collected, the absolute difference of each pixel block is accumulated in the epipolar direction and aggregated in multiple directions, the minimum matching cost index of each pixel point is recorded, and the uniqueness validity is determined by a threshold, and the discrete minimum disparity index is obtained;

[0040] According to the discrete minimum disparity index, the sub-pixel disparity value of the pixel point is calculated, and the calculation formula is:

[0041]

[0042] Wherein, d x，y is the sub-pixel disparity value of the pixel (x, y), d0 is the discrete minimum disparity index of the pixel (x, y), C x，y (d) is the multi-directional pixel block matching cost of the pixel (x, y) under the candidate disparity d, G x，y is the normalized sum of the gradient amplitude in the matching block of the pixel (x, y), λ is a regularization coefficient, and ∈ is a very small constant to prevent division by zero.

[0043] According to the sub-pixel disparity value, the three-dimensional space coordinates of each pixel point on the underground continuous wall surface are solved in combination with the pixel coordinates and the focal length, principal point coordinates, baseline length, rotation matrix and displacement vector of the camera internal and external parameters, and all three-dimensional space coordinates are aggregated to generate the reference wall space point set.

[0044] Specifically, according to the left and right camera images synchronously collected and the preset camera parameters, first, the hardware trigger is used to ensure that the left and right industrial cameras are exposed at the same time, for example, the collection frequency is set to 15 frames per second, and the time stamp is strictly corresponding to the left and right image pairs, then the epipolar correction is performed on each image pair, the preset camera internal and external parameters are used to reproject the image, so that the search range of the corresponding points in the left and right images is reduced from a two-dimensional plane to the same horizontal scanning line, that is, the epipolar direction, then for each pixel in the left image, a fixed size pixel block around it is selected, for example, a 9*9 window, along all possible pixel positions on the corresponding epipolar line of the right image, that is, the candidate disparity value range, the 9*9 window is moved, at each candidate position, the absolute value of the gray value difference between the corresponding pixels of the left image block and the right image block is calculated one by one, and all the 81 absolute difference values are added up to obtain the matching cost under the candidate disparity, that is, the absolute difference accumulation value (SAD), in order to overcome the problem of low matching accuracy in weak texture area of the SAD method, a multi-direction path aggregation strategy is introduced, specifically, for the initial matching cost of each pixel under each candidate disparity, dynamic programming type cost aggregation is performed along 16 preset directions (for example, horizontal, vertical, diagonal, etc.), the cost is propagated from the previous pixel point to the current pixel point, and a penalty term is added during the propagation, the penalty term is related to the disparity change amount of the previous and next two pixels, when the disparity change is 1 pixel, a smaller penalty value P1 is applied, and when the disparity change is greater than 1 pixel, a larger penalty value P2 is applied, the total cost aggregated in the 16 directions is taken as the final matching cost of the pixel under the disparity, after traversing all candidate disparities, the disparity value that causes the minimum final matching cost is recorded as the minimum matching cost index of the pixel, finally, in order to eliminate the wrong matching points, the uniqueness validity judgment is performed, the threshold of the judgment is calibrated according to the light and texture complexity of the test scene, the calibration process is carried out in a calibration board scene with a known three-dimensional structure, images are collected and disparity is calculated, the uniqueness threshold is adjusted to minimize the root mean square error between the calculated three-dimensional structure and the real structure, usually, when the ratio of the minimum matching cost value to the second minimum matching cost value is less than 0.8, the current matching is considered to be valid, for example, the minimum matching cost of a certain pixel is 1500, the second minimum matching cost is 1800, the ratio is 1500 / 1800≈0.83, which is greater than 0.8, and the matching is determined to be invalid, if the second minimum matching cost is 2000, the ratio is 1500 / 2000=0.75, which is less than 0.8, and the matching is determined to be valid, and the minimum matching cost index is recorded, the process is repeated for all pixels to obtain the discrete minimum disparity index.

[0045] Formula: The above formula is to introduce a regularization term based on image gradient, the traditional quadratic fitting method performs well in the area with rich image texture details, but in the weak texture or repetitive texture area, the matching cost curve shape is poor, leading to the extreme point of fitting deviating from the real disparity, resulting in large error, and the regularization term uses the gradient information G within the matching block where the pixel is located x，y When the regional texture is sparse, the gradient and G x，y is small, the value of the regularization term is large, the contribution to the denominator is increased, thereby inhibiting the excessive correction amount that may be caused by noise, making the sub-pixel disparity value tend to the discrete integer disparity value d0, enhancing the stability in the weak texture area, and in the rich texture area, the gradient and G x，y is large, the value of the regularization term is close to zero, the formula degenerates into the standard quadratic fitting, retaining the high precision characteristics in the high signal-to-noise ratio area.

[0046] d0 is the discrete minimum disparity index of the pixel (x, y), representing the best matching disparity value found at the integer level, which is the basis and center point for subsequent sub-pixel accuracy calculation, the value is obtained through the cost aggregation and minimum cost search process of the previous step, specifically, for each pixel, the matching cost under a series of discrete disparity candidates (for example, 0 to 128 pixels) is calculated, and the integer disparity value that makes the aggregated matching cost minimum is determined, this integer value is d0, for example, when matching the pixel (100, 250), it is found that when the disparity is 55 pixels, the aggregated matching cost is the lowest 1350 among all candidates, then the d0 of the pixel is recorded as 55, which indicates that in the right image, the most matching area with the pixel at (100, 250) in the left image is located near (100-55, 250) i.e. (45, 250), and the subsequent sub-pixel calculation will be around the integer point d0=55, using the cost value of its adjacent disparity for interpolation fitting.

[0047] C x，y (d) is the multi-directional pixel block matching cost of the pixel (x, y) under the candidate disparity d, which is a quantitative indicator reflecting the similarity of the corresponding pixel blocks in the left and right images under a given disparity, the smaller the cost, the higher the similarity, the calculation process is completed in the previous step, which integrates the initial pixel gray or color difference (such as SAD) and the smoothing constraint information propagated and aggregated from multiple directions, this cost is not a simple difference, but a result after complex aggregation operation, which encodes the structure information in the local neighborhood, making the matching result insensitive to noise and slight illumination changes, for example, for the pixel (100, 250), its discrete minimum disparity index d0 is 55, the corresponding minimum matching cost C 100,250(55) is 1350, in order to carry out sub-pixel calculation, also need to obtain its adjacent parallax under the matching cost, namely C 100，250 (54) and

[0048] C 100,250 (56), the value is 1382 and 1410 respectively, the three values of the cost {1382, 1350, 1410} will be used together to fit the local curve of the cost function.

[0049] G x，y is the normalized sum of gradient magnitude within the matching block where the pixel (x, y) is located, which is used to measure the texture richness of the region, the greater the gradient value, the more texture, the higher the reliability of the disparity estimation, the steps to obtain it are as follows: first, for each pixel within the matching block (such as 9x9 window), use Sobel operator to calculate its gradient in horizontal direction g x And the gradient in vertical direction g y Then, calculate the gradient magnitude of each pixel Sum up the gradient magnitude of all 81 pixels in the window to get the gradient magnitude sum S G Finally, in order to eliminate the influence of overall brightness difference between different image regions and map it to the uniform [0, 1] interval, normalization processing is needed, the maximum value S G，max And the minimum value S G，min Is determined by the distribution of gradient magnitude sum on a large number of image samples of foundation pit supporting structure, the normalization calculation formula is: G x，y =(S G -S G，min ) / (S G，max -S G，min ), for example, for the 9x9 matching block where the pixel (100, 250) is located, the calculated gradient magnitude sum S G Is 3500, by statistics on 10000 different wall region sample blocks, S G，min =500, S G，max =5500, then the normalized gradient sum of the matching block is: G 100,250 =(3500-500) / (5500-500) = 3000 / 5000 = 0.6.

[0050] is a regularization coefficient, which is a positive constant used to balance the weight of the quadratic fitting term and the gradient regularization term in sub-pixel disparity calculation, the setting of this coefficient is crucial to the performance of the algorithm, and the optimal value needs to be determined through experimental calibration, the calibration process is as follows: prepare a set of foundation pit scene data set containing three-dimensional ground true value, which can be point cloud data obtained by high-precision laser scanner, for the images in the data set, set a series of candidate values of lambda, for example, from 0.001 to 0.1, with a step of 0.001, for each lambda value, run the complete disparity calculation process, get the disparity map, and convert it into a three-dimensional point cloud, then, register the reconstructed point cloud with the ground true value point cloud, calculate the average distance error (such as root mean square error RMSE) between all corresponding points, draw the curve of RMSE changing with lambda, select the lambda that makes the RMSE minimum as the final regularization coefficient value, for example, through experiments, it is found that when lambda = 0.025, the RMSE of the reconstructed point cloud and the true point cloud is minimum, which is 1.2 millimeters, therefore, the value of lambda is finally determined as 0.025.

[0051] is a small constant to prevent division by zero, which is 1 x 10 ；9 .

[0052] According to the above parameters, the calculation is as follows:

[0053] Taking the above obtained parameters as an example, the sub-pixel disparity value of pixel (100, 250) is calculated.

[0054] The known parameters are as follows:

[0055] Discrete minimum disparity index: d0=55;

[0056] Matching cost under adjacent disparity: C 100,250 (54) = 1382, C 100,250 (55) = 1350,

[0057] C 100,250 (56) = 1410;

[0058] Normalized gradient sum: G 100,250 = 0.6;

[0059] Regularization coefficient: lambda = 0.025;

[0060] Prevent division by zero constant: epsilon = 1 x 10 ；9 ;

[0061] Substitute these values into the formula:

[0062]

[0063] Calculate the numerator part:

[0064] 1382 - 1410 = -28;

[0065] Compute the cost term part in the denominator:

[0066] 2 x [1382 - 2700 + 1410] = 2 x

[92] = 184;

[0067] Compute the regularization term part in the denominator:

[0068]

[0069] Compute the total denominator:

[0070] 184 + 0.0694 = 184.0694;

[0071] Compute the final sub-pixel disparity value:

[0072]

[0073] d 100,250 ≈ 54.8479;

[0074] This result shows that the accurate disparity value of pixel (100, 250) is about 54.8479 pixels, instead of the discrete integer 55, and this sub-pixel level correction value (-0.1521) is a fine tuning to the integer matching result, which reflects that in the real world, the projection of an object to a camera sensor does not strictly fall on the center of the pixel grid, and by calculating the sub-pixel disparity, the corresponding point can be located, so that higher depth accuracy can be obtained in subsequent three-dimensional reconstruction, and the value 54.8479 directly represents the horizontal coordinate difference of the spatial point in the left and right images in units of pixels.

[0075] According to the sub-pixel disparity value, that is, the disparity map composed of the high-precision disparity calculated for each pixel in the image, combined with the camera parameters calibrated once through standard processes such as Zhang Zhengyou calibration method, these parameters include the intrinsic matrix K L of the left camera (including focal length f x,L , f y,L and principal point coordinates c x,L , c y,L ), the intrinsic matrix K R of the right camera, and the relative pose relationship between the two cameras, that is, the rotation matrix R and the translation vector T from the left camera coordinate system to the right camera coordinate system, where the length of the translation vector T is the camera baseline length, for each valid pixel (x, y) in the disparity map and its corresponding sub-pixel disparity value d x，y , the three-dimensional space coordinates are solved, and the solving process follows the principle of triangulation. Specifically, for the pixel coordinates (x L , yL ), whose corresponding point coordinates in the right image are (x R ,y R ) = (x L -d x,y , y L ), first convert the two two-dimensional pixel coordinates to the respective camera's normalized image plane coordinates, the left camera is (u L , v L ) = ((x L -c x，L ) / f x，L , (y L -c y，L ) / f y，L ), the right camera is the same, so that the two three-dimensional space ray direction vectors from the two camera optical centers are obtained, pointing to the same point in space, by solving the intersection point of the two rays in space, the coordinates of the three-dimensional space point (X, Y, Z) corresponding to the pixel can be obtained, due to the existence of calculation error and noise, the two rays may not strictly intersect, in actual operation, a three-dimensional point with the smallest sum of squared distances from the two rays is found as the solution, this process is performed once for each pixel in the disparity map with an effective disparity value, each time the calculation generates a three-dimensional space point, all the calculated three-dimensional space point (X, Y, Z) coordinate data is aggregated together to form a large, unordered point set, this set is a digital description of the current moment of the underground diaphragm wall surface form, and a reference wall space point set is generated.

[0076] The acquisition step of the pose registration transformation parameter is:

[0077] Extract the two-dimensional coordinates of all pixel points on the wall surface through real-time images, convert the two-dimensional coordinates of each pixel point into three-dimensional space coordinates by calling the camera internal and external parameters, and accumulate point by point to form a real-time wall space point set;

[0078] According to the real-time wall space point set, match the corresponding points in the reference wall space point set point by point, calculate the Euclidean distance between the two sets of points round by round in an iterative manner and take the minimum distance, take the distance sum of all point pairs as the iterative target value, and update the rotation matrix and translation vector until the iterative target value converges, to obtain the pose registration transformation parameter.

[0079] Specifically, by real-time images, i.e. the left and right two digital images synchronously captured by the binocular camera system in the current monitoring period, first, the images are pre-processed, including applying the distortion coefficients obtained in the camera calibration process to correct the radial and tangential distortion, then the corrected images are extracted for the region of interest (ROI), the non-wall parts in the image, such as the sky, adjacent buildings and other areas are excluded by the preset mask, only the pixel of the underground continuous wall surface area is retained, then, for each pixel point on the wall surface, the same stereo matching algorithm as when establishing the reference wall space point set is adopted, i.e. the absolute difference accumulation is performed on each pixel block along the epipolar direction and is aggregated in multiple directions, the minimum matching cost index of each pixel point is recorded and the uniqueness validity is determined by a threshold, the discrete minimum disparity index is obtained, and then the sub-pixel accuracy disparity map is calculated by the quadratic polynomial fitting and gradient regularization method, the gray value of each pixel in the disparity map represents a sub-pixel disparity value, then, the pre-calibrated camera internal and external parameters are called, specifically including the focal length f x , y , the principal point coordinates c x , c y and the baseline length B between the binocular cameras, for each pixel (x, y) in the disparity map and its corresponding sub-pixel disparity value d x，y , the three-dimensional space coordinates (X, Y, Z) in the camera coordinate system are solved by using the triangulation principle, the calculation process is, first, the depth value Z = (f x · B) / d x，y is calculated, then the X and Y coordinates are calculated according to the pinhole camera model back projection, i.e. X = (x-c x ) · Z / f x and Y = (y-c y ) · Z / f y , the three-dimensional solving process is repeated for all pixel points with valid disparity values in the disparity map, and each three-dimensional space coordinate (X, Y, Z) calculated is stored as a data point, finally, all calculated three-dimensional space coordinate points are accumulated point by point to form the real-time wall space point set.

[0080] According to the real-time wall space point set, the point set is taken as a source point cloud, and is registered with the previously established reference wall space point set, that is, a target point cloud. This process adopts an iterative closest point (ICP) algorithm. First, a rigid transformation parameter, that is, a 3x3 rotation matrix and a 3x1 translation vector, is initialized. Generally, the initial rotation matrix is set as a unit matrix, and the translation vector is set as a zero vector. Then, an iterative loop is entered. In each iteration, for each point in the real-time wall space point set, a nearest corresponding point in the reference wall space point set is found. To accelerate the finding process, a K-D tree spatial index structure is pre-constructed for the reference wall space point set. Then, all found corresponding point pairs are screened, and unreasonable matching pairs are removed. The removal is based on the Euclidean distance between the point pairs. A distance threshold is set. The threshold is dynamically determined by multiplying a constant factor (for example, 2.5) by the median of all matching distances in the previous iterations. Any point pair with a distance exceeding the threshold is regarded as an outlier and is removed from the current round calculation. After obtaining the screened effective corresponding point set, a rigid transformation that can minimize the sum of squares of distances between the corresponding points is calculated. This solving process obtains a closed solution through singular value decomposition (SVD) method, so as to obtain the optimal rotation matrix and translation vector of the current round. Then, the newly calculated transformation matrix is applied to the entire real-time wall space point set to update the coordinate positions of all points in the real-time wall space point set. After that, the convergence state of the iteration is calculated. The convergence judgment criteria include two conditions: one is that the change amount of the transformation matrix calculated in the previous and current iterations is less than a preset threshold, for example, the norm of the change of the rotation matrix is less than 1x10 ；6 and the norm of the change of the translation vector is less than 1x10 ；6 ; the other is that the change rate of the average sum of squares of distances of the corresponding point pairs in continuous multiple iterations is less than a preset minimum value, for example, 0.01%. When any convergence condition is met or the preset maximum number of iterations (for example, 100 times) is reached, the loop is terminated, and the pose registration transformation parameter is obtained.

[0081] The obtaining step of the envelope surface normal distance set is as follows:

[0082] The pose registration transformation parameter is used to correct the three-dimensional coordinates of the real-time wall space point set as a whole, to calculate the normal distance from the corrected coordinates to the surface of the reference wall space point set point by point, and to record and generate the envelope surface normal distance set.

[0083] Specifically, the pose registration transformation parameter obtained in the previous step, that is, the final converged rotation matrix R and translation vector T, is used for one-time global coordinate transformation of the three-dimensional coordinates of each point in the original real-time wall space point set. For any point p i =(X i ,Y i ,Z i), its corrected coordinates p i ′ =(X i ′ Y i ′ Z i ′ ) through matrix operations p i ′ =R·p i After obtaining +T and completing the coordinate correction of all points, a corrected point set in the same coordinate system as the reference wall spatial point set is generated. Next, the normal distance from each point in the corrected point set to the surface formed by the reference wall spatial point set is calculated point by point. This process first requires estimating the normal vector of the surface of the reference wall spatial point set. For each point in the reference point cloud, a local neighborhood is formed by querying its nearest points (e.g., the 20 nearest points) in the KD tree. Then, principal component analysis (PCA) is performed on this neighborhood point set. The eigenvector corresponding to the minimum eigenvalue is the surface normal vector of that point. After calculating and storing the normal vectors for all points in the reference point cloud, for each point p in the corrected point set... i ′ The nearest neighbor q is found again by using the KD tree in the spatial point set of the reference wall. i and read q i normal vector n i , point p i ′ The normal distance to the reference surface is the vector (p) i ′ -q i ) in normal vector n i The projected length along the direction is obtained by calculating the dot product of the two vectors, i.e., distance. i =(p i ′ -q i )·n i This distance value has a positive or negative sign, and the sign indicates the direction of deformation. For example, if the normal vector points to the outside of the wall, a positive value represents outward convexity and a negative value represents inward concavity. The three-dimensional coordinates of each corrected point and its calculated normal distance value are recorded as a data pair. The records of all points are summarized to generate the normal distance set of the enclosure structure surface.

[0084] The steps for obtaining the three-dimensional coordinates of structural feature points are as follows:

[0085] The pixel points of the intersection of the crossbeam and the column and the anchor head position on the lattice anchor wall are located frame by frame in the real-time image, the pixel row and column coordinates are extracted, the camera internal parameter and the camera external parameter are read, the pixel row and column coordinates are projected to the imaging plane after distortion correction and principal point translation correction, the spatial intersection position is solved combining the baseline and the attitude, and the three-dimensional coordinates of the structural feature points are obtained.

[0086] Specifically, on the real-time image, that is, the current frame left and right image pair captured by the binocular camera, a pre-trained deep learning target detection model, specifically a model based on YOLOv5 architecture and optimized for the features of the lattice anchor wall, is used to locate the intersection of the crossbeam and the column and the position of the anchor head frame by frame. The training data set of the model contains more than 1000 images of the foundation pit site taken under different light, angle and distance, and the positions of the two types of feature points in the images are manually annotated. The model training process uses the Adam optimizer, the initial learning rate is set to 0.001, and the learning rate decay is performed using the cosine annealing strategy. The model is trained for a total of 300 cycles. The output of the model is the category, confidence score and pixel bounding box coordinates of each detected feature point in the image. The non-maximum suppression algorithm is applied to the output of the model to remove overlapping detection boxes. Then a confidence threshold is set. The setting of the threshold is based on the precision-recall curve drawn on a validation set containing 1000 images. The confidence corresponding to the highest F1 score point is selected, for example, 0.92. Only the detection results with a score higher than 0.92 are considered as valid pixel points. The pixel row and column coordinates of the centers of these valid detection boxes are extracted. Then the pre-calibrated camera internal parameters (focal length, principal point coordinates) and distortion coefficients (radial and tangential distortion parameters) are called. The extracted pixel row and column coordinates are corrected for distortion. Then the pixel coordinates are translated to the imaging plane coordinate system with the image principal point as the origin. For each feature point detected in the left image, its matching point is searched on the corresponding epipolar line in the right image using the epipolar constraint. The matching is based on the normalized cross-correlation (NCC) score of the image blocks in the two detection boxes. The one with the highest NCC score is selected as the matching pair. Finally, the baseline length and the rotation and translation relationship between the two cameras defined in the camera extrinsic parameters are combined to apply the triangulation method to each pair of matched imaging plane coordinate points to solve their intersection position in the three-dimensional space, and the three-dimensional coordinates of the structural feature points are obtained.

[0087] The acquisition step of the topological node set with shape variables is:

[0088] According to the three-dimensional coordinates of the structural feature points, a spatial mapping is established with the three-dimensional coordinates of the corrected points in the enclosing structure surface normal distance set as the retrieval object. The first matching point is retrieved from small to large in the Euclidean distance for each three-dimensional coordinate of the structural feature points, and the normal distance recorded in the matching point is read to generate a corresponding record of the three-dimensional coordinates of the structural feature points and the normal distance.

[0089] According to the corresponding record of the structural feature point three-dimensional coordinate and the normal distance, a topology node entry is created for each structural feature point three-dimensional coordinate, and the read normal distance is written as a deformation variable. All topology node entries are summarized and the one-to-one correspondence between the coordinates and deformation variables is maintained to generate a topology node set with deformation variables.

[0090] Specifically, according to the structural feature point three-dimensional coordinates obtained in the previous step, which is a sparse point set, and the envelope surface normal distance set generated in the foregoing, which is a dense data set containing wall surface dense three-dimensional points and their corresponding normal distances, in order to accurately assign deformation information to the structural feature points, a high-efficiency query mechanism needs to be established first. The specific operation is to use the three-dimensional coordinates of all corrected points in the envelope surface normal distance set as input data to construct a K-D tree (K-dimensional tree) spatial index structure. This K-D tree recursively divides the three-dimensional space into a series of hyper-rectangular regions, so that the subsequent nearest neighbor search operation can be completed in logarithmic time instead of linearly traversing all points. After the construction is completed, each point in the structural feature point three-dimensional coordinate set is traversed, and the three-dimensional coordinates of each structural feature point are used as query points to perform a nearest neighbor search (Nearest Neighbor Search) operation in the K-D tree that has been constructed. This search algorithm starts from the root node of the tree and recursively moves down according to the current node's split dimension and split value to decide whether to enter the left or right subtree. When a leaf node is reached, the algorithm backtracks to check whether there are closer points on other branches of the path, thereby finding the first matching point that is closest to the query point in Euclidean distance, i.e., the point in the dense wall point cloud that is physically closest to the structural feature point. After finding this first matching point, the associated attribute recorded in the envelope surface normal distance set for this matching point, i.e., the normal distance value, is read. This value directly reflects the local deformation size and direction of the wall surface where the point is located. Finally, the three-dimensional coordinates of each structural feature point are paired with the normal distance value retrieved by it to form a binary record, and all these records are summarized to generate a corresponding record of the structural feature point three-dimensional coordinate and the normal distance.

[0091] According to the corresponding record of the structural feature point three-dimensional coordinates and normal distance generated in the previous step, a structured data entry, i.e. a topology node entry, is created for each structural feature point, which defines the complete state information of the node, specifically including the following fields: node unique identifier (Node ID), initial three-dimensional coordinates, current three-dimensional coordinates, and deformation variable. First, the current structural feature point needs to be associated with a predefined initial topology node set representing the initial state of the foundation pit support structure. This initial topology node set is established by detecting feature points and recording their three-dimensional coordinates when the foundation pit is just excavated and has not yet deformed. The association process is completed by performing a nearest neighbor search on the initial topology node set for each current structural feature point three-dimensional coordinate, thereby assigning a unique and persistent node ID to each current node and obtaining its corresponding initial three-dimensional coordinates. Next, the fields of the topology node entry are filled in. For each entry read from the corresponding record of the structural feature point three-dimensional coordinates and normal distance, a new topology node entry is created, the node ID and initial three-dimensional coordinates associated with it are written to the corresponding fields, the three-dimensional coordinates in its record are written as the current three-dimensional coordinates, and the normal distance value in its record is directly written as the deformation variable in the deformation variable field. This deformation variable accurately quantifies the normal displacement of the wall surface at the structural node. This creation and filling process is repeated for all records. Finally, all generated topology node entries are aggregated to form a set. Each element in this set completely describes the spatial position change and surface deformation of a key structural node from the initial state to the current state, and strictly maintains the one-to-one correspondence between the node's identity, coordinates, and deformation variable. The topology node set with deformation variables is generated.

[0092] The network geometry parameter acquisition step is:

[0093] The topology node set with deformation variables and the initial topology node set are traversed one by one for each connection line in the preset connection relationship. The current three-dimensional coordinates and initial three-dimensional coordinates of the two ends of the connection line are extracted. The current connection line length and initial connection line length are calculated according to the Euclidean distance, and the current included angle and initial included angle are calculated by the inner product of adjacent connection line vectors. The network geometry parameters are obtained.

[0094] Specifically, according to the topology node set with deformation variables and the initial topology node set, first, a predefined data file describing the lattice anchor wall topology structure is loaded. This file is generated based on engineering design drawings, and all nodes physically connected by beams or columns are explicitly listed in the form of node ID pairs, forming a complete connection relationship list. Then, each connection line in this connection relationship list is traversed one by one. For each connection line, such as the connection line with node IDs u and v, the current three-dimensional coordinates of nodes u and v are queried and extracted from the topology node set with deformation variables, denoted as

[0095] P u，current and P v，current , and the initial three-dimensional coordinates of the two nodes are extracted from the initial topology node set, denoted as P u，initial and P v，initial , then the length of the connecting line is determined by calculating the Euclidean distance between the two points in the three-dimensional space, the current connecting line length is the distance between P u，current and P v，current , and the initial connecting line length is the distance between P u，initial and P v，initial . After the calculation of the length of all connecting lines, each node is traversed to calculate the included angle at the node. For any node j, first find all other nodes connected to node j from the connection relationship list to form an adjacent node list, then combine the nodes in the adjacent node list two by two to form all adjacent connecting line pairs with node j as the vertex, for example, connecting line j-i and connecting line j-k. For each pair of adjacent connecting lines, the vector representing the direction of the connecting line is obtained by subtracting the coordinates, for example, the vector V ji，current in the current state is P i,current -P j,current and V jk,current is P k,current -P j,current . The current included angle can be calculated by taking the inverse cosine of the dot product of the two vectors divided by the product of their modules. Similarly, the initial included angle is calculated using the initial coordinates. All calculated current connecting line lengths, initial connecting line lengths, current included angles and initial included angles are collected to obtain the network geometric parameters.

[0096] The steps for obtaining the structural distortion energy quantization value are as follows:

[0097] According to the network geometric parameters, the length change rate of each connecting line is calculated, the difference between the current connecting line length and the initial connecting line length is divided by the initial connecting line length to obtain the relative length change rate of each connecting line, and the relative angle change rate of each node is calculated, the difference between the current included angle and the initial included angle is divided by the initial included angle to obtain the dimensionless angle change rate of each node.

[0098] The relative length change rate and the relative angle change rate are used to calculate the structural distortion energy quantization value, and the calculation formula is:

[0099]

[0100] where E g is the structural distortion energy quantization value, K is the total number of connecting lines, M is the total number of nodes, p e is the relative length change rate of the e-th connecting line, R u(e) is the relative distance value of the first end point of the e-th connecting line, and Rv(e) R is the relative length change rate of the e-th link m α is the average of all relative length change rates j R is the relative angle change rate of the j-th node j R is the relative length change rate of the j-th link and R is the initial length of the two links that form the j-th node angle β is the average of all initial lengths of the links

[0101] Specifically, according to the network geometry parameters obtained in the previous step, which contain the initial and current geometry states of each link and each node angle, these data are processed to quantify the degree of change. First, the link length is processed. Each link recorded in the network geometry parameters is traversed to extract its current link length and initial link length. For the e-th link, the calculation method of the relative length change rate is as follows: the difference between the current link length and the initial link length is obtained, and then the difference is divided by the initial link length of the link. For example, the initial length of a link is 3000 mm, and the current monitored length is 3003 mm. The length change is +3 mm, and the relative length change rate is 3 / 3000=+0.001, indicating that the member has stretched by 0.1%. If the current length is 2998.5 mm, the change is -1.5 mm, and the relative length change rate is -1.5 / 3000=-0.0005, indicating that it has been compressed by 0.05%. Each calculated relative length change rate is recorded together with the corresponding link ID. Then, the node angle is processed in the same way. Each angle at each node recorded in the network geometry parameters is traversed to extract its current angle and initial angle. For a certain angle of the j-th node, the calculation method of the relative angle change rate is as follows: the difference between the current angle and the initial angle is obtained, and then the difference is divided by the initial value of the angle. For example, the initial angle of a node is 90.0 degrees (π / 2 in radian), and the current angle is 90.18 degrees. The angle change is +0.18 degrees, and the relative angle change rate is 0.18 / 90.0=+0.002, indicating that the angle has increased by 0.2%. Each calculated relative angle change rate is recorded together with the corresponding node ID and angle information to obtain the dimensionless angle change rate of each node.

[0102] Formula: The overall distortion of the structure is decomposed into two main parts: the distortion caused by the change in the length of the member (stretching or compression) and the distortion caused by the change in the angle of the node, corresponding to the first and second terms in the formula, respectively. The weight factor related to the deformation amount is introduced. In the length distortion term, the relative length change rate ρe is adjusted by a weight factor , which represents the ratio of the average normal deformation of the two nodes of the link to the average normal deformation of all nodes, which means that in the area where the wall normal deformation is large, even if the axial strain of the component is small, its contribution to the overall distortion energy will be amplified, thus highlighting the risk of dangerous areas, and in the angular distortion term, the relative angle change rate α j is not only adjusted by the deformation weight of the node itself , but also by a geometric weight , which represents the ratio of the geometric mean of the initial lengths of the two links forming the angle to the average initial length of all links, which makes the distortion energy higher when the angle formed by the longer component changes, which is consistent with the principle that longer rods have a greater impact on instability in structural mechanics. Through this weighting method, the formula can more sensitively capture small changes in key parts of the structure.

[0103] K is the total number of links, representing the total number of edges in the topological network model, which is uniquely determined by the preset connection relationship, and is obtained by counting the total number of entries in the connection relationship list. For example, in a lattice anchor wall monitoring area containing 80 nodes, according to its engineering design drawing, a total of 150 connections between beams and columns are defined, so the value of K is 150.

[0104] M is the total number of nodes, representing the total number of nodes in the topological network model, which is obtained by counting the number of unique node IDs in the set of topological nodes with deformation variables. In a typical monitoring scenario, it may contain dozens to hundreds of structural feature points. For example, in a monitored facade, a total of 80 beam-column intersection points or anchor heads are identified and continuously tracked, so the value of M is 80.

[0105] ρ e is the relative length change rate of the e-th link, which is a dimensionless parameter representing the axial strain degree of the component, which is obtained by the calculation in the previous step, i.e. where L e and are the current length and initial length of the e-th link, respectively. For example, for link No. 25, its initial length is 3000 mm, and the current length is 3003 mm, so its relative length change rate is ρ 25 = (3003-3000) / 3000 = 0.001.

[0106] R u(e) and R v(e)R 12 is the associated distance value of the two end points u(e) and v(e) of the e-th link, which is the "deformation variable" of the node recorded in the topology node set, i.e. the normal distance of the wall surface, with the unit of millimeter, which directly reflects the deformation size of the wall at the node position, for example, for the link numbered 25, the end points are node 12 and node 13, it is obtained from the topology node set that the deformation variable of node 12 is R 12 = 5.2 mm, and the deformation variable of node 13 is R 13 = 4.8 mm.

[0107] R m is the average value of the associated distance values of all nodes, which represents the average normal deformation level of the wall surface of the entire monitoring area, and the calculation formula is wherein R j is the deformation variable of the j-th node, for example, in a system with 80 nodes, the sum of the deformation variables of all nodes is 320 mm, and the average value is R m = 320 / 80 = 4.0 mm.

[0108] α j is the relative angle change rate of the j-th node, which is a dimensionless parameter representing the change of the connection stiffness at the node, which is obtained by the calculation of the previous step, i.e. wherein θ j and are the current value and initial value of a certain angle at node j, for example, for a certain angle at node 5, the initial value is 90 degrees, and the current value is 89.55 degrees, and the relative angle change rate is α5= (89.55-90) / 90 = -0.005.

[0109] R j is the associated distance value of the j-th node, which is the same as the definition of R u(e) , i.e. the normal deformation variable of node j, for example, it is obtained from the topology node set that the deformation variable of node 5 is R5= 6.1 mm.

[0110] and are the initial lengths of the two links constituting the angle at the j-th node, with the unit of millimeter, which are obtained from the network geometric parameters, for example, the initial lengths of the two links constituting the angle at node 5 are mm and mm.

[0111] is the average value of the initial lengths of all links, representing the average level of the lattice size, and the calculation formula is wherein is the initial length of the e-th link, for example, in a system with 150 links, the sum of the initial lengths of all links is 450000 mm, then the average value is

[0112] β is a coefficient for balancing the relative influence of length distortion energy and angle distortion energy, the setting of this coefficient is based on engineering judgment of structure type and main failure mode, for lattice support structure, the change of node angle (shear deformation) often reflects the overall instability trend of the structure more than the axial stretching of the bar, so it is necessary to give greater weight to the angle distortion term, the value of which is determined as follows: construct a finite element model of the support structure, apply typical load cases (such as increase of soil pressure, overload, etc.), analyze and calculate the strain energy caused by axial force and the strain energy caused by shear force and bending moment at different deformation stages, statistically average the strain energy proportion of multiple working conditions, for example, if the analysis results show that the energy contribution related to angle deformation is generally 1.5 times that of length deformation, then β can be set to 1.5, so that the calculated structure distortion energy can better match the real structural mechanics response.

[0113] According to the above parameters, the calculation is as follows:

[0114] Taking a link (e = 25) and a node (j = 5) as an example for single item calculation:

[0115] Length distortion term (for e = 25):

[0116]

[0117] Angle distortion term (for j = 5):

[0118]

[0119] Calculate and sum the length distortion terms of all 150 links and the angle distortion terms of all 80 nodes to get the final structure distortion energy quantization value:

[0120]

[0121] For example, after calculating all the terms, the total length distortion energy is 0.0018 and the total angle distortion energy is 0.0045, then:

[0122] E g = 0.0018 + 1.5 * 0.0045 = 0.0018 + 0.00675 = 0.00855;

[0123] The result shows that the overall distortion energy quantization value of the foundation pit support structure at the current time is 0.00855, which condenses the deformation information of the entire structure into a single value, facilitating trend analysis and threshold judgment. The change trend and size can reflect the evolution of the structure safety state. The graded warning threshold is set to explain its meaning. For example, by combining historical data and simulation analysis, a warning threshold system is set: E g <0.01 is "safe" state, 0.01≤E g <0.05 is "attention" state, 0.05≤E g <0.1 is "warning" state, E g ≥0.1 is "reminder" state, the current calculation result 0.00855 is less than 0.01, so the current structure state is determined to be "safe".

Claims

1. A health management system for the real-time safety status of a foundation pit support structure, characterized in that, The system includes: The 3D scene point cloud construction module synchronously collects images from the left and right cameras and preset camera parameters, calculates the disparity value of each pixel, and calculates the 3D spatial coordinates of each pixel on the surface of the underground continuous wall based on the disparity value, pixel coordinates and camera internal and external parameters, and establishes a reference wall spatial point set. The wall deformation deviation calculation module acquires a real-time wall spatial point set through real-time images, minimizes the distance between corresponding points between the real-time wall spatial point set and the reference wall spatial point set through iterative calculation, solves for the pose registration transformation parameters, applies the pose registration transformation parameters to correct the real-time wall spatial point set, calculates the normal distance from each point in the corrected point set to the surface formed by the reference wall spatial point set, and generates a set of normal distances for the surface of the enclosure structure. The key node 3D positioning module detects the pixel points at the intersection of the beam and column and the anchor head position on the lattice anchor wall in the real-time image, calculates the 3D coordinates, obtains the 3D coordinates of the structural feature points, traverses each point in the 3D coordinates of the structural feature points, finds the distance value of the nearest neighbor point, and generates a set of topological nodes with shape variables. The topology network deformation energy calculation module calculates the length of the connection between nodes and the angle between the nodes based on the preset connection relationship to obtain the network geometric parameters. The network geometric parameters and the topology node set with deformation variables are used to calculate the quantification value of structural distortion energy.

2. The health management system for the real-time safety status of the foundation pit support structure according to claim 1, characterized in that, The steps for obtaining the reference wall spatial point set are as follows: Simultaneously acquired images from the left and right cameras and preset camera parameters are used to accumulate the absolute difference of each pixel block along the epipolar direction and aggregate them in multiple directions. The minimum matching cost index of each pixel is recorded and the uniqueness and validity are determined by a threshold to obtain the discrete minimum disparity index. Calculate the sub-pixel disparity value of the pixel based on the discrete minimum disparity index; Based on the subpixel parallax value, and combined with the pixel coordinates and the camera's intrinsic and extrinsic parameters such as focal length, principal point coordinates, baseline length, rotation matrix, and displacement vector, the three-dimensional spatial coordinates of each pixel on the surface of the underground continuous wall are calculated and all three-dimensional spatial coordinates are aggregated to generate a reference wall spatial point set.

3. The health management system for the real-time safety status of the foundation pit support structure according to claim 1, characterized in that, The steps for obtaining the pose registration transformation parameters are as follows: The two-dimensional coordinates of all pixels on the wall surface are extracted by real-time images. The camera's internal and external parameters are called to convert the two-dimensional coordinates of each pixel into three-dimensional spatial coordinates. The coordinates are accumulated point by point to form a real-time spatial point set of the wall. Based on the real-time wall space point set, the corresponding points in the reference wall space point set are matched point by point. The Euclidean distance between the two sets of points is calculated in an iterative manner and the minimum distance is taken. The sum of the distances of all point pairs is used as the iterative target value. The rotation matrix and translation vector are updated until the iterative target value converges, and the pose registration transformation parameters are obtained.

4. The health management system for the real-time safety status of the foundation pit support structure according to claim 1, characterized in that, The steps for obtaining the normal distance set of the enclosure structure surface are as follows: The pose registration transformation parameters are used to perform overall coordinate system correction on the three-dimensional coordinates of the real-time wall space point set, calculate and record the normal distance from the corrected coordinates to the surface of the reference wall space point set point by point, and generate a set of normal distances on the surface of the enclosure structure.

5. The health management system for the real-time safety status of the foundation pit support structure according to claim 1, characterized in that, The steps for obtaining the three-dimensional coordinates of the structural feature points are as follows: In real-time images, the pixel points at the intersection of the beams and columns on the lattice anchor wall and the anchor head position are located frame by frame. The pixel row and column coordinates are extracted and the camera intrinsic and extrinsic parameters are read. After distortion correction and principal point translation correction, the pixel row and column coordinates are projected onto the imaging plane. Combined with the intersection of the baseline and attitude calculation space, the three-dimensional coordinates of the structural feature points are obtained.

6. The health management system for the real-time safety status of the foundation pit support structure according to claim 1, characterized in that, The steps for obtaining the topological node set with shape variables are as follows: Based on the three-dimensional coordinates of the structural feature points, a spatial mapping is established with the three-dimensional coordinates of the corrected points corresponding to the normal distance set of the surface of the enclosure structure as the retrieval object. For each structural feature point, the first matching point is retrieved according to the Euclidean distance from smallest to largest, and the normal distance recorded by the matching point is read to generate a corresponding record of the three-dimensional coordinates and normal distance of the structural feature points. Based on the correspondence between the three-dimensional coordinates and normal distances of the structural feature points, a topological node entry is created for each structural feature point's three-dimensional coordinates, and the read normal distance is written as a deformation variable. All topological node entries are summarized, and the one-to-one correspondence between coordinates and deformation variables is maintained, generating a topological node set with deformation variables.

7. The health management system for the real-time safety status of the foundation pit support structure according to claim 1, characterized in that, The steps for obtaining the network geometric parameters are as follows: The topological node set with shape variables and the initial topological node set traverse each connection with the preset connection relationship one by one, extract the current three-dimensional coordinates and the initial three-dimensional coordinates at both ends of the connection, calculate the current connection length and the initial connection length according to Euclidean distance, and calculate the current angle and the initial angle by the inner product of the adjacent connection vectors to obtain the network geometric parameters.

8. The health management system for real-time safety status of foundation pit support structure according to claim 1, characterized in that, The steps for obtaining the quantized value of the structural distortion energy are as follows: Based on the network geometric parameters, calculate the length change rate of each connection. Divide the difference between the current connection length and the initial connection length by the initial connection length to obtain the relative length change rate of each connection. At the same time, calculate the relative angle change rate of each node. Divide the difference between the current included angle and the initial included angle by the initial included angle to obtain the dimensionless angle change rate of each node. The relative length change rate and the relative angle change rate are used to calculate the quantified value of structural distortion energy.

Images