A method for detecting defects of a core rod based on a three-dimensional point cloud

Abstract

The application provides a core rod defect detection method based on a three-dimensional point cloud. The method realizes high-precision and automatic identification of surface defects by pre-processing, multi-stage filtering, weighted fitting and residual analysis on three-dimensional point cloud data collected on a production site, in combination with prior geometric constraints and dynamic threshold criteria. The method does not need to rely on standard sample templates or deep learning models, has the advantages of high calculation efficiency, strong noise resistance, less manual intervention, etc., and can meet the real-time detection needs in an online high-speed production environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial surface defect detection, and relates to a mandrel defect detection method based on three-dimensional point cloud. BACKGROUND

[0002] As a key basic material in the modern industrial system, seamless steel pipes are widely used in major fields such as oil and gas exploration, high-pressure boilers, nuclear power equipment, aerospace and transportation, and the quality thereof is directly related to energy transportation safety, major equipment operation reliability and public infrastructure durability. Especially in deep well drilling and high temperature and high pressure service environments, seamless steel pipes need to have excellent strength, toughness and sealing performance, and any small surface or internal defect may evolve into a crack propagation source in long-term use, causing leakage, pipe explosion accidents, and huge economic losses and safety risks. Therefore, high-precision and high-reliability seamless steel pipe quality detection is not only the core link of the manufacturing process, but also a strategic demand to ensure the safety of national energy and major engineering operation.

[0003] In the production process of seamless steel pipes, the mandrel, as one of the core forming dies, directly participates in the shaping process of the inner wall of the steel pipe, and the geometric precision, surface integrity and wear state of the outer surface thereof will directly affect the quality consistency and surface finish of the inner wall of the steel pipe. With the repeated action of high temperature, high contact stress and cyclic cooling medium in the continuous rolling process, fatigue cracks, pits, scratches or local peeling defects are easily generated on the surface of the mandrel. These surface damages will be "copied" to the inner wall of the steel pipe in the subsequent rolling process, forming corresponding protrusions, depressions or scratches, which not only reduces the surface quality of the steel pipe, but also may become stress concentration points, weakening the pressure-bearing capacity and service life thereof.

[0004] Therefore, high-precision, online or quasi-online surface defect detection of the mandrel is not only an important means to extend its service life, optimize the replacement cycle and reduce production costs, but also a key link to control the quality of the inner wall of seamless steel pipes from the source and improve product consistency. However, the detection of the mandrel faces many real challenges: its surface is in a high-temperature oxidation environment for a long time, often covered with loose oxide skin, oil film or cooling residues, resulting in uneven laser reflection, dense and scattered noise points in point cloud data; at the same time, the limited detection space, visual angle obstruction and background stray light interference further exacerbate the decline in imaging quality, making it easy for small defect signals to be overwhelmed by noise. Traditional manual visual or simple measurement methods are not only inefficient, subjective and poorly reproducible, but also difficult to meet the requirements of modern high-speed production lines. Although there are currently many three-dimensional detection techniques for the surface defects of steel pipes, there are still few systematic detection schemes specifically for the mandrel, especially how to effectively suppress interference and improve the detectability of weak anomalies under complex working conditions, and there is still a lack of effective solutions that take into account precision, robustness and real-time performance. Therefore, it is urgent to develop an intelligent detection method that can adapt to high-temperature residues, surface contamination and strong noise environments to realize automatic identification and quantitative evaluation of the surface defects of the mandrel and provide solid technical support for high-quality and efficient production of seamless steel pipes. SUMMARY

[0005] To solve the above technical problems, the purpose of the present application is to provide a mandrel defect detection method based on three-dimensional point cloud.

[0006] The mandrel defect detection method based on three-dimensional point cloud of the present application comprises:

[0007] Step 1: Use multiple three-dimensional line array cameras to surround and arrange, synchronously scan the surface of the moving mandrel, obtain three-dimensional point cloud data and laser intensity information of the mandrel under the real world coordinate system, and form a structured point cloud data set;

[0008] Step 2: Preprocess and format the point cloud data set, check the data integrity and convert it into a tensor structure suitable for GPU parallel computing, provide input for subsequent high-precision fitting and anomaly detection;

[0009] Step 3: Perform weighted least squares circle fitting on each circular arc contour line, combine laser intensity mask and local density anomaly analysis method to preliminarily identify and remove abnormal noise points;

[0010] Step 4: Calculate the real radial distance of each point based on the first fitting result, introduce prior geometric constraints to construct a radius tolerance band-shaped region, and filter out abnormal points that deviate significantly from the theoretical circular arc contour line;

[0011] Step 5: Based on the geometric filtering in Step 4, apply the statistical 3σ principle to fine-screen the residuals, further improving the quality and consistency of the effective point set;

[0012] Step 6: Based on the high-quality point set after multi-level filtering, perform a second weighted least squares circle fitting on each contour line to obtain high-precision center, radius, and residual information;

[0013] Step 7: Calculate the radial deviation according to the results of the second fitting. If the radial deviation is less than the threshold, output the core rod defect-free result. Otherwise, generate an abnormal mask binary matrix, and through morphological opening and closing operation, repair the broken area, eliminate isolated noise points, and enhance the spatial connectivity of the abnormal area;

[0014] Step 8: Construct a k-d tree index on the abnormal point set after morphological processing, and perform fast density clustering. If the number of points in all clusters is less than the threshold, output the core rod defect-free result. Otherwise, only keep the clusters with a number greater than the threshold, and realize automatic segmentation and group labeling of the abnormal area;

[0015] Step 9: Output the defect detection results, including the size, depth, and position information of the defects, supporting the visualization presentation, quantitative analysis, and quality traceability of the defects.

[0016] The core rod defect detection method based on three-dimensional point cloud has the following beneficial effects: the present application introduces laser intensity mask and local density anomaly analysis mechanism, constructs multi-level collaborative filtering strategy, significantly reduces the interference of noise points and edge weak signal points on the fitting process, and improves the robustness of initial contour fitting. On this basis, a quadratic iterative weighted least squares circle fitting framework is adopted, the first fitting is used to generate a priori geometric model, and the second fitting is based on the high-quality point set filtered by the multi-level filtering to perform high-precision parameter estimation, effectively suppressing the deviation accumulation of the abnormal points on the calculation of the center and the radius. Further, the priori geometric constraint based radius tolerance band shape region filtering method is proposed by using the priori knowledge that the core rod physical structure has a stable radius, and the points with a real radial distance deviating from the theoretical radius too much are removed, and the identification ability of non-ideal circular arc points is strengthened. In order to realize the connectivity repair and boundary integrity enhancement of the abnormal area, the morphological opening and closing operation is introduced to perform topological optimization on the preliminary residual error abnormal graph, effectively solving the fracture problem of adjacent abnormal areas caused by sparse sampling or local shielding, and preliminarily realizing the fusion connection of abnormal fragments. In the abnormal area segmentation stage, the k-d tree is constructed to accelerate the spatial neighborhood query, which significantly improves the efficiency of the neighborhood search of abnormal points in large-scale point cloud, and meets the real-time demand in the online high-speed production scene; then, the fast density clustering algorithm is used to generate initial defect clusters, and then all the clustering clusters are subjected to significance screening to exclude non-substantial point groups formed by residual noise or slight disturbance. The method realizes high-precision, high-efficiency and automatic detection of small defects on the surface of the core rod or pipe, can timely identify and locate typical defects such as cracks and pits, and provides a reliable and traceable intelligent quality control solution for enterprises. The method can also be applied to the detection of other pipe and rod surface defects. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of the core rod defect detection method based on three-dimensional point cloud of the present application. DETAILED DESCRIPTION

[0018] As shown in Figure 1 , the core rod defect detection method based on three-dimensional point cloud of the present application comprises:

[0019] Step 1: A plurality of three-dimensional linear array cameras are arranged around to synchronously scan the surface of the moving core rod, three-dimensional point cloud data and laser intensity information of the core rod in the real world coordinate system are obtained, and a structured point cloud data set is formed.

[0020] In specific implementation, four three-dimensional linear array cameras are used in hardware, the cameras are arranged at intervals of about 90 degrees, the moving core rod is surrounded in the middle, the depth of field and other parameters are adjusted, the cameras are fixed, and then the camera parameters are calibrated, so that the camera obtains the point cloud data set of the outer surface of the core rod.

[0021] Step 2: Preprocessing and formatting of point cloud data set, checking data integrity and converting to tensor structure suitable for GPU parallel computing, providing input for subsequent high-precision fitting and anomaly detection, specifically:

[0022] For the point cloud data set collected by the camera, data type checking, null value checking and critical value checking based on prior knowledge are performed, and then the data is formatted into a tensor matrix that can be directly processed by the GPU.

[0023] Step 3: Weighted least squares fitting of each circular arc contour line, combined with laser intensity mask and local density anomaly analysis method, to preliminarily identify and remove abnormal noise points, specifically:

[0024] Step 3.1: For a circular arc contour line in the point cloud data set, the weighted least squares problem can be represented by equation (1):

[0025] (1)

[0026] where is the coordinate of the mth point on the circular arc, represents the weight factor of the point, a, b, and c are the parameters of the circular arc to be fitted, and C is the number of points on the circular arc contour line.

[0027] Let , , , , then equation (1) is rewritten as:

[0028] (2)

[0029] Solve the matrix in equation (2), since is a diagonal weight matrix, let , then equation (3) is obtained as follows:

[0030] (3)

[0031] In order to obtain the fitted circular arc parameters , solve equation (3) to get:

[0032] (4)

[0033] Expand and calculate the elements on both sides of equation (4) and , then get equations (5) and (6):

[0034] (5)

[0035] (6)

[0036] By substituting the points on the circular arc, the circular arc parameters a, b, c can be obtained, and then the fitting center o, the fitting radius r and the residual res of each point can be obtained m :

[0037] (7)

[0038] Step 3.2: In this fitting, considering the influence of laser line intensity on the reliability of data source, in order to solve the problem of unreliable data caused by noise points on both sides of the core rod in the camera field of view with weak laser intensity, a laser intensity mask matrix is introduced in the fitting process, the value of the matrix is:

[0039] (8)

[0040] Wherein, is the laser intensity of the mth point on the improved circular arc, is the preset laser intensity threshold, that is, only when the laser intensity of the collected point cloud data is greater than the value, the data of the point is considered to be effective, and the laser intensity mask matrix is effective by taking the union of the weight matrix in the least square fitting.

[0041] Step 3.3: In order to solve the problem of abnormal noise points, a local density analysis anomaly detection method based on prior knowledge is adopted, and one-dimensional convolution is used to count the field in the local sliding window of each time point to identify potential abnormal noise points in the effective data points whose neighborhood effective sampling density is significantly lower than the prior preset threshold; Specifically, for a circular arc profile line to be fitted, an effective height mask matrix Vz is constructed according to the height z value of the fitting point of the circular arc profile line:

[0042] (9)

[0043] (10)

[0044] Wherein, represents the height value of the mth point on the circular arc to be fitted, a sliding window with a size of w_size is used to perform local density statistics on Vz, and a one-dimensional convolution kernel k with a width of w_size is defined:

[0045] (11)

[0046] Vz is calculated by one-dimensional convolution, wherein zero padding is used during convolution to ensure that the matrix dimension remains unchanged after convolution:

[0047] (12)

[0048] The convolution result D is normalized to represent the density ratio :

[0049] (13)

[0050] A determination rule for abnormal points is defined, assuming that when a certain point on the density ratio matrix is less than the density threshold , the point is considered to be an abnormal point, that is, an effective point mask matrix based on local density statistics is obtained :

[0051] (14)

[0052] (15)

[0053] The effective point mask matrix is activated by taking the union with the weight matrix W in the least squares fitting.

[0054] Step 4: Perform radius mask filtering based on prior geometric constraints, specifically: calculate the true radial distance of each point based on the first fitting result, introduce prior geometric constraints to construct a radius tolerance band-shaped region, and filter out abnormal points that significantly deviate from the theoretical circular arc profile line, specifically:

[0055] Based on the result obtained by the weighted least squares fitting in step 3, for each circular arc profile line to be fitted, its fitting center and fitting radius can be obtained, and the effective point cloud data participating in the fitting is re-substituted into the implicit circle equation composed of the fitting center and the fitting radius, so that the Euclidean distance of each point to the center can be calculated, denoted as the true radial distance d m , using the prior knowledge that the mandrel structure has a stable radius in physics, a reasonable radius tolerance threshold r threshold is set to construct a dynamic radius band-shaped region:

[0056] (16)

[0057] wherein represents the radius mask matrix of a certain circular arc profile line, represents the fitting radius of a certain circular arc to be fitted, and only when the true distance is within , , it will be considered as an effective point participating in the subsequent fitting operation, otherwise it will be removed as an abnormal noise point.

[0058] Step 5: On the basis of geometric filtering in step 4, apply the statistical 3σ principle to fine-screen the residuals, further improve the quality and consistency of the effective point set, specifically:

[0059] According to the effective point set screened in step 4, obtain the residual set res corresponding to each point m , calculate the mean and standard deviation of the residual set, according to the 3σ principle under the assumption of normal distribution, about 99.7% of the data should fall within the mean ± 3 times the standard deviation range, therefore, define the statistical effectiveness criterion as follows:

[0060] (17)

[0061] Wherein, indicates the effective mask matrix on all profiles that meet the normal distribution, only the points that meet the condition are retained as the final high-quality input point set, which is used for the next round of iterative fitting.

[0062] Step 6: Based on the high-quality point set after multi-level filtering, perform a second weighted least squares circle fitting on each contour line to obtain high-precision center, radius and residual information, specifically:

[0063] Step 6.1: According to the mask matrix obtained by multi-level filtering in steps 4 and 5, design the weight matrix of the second least squares circle fitting:

[0064] (18)

[0065] Wherein, indicates the radius mask matrix based on prior geometric constraints in step 4, indicates the effective mask matrix on all profiles that meet the 3σ criterion in step 5, then indicates the weight matrix in the second weighted least squares fitting process.

[0066] Step 6.2: Participate in the second weighted least squares fitting process , to obtain high-precision center, radius and residual information.

[0067] Step 7: Calculate the radial deviation according to the results of the second fitting, if the radial deviation is less than the threshold, output the core rod defect-free result; otherwise, generate an abnormal mask binary matrix, and repair the broken area, eliminate isolated noise points, and enhance the spatial connectivity of the abnormal area through morphological opening and closing operation, specifically:

[0068] Step 7.1: For a certain arc profile to be fitted, when the second fitting is completed, the radial deviation res2 of all points is recalculated based on the fitting parameters m , i.e. the distance from the real point p m to the center of each fitted circle o2 m minus the radius of each fitted circle r2 m , which reflects the degree of deviation of each point from the ideal arc:

[0069] (19)

[0070] Determine whether the radial deviation is less than the set threshold value, if it is less than the threshold value, output the core rod defect-free result; otherwise, generate an abnormal mask binary matrix B, as shown in the following formula:

[0071] (20)

[0072] (21)

[0073] wherein, is a depth anomaly detection threshold value, is the abnormal mask value of a certain point on the arc profile to be fitted.

[0074] Step 7.2: Based on the abnormal mask binary matrix, a binary morphological opening and closing operation sequence based on a structure element is used for topological repair and boundary smoothing processing, first the morphological opening operation is performed, which is defined as erosion followed by dilation:

[0075] (22)

[0076] wherein, represents an erosion operation, only when the structure element K completely covers the foreground area, the center pixel is retained as 1, and as the structure element K slides on the binary matrix B, isolated points smaller than the structure element K can be effectively eliminated; represents a dilation operation, which is opposite to the erosion operation, as the structure element K slides on the binary matrix B, as long as there is an overlapping part, the center pixel will be set to 1, which can fill small holes or link broken objects, realize the linking between adjacent abnormal points, in the embodiment, the K structure matrix uses a 5x5 square matrix.

[0077] Step 8: Construct a k-d tree index for the abnormal point set after morphological processing, perform fast density clustering, and perform significance screening on the obtained multiple independent clusters, if the number of points in all clusters is less than the threshold value, output the core rod defect-free result; otherwise, only keep the clusters with a number greater than the threshold value, realize automatic segmentation and group labeling of abnormal areas;

[0078] In implementation, the step 8 of constructing k-d tree index for the set of abnormal points after morphological processing, performing fast density clustering, specifically comprises:

[0079] Step 8.1: the set of abnormal points Q obtained in the above opening operation, each point of which contains three-dimensional coordinates x, y, z, is constructed into a k-d tree to realize efficient neighborhood query; for a point q k , the index set of all points in the spherical region with q as the center and radius R is denoted as:

[0080] (23)

[0081] , wherein, denotes the Euclidean distance between the point q k and the point .

[0082] Step 8.2: on the basis of completing the construction of k-d tree and realizing efficient neighborhood query, a fast clustering process based on spatial connectivity is performed, which sequentially selects a starting point from all unprocessed abnormal points, uses the k-d tree to quickly search all neighboring points within a specified radius range of the starting point, and forms an initial point set.

[0083] Step 8.3: it is judged whether the point set intersects with any existing cluster that has been generated, that is, whether there are common contained points.

[0084] Step 8.4: if there is overlap, the current point set is merged with all intersecting clusters to form a larger connected region; if there is no intersection, it is added to the cluster set as a new independent cluster; in the whole process, all points that have been included in a cluster are marked as “visited” to prevent repeated processing, thereby ensuring that each point participates in clustering decision only once.

[0085] The step 8 of merging neighboring clusters and screening significant point groups, specifically comprises:

[0086] Step 8.5: a minimum point number threshold is set, if the number of points in all clusters is less than the threshold, the core rod is output as a defect-free result; otherwise, only the independent clusters with the number of members exceeding the threshold are reserved as effective abnormal regions, and the remaining clusters are regarded as invalid noise and removed, so as to exclude non-substantial point groups formed by residual noise or small disturbances.

[0087] Step 8.6: for each effective cluster reserved, a unique group identifier is assigned to it, and the identifier is attached to the feature data of the corresponding point.

[0088] Step 9: output the defect detection result, including the size, depth and position information of the defect, support the visual presentation, quantitative analysis and quality traceability of the defect.

[0089] The above merely describes the preferred embodiments of the present application and is not intended to limit the idea of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting a core rod defect based on a three-dimensional point cloud, characterized in that, The method comprises the following steps: Step 1: synchronously scan the surface of the moving mandrel by using a plurality of three-dimensional linear array cameras arranged in a ring shape, acquire three-dimensional point cloud data and laser intensity information of the mandrel in a real world coordinate system, and form a structured point cloud data set; Step 2: pre-process and format the point cloud data set, check data integrity and convert it into a tensor structure suitable for GPU parallel computing, to provide input for subsequent high-precision fitting and anomaly detection; Step 3: perform weighted least squares circle fitting on each circular arc contour line, and preliminarily identify and remove abnormal noise points by combining laser intensity masks and local density anomaly analysis methods; Step 4: calculate the real radial distance of each point based on the first fitting result, introduce prior geometric constraints to construct a radius tolerance band-shaped region, and filter out abnormal points that significantly deviate from the theoretical circular arc contour line; Step 5: on the basis of the geometric filtering of step 4, apply the statistical 3σ principle to finely screen the residuals, and further improve the quality and consistency of the effective point set; Step 6: based on the high-quality point set filtered by multiple levels, perform second weighted least squares circle fitting on each contour line to obtain high-precision center, radius and residual information; Step 7: calculate the radial deviation according to the results of the second fitting, if the radial deviation is less than a threshold value, output the mandrel defect-free result; otherwise, generate an abnormal mask binary matrix, and repair the broken area and eliminate isolated noise points through morphological opening and closing operation to enhance the spatial connectivity of the abnormal area; Step 8: construct a k-d tree index for the abnormal point set after morphological processing, and perform fast density clustering, if the number of points in all clusters is less than a threshold value, output the mandrel defect-free result; otherwise, only keep the clusters with a number greater than the threshold value, to realize automatic segmentation and group labeling of the abnormal area; Step 9: output the defect detection result, including the size, depth and position information of the defect, to support the visualization, quantitative analysis and quality traceability of the defect.

2. The method of claim 1, wherein, The pre-processing and formatting in step 2 specifically comprises: The point cloud data set collected by the camera is checked for data type, null value and critical value based on prior knowledge, and then formatted into a tensor matrix that can be directly processed by the GPU.

3. The three-dimensional point cloud based method of claim 1, wherein, Step 3 specifically comprises: Step 3.1: for a circular arc contour line in the point cloud data set, the weighted least squares problem can be represented by equation (1): (1) wherein, is the coordinate of the mth point on the circle arc, denotes the weight factor of the point, a, b, c are the parameters of the circle arc to be fitted, and C is the number of points on the circle arc contour line. Let , , , Equation (1) is rewritten as: (2) Solving the matrix in equation (2) gives is a diagonal weight matrix, then let Equation (3) is then given by (3) To find the parameters of the circle arc to be fitted Solving equation (3) gives: (4) Expanding the calculation by elements, we obtain the following equations (5) and (6): and Expanding the calculation by elements, we obtain the following equations (5) and (6): (5) (6) By substituting the points on the circle arc, the circle arc parameters a, b, c can be obtained, and then the fitting circle center o, the fitting radius r and the residual res of each point can be obtained m : (7) Step 3.2: in this fitting, considering the influence of laser intensity on the reliability of data source, the laser intensity mask matrix is introduced in the fitting process to solve the problem of unreliable data caused by weak laser intensity on both sides of the mandrel in the camera field of view. The value of the matrix is: (8) wherein, is the laser intensity at the mth point on the arc, is a preset laser intensity threshold, that is, only when the laser intensity of the collected point cloud data is greater than the value, the data of the point is considered valid, and the laser intensity mask matrix by the weight matrix in the least squares fitting takes effect in the form of union; Step 3.3: for the problem of abnormal noise points, a local density analysis anomaly detection method based on prior knowledge is used. One-dimensional convolution is used to statistically analyze the domain in the local sliding window of each time point to identify potential abnormal noise points whose neighborhood effective sampling density is significantly lower than the preset threshold in the effective data points. Specifically, for a circular arc contour line to be fitted, an effective height mask matrix Vz is constructed according to the height z value of the fitting points: (9) (10) wherein, represents the height value of the mth point on the to-be-fitted circular arc, a sliding window with a size of w_size is adopted to perform local density statistics on VZ, and a one-dimensional convolution kernel k with a width of w_size is defined. (11) The Vz is calculated by one-dimensional convolution, and the zero padding is used in the convolution to ensure that the matrix dimension remains unchanged after the convolution: (12) The convolution result D is normalized to represent the density ratio : (13) The decision rule to define outliers is that a point is considered as an outlier if the density ratio at that point is smaller than a density threshold :​​ (14) (15) Effective point mask matrix Effective by taking the union with the weight matrix W in the least squares fit.

4. The three-dimensional point cloud based method of claim 1, wherein, The step 4 is specifically: Based on the results obtained by the weighted least squares fitting in step 3, for each to-be-fitted circular arc profile line, its fitting center and fitting radius can be obtained, and the effective point cloud data involved in the fitting is re-substituted into the implicit circle equation composed of the fitting center and the fitting radius, so that the Euclidean distance of each point to the center can be calculated, which is recorded as the true radial distance d m , using the prior knowledge that the mandrel structure has a stable radius in physics, a reasonable radius tolerance threshold r threshold , a dynamic radius strip region is constructed: (16) wherein, represents a radius mask matrix on a certain arc contour line, represents a fitting radius of a certain arc to be fitted, only when the real distance is within [ , ], it will be considered as a valid point to participate in the subsequent fitting operation, otherwise it will be treated as an abnormal noise point and eliminated.

5. The three-dimensional point cloud based method of claim 4, wherein, The step 5 is specifically: According to the effective point set screened in step 4, obtain the residual set res corresponding to each point m , the mean of the residual set is counted and the standard deviation , according to the 3σ principle under the normal distribution assumption, about 99.7% of the data should fall within the mean ± 3 times the standard deviation range, therefore, the statistical effectiveness criterion is defined as follows: (17) wherein, represents an effective mask matrix over all profiles that satisfies a normal distribution, only keeping the points satisfying this condition as the final high-quality input point set for the next round of iteration fitting.

6. The three-dimensional point cloud-based method of claim 5, wherein, The step 6 is specifically: Step 6.1: According to the mask matrix obtained by the multi-level filtering calculation of steps 4 and 5, a weight matrix for the second least square circle fitting is designed: (18) wherein, represents the radius mask matrix based on prior geometric constraints in step 4, represents the valid mask matrix that satisfies the 3σ criterion on all profiles in step 5, represents the weight matrix in the second weighted least squares fitting process. Step 6.2: The Participate in the second weighted least squares fitting process, get high-precision center, radius and residual information.

7. The three-dimensional point cloud based method of claim 5, wherein, The step 7 is specifically: Step 7.1: For a certain arc profile to be fitted, when the second fitting is done, the radial deviation res2 of all points is recalculated based on the fitting parameters m i.e. the distance of the real point p m to the center of the fitting circle o2 m minus the fitting radius r2 m , this value reflects the degree of deviation of each point from the ideal arc: (19) If the radial deviation is less than the threshold, a result of no defects of the mandrel is output; otherwise, an abnormal mask binary matrix B is generated, as shown in the following formula: (20) (21) wherein, is a depth anomaly detection threshold value, is an anomaly mask value for a point on the segment of the arc profile to be fitted. Step 7.2: Based on the abnormal mask binary matrix, a binary morphological opening and closing operation sequence based on a structure element is used for topological repair and boundary smoothing processing. First, the morphological opening operation is performed, which is defined as erosion followed by dilation: (22) wherein, represents the erosion operation, only when the structure element K completely covers the foreground region, the center pixel is retained as 1, and as the structure element K slides on the binary matrix B, the isolated points smaller than the structure element K can be effectively eliminated; represents the dilation operation, contrary to the erosion operation, as the structure element K slides on the binary matrix B, as long as there is an overlapping part, the center pixel will be set to 1, which can fill small holes or link broken objects, realize the linkage between adjacent abnormal points, and the K structure matrix adopts a 5*5 square matrix.

8. The three-dimensional point cloud based method of claim 1, wherein, In the step 8, the k-d tree index is constructed for the abnormal point set after the morphological processing, and the fast density clustering is performed, which is specifically: Step 8.1: The set of outliers Q obtained in the above operation, each of which contains three-dimensional coordinates x, y, z, is constructed into a k-d tree to achieve efficient neighborhood query; for a point q k The index set of all points in the spherical region centered at q with radius R is denoted as: (23) wherein denotes the Euclidean distance between the point q k and the point ; Step 8.2: Based on the construction of the k-d tree and the implementation of the efficient neighborhood query, a fast clustering process based on spatial connectivity is performed. The process selects a starting point from all unprocessed abnormal points, and uses the k-d tree to quickly search all neighboring points within a specified radius to form an initial point set; Step 8.3: It is judged whether the point set intersects with any existing cluster that has been generated, that is, whether there are common contained points; Step 8.4: If there is overlap, the current point set and all intersecting clusters are merged to form a larger connected region; if there is no intersection, it is added as a new independent cluster to the cluster set; in the whole process, all points that have been included in a cluster are marked as "visited" to prevent repeated processing, thereby ensuring that each point participates in the clustering decision only once.

9. The three-dimensional point cloud based method of claim 8, wherein, In the step 8, the neighboring clusters are merged and the significant point groups are screened, which is specifically: Step 8.5: A minimum point number threshold is set. If the number of points in all clusters is less than the threshold, a result of no defects of the mandrel is output; otherwise, only the independent clusters with the number of members exceeding the threshold are retained as effective abnormal regions, and the remaining clusters are regarded as invalid noise and are removed to exclude non-substantial point groups formed by residual noise or small disturbances; Step 8.6: For each effective cluster retained, a unique group identifier is assigned, and the identifier is attached to the feature data of the corresponding point.

Images