A non-contact detection method and system for hollow fiber membrane

Abstract

The present invention provides a non-contact detection method and system for hollow fiber membranes, belonging to the field of joint visual inspection technology. The detection method includes: tightly covering the detection end of the hollow fiber membrane with a flexible film; two cameras arranged in front of the flexible film respectively collect point clouds from the detection end uncovered by the flexible film and the detection end covered by the flexible film, thereby obtaining a target point cloud and a source point cloud; extracting SIFT feature points from the source point cloud and the target point cloud respectively, and describing them using a 3D Scale (DSC) descriptor with a normal constraint; based on the SIFT feature points of the source and target point clouds and their corresponding 3D Scale (DSC) descriptors, performing coarse registration using a sampling consistency initial registration algorithm to obtain an initial transformation matrix; constructing a generalized iterative closest point optimization model with a normal vector constraint, and designing an objective function related to the initial transformation matrix to obtain an optimal transformation matrix, thereby locating the defect area. The present invention can effectively improve detection efficiency and detection rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of combined visual detection, and in particular relates to a non-contact detection method and system for hollow fiber membranes. Background Art

[0002] Hollow fiber membrane is a type of ultrafiltration membrane widely used in water treatment, biomedicine and other fields. Its integrity directly affects filtration performance, so its integrity needs to be tested. Traditional testing methods are basically wet testing or offline testing. Both methods are contact testing. Wet testing requires the hollow fiber membrane to be immersed in a specific liquid environment, which can easily affect the subsequent performance of the membrane due to liquid residue, and requires additional drying and other treatments after testing, which is a cumbersome process. Offline testing requires the hollow fiber membrane to be removed from the actual operating system, which not only interrupts the production process and affects the efficiency of continuous operation, but may also cause physical damage to the membrane due to disassembly and assembly operations. At the same time, contact testing is affected by the uniformity of the test medium, contact pressure, etc., and the detection accuracy is difficult to meet high-precision requirements. It is also not suitable for membrane material testing that is sensitive to the test environment. Therefore, non-contact measurement methods such as those described in patents with application numbers 202311719585.8 and 202411162166.3 have emerged, but they also have defects: the measurement time of the former is still long, the operation process is relatively cumbersome, and the uneven size of the gas molecules used is difficult to control; the latter uses three sets of binocular vision systems with high requirements for calibration accuracy. The calibration errors of the rotation matrix and translation vector will be amplified with the measurement distance, resulting in a decrease in defect positioning accuracy. It also relies on point cloud centroid clustering, does not integrate the geometric features of the membrane wire, and has insufficient discrimination for low-contrast defects. Summary of the Invention

[0003] The purpose of the present invention is to provide a non-contact detection method and system for hollow fiber membranes, which can effectively improve the detection efficiency and detection rate.

[0004] The present invention is achieved through the following technical solutions:

[0005] A non-contact detection method for hollow fiber membranes comprises the following steps:

[0006] Step S1: tightly covering the flexible film on the detection end of the hollow fiber membrane, and inputting gas from the air inlet end of the hollow fiber membrane to deform the flexible film;

[0007] Step S2: Two cameras disposed in front of the flexible film respectively collect point clouds from the detection end uncovered by the flexible film and the detection end covered by the flexible film, fuse the collected point clouds to obtain a target fused point cloud and a source fused point cloud, and perform streamlined preprocessing on the target fused point cloud and the source fused point cloud to obtain a target point cloud and a source point cloud, wherein the two cameras operate synchronously;

[0008] Step S3, extracting SIFT feature points of the source point cloud and the target point cloud respectively, and describing the SIFT feature points using a 3DSC descriptor with normal constraints;

[0009] Step S4: Based on the SIFT feature points of the source point cloud and the target point cloud and their corresponding 3DSC descriptors, a sampling consistency initial registration algorithm is used to perform coarse registration to obtain an initial transformation matrix;

[0010] Step S5: Construct a generalized iterative closest point optimization model including normal vector constraints, design an objective function related to the initial transformation matrix to obtain the optimal transformation matrix, use the optimal transformation matrix to transform the source point cloud, and locate the defect area based on the difference between the transformed result and the target point cloud.

[0011] Furthermore, in step S1, the air inlet end of the hollow fiber membrane is connected to an air pressure control module, which includes an air compressor and a pressure stabilizing device connected to the air compressor to maintain a constant air pressure in the hollow fiber membrane cavity, and the flexible film is made of a stretchable material.

[0012] Furthermore, in step S2, the two cameras are located in the same horizontal plane and form a set angle. After the two cameras are calibrated, the disparity map is calculated based on the calibration parameters of the two cameras and the collected point cloud through a stereo matching algorithm, and the three-dimensional point cloud of the detection end is reconstructed and generated. The three-dimensional point cloud is subjected to redundant point elimination and overlapping area optimization to obtain a source fusion point cloud.

[0013] Furthermore, in step S2, the simplified pre-processing of the source fusion point cloud specifically includes: establishing a kd-tree for the source fusion point cloud, and N Calculate the minimum bounding box containing the source fused point cloud and divide the minimum bounding box into of For each voxel, use kd-tree to traverse each voxel, calculate the center of gravity of each non-empty voxel, find the nearest neighbor point cloud of the center of gravity, and replace all point clouds in the voxel with the nearest neighbor point cloud to obtain the source point cloud, where, To adjust the scale factor of the side length according to the number of point clouds, s is the proportionality coefficient, L x 、 L y 、 L z For the minimum bounding box x 、 y 、 z The length in the direction, , , , is the ceiling function.

[0014] Furthermore, in step S3, for the source point cloud, a multi-resolution voxel pyramid scale space is constructed. s The layer scale space is represented as , according to the formula Calculate the s Voxels in layer-scale space The Gaussian difference response value of the voxel The Gaussian difference response values ​​of the scale layer are greater than those of the scale layer and the upper and lower adjacent scale layers. When the Gaussian difference response value of the neighborhood point is used to determine the point cloud p k is a SIFT feature point, where For the s The three-dimensional Gaussian filter function used in the layer scale space, From point cloud p k Three-dimensional Gaussian filter function After processing, For the s The Gaussian kernel standard deviation corresponding to the layer scale space, For the s The voxel side length corresponding to the layer scale space, l 0 is the initial voxel side length, N' Voxel The number of valid neighborhood voxels of is the neighborhood curvature weight, , and is the mean and standard deviation of the neighborhood curvature at the current scale layer, and For the i The minimum and maximum eigenvalues ​​of the covariance matrix of the valid neighborhood voxels.

[0015] Furthermore, in step S3, the corresponding radius is constructed with each SIFT feature point as the center. Each layer of spherical shell is divided into 72 orientation units with a step length of 30° in the longitude direction and 30° in the latitude direction. For the points in each orientation unit, the normal direction distribution histogram weighted by Gaussian weighted statistics is used, and the spherical shell radius is adjusted according to the scale parameter of the SIFT feature point. , to generate the 3DSC descriptor, where the weighting coefficient of Gaussian weighting is expressed as , represents the standard deviation of the Gaussian distribution in the main direction, Indicates the first i' The normal direction of a point, Indicates the main direction, which is determined by the distribution extreme value of the normal direction of the point cloud in the statistical field with the center of the SIFT feature point as the center.

[0016] Furthermore, in step S4, a feature matching model based on SAC-IA sampling consistency is constructed, SIFT feature points and corresponding 3DSC descriptors are used as inputs of the feature matching model, a minimum sampling distance threshold is set, and a similarity measurement function of the mixed features is designed as , and through the RANSAC algorithm, based on the confidence Calculate the sampling probability, iteratively eliminate the mismatches and output the initial transformation matrix, where: and is a coefficient that is dynamically adjusted according to the point cloud density, p i The source point cloud i SIFT feature points and corresponding 3DSC descriptors, The target point cloud j SIFT feature points and corresponding 3DSC descriptors, p i and Corresponding to each other, represents the chi-square test function, Represents the histogram of the source point cloud feature descriptor, Represents the histogram of the target point cloud feature descriptor, To find the 2-norm.

[0017] Furthermore, in step S5, the objective function designed with the transformation matrix T1 as the optimization target is expressed as , iterate with the initial transformation matrix as the initial value to obtain the optimal transformation matrix Tfinal, where, is the normal vector of the i-th SIFT feature point of the source point cloud, is the normal vector of the j-th SIFT feature point of the target point cloud, represents the balance coefficient, represents the standard deviation of the geometric distance error, Represents the standard deviation of the normal vector error.

[0018] The present invention is also achieved through the following technical solutions:

[0019] A detection system based on any of the hollow fiber membrane integrity detection methods described above includes a bracket, a flexible film, an air pressure control module, a binocular vision module and a detection module. The hollow fiber membrane is horizontally arranged on the bracket, and the flexible film tightly covers the detection end of the hollow fiber membrane. The air pressure control module is connected to the air inlet end of the hollow fiber membrane to provide stable air pressure. The binocular vision module includes two cameras arranged in front of the detection end of the hollow fiber membrane. The point clouds collected by the two cameras are fed back to the detection module, and the defect area is determined by the detection module.

[0020] The present invention has the following beneficial effects:

[0021] 1. The present invention first tightly covers the detection end of the hollow fiber with a flexible film, and uses two cameras set in front of the flexible film to collect point clouds of the detection end not covered with the flexible film and the detection end covered with the flexible film respectively, and fuses the collected point clouds to obtain a target fused point cloud and a source fused point cloud respectively, and performs simplified preprocessing on the target fused point cloud and the source fused point cloud respectively to obtain a target point cloud and a source point cloud, and then extracts SIFT feature points of the source point cloud and the target point cloud respectively, and uses a 3DSC descriptor with normal constraints to describe the SIFT feature points, and based on the SIFT feature points of the source point cloud and the target point cloud and their corresponding The 3DSC descriptor is then coarsely aligned using the sampled consistency initial registration algorithm (SAC-IA) to obtain the initial transformation matrix. Finally, a generalized iterative closest point optimization model with normal vector constraints is constructed, and an objective function related to the initial transformation matrix is ​​designed to obtain the optimal transformation matrix. The source point cloud is transformed using the optimal transformation matrix, and the defect area is located based on the difference between the transformed result and the target point cloud. This enables rapid online positioning of membrane filament defects under dry conditions, effectively improving detection efficiency and detection rate. Synchronous acquisition by two cameras is used to reduce multi-sensor calibration errors, effectively reducing system complexity and operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention will be described in further detail below with reference to the accompanying drawings.

[0023] Figure 1 Schematic diagram of the structure of the detection system of the present invention.

[0024] Figure 2 Flowchart of the present invention.

[0025] Figure 3 Flowchart of the rough matching of the detection method of the present invention.

[0026] Figure 4 The present invention is a flow chart for obtaining the optimal transformation matrix for the detection method.

[0027] Among them, 1. bracket; 2. binocular vision module; 3. detection module; 4. air pressure control module; 5. hollow fiber membrane; 6. flexible film. DETAILED DESCRIPTION

[0028] like Figure 1 As shown, the non-contact hollow fiber membrane inspection system specifically includes a support 1, a flexible membrane 6, an air pressure control module 4, a binocular vision module 2, and a detection module 3. The hollow fiber membrane 5 is horizontally arranged on the support 1, with the flexible membrane 6 tightly covering the detection end of the hollow fiber membrane 5. The air pressure control module 4 is connected to the air inlet end of the hollow fiber membrane 5 to provide stable air pressure. The binocular vision module 2 includes two cameras positioned in front of the detection end of the hollow fiber membrane 5. The point cloud captured by the two cameras is fed back to the detection module 3, which then determines the defect area. The air pressure control module 4 includes an air compressor and a pressure stabilizing device connected to the air compressor to maintain a constant air pressure in the cavity of the hollow fiber membrane 5. The pressure stabilizing device is conventional. The flexible membrane 6 is made of a stretchable material. Before ventilation, the flexible membrane 6 is tightly attached to the detection end of the hollow fiber membrane 5 assembly. When gas molecules penetrate the interstices between the membrane fibers, they lift the flexible membrane 6, causing it to deform. The detection module 3 is a computer equipped with point cloud processing software. The two cameras are specifically high-resolution industrial cameras. The two cameras are located in the same horizontal plane and the angle between the two cameras is set at 15°-30°.

[0029] like Figure 2 As shown, the hollow fiber membrane non-contact detection method includes the following steps:

[0030] Step S1: tightly covering the flexible film 6 on the detection end of the hollow fiber membrane 5, and inputting gas from the air inlet end of the hollow fiber membrane 5 to deform the flexible film 6;

[0031] Step S2: Two cameras disposed in front of the flexible film 6 respectively collect point clouds from the detection end uncovered by the flexible film 6 and the detection end covered by the flexible film 6, fuse the collected point clouds to obtain a target fused point cloud and a source fused point cloud, and perform simplified preprocessing on the target fused point cloud and the source fused point cloud to obtain a target point cloud and a source point cloud, wherein the two cameras operate synchronously;

[0032] After calibrating the two cameras separately, a disparity map is calculated using a stereo matching algorithm based on the calibration parameters of the two cameras and the collected point cloud. A 3D point cloud is then reconstructed to generate the detection end. Redundant points are removed from the 3D point cloud and overlapping areas are optimized to obtain a source fused point cloud. The stereo matching algorithm used in this embodiment is the SGBM algorithm, and the camera calibration and 3D point cloud reconstruction process using the SGBM algorithm are conventional techniques.

[0033] The simplified pre-processing of the source fusion point cloud specifically includes: establishing a kd-tree for the source fusion point cloud, andN Calculate the minimum bounding box containing the source fused point cloud and divide the minimum bounding box into of For each voxel, use kd-tree to traverse each voxel, calculate the center of gravity of each non-empty voxel, find the nearest neighbor point cloud of the center of gravity, and replace all point clouds in the voxel with the nearest neighbor point cloud to obtain the source point cloud, where, To adjust the scale factor of the side length according to the number of point clouds, s is the proportionality coefficient, L x 、 L y 、 L z For the minimum bounding box x 、 y 、 z The length in the direction, , , , In this embodiment, replacing all point clouds in a voxel with a nearest neighbor point cloud at the center can significantly reduce computational redundancy and improve detection speed.

[0034] The process of streamlining preprocessing the target fused point cloud is the same as the process of streamlining preprocessing the source fused point cloud.

[0035] Step S3: extract SIFT feature points (3D scale-invariant feature transform feature points) of the source point cloud and the target point cloud respectively, and describe the SIFT feature points using a 3DSC descriptor (3D shape context descriptor) with normal constraints;

[0036] This step has the same processing steps for the source point cloud and the target point cloud. We will only use the source point cloud as an example to illustrate: For the source point cloud, construct a multi-resolution voxel pyramid scale space. s The layer scale space is represented as Based on the traditional Gaussian difference (DoG), the curvature weight is introduced to enhance the sensitivity of low contrast defects. According to the formula Calculate the s Voxels in layer-scale space The Gaussian difference response value of the voxel The Gaussian difference response values ​​of the scale layer are greater than those of the scale layer and the upper and lower adjacent scale layers. When the Gaussian difference response value of the neighborhood point is used to determine the point cloud p k is a SIFT feature point; among them, For the s The three-dimensional Gaussian filter function used in the layer scale space, The point cloud in the source point cloud is filtered by a three-dimensional Gaussian filter function After processing, we get the three-dimensional Gaussian filter function is prior art; For the s The Gaussian kernel standard deviation corresponding to the layer scale space, For the s The voxel side length corresponding to the layer scale space, l 0 is the set initial voxel side length; N' Voxel The number of valid neighborhood voxels is used to limit the local feature calculation range. In this embodiment, ; is the neighborhood curvature weight, is the local curvature of the neighborhood voxel, and is the mean and standard deviation of the neighborhood curvature at the current scale layer, and For the i The minimum eigenvalue and maximum eigenvalue of the covariance matrix of the valid neighborhood voxels reflect the local geometric curvature;

[0037] The corresponding radius is constructed with each SIFT feature point as the center Each layer of spherical shell is divided into 72 orientation units with a step length of 30° in the longitude direction and 30° in the latitude direction. For the points in each orientation unit, the normal direction distribution histogram weighted by Gaussian weighted statistics is used, and the spherical shell radius is adjusted according to the scale parameter of the SIFT feature point. , to generate 3DSC descriptors; where, Ensure that the descriptor adapts to features of different scales, , Take 0.5 times the scale parameter of SIFT feature points, Take 2.5 times of SIFT feature point scale parameter; k SIFT feature points, corresponding to t Layer of spherical shell ( u , v ) The normal direction distribution histogram value after weighting of the orientation unit is defined as , is the weighting coefficient of Gaussian weighting, Indicates the first position in the orientation unit i' The normal direction of a point, is the indicator function, when the normal direction of the point fall into( u , v ) unit corresponds to the longitude and latitude interval, , otherwise , is the standard deviation of the Gaussian distribution in the main direction, Indicates the main direction, which is the "most representative geometric orientation" of the local SIFT feature point. It is specifically determined by the center of the SIFT feature point and the extreme value of the distribution of the normal direction of the point cloud in the statistical field. Its function is to make the 3DSC descriptor have "direction invariance", which not only assists feature matching, but also indirectly reflects the defect characteristics of the membrane surface.

[0038] When constructing the 3DSC descriptor, the geometric characteristics of the membrane surface normal vector are weighted by the deviation between the normal direction and the principal direction to generate a local shape description, which can enhance the ability to capture subtle deformations of the membrane. Step S4: Based on the SIFT feature points of the source and target point clouds and their corresponding 3DSC descriptors, a sampling consistency initial registration algorithm (SAC-IA) is used to perform coarse registration to obtain an initial transformation matrix;

[0039] Construct a feature matching model based on SAC-IA sampling consistency, use SIFT feature points and corresponding 3DSC descriptors as input of the feature matching model, set the minimum sampling distance threshold, and design the similarity measurement function of the mixed features as follows: , and through the RANSAC algorithm, based on the confidence Calculate the sampling probability, iteratively eliminate false matches and output the initial transformation matrix. False matches refer to matching pairs whose similarity metric function exceeds the set threshold. The set threshold is a value greater than the mean similarity of normal matches, where: and is a coefficient that is dynamically adjusted according to the point cloud density; p i The source point cloud i SIFT feature points and corresponding 3DSC descriptors, The target point cloud j SIFT feature points and corresponding 3DSC descriptors, p i and correspond to each other; Represents the histogram of the source point cloud feature descriptor, The histogram representing the target point cloud feature descriptor. The specific value of the histogram is calculated according to the above weighted normal direction distribution histogram value definition formula; Represents the chi-square test function, which is used to measure the histogram of the source point cloud feature descriptor Histogram of target point cloud feature descriptor The statistical difference between them is smaller, indicating that the histograms are more similar; To find the 2-norm.

[0040] In this embodiment, the minimum sampling distance threshold is set to 0.028, which can be adjusted based on the point cloud density so that the sampling points can cover the feature distribution while avoiding "close-range repeated sampling".

[0041] Eliminating mismatches specifically includes: RANSAC iteratively randomly selects a set of feature matching pairs, calculates the transformation matrix, and then counts the fit between other matching pairs and the transformation matrix of the feature matching pair, that is, if a matching pair After substituting the transformation matrix, the similarity measure Exceeding the set threshold ( S is much larger than the average of normal matches), it is determined to be a false match and the matching pair is removed. This process is the existing technology.

[0042] Step S5: construct a generalized iterative closest point optimization model with normal vector constraints, design an objective function related to the initial transformation matrix to obtain the optimal transformation matrix, use the optimal transformation matrix to transform the source point cloud, and locate the defect area based on the difference between the transformed result and the target point cloud;

[0043] Transformation matrix T 1 The objective function designed for the optimization goal is expressed as , iterate with the initial transformation matrix as the initial value to obtain the optimal transformation matrix T final ,in: The source point cloud i Normal vectors of SIFT feature points, The target point cloud j Normal vector of SIFT feature points; Indicates the balance coefficient, which adjusts the "point-to-geometric distance error term" ” and “normal vector error term The weight of the objective function controls the degree of influence of the two on the optimization of the transformation matrix (e.g., a larger λ means a stronger normal constraint); Represents the standard deviation of the geometric distance error, which is used to normalize the geometric distance error between point pairs, so that errors of different magnitudes account for a more reasonable proportion in the objective function. Indicates the standard deviation of the normal vector error, which is used to normalize the angular error (or vector difference modulus) between normal vectors. Same, improve the stability of the objective function.

[0044] The present invention reduces multi-sensor calibration errors through a dual-camera synchronous triggering architecture. Combining feature fusion with hierarchical optimization registration, the system significantly improves detection speed while ensuring submillimeter positioning accuracy, achieving an efficiency improvement of over 70% compared to existing technologies and reducing system hardware costs by 60%. The present invention integrates geometric features with scale invariance, increasing the detection rate of low-contrast defects to 95%, overcoming the limitations of traditional point cloud centroid clustering methods. The present invention optimizes registration accuracy by constraining the normal vector consistency of the source and target point clouds and utilizing the geometric continuity characteristics of the membrane wire surface.

[0045] The above description is merely a preferred embodiment of the present invention and therefore cannot be used to limit the scope of the present invention. In other words, equivalent changes and modifications made according to the scope of the patent application and the contents of the specification should still fall within the scope of the patent of the present invention.

Claims

1. A non-contact detection method for hollow fiber membranes, characterized by: The steps include: Step S1: tightly covering the flexible film on the detection end of the hollow fiber membrane, and inputting gas from the air inlet end of the hollow fiber membrane to deform the flexible film; Step S2: Two cameras disposed in front of the flexible film respectively collect point clouds from the detection end uncovered by the flexible film and the detection end covered by the flexible film, fuse the collected point clouds to obtain a target fused point cloud and a source fused point cloud, and perform streamlined preprocessing on the target fused point cloud and the source fused point cloud to obtain a target point cloud and a source point cloud, wherein the two cameras operate synchronously; Step S3, for the source point cloud and the target point cloud respectively, by constructing a multi-resolution voxel pyramid scale space, extracting SIFT feature points based on the relationship between each voxel and its scale layer and the Gaussian difference response value of the neighborhood points in the upper and lower adjacent scale layers, constructing a multi-layer radial spherical shell with each SIFT feature point as the center, dividing each layer of the spherical shell into multiple orientation units, and using a Gaussian weighted statistical weighted normal direction distribution histogram for the points in each orientation unit, and adjusting the spherical shell radius according to the scale parameter of the SIFT feature point to generate a 3DSC descriptor with normal constraint to describe the SIFT feature points, wherein the source point cloud and the target point cloud are segmented to obtain corresponding voxels, the Gaussian difference response value is related to the neighborhood curvature weight, the neighborhood curvature weight is related to the eigenvalue of the covariance matrix of the neighborhood voxel, and the weighting coefficient of the Gaussian weighting is related to the normal direction of each point in each orientation unit; Step S4: Based on the SIFT feature points of the source point cloud and the target point cloud and their corresponding 3DSC descriptors, a sampling consistency initial registration algorithm is used to perform coarse registration to obtain an initial transformation matrix; Step S5: Construct a generalized iterative closest point optimization model including normal vector constraints, design an objective function related to the initial transformation matrix to obtain the optimal transformation matrix, use the optimal transformation matrix to transform the source point cloud, and locate the defect area based on the difference between the transformed result and the target point cloud.

2. A non-contact detection method for hollow fiber membranes according to claim 1, characterized in that: In step S1, the air inlet end of the hollow fiber membrane is connected to an air pressure control module, which includes an air compressor and a pressure stabilizing device connected to the air compressor to maintain a constant air pressure in the hollow fiber membrane cavity. The flexible film is made of a stretchable material.

3. The non-contact detection method for hollow fiber membranes according to claim 1, characterized in that: In step S2, the two cameras are located in the same horizontal plane and form a set angle. After the two cameras are calibrated, a disparity map is calculated based on the calibration parameters of the two cameras and the collected point cloud using a stereo matching algorithm, and a three-dimensional point cloud of the detection end is reconstructed. The three-dimensional point cloud is subjected to redundant point elimination and overlapping area optimization to obtain a source fused point cloud.

4. A non-contact detection method for hollow fiber membranes according to claim 1, 2 or 3, characterized in that: In step S2, the simplified pre-processing of the source fusion point cloud specifically includes: establishing a kd-tree for the source fusion point cloud, and N Calculate the minimum bounding box containing the source fused point cloud and divide the minimum bounding box into of For each voxel, use kd-tree to traverse each voxel, calculate the center of gravity of each non-empty voxel, find the nearest neighbor point cloud of the center of gravity, and replace all point clouds in the voxel with the nearest neighbor point cloud to obtain the source point cloud, where, To adjust the scale factor of the side length according to the number of point clouds, s is the proportionality coefficient, L x 、 L y 、 L z For the minimum bounding box x 、 y 、 z The length in the direction, , , , is the ceiling function.

5. A non-contact detection method for hollow fiber membranes according to claim 1, 2 or 3, characterized in that: In step S3, for the source point cloud, a multi-resolution voxel pyramid scale space is constructed. s The layer scale space is represented as , according to the formula Calculate the s Voxels in layer-scale space The Gaussian difference response value of the voxel The Gaussian difference response values ​​of the scale layer are greater than those of the scale layer and the upper and lower adjacent scale layers. When the Gaussian difference response value of the neighborhood point is used to determine the point cloud p k is a SIFT feature point, where For the s The three-dimensional Gaussian filter function used in the layer scale space, From point cloud p k Three-dimensional Gaussian filter function After processing, For the s The Gaussian kernel standard deviation corresponding to the layer scale space, For the s The voxel side length corresponding to the layer scale space, l 0 is the initial voxel side length, N' Voxel The number of valid neighborhood voxels of is the neighborhood curvature weight, , and is the mean and standard deviation of the neighborhood curvature at the current scale layer, and For the i The minimum and maximum eigenvalues ​​of the covariance matrix of the valid neighborhood voxels.

6. A non-contact detection method for hollow fiber membranes according to claim 5, characterized in that: In step S3, the corresponding radius is constructed with each SIFT feature point as the center. Each layer of spherical shell is divided into 72 orientation units with a step length of 30° in the longitude direction and 30° in the latitude direction. For the points in each orientation unit, the normal direction distribution histogram weighted by Gaussian weighted statistics is used, and the spherical shell radius is adjusted to 1 according to the scale parameter of the SIFT feature point. , to generate the 3DSC descriptor, where the weighting coefficient of Gaussian weighting is expressed as , represents the standard deviation of the Gaussian distribution in the main direction, Indicates the first i' The normal direction of a point, Indicates the main direction, which is determined by the distribution extreme value of the normal direction of the point cloud in the statistical field with the center of the SIFT feature point as the center.

7. The non-contact detection method for hollow fiber membranes according to claim 6, characterized in that: In step S4, a feature matching model based on SAC-IA sampling consistency is constructed, SIFT feature points and corresponding 3DSC descriptors are used as inputs of the feature matching model, the minimum sampling distance threshold is set, and the similarity measurement function of the mixed features is designed as , and through the RANSAC algorithm, based on the confidence Calculate the sampling probability, iteratively eliminate the mismatches and output the initial transformation matrix, where: and is a coefficient that is dynamically adjusted according to the point cloud density, p i The source point cloud i SIFT feature points and corresponding 3DSC descriptors, The target point cloud j SIFT feature points and corresponding 3DSC descriptors, p i and Corresponding to each other, represents the chi-square test function, Represents the histogram of the source point cloud feature descriptor, Represents the histogram of the target point cloud feature descriptor, To find the 2-norm.

8. The non-contact detection method for hollow fiber membranes according to claim 7, characterized in that: In step S5, the transformation matrix T 1 The objective function designed for the optimization goal is expressed as , iterate with the initial transformation matrix as the initial value to obtain the optimal transformation matrix T final ,in, The source point cloud i Normal vectors of SIFT feature points, The target point cloud j Normal vectors of SIFT feature points, represents the balance coefficient, represents the standard deviation of the geometric distance error, Represents the standard deviation of the normal vector error.

9. A detection system based on the non-contact detection method for hollow fiber membranes according to any one of claims 1 to 8, characterized in that: It includes a bracket, a flexible film, an air pressure control module, a binocular vision module, and a detection module. The hollow fiber membrane is horizontally arranged on the bracket, and the flexible film tightly covers the detection end of the hollow fiber membrane. The air pressure control module is connected to the air inlet end of the hollow fiber membrane to provide stable air pressure. The binocular vision module includes two cameras arranged in front of the detection end of the hollow fiber membrane. The point cloud collected by the two cameras is fed back to the detection module, and the detection module determines the defect area. The detection module includes: Fusion module: fuses the collected point clouds to obtain target fused point cloud and source fused point cloud, and performs simplified preprocessing on the target fused point cloud and source fused point cloud to obtain target point cloud and source point cloud respectively; SIFT feature point and 3DSC descriptor acquisition module: for the source point cloud and target point cloud respectively, by constructing a multi-resolution voxel pyramid scale space, SIFT feature points are extracted based on the relationship between the Gaussian difference response value of each voxel and its scale layer and the neighboring points of the upper and lower adjacent scale layers, and a multi-layer radial spherical shell is constructed with each SIFT feature point as the center. Each layer of the spherical shell is divided into multiple orientation units, and the normal direction distribution histogram weighted by Gaussian weighted statistics is used for the points in each orientation unit. The spherical shell radius is adjusted according to the scale parameter of the SIFT feature point to generate a 3DSC descriptor with normal constraints to describe the SIFT feature points. Among them, the source point cloud and the target point cloud are segmented to obtain the corresponding voxels, the Gaussian difference response value is related to the neighborhood curvature weight, the neighborhood curvature weight is related to the eigenvalue of the covariance matrix of the neighborhood voxel, and the weighting coefficient of Gaussian weighting is related to the normal direction of each point in each orientation unit; Coarse registration module: Based on the SIFT feature points of the source point cloud and the target point cloud and their corresponding 3DSC descriptors, a sampling consistency initial registration algorithm is used to perform coarse registration to obtain an initial transformation matrix; Positioning module: Construct a generalized iterative closest point optimization model with normal vector constraints, design an objective function related to the initial transformation matrix to obtain the optimal transformation matrix, use the optimal transformation matrix to transform the source point cloud, and locate the defect area based on the difference between the transformed result and the target point cloud.

Images