A point cloud registration detection method for hollow fiber membranes

Abstract

This invention discloses a point cloud registration and detection method for hollow fiber membranes. The method scans and acquires data on the surface of the hollow fiber membrane using structured light imaging, obtains the three-dimensional coordinates of the membrane surface based on triangulation principles, and generates point cloud data to achieve three-dimensional reconstruction of the hollow fiber membrane. To address the issues of initial value sensitivity and numerous outliers in point cloud registration, a point cloud registration algorithm based on NPFC+RANSAC+DBO is proposed: NPFC is used to construct neighborhood feature descriptors and perform feature matching; RANSAC is used to filter outliers and estimate the initial pose; then, dung beetle optimization is used to globally search and refine the registration parameters and pose, thereby improving registration accuracy and robustness. Finally, defect detection is performed on the reconstructed model to detect membrane fiber damage. Experiments show that this method can achieve stable registration and high-precision reconstruction under complex conditions such as occlusion, noise, and repetitive textures, providing reliable data support for the quality inspection of hollow fiber membranes.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing technology, and specifically designs a point cloud registration and testing method for hollow fiber membranes. Background Technology

[0002] Hollow fiber membranes, as an important separation material, are widely used in water treatment, air purification, drug separation, gas separation, and bioengineering. Their excellent separation performance and high mass transfer rate make them particularly important in water treatment and medical fields. With increasing demand, especially in seawater desalination, wastewater treatment, and biopharmaceutical industries, the production and application of hollow fiber membranes are constantly developing, making quality control and performance monitoring especially crucial. The quality of hollow fiber membranes directly affects the overall performance and service life of equipment; therefore, efficient and accurate testing of hollow fiber membranes is of significant practical importance. Traditional testing methods, such as visual inspection, chemical reaction testing, and destructive testing, often cannot comprehensively and accurately assess the quality of hollow fiber membranes and are prone to damaging the membrane material. Therefore, non-destructive testing technology has become an important means of quality assessment for hollow fiber membranes, especially for defect detection. This invention proposes a three-dimensional reconstruction detection method for hollow fiber membranes. By acquiring high-quality three-dimensional point cloud data, it achieves accurate reconstruction of the surface morphology of hollow fiber membranes, providing reliable data support for subsequent defect detection. This method effectively improves the accuracy and efficiency of hollow fiber membrane testing, providing a new technical means for the quality control and performance optimization of membrane materials. Summary of the Invention

[0003] 1. A method for point cloud registration and detection of hollow fiber membranes, characterized by comprising the following steps: Step S1: 3D data acquisition and point cloud construction; A laser is used to project light stripes onto the surface of the hollow fiber membrane, and a camera acquires images of the deformation of the light stripes on the fiber surface from different angles. Based on the principle of triangulation, the 3D coordinates of each point on the end face of the hollow fiber membrane are calculated to generate multi-view point cloud data. Step S2: For the multi-view point cloud data obtained in Step 1, only local point cloud information can be obtained each time. Therefore, it is necessary to obtain point cloud information from different perspectives. However, the point cloud data obtained each time is independent in a coordinate system and cannot be directly stitched together. To address this, an NPFC+RANSAC+DBO point cloud registration algorithm is used to align the point clouds from each perspective and unify them to the same coordinate system for coarse registration. Then, the ICP algorithm is used for fine registration. Step S3: For the registered point cloud obtained in Step 2, the local weighted least squares method is used to detect the end face of the hollow fiber membrane. First, a plane is fitted, and the point cloud coordinates are obtained based on the distance between the fitted plane and the top of the point cloud to determine the location of the defect coordinates.

[0004] 2. The point cloud registration and detection method for hollow fiber membranes according to claim 1, characterized in that: the point cloud registration algorithm of step S2 is NPFC+RANSAC+DBO: S2.1 The scanned uncoated hollow fiber membrane end face is taken as the target point cloud, and the coated hollow fiber membrane end face is taken as the source point cloud; S2.2 Performs statistical filtering, downsampling, and normal estimation on the source and target point clouds to reduce the impact of noise and outliers on subsequent point cloud registration; After S2.3, key point extraction is performed to remove point clouds with inconspicuous features and unstable edges; After S2.4, NPFC feature construction and matching are performed. NPFC neighborhood feature descriptors are constructed for each candidate point in the source point cloud and the target point cloud, and nearest neighbor matching is performed through feature distance metric to obtain an initial set of matching point pairs. S2.5 uses the RANSAC algorithm to perform coarse registration of matching point pairs, and estimates the initial transformation matrix by randomly selecting the set of minimum point pairs; After coarse registration, S2.6 uses the DBO algorithm to globally optimize the initial transformation matrix. It performs global search and optimization by simulating the behavior of a dung beetle. The optimization objective is to minimize the registration error between the source point cloud and the target point cloud. After iteratively updating the transformation parameters, the optimal transformation matrix is ​​finally obtained.

[0005] 3. The point cloud registration and detection method for hollow fiber membranes according to claim 1, characterized in that: step s2 involves a fine ICP registration method: S3.1 By iteratively calculating the rigid transformation matrix between the source point cloud and the target point cloud, the error function between the model point set and the target point set is minimized, thereby obtaining the optimal registration result; S3.2 First, perform point-to-point matching. In each iteration, for each point P in the target point cloud, perform point-to-point matching. Find the closest point in the model point cloud Q. Obtain the corresponding point pairs between the source point cloud and the target point cloud. ; S3.3 Using these point pairs, the transformation matrix from the source point cloud to the target point cloud is calculated. The goal is to minimize the error between the point pairs, thereby obtaining the rigid transformation matrix, as follows: Where R is the rotation matrix and t is the translation vector, the optimal rotation matrix R and translation vector t are calculated by minimizing this error function. S3.4 uses the calculated transformation matrix to update the source point cloud. After this transformation, the source point cloud is moved to a position closer to the target point cloud. Then, an error function is calculated to determine whether the iteration has converged; the error function is defined as: in In the model point cloud and the target point cloud The corresponding points; n and m are the number of P and Q, respectively; It is the threshold for the minimum distance between two point clouds. dRMS is the error equation of the ICP algorithm, which determines whether the ICP algorithm continues or stops iterating. S3.5 If the registration error meets the requirements, the registration process ends and the final transformation matrix is ​​obtained. Finally, the updated source point cloud and the target point cloud are accurately aligned, providing reliable point cloud data for subsequent hollow fiber membrane defect detection.

[0006] 4. The point cloud registration and detection method for hollow fiber membranes according to claim 1, characterized in that: the depth information of the hollow fiber membrane end face is obtained by a locally weighted least squares fitting method in step S3: S4.1 By calculating the local geometric features of each point in the point cloud of the film, such as the point's normal and neighborhood density, the local structural information of each point in three-dimensional space is obtained; S4.2 For each point's local structural information in 3D space, assign a weighting coefficient to each point in the point cloud. This weighting coefficient can be adjusted based on factors such as the point's depth, surface roughness, and normal variation. S4.3 Finally, a locally weighted least squares fitting method is used to fit a planar model of the hollow fiber membrane end face based on the weighted point cloud data; the specific fitting error function is as follows: Where: E is the total fitting error, N is the number of points in the point cloud, and is the weight of the i-th point in the point cloud. These are points in the fitted point cloud. These are points in the original point cloud. S4.4 Based on the fitting results, calculate the distance between each point and the fitting plane, using the information from the point cloud on the end face. By calculating the vertical distance between each point in the point cloud data and the fitting plane, obtain the depth value of that point, and thus determine whether there is a defect on the end face of the hollow fiber membrane based on the distance.

[0007] 5. The NPFC algorithm according to claim 2, characterized in that: the NPFC algorithm extracts local geometric feature descriptors such as normal vectors and curvature from point cloud data, and performs matching by calculating the feature similarity between point clouds: S5.1 Definition Point Its neighboring points The cosine of the normal angle: S5.2 Define neighborhood curvature (given by PCA eigenvalues): S5.3 in constructing NPFC descriptors: S5.4 Finally, nearest neighbor matching is used, and a ratio test is used to filter ambiguous matches: Then retain the matching pairs The initial set of matching point pairs is obtained by summarizing these pairs.

[0008] 6. The RANSAC algorithm according to claim 2, characterized in that: the RANSAC algorithm performs preliminary transformation estimation by randomly selecting the smallest sample set, and filters interior points through consistency verification, finally selecting the transformation matrix with the most interior points as the initial transformation matrix: S6.1 uses RANSAC to perform consistency verification on the set of matching pairs C, removes outlier matching points, and estimates the initial rigid body transformation matrix; in For rotation matrix, It is a translation vector. S6.2 Each time, the minimum set of point pairs is randomly sampled. Solving rigid body transformations using SVD: After S6.3, the error is calculated for all matching pairs: S6.4 Final set of interior points: Select to make The largest transformation as .

[0009] 7. The DBO algorithm according to claim 2, characterized in that: sDOB global refinement optimization uses the initial value obtained from RANSAC. Centered on the target matrix, the optimal transformation and key registration parameters are searched within its neighborhood using dung beetle optimization to obtain the optimal transformation matrix T. S7.1 Let the DBO individual position vector be a combination of pose increment and threshold parameter: S7.2 constructs the incremental transformation: The overall transformation is: S7.3 Constructs a fitness function based on the ratio of registration residuals to interior points: For the source point Transformed to Its nearest neighbor correspondence is: S7.4 performs point-to-point residual analysis: If the normal of the target point cloud is Then the point-to-surface residual is: Set of interior points: fitness: in The weights are the threshold values ​​for the lowest interior point. After S7.5, DBO iterative updates are performed and the following is retrieved: S7.6 finally uses T to transform the source point cloud to the target coordinate system and fuse them: The registration quality was assessed using RMSE or overlap region residuals. Attached Figure Description

[0010] Figure 1 This is a flowchart of the method of the present invention; Figure 2 Here is the flowchart for the NPFC algorithm; Figure 3 Here is a flowchart of the RANSAC algorithm; Figure 4 Here is a flowchart of the DBO algorithm; Detailed Implementation

[0011] This invention proposes a point cloud registration and detection method for hollow fiber membranes, employing multi-view point cloud data acquisition, point cloud registration, and defect detection techniques. The specific implementation includes the following steps: Step s1: Use a laser to project light stripes onto the surface of the hollow fiber membrane. A camera collects images of the deformation of the light stripes on the end face of the hollow fiber membrane from different angles. Based on the principle of triangulation, the three-dimensional coordinates of each point on the surface of the hollow fiber membrane are calculated from the images taken by the camera at different angles, generating multi-view point cloud data. Each collection of data only contains local point cloud information and cannot be directly stitched together. Therefore, it is necessary to collect data from different perspectives multiple times and merge these point cloud data into a complete point cloud model.

[0012] Step s2: For the multi-view point cloud data obtained in step S1, since the data collected each time is in a different coordinate system, they cannot be directly stitched together. Therefore, registration is required. The specific steps are as follows: S2.1 The uncoated hollow fiber membrane end face obtained from the scan is used as the target point cloud, and the coated hollow fiber membrane end face is used as the source point cloud; S2.2 performs statistical filtering to denoise the source point cloud and the target point cloud, downsampling and normal estimation to reduce the impact of noise and outliers on subsequent point cloud registration. First, statistical filtering is used to remove outliers in the point cloud. Then, the point cloud is downsampled using the voxel grid method to reduce the point cloud density and reduce the computational burden. Finally, the normal vector of each point is calculated using the neighborhood points to provide the necessary geometric information. S2.3 Extract key points from the source point cloud and the target point cloud, and remove point clouds with indistinct features and points with unstable edges. This step mainly uses geometric features such as normal vector changes and curvature changes to select stable and representative point clouds. S2.4 Constructs an NPFC neighborhood feature descriptor for each candidate point in the source point cloud and the target point cloud. By calculating local geometric features such as normal vector and curvature, a feature vector for each point is constructed. Feature matching is performed using Euclidean distance metric to obtain an initial set of matching point pairs. S2.5 uses RANSAC to perform coarse registration on the matching point pairs, randomly selects the smallest set of point pairs, and calculates the preliminary transformation matrix between the source point cloud and the target point cloud. The optimal preliminary transformation matrix is ​​selected by calculating the error value and the proportion of interior points. After coarse registration, S2.6 uses the DBO algorithm to globally optimize the initial transformation matrix. The DBO algorithm performs a global search and optimization in the neighborhood of the transformation matrix by simulating the behavior of a dung beetle. The goal is to minimize the registration error between the source point cloud and the target point cloud. After multiple iterations, the optimal transformation matrix T is finally obtained.

[0013] Step s3: After coarse registration, fine registration is performed using the ICP algorithm. S3.1 First, perform point-to-point matching, matching each point P in the target point cloud with each point P. Find the point that is closest to the model point cloud Q. This yields the corresponding point pairs between the source point cloud and the target point cloud. ; S3.2 Using these point pairs, calculate the transformation matrix from the source point cloud to the target point cloud. The goal is to minimize the error between the point pairs, thereby obtaining the rigid transformation matrix, as follows: Where R is the rotation matrix and t is the translation vector, the optimal rotation matrix R and translation vector t are calculated by minimizing this error function; S3.3 uses the calculated transformation matrix to update the source point cloud. After this transformation, the source point cloud is moved to a position closer to the target point cloud. Then, an error function is calculated to determine whether the iteration has converged; the error function is defined as: in In the model point cloud and the target point cloud The corresponding points; n and m are the number of P and Q, respectively; It is the threshold of the minimum distance between two point clouds, and dRMS is the error equation of the ICP algorithm, which determines whether the iteration of the ICP algorithm continues or stops. S3.4 If the registration error meets the requirements, the registration process ends and the final transformation matrix is ​​obtained. Finally, the updated source point cloud and the target point cloud are accurately aligned, providing reliable point cloud data for subsequent hollow fiber membrane defect detection.

[0014] Step s4: Extract depth information and detect defects in the registered point cloud data; S4.1 First, local geometric features are calculated. By calculating the local geometric features such as the normal vector and neighborhood density of each point in the point cloud, the local structural information of each point in the three-dimensional space is obtained. Then, based on these local structural information, weighting coefficients are assigned to each point in the point cloud. S4.2 The planar model of the hollow fiber membrane end face is fitted using the weighted least squares method. The specific fitting error function is as follows: Where: E is the total fitting error, and N is the number of points in the point cloud. It is the weight of the i-th point in the point cloud. These are points in the fitted point cloud. These are points in the original point cloud; S4.3 Based on the distance between the fitting plane and the point cloud, calculate the vertical distance from each point to the fitting plane, thereby extracting the defect information on the end face of the hollow fiber membrane. By analyzing the distance between the point cloud and the fitting plane, determine whether there are defects in the membrane filaments on the end face of the hollow fiber membrane.

[0015] Step s5: To verify the effectiveness of the algorithm of this invention, we compared it with other algorithms: NPFC+RANSAC algorithm: registration error average 0.7mm, maximum 1.1mm, processing time: 12 minutes; FFPH+RANSAC algorithm: registration error average 0.9mm, maximum 1.5mm, processing time: 14 minutes; Algorithm of this invention: registration error average 0.4mm, maximum 0.8mm, processing time: 8 minutes; Through experimental comparison, the algorithm of this invention significantly reduces the registration error and the processing time. These results show that the algorithm of this invention not only has a significant advantage in accuracy, but also has good real-time performance in practical applications, making it suitable for the detection of defects on the end face of hollow fiber membranes.

Claims

1. A method for point cloud registration and detection of hollow fiber membranes, characterized in that, Includes the following steps: Step S1: 3D data acquisition and point cloud construction: A laser is used to project light stripes onto the surface of the hollow fiber membrane. A camera acquires images of the deformation of the light stripes on the fiber surface from different angles. Based on the principle of triangulation, the 3D coordinates of each point on the end face of the hollow fiber membrane are calculated to generate multi-view point cloud data. Step S2: For the multi-view point cloud data obtained in Step 1, only local point cloud information can be obtained each time. Therefore, it is necessary to obtain point cloud information from different perspectives. However, the point cloud data obtained each time is independent in a coordinate system and cannot be directly stitched together. To address this, an NPFC+RANSAC+DBO point cloud registration algorithm is used to align the point clouds from each perspective and unify them to the same coordinate system for coarse registration. Then, the ICP algorithm is used for fine registration. Step S3: For the registered point cloud obtained in Step 2, the local weighted least squares method is used to detect the end face of the hollow fiber membrane. First, a plane is fitted, and the point cloud coordinates are obtained based on the distance between the fitted plane and the top of the point cloud to determine the location of the defect coordinates.

2. The point cloud registration and detection method for hollow fiber membranes according to claim 1, characterized in that: A point cloud registration algorithm based on NPFC+RANSAC+DBO is described in step S2: S2.1 The scanned uncoated hollow fiber membrane end face is taken as the target point cloud, and the coated hollow fiber membrane end face is taken as the source point cloud; S2.2 Performs statistical filtering, downsampling, and normal estimation on the source and target point clouds to reduce the impact of noise and outliers on subsequent point cloud registration; After S2.3, key point extraction is performed to remove point clouds with inconspicuous features and unstable edges; After S2.4, NPFC feature construction and matching are performed. NPFC neighborhood feature descriptors are constructed for each candidate point in the source point cloud and the target point cloud, and nearest neighbor matching is performed through feature distance metric to obtain an initial set of matching point pairs. S2.5 uses the RANSAC algorithm to perform coarse registration of matching point pairs, and estimates the initial transformation matrix by randomly selecting the set of minimum point pairs; After coarse registration, S2.6 uses the DBO algorithm to globally optimize the initial transformation matrix. It performs global search and optimization by simulating the behavior of a dung beetle. The optimization objective is to minimize the registration error between the source point cloud and the target point cloud. After iteratively updating the transformation parameters, the optimal transformation matrix is ​​finally obtained.

3. The point cloud registration and detection method for hollow fiber membranes according to claim 1, characterized in that: A fine registration method for ICP in step s2: S3.1 By iteratively calculating the rigid transformation matrix between the source point cloud and the target point cloud, the error function between the model point set and the target point set is minimized, thereby obtaining the optimal registration result; S3.2 First, perform point-to-point matching. In each iteration, for each point P in the target point cloud, perform point-to-point matching. Find the point in the model point cloud Q that is closest to it. This yields the corresponding point pairs between the source point cloud and the target point cloud. ; S3.3 Using these point pairs, the transformation matrix from the source point cloud to the target point cloud is calculated. The goal is to minimize the error between the point pairs, thereby obtaining the rigid transformation matrix, as follows: Where R is the rotation matrix and t is the translation vector; by minimizing this error function, the optimal rotation matrix R and translation vector t can be calculated. S3.4 uses the calculated transformation matrix to update the source point cloud. After this transformation, the source point cloud is moved to a position closer to the target point cloud. Then, an error function is calculated to determine whether the iteration has converged. The error function is defined as: in In the model point cloud and the target point cloud The corresponding points; n and m are the number of P and Q, respectively; It is the threshold for the minimum distance between two point clouds. dRMS is the error equation of the ICP algorithm, which determines whether the ICP algorithm continues or stops iterating. S3.5 If the registration error meets the requirements, the registration process ends and the final transformation matrix is ​​obtained. Finally, the updated source point cloud and the target point cloud are accurately aligned, providing reliable point cloud data for subsequent hollow fiber membrane defect detection.

4. The point cloud registration and detection method for hollow fiber membranes as described in claim 1, characterized in that: Step S3 describes a method for obtaining depth information of the hollow fiber membrane end face using locally weighted least squares fitting: S4.1 By calculating the local geometric features of each point in the point cloud of the film, such as the point's normal and neighborhood density, the local structural information of each point in three-dimensional space is obtained; S4.2 For the local structural information of each point in three-dimensional space, assign a weighting coefficient to each point in the point cloud. This weighting coefficient can be adjusted according to factors such as the depth of the point, surface roughness, and normal variation. S4.3 Finally, a locally weighted least squares fitting method is used to fit a planar model of the hollow fiber membrane end face based on the weighted point cloud data; the specific fitting error function is as follows: Where: E is the total fitting error, and N is the number of points in the point cloud. It is the weight of the i-th point in the point cloud. These are points in the fitted point cloud. These are points in the original point cloud. S4.4 Based on the fitting results, calculate the distance between each point and the fitting plane. Based on the information of the point cloud of the end face protrusion, calculate the depth value of each point by calculating the vertical distance between each point in the point cloud data and the fitting plane. Then, determine whether there is a defect on the end face of the hollow fiber membrane based on the distance.

5. The NPFC algorithm according to claim 2, characterized in that: The NPFC algorithm extracts local geometric feature descriptors such as normal vectors and curvature from point cloud data, and performs matching by calculating the feature similarity between point clouds; S5.1 Definition Point Its neighboring points The cosine of the normal angle: S5.2 Define neighborhood curvature (given by PCA eigenvalues): S5.3 in constructing NPFC descriptors: S5.4 Finally, nearest neighbor matching is used, and a ratio test is used to filter ambiguous matches: Then retain the matching pairs The initial set of matching point pairs is obtained by summarizing.

6. The RANSAC algorithm according to claim 2, characterized in that: The RANSAC algorithm performs initial transformation estimation by randomly selecting the smallest sample set, filters inliers through consistency verification, and finally selects the transformation matrix with the most inliers as the initial transformation matrix. S6.1 uses RANSAC to verify the consistency of the matching pair set C, removes outlier matching points, and estimates the initial rigid body transformation matrix; in For rotation matrix, It is a translation vector; S6.2 Each time, the minimum set of point pairs is randomly sampled. Solving rigid body transformations using SVD After S6.3, the error is calculated for all matching pairs. S6.4 Last set of interior points Select to make The largest transformation as 7. The DBO algorithm according to claim 2, characterized in that: sDOB global refinement optimization, using initial values ​​obtained from RANSAC Centered on the dung beetle, the optimal transformation and key registration parameters are searched within its neighborhood using dung beetle optimization to obtain the optimal transformation moments. S7.1 Let the DBO individual position vector be a combination of pose increment and threshold parameter: S7.2 constructs the incremental transformation: The overall transformation is: S7.3 Constructs a fitness function based on the ratio of registration residuals to interior points: For the source point Its nearest neighbor correspondence is: S7.4 performs point-to-point residual analysis: If the normal of the target point cloud is The point-to-surface residual is: Set of interior points: fitness: in The weights are the threshold values ​​for the lowest interior point. After S7.5, DBO iterative updates are performed and the following is retrieved: S7.6 finally uses T to transform the source point cloud to the target coordinate system and fuse them: The registration quality was assessed using RMSE or overlap region residuals.

Images