Elevator guide rail member size high-precision measurement method based on visual intelligence

Abstract

The application discloses a high-precision measurement method for elevator guide rail component size based on visual intelligence, and relates to the technical field of visual detection of elevator guide rails, which comprises the following steps: a calibration target is used to complete joint calibration on a multi-source visual acquisition channel, a unified coordinate mapping is established, and multi-modal visual data is synchronously acquired; a guide rail geometric atomic dictionary is constructed, a section parameter is decoded by sparse decomposition based on a joint observation vector with a minimum residual criterion, and cross-modal verification is performed in combination with a texture image; a graph optimization model is constructed to solve a global optimal pose of each section, thermal deformation compensation is completed in combination with a non-uniform temperature field model; and full-size parameters are calculated based on compensated data, and a measurement report is generated. The application has the advantages that: multi-modal data is accurately spatially aligned, the robustness of parameter decoding under complex working conditions is improved, splicing cumulative error and thermal deformation interference are effectively inhibited, and the high precision and long-term stability of full-size detection of long guide rails are ensured.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology for elevator guide rails, specifically to a high-precision measurement method for the dimensions of elevator guide rail components based on visual intelligence. Background Technology

[0002] Elevator guide rails are the core guiding components of an elevator operating system, and their dimensional accuracy and geometric tolerances directly determine the smoothness, safety, and ride comfort of the elevator. With the elevator industry's continuous increasing demands for product quality and production efficiency, high-precision full-dimensional inspection of guide rails has become a core aspect of production quality control. Vision-based non-contact measurement technology, with its advantages of high inspection efficiency and comprehensive parameter coverage, is gradually becoming an important technological direction for inspecting guide rail geometric parameters, possessing significant engineering value in ensuring product consistency and supporting process optimization.

[0003] Existing vision-based guide rail measurement methods generally suffer from insufficient calibration accuracy of multi-source channels and are prone to drift, making it difficult to guarantee detection stability under long-term operation. Cross-sectional parameter decoding often relies on single-modal data, resulting in poor robustness when interference such as reflections or oil stains are present on the workpiece surface. Segmental splicing of multiple cross-sections easily leads to cumulative errors and does not fully consider the influence of thermal deformation caused by the ambient temperature field, making it difficult to meet the high-precision control requirements for measuring the straightness, torsion, and other form and position parameters of long guide rails. Therefore, it is necessary to provide a high-precision measurement method for elevator guide rail component dimensions based on vision intelligence to solve the above-mentioned problems. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides a high-precision measurement method for elevator guide rail component dimensions based on visual intelligence. This technical solution solves the problems of existing visual intelligence guide rail measurement methods mentioned in the background, such as insufficient accuracy and easy drift in multi-source channel calibration, poor anti-interference robustness of single-mode decoding, easy accumulation of errors in multi-section splicing, and failure to consider the influence of thermal deformation, making it difficult to meet the high-precision control requirements of long guide rail shape and position parameters.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A high-precision measurement method for elevator guide rail component dimensions based on visual intelligence includes: S1. Use a calibration target to perform joint calibration on the multi-source visual acquisition channels, establish a unified spatial coordinate mapping between the two-dimensional area array image channel, the three-dimensional line laser contour channel and the telecentric end face image channel, and generate a joint calibration parameter matrix; S2. Synchronously acquire two-dimensional texture image data, three-dimensional contour point cloud data and telecentric end face image data of the same cross section of the elevator guide rail to form a multimodal synchronous visual dataset; S3. Extract the cross-sectional profile point set perpendicular to the guide rail axis, project it onto the calibration coordinate system to generate an observation vector, and use the guide rail T-section width, height, web thickness, guide shoe surface inclination angle, and guide surface inclination angle as parameter variables. Discretely sample within the manufacturing tolerance to construct a guide rail geometric atom dictionary composed of cross-sectional profile point sets. Based on the observation vector and using the minimum residual as the criterion, perform sparse decomposition on the dictionary to produce a coefficient vector. When the coefficient vector has only one non-zero coefficient, the parameter is decoded by index. When there are multiple non-zero coefficients, the parameter is decoded by coefficient magnitude weighted fusion. The decoded parameter is verified across modes using two-dimensional texture image data. If the verification fails, a review is triggered. S4. Repeat S2 to S3 along the guide rail length direction with a preset step size to construct a graph optimization model for the decoding parameters, with the cross-section pose as the node, the adjacent cross-section shape and position as the local edge, and the standard cross-section size as the prior constraint. Iteratively solve the global optimal pose of each cross-section with the coarse registration pose as the initial value. S5. Arrange temperature acquisition arrays on the guide rail surface with denser spacing at both ends and sparser spacing in the middle to construct a non-uniform temperature field model, calculate the thermal deformation correction amount for each section and perform temperature compensation. S6. After fusing the compensated geometric data and texture image, calculate the full-size parameters of the guide rail, compare them with the standard tolerance, and generate a full-size measurement report.

[0006] Preferably, step S1 specifically includes: S101. Fabricate a three-dimensional calibration target that is geometrically complementary to the T-section of the guide rail. Each working surface of the target is machined with a feature point array with known precise coordinates. The feature point array includes a checkerboard corner point array and a circular hole center point array. S102. Fix the calibration target to the guide rail support platform, and collect target images and point clouds through the two-dimensional area array image channel, the three-dimensional line laser contour channel and the telecentric end face image channel respectively, and extract the coordinates of feature points in each channel; S103. Using the known coordinates of the target feature point array as a reference, solve the homography mapping between the two-dimensional channel image coordinate system and the world coordinate system, the rigid body transformation between the three-dimensional channel coordinate system and the world coordinate system, and the scaling mapping between the telecentric channel image coordinate system and the world coordinate system respectively, and combine them into a joint calibration parameter matrix. S104. A reference block with known precise dimensions is fixedly set on the guide rail bearing platform. Every preset number of cross sections or preset time intervals, a single measurement of the reference block is automatically triggered. The measured value is compared with the known dimensions of the reference block to obtain the calibration drift amount. When the drift exceeds the preset threshold, the online recalibration process from S102 to S103 is automatically triggered.

[0007] Preferably, the construction of the guide rail geometry atom dictionary in step S3 specifically includes: S301. Determine if the current guide rail model is known; If the model number is known, obtain the standard cross-sectional dimensions and manufacturing tolerance range corresponding to that model number; If the model is unknown, the initial geometric parameters of the T-shaped section are directly fitted from the three-dimensional point cloud of several sections of the first section of the guide rail. The initial parameter spatial range is determined with the fitted value as the center and the preset loose tolerance as the radius. S302. Uniformly discretely sample each parameter variable in the parameter space according to a preset initial resolution, combine the discrete values ​​of each variable by Cartesian product to generate an initial parameter vector set, and analyze the corresponding complete cross-sectional contour point set according to the cross-sectional geometric topology relationship for each parameter vector in the set to construct an initial geometric atom dictionary; S303. During the measurement process of the subsequent cross-section of the guide rail, based on the statistical mean and variance of the decoded parameters, the sampling range of each parameter variable is dynamically narrowed and the sampling resolution is improved, the geometric atom dictionary is updated, and an iterative update mechanism in which the dictionary adaptively shrinks with the measurement process is formed.

[0008] Preferably, in step S3, the set of cross-sectional profile points perpendicular to the guide rail axis is extracted and projected onto the calibration coordinate system to generate an observation vector, specifically including: S311. Extract a slice point cloud of a preset thickness from the three-dimensional contour point cloud along the guide rail axis, and project the point cloud within the slice onto a calibration coordinate plane perpendicular to the guide rail axis to obtain a two-dimensional cross-sectional contour point set. S312. Uniformly sample the two-dimensional cross-sectional contour point set according to the arc length to generate an N-dimensional cross-sectional contour sampling sequence; S313. Extract the gray-level gradient magnitude sequence from the two-dimensional texture image at the image position corresponding to the cross-sectional contour sampling point, and concatenate the gradient magnitude sequence with the cross-sectional contour sampling sequence to form a joint observation vector; S314. Perform integrity assessment on the joint observation vector: Calculate the average distance from each sampling point to its neighboring point set. If the average distance exceeds a preset threshold, the sampling point is marked as a missing point. The integrity test fails when the number of missing points exceeds N × 30%. S315. When the integrity is not passed, obtain the decoded geometric parameters of the first k adjacent sections, use the kinematic interpolation results of the adjacent section parameters as the predicted profile of the current section, replace the missing points with the corresponding sampling points in the predicted profile, and reconstruct the joint observation vector.

[0009] Preferably, in step S3, sparse decomposition is performed on the dictionary based on the observation vector and with the minimum residual as the criterion, to produce a coefficient vector. Specifically, this involves solving the following optimization problem: ; in, A dictionary of geometric atoms. For joint observation vectors, Let be the coefficient vector to be solved. It is the L2 norm. Let L0 norm represent the number of non-zero elements. This is the preset upper limit for sparsity; The obtained coefficient vector There is only one non-zero coefficient. At that time, use dictionary indexes Directly decode cross-sectional geometric parameters ,in Indexing when building a dictionary The corresponding parameter vector; The obtained coefficient vector When there are m non-zero coefficients, the amplitude of each non-zero coefficient is used as the weight for weighted fusion decoding: ; in, Non-zero coefficients Corresponding dictionary index The associated cross-sectional geometric parameter vector.

[0010] Preferably, in step S3, the decoding parameters are verified across modalities using two-dimensional texture image data. If the verification fails, a review is triggered. Specifically, this includes: S331. Obtain the decoded cross-sectional geometric parameters The theoretical cross-sectional profile is obtained by inverse calculation of geometric topological relationships, and then the theoretical cross-sectional profile is projected onto the two-dimensional texture image coordinate system through the calibration parameter matrix to obtain the theoretical edge position curve. ; S332. In a two-dimensional texture image, the curve at the theoretical edge position... In the direction of the normal, with Pixels are used to extract local maxima of the actual grayscale gradient, forming a set of actual edge points. ; S333. Calculate the theoretical edge curve With actual edge point set Average Hausdorff distance between them: ; S334. Will With preset verification threshold In comparison, if If the verification passes, then... Confirmed as valid decoding parameters for the current section; like If the verification fails, the current section's review and data collection process will be triggered.

[0011] Preferably, in step S4, constructing a graph optimization model for the decoding parameters, with cross-sectional poses as nodes, adjacent cross-sectional shapes as local edges, and standard cross-sectional dimensions as prior constraints, specifically includes: S401. Construct a node set V, where each node in node set V... For the six-degree-of-freedom pose corresponding to the i-th cross section, the initial value of the node is taken as the coarse registration pose of the point cloud of the adjacent cross section; S402. Constructing Local Constraint Boundary Sets ,side Connect adjacent cross-section nodes and The local error term is: ; in , Let i and j be the six-degree-of-freedom poses of the i-th and j-th sections. The relative pose observations from i to j obtained from ICP registration. For Lie algebra mappings; S403. Constructing the Prior Constraint Boundary Set The edges connect each cross-section node to the virtual reference node of the standard cross-section, and the prior error term... For nodes The deviation vector between the corresponding decoding parameters and the standard cross-sectional dimensions, and the weight factor ω of the prior constraint edge are determined according to the guide rail accuracy level; S404. Construct and solve the total cost function: ; in This is the information matrix for local observations; Using the coarse registration pose as the initial value, iteratively minimize the total cost function, and output the global optimal pose of each cross section node after convergence.

[0012] Preferably, step S5 specifically includes: S501. Longitudinal temperature acquisition points are arranged along the length of the guide rail surface with denser spacing at both ends and sparser spacing in the middle. At the same time, at least one circumferential temperature acquisition point is arranged on the guide shoe surface, guide surface, bottom surface and back surface of the guide rail cross section, forming a two-dimensional temperature acquisition array with one-dimensional longitudinal and one-dimensional circumferential cross section. S502. Based on the temperature measurements from the longitudinal temperature acquisition points, solve the one-dimensional heat conduction equation along the length of the guide rail to obtain the longitudinal temperature field distribution function. ; S503. Based on the temperature measurements from the circumferential temperature acquisition points of the cross-section, calculate the temperature difference between the working surfaces within the cross-section to obtain the cross-sectional temperature gradient: ; in and Positions Temperature values ​​of the cross section relative to the two working surfaces; S504. Based on the longitudinal temperature field T l ( Given the linear expansion coefficient α of the guide rail material, calculate the thermal deformation correction in the length direction: ; in For calibration reference temperature; S505. Based on the cross-sectional temperature gradient Given the guide rail section height h and the material linear expansion coefficient α, calculate the correction amount for the section warpage angle: ; in The arc length distance between the circumferential temperature sampling points of the cross section; S506. The thermal deformation correction in the length direction is superimposed on the translation component of the global optimal pose of the cross section, and the warpage correction in the cross section is superimposed on the rotation component to complete the temperature compensation.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. The proposed high-precision measurement method for elevator guide rail component dimensions based on visual intelligence uses a calibration target to perform joint calibration on multiple visual acquisition channels, including two-dimensional area array images, three-dimensional line laser contours, and telecentric end face images. A reference block is set to realize online detection of calibration drift and automatic recalibration. A unified spatial coordinate mapping for multiple channels is established, which realizes accurate spatial alignment of multimodal visual data and ensures the calibration accuracy and detection stability of the system during long-term operation. 2. The proposed method for high-precision measurement of elevator guide rail component dimensions based on visual intelligence constructs a guide rail geometric atom dictionary and performs sparse decomposition based on multimodal joint observation vectors to complete parameter decoding. Combined with a two-dimensional texture image cross-modal verification and anomaly verification mechanism, it achieves high-precision calculation of cross-sectional geometric parameters under complex working conditions, effectively improving the robustness and reliability of parameter decoding when there is interference such as reflection and oil on the workpiece surface. 3. The proposed high-precision measurement method for elevator guide rail components based on visual intelligence solves the global optimal pose by constructing a graph optimization model with cross-sectional pose as nodes and adjacent cross-sectional shape and position as constraints. It also calculates the cross-sectional thermal deformation correction amount by combining a non-uniform temperature field model to complete temperature compensation, thereby achieving high-precision reconstruction of the full-size parameters of long guide rails. This effectively suppresses the influence of multi-section splicing cumulative error and environmental thermal deformation on the measurement accuracy of shape and position tolerances. Attached Figure Description

[0014] Figure 1This is a schematic diagram of the overall process of the high-precision measurement method for elevator guide rail components based on visual intelligence proposed in this invention. Figure 2 This is a schematic diagram of the steps for constructing the guide rail geometry atom dictionary, sparse decoding of cross-sectional parameters, and cross-modal verification in this invention; Figure 3 This is a schematic diagram of the steps for optimizing the global pose diagram of the cross section and compensating for thermal deformation temperature in this invention. Detailed Implementation

[0015] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0016] Reference Figure 1-3 As shown, the high-precision measurement method for elevator guide rail component dimensions based on visual intelligence proposed in this invention includes: S1. Use a calibration target to perform joint calibration on the multi-source visual acquisition channels, establish a unified spatial coordinate mapping between the two-dimensional area array image channel, the three-dimensional line laser contour channel and the telecentric end face image channel, and generate a joint calibration parameter matrix; Furthermore, step S1 specifically includes: S101. Fabricate a three-dimensional calibration target that is geometrically complementary to the T-section of the guide rail. Each working surface of the target is machined with a feature point array of known precise coordinates. The feature point array includes a checkerboard corner point array and a circular hole center point array. S102. Fix the calibration target to the guide rail support platform, and collect target images and point clouds through the two-dimensional area array image channel, the three-dimensional line laser contour channel and the telecentric end face image channel respectively, and extract the coordinates of feature points in each channel; S103. Using the known coordinates of the target feature point array as a reference, solve the homography mapping between the two-dimensional channel image coordinate system and the world coordinate system, the rigid body transformation between the three-dimensional channel coordinate system and the world coordinate system, and the scaling mapping between the telecentric channel image coordinate system and the world coordinate system respectively, and combine them into a joint calibration parameter matrix. S104. A reference block with known precise dimensions is fixedly set on the guide rail bearing platform. Every preset number of cross sections or preset time intervals, a single measurement of the reference block is automatically triggered. The measured value is compared with the known dimensions of the reference block to obtain the calibration drift amount. When the drift exceeds the preset threshold, the online recalibration process from S102 to S103 is automatically triggered.

[0017] The following describes the steps of the present invention using the commonly used elevator guide rail model T89 / B as a specific embodiment. This model has a standard cross-section width of 89mm, a height of 62mm, a web thickness of 10mm, and both the guide shoe surface inclination angle and the guide surface inclination angle are 90°. The material is Q235 carbon steel. In this embodiment, the three-dimensional calibration target is made of 7075 aviation aluminum alloy through CNC precision machining, and its T-shaped groove contour is precisely complementary to the cross-section of the T89 / B guide rail. A checkerboard corner point array is machined on the target guide shoe surface, with a corner point machining accuracy of ±0.002mm, used for two-dimensional area array image channel calibration. A triangular array of circular holes is machined on the target guide surface and bottom surface, with a hole center coordinate machining accuracy of ±0.003mm, used for three-dimensional line laser contour channel calibration. The calibration process involves horizontally mounting the calibration target on a guide rail platform, simultaneously triggering three channels to acquire multiple frames of target data, extracting the sub-pixel coordinates of feature points in each frame and averaging them to suppress random noise, and solving the mapping relationship between the image coordinate system and the world coordinate system for each channel. The 2D area image channel uses a homography mapping matrix, the 3D line laser contour channel uses a rigid body transformation matrix, and the telecentric end face image channel uses a scaling mapping matrix; these are combined to form a joint calibration parameter matrix. After joint calibration, the feature points of each channel are back-projected to the world coordinate system using the calibration parameter matrix to verify the calibration accuracy.

[0018] For online drift detection, in this embodiment, the fixed reference block is a zero-expansion microcrystalline glass block, which is fixedly installed in the non-moving area of ​​the guide rail bearing platform, and the dimensional accuracy of each working surface is ±0.001mm. The system automatically triggers a reference block measurement every time a certain number of sections are measured or at a preset time interval. The measured value is compared with the standard value to obtain the calibration drift amount. When the drift amount exceeds a preset threshold, the online recalibration process is automatically triggered. After completion, the normal measurement queue continues.

[0019] S2. Synchronously acquire two-dimensional texture image data, three-dimensional contour point cloud data and telecentric end face image data of the same cross section of the elevator guide rail to form a multimodal synchronous visual dataset; S3. Extract the cross-sectional profile point set perpendicular to the guide rail axis, project it onto the calibration coordinate system to generate an observation vector, and use the guide rail T-section width, height, web thickness, guide shoe surface inclination angle, and guide surface inclination angle as parameter variables. Discretely sample within the manufacturing tolerance to construct a guide rail geometric atom dictionary composed of cross-sectional profile point sets. Based on the observation vector and using the minimum residual as the criterion, perform sparse decomposition on the dictionary to produce a coefficient vector. When the coefficient vector has only one non-zero coefficient, the parameter is decoded by index. When there are multiple non-zero coefficients, the parameter is decoded by coefficient magnitude weighted fusion. The decoded parameter is verified across modes using two-dimensional texture image data. If the verification fails, a review is triggered. Furthermore, step S3 involves constructing a dictionary of guide rail geometry atoms, specifically including: S301. Determine if the current guide rail model is known; If the model number is known, obtain the standard cross-sectional dimensions and manufacturing tolerance range corresponding to that model number; If the model is unknown, the initial geometric parameters of the T-shaped section are directly fitted from the three-dimensional point cloud of several sections of the first section of the guide rail. The initial parameter spatial range is determined with the fitted value as the center and the preset loose tolerance as the radius. S302. Uniformly discretely sample each parameter variable in the parameter space according to a preset initial resolution, combine the discrete values ​​of each variable by Cartesian product to generate an initial parameter vector set, and analyze the corresponding complete cross-sectional contour point set according to the cross-sectional geometric topology relationship for each parameter vector in the set to construct an initial geometric atom dictionary; S303. During the measurement process of the subsequent cross-section of the guide rail, based on the statistical mean and variance of the decoded parameters, the sampling range of each parameter variable is dynamically narrowed and the sampling resolution is improved, the geometric atom dictionary is updated, and an iterative update mechanism in which the dictionary adaptively shrinks with the measurement process is formed.

[0020] In this embodiment, the manufacturing tolerances of each parameter of the T89 / B guide rail are determined according to the GB / T 22562 standard. The discrete sampling resolutions for each parameter are 0.02 mm for width, 0.02 mm for height, 0.01 mm for web thickness, 0.01° for guide shoe surface inclination, and 0.01° for guide surface inclination. The discrete sampled values ​​of the five parameter variables are combined by Cartesian product, resulting in a dictionary of approximately tens of millions of atoms. The parameter vector corresponding to each atom in the dictionary is determined by linear mapping of the discrete indices of its five parameter components. The method for analytically calculating the complete cross-sectional profile point set from the parameter vectors is as follows: taking the lower left corner of the cross-section as the origin, the coordinates of the endpoints of each working surface are calculated sequentially according to the geometric topological relationship of the guide rail T-shaped cross-section, forming a closed cross-sectional profile polygon. The points are sampled uniformly along each edge with an arc length to obtain the cross-sectional profile point set corresponding to that atom. The dictionary construction is a one-time offline preprocessing operation. After construction, a bidirectional mapping table from the dictionary index to the parameter vector is established. During actual online measurement, parameter decoding is completed by looking up the table through the index, without affecting the real-time measurement cycle.

[0021] When the model of the guide rail to be tested is unknown, such as during mixed material production or the first testing of a new model, the system extracts the four working surfaces of the T-section from the 3D point cloud of several cross-sections of the first section of the guide rail using least-squares linear fitting. Then, it calculates the fitted values ​​of the cross-section width, height, web thickness, and inclination angle as the estimation center for the initial parameters. The initial parameter space is constructed with the fitted values ​​as the center and a preset loose tolerance as the radius. The initial sampling resolution is set to a relatively coarse value to control the initial dictionary size. As the measurement progresses, the geometric parameters of the decoded cross-sections gradually accumulate. When the number of accumulated cross-sections reaches a preset number, the mean and standard deviation of each parameter are statistically analyzed. The sampling range of each parameter is narrowed to the mean plus or minus several times the standard deviation, and the sampling resolution is increased, updating the geometric atom dictionary. As the measurement progresses, the dictionary gradually shrinks from the loose initial range to the precise parameter space, and the number of dictionary atoms decreases significantly, with the decoding accuracy gradually converging to the guide rail standard tolerance requirement range.

[0022] Furthermore, in step S3, the set of cross-sectional profile points perpendicular to the guide rail axis is extracted and projected onto the calibration coordinate system to generate the observation vector, specifically including: S311. Extract a slice point cloud of a preset thickness from the three-dimensional contour point cloud along the guide rail axis, and project the point cloud within the slice onto a calibration coordinate plane perpendicular to the guide rail axis to obtain a two-dimensional cross-sectional contour point set. S312. Uniformly sample the two-dimensional cross-sectional contour point set according to the arc length to generate an N-dimensional cross-sectional contour sampling sequence; S313. Extract the gray-level gradient magnitude sequence from the two-dimensional texture image at the image position corresponding to the cross-sectional contour sampling point, and concatenate the gradient magnitude sequence with the cross-sectional contour sampling sequence to form a joint observation vector; S314. Perform integrity assessment on the joint observation vector: Calculate the average distance from each sampling point to its neighboring point set. If the average distance exceeds a preset threshold, the sampling point is marked as a missing point. The integrity test fails when the number of missing points exceeds N × 30%. S315. When the integrity is not passed, obtain the decoded geometric parameters of the first k adjacent sections, use the kinematic interpolation results of the adjacent section parameters as the predicted profile of the current section, replace the missing points with the corresponding sampling points in the predicted profile, and reconstruct the joint observation vector.

[0023] In this embodiment, the specific process of observation vector extraction is as follows: A slice of point cloud with a preset thickness, such as 2mm, is extracted from the 3D contour point cloud along the guide rail axis. The point cloud within the slice is projected onto a calibration coordinate plane perpendicular to the guide rail axis to obtain a 2D cross-sectional contour scatter set. This scatter set is uniformly sampled according to the arc length. The number of sampling points N is determined based on the cross-sectional perimeter and the required accuracy. For example, for the T89 / B cross-section with a perimeter of approximately 260mm, N is set to 200, resulting in a cross-sectional contour sampling sequence. For each sampling point, its corresponding pixel coordinates in the 2D texture image are calculated using the homography mapping in the calibration parameter matrix. The grayscale gradient magnitude is extracted within the neighborhood of these pixel coordinates. The gradient magnitude sequence is concatenated with the cross-sectional contour sampling sequence to form a joint observation vector. The gradient magnitude sequence is weighted by a normalized weighting coefficient λ to balance the contribution ratio of geometric and texture information.

[0024] The specific process of integrity assessment is as follows: Calculate the average Euclidean distance from each sampling point to its neighborhood point set. If the average distance exceeds a preset distance threshold, the point is marked as a missing point. In this embodiment, the neighborhood point set consists of five adjacent sampling points before and after each sampling point. When the proportion of missing points exceeds a preset ratio, such as 30%, the integrity of the current cross-sectional contour is determined to be unsatisfactory, triggering the reconstruction process. During reconstruction, the decoded parameters of the first k adjacent cross-sections (e.g., k = 3) are obtained by linear interpolation to obtain the predicted contour. The missing points are replaced with the sampling points at the corresponding positions in the predicted contour. After reconstructing the joint observation vector, sparse decomposition is re-executed.

[0025] Furthermore, in step S3, a sparse decomposition is performed on the dictionary based on the observation vector and with the minimum residual as the criterion, to produce a coefficient vector. Specifically, this involves solving the following optimization problem: ; in, A dictionary of geometric atoms. For joint observation vectors, Let be the coefficient vector to be solved. It is the L2 norm. Let L0 norm represent the number of non-zero elements. This is the preset upper limit for sparsity; The obtained coefficient vector There is only one non-zero coefficient. At that time, use dictionary indexes Directly decode cross-sectional geometric parameters ,in Indexing when building a dictionary The corresponding parameter vector; The obtained coefficient vector When there are m non-zero coefficients, the amplitude of each non-zero coefficient is used as the weight for weighted fusion decoding: ; in, Non-zero coefficients Corresponding dictionary index The associated cross-sectional geometric parameter vector.

[0026] In this embodiment, the sparsity upper limit K of the sparse decomposition is set to a preset value, such as 3, which means that it is assumed that each cross-sectional observation vector can be represented by a linear combination of no more than K atoms in the geometric atom dictionary. The solution adopts a greedy iterative strategy based on residual minimization. In each iteration, the dictionary atom with the largest absolute value of the inner product with the current residual is added to the support set, the coefficient vector is updated, and the residual is recalculated. The iteration stops when the residual is lower than the preset convergence threshold or the size of the support set reaches K.

[0027] The following two specific cross-section decoding examples illustrate the parameter decoding process. Example 1: After sparse decomposition, the coefficient vector of a certain cross-section has only one non-zero coefficient. Its dictionary index corresponds to the parameter vector with a width of 89.02 mm, a height of 62.00 mm, a web thickness of 10.01 mm, a guide shoe surface inclination angle of 90.00°, and a guide surface inclination angle of 89.99°. The directly decoded cross-section geometric parameters are the above values. The cross-modal verification Hausdorff distance is 0.22 pixels, and the verification passes. Example 2: Due to local blurring at the edges, a cross-section exhibits two non-zero coefficients after sparse decomposition. The first non-zero coefficient, 0.72, corresponds to parameters of width 89.00 mm, height 62.01 mm, web thickness 9.99 mm, guide shoe inclination angle 89.99°, and guide surface inclination angle 90.00°. The second non-zero coefficient, 0.31, corresponds to parameters of width 89.04 mm, height 61.98 mm, web thickness 10.02 mm, guide shoe inclination angle 90.01°, and guide surface inclination angle 90.00°. Weighted fusion decoding using the amplitudes of each non-zero coefficient yields a cross-section with a width of 89.01 mm, a height of 62.00 mm, a web thickness of 10.00 mm, a guide shoe inclination angle of 90.00°, and a guide surface inclination angle of 90.00°. The Hausdorff distance in the cross-modal verification is 0.41 pixels, and the verification passes.

[0028] Furthermore, in step S3, the decoding parameters are verified across modalities using two-dimensional texture image data. If the verification fails, a review is triggered, specifically including: S331. Obtain the decoded cross-sectional geometric parameters The theoretical cross-sectional profile is obtained by inverse calculation of geometric topological relationships, and then the theoretical cross-sectional profile is projected onto the two-dimensional texture image coordinate system through the calibration parameter matrix to obtain the theoretical edge position curve. ; S332. In a two-dimensional texture image, the curve at the theoretical edge position... In the direction of the normal, with Pixels are used to extract local maxima of the actual grayscale gradient, forming a set of actual edge points. ; S333. Calculate the theoretical edge curve With actual edge point set Average Hausdorff distance between them: ; S334. Will With preset verification threshold In comparison, if If the verification passes, then... Confirmed as valid decoding parameters for the current section; like If the verification fails, the current section's review and data collection process will be triggered.

[0029] In this embodiment, the search bandwidth Δ and verification threshold ε for cross-modal verification are set according to the imaging resolution and measurement accuracy requirements of the guide rail surface. For example, Δ is set to ±5 pixels, corresponding to a range of approximately ±0.15 mm on the guide rail surface, and ε is set to 0.5 pixels. A complete measurement of the entire length of a T89 / B guide rail includes 250 cross-sections, most of which pass verification on the first attempt. A few cross-sections, due to severe local reflections or oil contamination, cause the Hausdorff distance to exceed the threshold, automatically triggering a re-verification acquisition. During the re-verification acquisition process, the system fine-tunes the lighting parameters or changes the camera exposure time before re-acquiring and verifying, and all re-verified cross-sections pass the second verification.

[0030] S4. Repeat S2 to S3 along the guide rail length direction with a preset step size to construct a graph optimization model for the decoding parameters, with the cross-section pose as the node, the adjacent cross-section shape and position as the local edge, and the standard cross-section size as the prior constraint. Iteratively solve the global optimal pose of each cross-section with the coarse registration pose as the initial value. Furthermore, in step S4, a graph optimization model is constructed for the decoding parameters, with cross-sectional poses as nodes, adjacent cross-sectional shapes as local edges, and standard cross-sectional dimensions as prior constraints. This specifically includes: S401. Construct a node set V, where each node in node set V... For the six-degree-of-freedom pose corresponding to the i-th cross section, the initial value of the node is taken as the coarse registration pose of the point cloud of the adjacent cross section; S402. Constructing Local Constraint Boundary Sets ,side Connect adjacent cross-section nodes and The local error term is: ; in , Let i and j be the six-degree-of-freedom poses of the i-th and j-th sections. The relative pose observations from i to j obtained from ICP registration. The pose deviation is transformed into an error vector using Lie algebra mapping. S403. Constructing the Prior Constraint Boundary Set The edges connect each cross-section node to the virtual reference node of the standard cross-section, and the prior error term... For nodes The deviation vector between the corresponding decoding parameters and the standard cross-sectional dimensions, and the weight factor ω of the prior constraint edge are determined according to the guide rail accuracy level; S404. Construct and solve the total cost function: ; in This is the information matrix of local observations. This represents the six-degree-of-freedom pose transformation matrix corresponding to the 1st to nth guide rail sections. Let i and j be the nodes of the i-th and j-th cross sections in the graph model, respectively, and let j be the corresponding pose matrices. , ; Using the coarse registration pose as the initial value, iteratively minimize the total cost function, and output the global optimal pose of each cross section node after convergence.

[0031] In this embodiment, the guide rail is 5000mm long, the scanning step size is 20mm, and a total of 250 cross sections are obtained. The graph optimization model has 250 nodes. There are 249 local constraint edges connecting the nodes of the i-th cross section to the (i+1)-th cross section, and 250 prior constraint edges connecting the nodes of each cross section to the virtual reference nodes of the standard cross section. First, the ICP algorithm is used to coarsely register the point clouds of adjacent cross sections to obtain the initial pose of each cross section node. The root mean square error of the ICP registration of adjacent cross section point clouds is approximately 0.015mm to 0.025mm. Using the coarsely registered pose as the initial value, the total cost function is iteratively minimized until convergence.

[0032] The comparison results before and after iterative optimization show the effectiveness of global graph optimization: before optimization, the 250 sections were directly chained together using ICP coarse registration poses, and the cumulative splicing error of the first and last sections reached 0.18 mm; after optimization, the cumulative error of the first and last sections decreased to 0.03 mm. The largest pose correction of each section node occurred in the middle region of the guide rail, which is consistent with the magnitude of the natural deflection deformation of the guide rail due to its own weight. This indicates that global graph optimization successfully decoupled the cumulative splicing error from the flexible deformation of the guide rail itself, avoiding misjudging the natural deflection as a geometric deviation.

[0033] S5. Arrange temperature acquisition arrays on the guide rail surface with denser spacing at both ends and sparser spacing in the middle to construct a non-uniform temperature field model, calculate the thermal deformation correction amount for each section and perform temperature compensation. Furthermore, step S5 specifically includes: S501. Longitudinal temperature acquisition points are arranged along the length of the guide rail surface with denser spacing at both ends and sparser spacing in the middle. At the same time, at least one circumferential temperature acquisition point is arranged on the guide shoe surface, guide surface, bottom surface and back surface of the guide rail cross section, forming a two-dimensional temperature acquisition array with one-dimensional longitudinal and one-dimensional circumferential cross section. S502. Based on the temperature measurements from the longitudinal temperature acquisition points, solve the one-dimensional heat conduction equation along the length of the guide rail to obtain the longitudinal temperature field distribution function. ; S503. Based on the temperature measurements from the circumferential temperature acquisition points of the cross-section, calculate the temperature difference between the working surfaces within the cross-section to obtain the cross-sectional temperature gradient: ; in and Positions Temperature values ​​of the cross section relative to the two working surfaces; S504. Based on the longitudinal temperature field T l ( Given the linear expansion coefficient α of the guide rail material, calculate the thermal deformation correction in the length direction: ; in For calibration reference temperature; S505. Based on the cross-sectional temperature gradient Given the guide rail section height h and the material linear expansion coefficient α, calculate the correction amount for the section warpage angle: ; in The arc length distance between the circumferential temperature sampling points of the cross section; S506. The thermal deformation correction in the length direction is superimposed on the translation component of the global optimal pose of the cross section, and the warpage correction in the cross section is superimposed on the rotation component to complete the temperature compensation.

[0034] In this embodiment, taking the T89 / B guide rail made of Q235 carbon steel as an example, the coefficient of linear expansion α is 11.7 × 10⁻⁶. -6 / °C, with the calibrated reference temperature T0 set to 20°C. Longitudinal temperature acquisition points are arranged along a 5m length of the guide rail, with a spacing of 200mm at the ends and 500mm in the middle, for a total of 15 acquisition points. Simultaneously, one circumferential temperature acquisition point is placed on each cross-section on the guide shoe surface, right guide surface, bottom surface, and left guide surface, forming a two-dimensional temperature acquisition array consisting of a one-dimensional longitudinal dimension and a one-dimensional circumferential dimension across the cross-sections.

[0035] In one actual measurement, the ambient temperature in the workshop was higher than the reference temperature, resulting in a temperature difference between the two ends of the guide rail. The longitudinal temperature field distribution was obtained by solving the one-dimensional heat conduction equation. At the end of the guide rail furthest from the heat source, the accumulated longitudinal thermal deformation correction amount could reach more than 0.5 mm. If not compensated, it would be misjudged as the guide rail straightness being out of tolerance.

[0036] Regarding cross-sectional warpage compensation, taking a cross-section located in the middle of the guide rail as an example, there is a cross-sectional temperature gradient of approximately 2°C between the guide shoe surface temperature and the bottom surface temperature. The cross-sectional height is approximately 62 mm, and the arc distance between circumferential temperature sampling points is approximately 130 mm. The calculated cross-sectional warpage angle correction is on the order of approximately 0.003°. Although the displacement deviation caused in the guide rail height direction is small in magnitude for a single cross-section, it can produce a considerable systematic deviation in the overall straightness assessment of the guide rail when accumulated across multiple cross-sections along the entire length. Two-dimensional temperature compensation is achieved by superimposing the thermal deformation correction in the length direction onto the translation component of the globally optimal pose of the cross-section, and by superimposing the cross-sectional warpage angle correction onto the rotation component.

[0037] S6. After fusing the compensated geometric data and texture image, calculate the full-size parameters of the guide rail, compare them with the standard tolerance, and generate a full-size measurement report.

[0038] Furthermore, step S6 includes the following: S7. Compare the measured values ​​of each dimensional parameter in the full-size measurement report with the corresponding standard tolerance thresholds to determine the pass / fail status of each dimensional parameter; For dimensional parameters that are determined to be out of tolerance, the correction direction is determined based on the sign of the out-of-tolerance amount, and the correction amount is determined based on the out-of-tolerance amplitude, generating a correction work order that includes the out-of-tolerance location, out-of-tolerance amount, correction direction, and correction amount. The calibration work order is fed back to the guide rail straightening or milling process, forming a closed-loop quality control link of measurement-judgment-feedback-calibration.

[0039] In this embodiment, the full-dimensional measurement report for the T89 / B guide rail includes seven parameters: guide rail width, height, web thickness, guide shoe surface straightness, guide surface straightness, bottom surface straightness, and guide rail torsion. Each parameter is compared one by one with the corresponding tolerance threshold specified in the GB / T 22562 standard. In a certain test, all dimensional parameters fell within the tolerance range, and the guide rail was deemed a qualified product. The report also generates a color chromatogram of the deviation curves for each dimensional parameter along the guide rail's length, visually indicating the direction and magnitude of the deviation at each cross-sectional position using color gradients.

[0040] The straightness of each working surface of the guide rail is evaluated using the least squares method: based on the globally optimal pose of each section, the feature point sequence of the corresponding working surface is extracted, a full-length reference straight line is fitted, the normal distance from each feature point to the reference straight line is calculated, and the maximum value is taken as the straightness error of the working surface. The guide rail torsion is calculated by the torsion angle of each section around the guide rail axis, and the maximum difference of the torsion angle over the entire length is taken, combined with the guide rail width, to convert into the torsion evaluation value at the corresponding position.

[0041] Statistical analysis of measurement data from multiple guide rails in the same batch reveals a systematic deviation in a certain dimensional parameter. For example, if the mean deviation of the straightness of the guide shoe surface is positive and exceeds half of the standard tolerance, it indicates a systematic process deviation in the upstream straightening process. After generating a correction work order based on this statistical conclusion and feeding it back to the straightening process to adjust the process parameters, the deviation of this parameter in the next batch of guide rails of the same model significantly decreases, and the batch pass rate is significantly improved.

[0042] The advantages of this invention are as follows: It systematically improves the comprehensive performance of elevator guide rail vision measurement from three dimensions: calibration accuracy, decoding robustness, and error suppression. Through multi-source vision channel joint calibration combined with online drift detection and automatic recalibration mechanisms, it ensures the spatial coordinate alignment accuracy and detection stability of the system during long-term operation. Relying on a sparse decoding method based on a geometric atom dictionary and multi-modal joint observation, combined with a cross-modal verification and review mechanism, it effectively improves the reliability of cross-sectional parameter calculation when interference exists on the workpiece surface. The introduction of global graph optimization to eliminate accumulated errors from multi-section splicing, coupled with thermal deformation compensation for non-uniform temperature fields, further improves the measurement accuracy of form and position tolerances for long guide rails, enabling it to meet the high-precision, high-efficiency, full-size quality control requirements of industrial sites.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A high-precision measurement method for elevator guide rail component dimensions based on visual intelligence, characterized in that, include: S1. Use a calibration target to perform joint calibration on the multi-source visual acquisition channels, establish a unified spatial coordinate mapping between the two-dimensional area array image channel, the three-dimensional line laser contour channel and the telecentric end face image channel, and generate a joint calibration parameter matrix; S2. Synchronously acquire two-dimensional texture image data, three-dimensional contour point cloud data and telecentric end face image data of the same cross section of the elevator guide rail to form a multimodal synchronous visual dataset; S3. Extract the cross-sectional profile point set perpendicular to the guide rail axis, project it onto the calibration coordinate system to generate an observation vector, and use the guide rail T-section width, height, web thickness, guide shoe surface inclination angle, and guide surface inclination angle as parameter variables. Discretely sample within the manufacturing tolerance to construct a guide rail geometric atom dictionary composed of cross-sectional profile point sets. Based on the observation vector and using the minimum residual as the criterion, perform sparse decomposition on the dictionary to produce a coefficient vector. When the coefficient vector has only one non-zero coefficient, the parameter is decoded by index. When there are multiple non-zero coefficients, the parameter is decoded by coefficient magnitude weighted fusion. The decoded parameter is verified across modes using two-dimensional texture image data. If the verification fails, a review is triggered. S4. Repeat S2 to S3 along the guide rail length direction with a preset step size to construct a graph optimization model for the decoding parameters, with the cross-section pose as the node, the adjacent cross-section shape and position as the local edge, and the standard cross-section size as the prior constraint. Iteratively solve the global optimal pose of each cross-section with the coarse registration pose as the initial value. S5. Arrange temperature acquisition arrays on the guide rail surface with denser spacing at both ends and sparser spacing in the middle to construct a non-uniform temperature field model, calculate the thermal deformation correction amount for each section and perform temperature compensation. S6. After fusing the compensated geometric data and texture image, calculate the full-size parameters of the guide rail, compare them with the standard tolerance, and generate a full-size measurement report.

2. The high-precision measurement method for elevator guide rail component dimensions based on visual intelligence according to claim 1, characterized in that, Step S1 specifically includes: S101. Fabricate a three-dimensional calibration target that is geometrically complementary to the T-section of the guide rail. Each working surface of the target is machined with a feature point array with known precise coordinates. The feature point array includes a checkerboard corner point array and a circular hole center point array. S102. Fix the calibration target to the guide rail support platform, and collect target images and point clouds through the two-dimensional area array image channel, the three-dimensional line laser contour channel and the telecentric end face image channel respectively, and extract the coordinates of feature points in each channel; S103. Using the known coordinates of the target feature point array as a reference, solve the homography mapping between the two-dimensional channel image coordinate system and the world coordinate system, the rigid body transformation between the three-dimensional channel coordinate system and the world coordinate system, and the scaling mapping between the telecentric channel image coordinate system and the world coordinate system respectively, and combine them into a joint calibration parameter matrix. S104. A reference block with known precise dimensions is fixedly set on the guide rail bearing platform. Every preset number of cross sections or preset time intervals, a single measurement of the reference block is automatically triggered. The measured value is compared with the known dimensions of the reference block to obtain the calibration drift amount. When the drift exceeds the preset threshold, the online recalibration process from S102 to S103 is automatically triggered.

3. The high-precision measurement method for elevator guide rail component dimensions based on visual intelligence according to claim 1, characterized in that, Step S3 involves constructing the guide geometry atom dictionary, specifically including: S301. Determine if the current guide rail model is known; If the model number is known, obtain the standard cross-sectional dimensions and manufacturing tolerance range corresponding to that model number; If the model is unknown, the initial geometric parameters of the T-shaped section are directly fitted from the three-dimensional point cloud of several sections of the first section of the guide rail. The initial parameter spatial range is determined with the fitted value as the center and the preset loose tolerance as the radius. S302. Uniformly discretely sample each parameter variable in the parameter space according to a preset initial resolution, combine the discrete values ​​of each variable by Cartesian product to generate an initial parameter vector set, and analyze the corresponding complete cross-sectional contour point set according to the cross-sectional geometric topology relationship for each parameter vector in the set to construct an initial geometric atom dictionary; S303. During the measurement process of the subsequent cross-section of the guide rail, based on the statistical mean and variance of the decoded parameters, the sampling range of each parameter variable is dynamically narrowed and the sampling resolution is improved, the geometric atom dictionary is updated, and an iterative update mechanism in which the dictionary adaptively shrinks with the measurement process is formed.

4. The high-precision measurement method for elevator guide rail component dimensions based on visual intelligence according to claim 1, characterized in that, Step S3 involves extracting the set of cross-sectional profile points perpendicular to the guide rail axis and projecting them onto the calibration coordinate system to generate an observation vector. Specifically, this includes: S311. Extract a slice point cloud of a preset thickness from the three-dimensional contour point cloud along the guide rail axis, and project the point cloud within the slice onto a calibration coordinate plane perpendicular to the guide rail axis to obtain a two-dimensional cross-sectional contour point set. S312. Uniformly sample the two-dimensional cross-sectional contour point set according to the arc length to generate an N-dimensional cross-sectional contour sampling sequence; S313. Extract the gray-level gradient magnitude sequence from the two-dimensional texture image at the image position corresponding to the cross-sectional contour sampling point, and concatenate the gradient magnitude sequence with the cross-sectional contour sampling sequence to form a joint observation vector; S314. Perform integrity assessment on the joint observation vector: Calculate the average distance from each sampling point to its neighboring point set. If the average distance exceeds a preset threshold, the sampling point is marked as a missing point. The integrity test fails when the number of missing points exceeds N × 30%. S315. When the integrity is not passed, obtain the decoded geometric parameters of the first k adjacent sections, use the kinematic interpolation results of the adjacent section parameters as the predicted profile of the current section, replace the missing points with the corresponding sampling points in the predicted profile, and reconstruct the joint observation vector.

5. The high-precision measurement method for elevator guide rail component dimensions based on visual intelligence according to claim 4, characterized in that, In step S3, sparse decomposition is performed on the dictionary based on the observation vector and with the minimum residual as the criterion, to produce a coefficient vector. Specifically, the following optimization problem is solved: ； in, A dictionary of geometric atoms. For joint observation vectors, Let be the coefficient vector to be solved. It is the L2 norm. Let L0 norm represent the number of non-zero elements. This is the preset upper limit for sparsity; The obtained coefficient vector There is only one non-zero coefficient. At that time, use dictionary indexes Directly decode cross-sectional geometric parameters ,in Indexing when building a dictionary The corresponding parameter vector; The obtained coefficient vector When there are m non-zero coefficients, the amplitude of each non-zero coefficient is used as the weight for weighted fusion decoding: ； in, Non-zero coefficients Corresponding dictionary index The associated cross-sectional geometric parameter vector.

6. The high-precision measurement method for elevator guide rail component dimensions based on visual intelligence according to claim 5, characterized in that, Step S3 involves cross-modal verification of decoding parameters using 2D texture image data. Verification failure triggers a review, specifically including: S331. Obtain the decoded cross-sectional geometric parameters The theoretical cross-sectional profile is obtained by inverse calculation of geometric topological relationships, and then the theoretical cross-sectional profile is projected onto the two-dimensional texture image coordinate system through the calibration parameter matrix to obtain the theoretical edge position curve. ; S332. In a two-dimensional texture image, the curve at the theoretical edge position... In the direction of the normal, with Pixels are used to extract local maxima of the actual grayscale gradient, forming a set of actual edge points. ; S333. Calculate the theoretical edge curve With actual edge point set Average Hausdorff distance between them: ； S334. Will With preset verification threshold In comparison, if If the verification passes, then... Confirmed as valid decoding parameters for the current section; like If the verification fails, the current section's review and data collection process will be triggered.

7. The high-precision measurement method for elevator guide rail component dimensions based on visual intelligence according to claim 1, characterized in that, Step S4 involves constructing a graph optimization model for the decoding parameters, using cross-sectional poses as nodes, adjacent cross-sectional shapes as local edges, and standard cross-sectional dimensions as prior constraints. This specifically includes: S401. Construct a node set V, where each node in node set V... For the six-degree-of-freedom pose corresponding to the i-th cross section, the initial value of the node is taken as the coarse registration pose of the point cloud of the adjacent cross section; S402. Constructing Local Constraint Boundary Sets ,side Connect adjacent cross-section nodes and The local error term is: ； in , Let i and j be the six-degree-of-freedom poses of the i-th and j-th sections. The relative pose observations from i to j obtained from ICP registration. For Lie algebra mappings; S403. Constructing the Prior Constraint Boundary Set The edges connect each cross-section node to the virtual reference node of the standard cross-section, and the prior error term... For nodes The deviation vector between the corresponding decoding parameters and the standard cross-sectional dimensions, and the weight factor ω of the prior constraint edge are determined according to the guide rail accuracy level; S404. Construct and solve the total cost function: ； in This is the information matrix for local observations; Using the coarse registration pose as the initial value, iteratively minimize the total cost function, and output the global optimal pose of each cross section node after convergence.

8. The high-precision measurement method for elevator guide rail component dimensions based on visual intelligence according to claim 1, characterized in that, Step S5 specifically includes: S501. Longitudinal temperature acquisition points are arranged along the length of the guide rail surface with denser spacing at both ends and sparser spacing in the middle. At the same time, at least one circumferential temperature acquisition point is arranged on the guide shoe surface, guide surface, bottom surface and back surface of the guide rail cross section, forming a two-dimensional temperature acquisition array with one-dimensional longitudinal and one-dimensional circumferential cross section. S502. Based on the temperature measurements from the longitudinal temperature acquisition points, solve the one-dimensional heat conduction equation along the length of the guide rail to obtain the longitudinal temperature field distribution function. ; S503. Based on the temperature measurements from the circumferential temperature acquisition points of the cross-section, calculate the temperature difference between the working surfaces within the cross-section to obtain the cross-sectional temperature gradient: ； in and Positions Temperature values ​​of the cross section relative to the two working surfaces; S504. Based on the longitudinal temperature field T l ( Given the linear expansion coefficient α of the guide rail material, calculate the thermal deformation correction in the length direction: ； in For calibrating the reference temperature; S505. Based on the cross-sectional temperature gradient Given the guide rail section height h and the material linear expansion coefficient α, calculate the correction amount for the section warpage angle: ； in The arc length distance between the circumferential temperature sampling points of the cross section; S506. The thermal deformation correction in the length direction is superimposed on the translation component of the global optimal pose of the cross section, and the warpage correction in the cross section is superimposed on the rotation component to complete the temperature compensation.

Images