Wind power pile foundation scouring detection method based on side scan sonar

Abstract

The invention discloses a wind power pile foundation scour detection method based on side-scan sonar. The method comprises the following steps: reading a wind power pile foundation side-scan sonar image to be scoured and detected; performing image processing on the wind power pile foundation side-scan sonar image to obtain a wind power pile foundation scouring pit grey-scale map; performing three-dimensional reconstruction on the wind power pile foundation scouring pit grey-scale map to obtain pile foundation scouring pit point cloud data; multiple pieces of independent scouring pit point cloud data are obtained from the pile foundation scouring pit point cloud data; and performing volume calculation on the point cloud data of each independent scouring pit to realize scouring detection. According to the method, accurate quantification of scour volume detection is achieved, and an advanced and effective technical means is provided for management and maintenance of the offshore wind power pile foundation.

Description

Technical Field

[0001] This invention belongs to the field of marine engineering technology, specifically relating to a method for detecting scour of wind turbine pile foundations based on side-scan sonar. Background Technology

[0002] In the operation and maintenance of offshore wind turbine foundations, how to efficiently and accurately obtain the underwater scouring status and scouring information of offshore wind turbine foundations has always been a technical problem that the offshore wind power industry urgently needs to solve.

[0003] Studies have shown that scour pits formed around pile foundations by ocean currents significantly reduce structural strength, thereby shortening the lifespan of the pile foundations. Traditional manual underwater measurements not only suffer from high visibility dependence and low quantification accuracy, but also involve significant underwater operational risks. With advancements in sonar technology, sonar detection for pile foundation scour has been widely applied in recent years. Chinese invention patent application number "202411659897.9" discloses a multi-beam sonar method for detecting pile foundation scour, achieving effective detection of wind turbine pile foundation scour. However, multi-beam sonar is expensive, and the detection requires slow, multi-directional data acquisition around the pile foundation, resulting in low operational efficiency. Furthermore, the AlphaShape optimization algorithm proposed in the patented multi-beam sonar pile foundation detection method mainly uses the shape control parameter α to iteratively update the volume to obtain the scour pit volume, but it cannot individually adjust each tetrahedron of the model, i.e., it cannot correct the volume contribution of each tetrahedron, thus affecting the accuracy of scour detection. Compared to multibeam sonar, side-scan sonar is not only cheaper but also has a wider coverage area per scan and higher operational efficiency, providing a new approach for wind turbine pile foundation scour detection. However, side-scan sonar obtains high-resolution two-dimensional acoustic images. For accurate quantitative detection of wind turbine pile foundation scour, it is necessary to reconstruct the scour pit sonar images in three dimensions to achieve the quantitative detection objective. Chinese invention patent application number "202411083322.7" discloses a three-dimensional reconstruction method for side-scan sonar images, proposing an SFS three-dimensional reconstruction method based on a hybrid reflection model. This method uses multiple adaptive parameters and complex illumination expressions for three-dimensional image reconstruction, but it is highly dependent on external information and difficult to obtain stable three-dimensional reconstruction results, making it difficult to meet the requirements for accurate quantitative detection of wind turbine pile foundation scour.

[0004] Against the backdrop of global energy structure transformation and increasing demands for sustainable development, offshore wind power has experienced rapid development as a clean and efficient renewable energy source. However, this rapid expansion in installed capacity is accompanied by a surge in operation and maintenance (O&M) requirements, with O&M costs being substantial, accounting for approximately 30% of offshore wind power generation costs—twice that of onshore wind power O&M. How to efficiently and accurately obtain the underwater scour status of offshore wind turbine foundations has remained a critical technical challenge for the offshore wind power industry. Summary of the Invention

[0005] Purpose of the invention: In order to solve the problem of how to efficiently and accurately obtain the underwater scour status and scour information of offshore wind turbine pile foundations, this invention proposes a wind turbine pile foundation scour detection method based on side-scan sonar to obtain high-precision scour pit volume data, thereby significantly improving the reliability of wind turbine pile foundation operation and maintenance.

[0006] Technical solution: This invention proposes a method for detecting scour of wind turbine pile foundations based on side-scan sonar, comprising the following steps:

[0007] Read the side-scan sonar images of the wind turbine pile foundation to be scoured;

[0008] Image processing was performed on the side-scan sonar images of the wind turbine pile foundation to obtain a grayscale image of the scour pits of the wind turbine pile foundation.

[0009] The grayscale image of the wind turbine pile foundation scour pit is reconstructed in three dimensions to obtain the point cloud data of the pile foundation scour pit;

[0010] Multiple individual scour pit point cloud data are obtained from the pile foundation scour pit point cloud data;

[0011] Volume calculation is performed on the point cloud data of each individual scour pit to achieve scour detection.

[0012] Furthermore, the image processing of the side-scan sonar images of the wind turbine pile foundation to obtain a grayscale image of the wind turbine pile foundation scour pits specifically includes:

[0013] The side-scan sonar images of wind turbine pile foundations are converted to grayscale to obtain grayscale images;

[0014] The grayscale image is binarized to obtain a binary image containing the scour pit region;

[0015] Connectivity component identification is performed on the binary graph containing scour pits. Based on the size, centroid, and boundary features of each connected component, the identified connected components are filtered to obtain the retained connected components.

[0016] After performing neighborhood expansion on all retained connected domains, merging the regions of the retained connected domains after neighborhood expansion, and removing the pile shadows of non-real scour pits, a binary map with the pile shadows of non-real scour pits removed is obtained.

[0017] The binary image of the pile foundation with the non-real scour pits removed is merged with the grayscale image to obtain the grayscale image of the wind turbine pile foundation scour pits.

[0018] Furthermore, the process of identifying connected components in the binary graph containing the scour pit area involves filtering the identified connected components based on their size, centroid, and boundary features to obtain the retained connected components. Specific operations include:

[0019] Suppose that in the binary image containing the scour pit region there exists a path (p1, p2, ..., p...) consisting of n points with a pixel value of 255. n If the distance between adjacent points is less than the neighborhood distance threshold, then points p1 and p2 are defined. n It is connected, represented as:

[0020]

[0021] In the formula, ‖·‖ represents the Euclidean distance, and ϵ is the neighborhood distance threshold;

[0022] After obtaining multiple connected components, let the sum of the gray values ​​of all pixels in the i-th connected component be . If it satisfies the following formula, then the connected component is retained;

[0023]

[0024] Let W be the width of the side-scan sonar image of the wind turbine pile foundation, then the x-coordinate of the middle column of the side-scan sonar image of the wind turbine pile foundation is c:

[0025]

[0026] Let c be the x-coordinate of the centroid of the connected components retained after filtering by the sum of gray values. x Connected regions that satisfy the following equation are preserved:

[0027]

[0028] In the formula, δ represents the centroid distance threshold;

[0029] For all connected components retained after centroid filtering, based on the maximum coordinate (x) of each connected component... max , y max Minimum coordinates (x) min , y min Perform bounding box filtering and ultimately retain connected components that meet the following conditions:

[0030]

[0031] In the formula, w is the width of the bounding box, h is the length of the bounding box, and w min h is the minimum width threshold. min This is the minimum length threshold.

[0032] Furthermore, the process of expanding the neighborhood of all retained connected components, merging the regions of the expanded retained connected components, and removing the pile shadows of non-real scour pits to obtain a binary image with the pile shadows of non-real scour pits removed specifically includes:

[0033] Suppose that m connected components are preserved, forming a set of connected components R(m);

[0034] For each connected component, search for neighboring pixels with a value of 0 around each pixel in the connected component, but whose distance from each other does not exceed the expansion threshold, and perform the following k iterations of expansion:

[0035]

[0036] In the formula, (x', y') and (x, y) represent the pixel coordinates before and after expansion, respectively; R k (m) 、 R k-1 (m) represent the set of m connected components after k and k-1 iterations, respectively, and ω is the expansion threshold;

[0037] The iterative expansion stops and a new set of connected components R'(m) is formed when the following conditions are met:

[0038]

[0039] In the formula, d max Indicates the maximum expansion distance;

[0040] Calculate the minimum boundary distance d(R'(i), R'(j)) between any two connected components R'(i) and R'(j) in the new set of connected components R'(m):

[0041]

[0042] In the formula, ∂R'(i) and ∂R'(j) are the boundary point sets of the connected domains R'(i) and R'(j), respectively;

[0043] Determine whether d(R'(i),R'(j)) is less than or equal to the maximum allowed merging distance threshold l; if so, it means that R'(i) and R'(j) meet the merging condition, and merge R'(i) and R'(j) according to the following formula to obtain a new connected component R'(ij); otherwise, R'(i) and R'(j) belong to two separate connected components.

[0044]

[0045] Identify and remove pile shadows that are not real scour pits from all the individual connected domains obtained, and obtain a binary map of pile shadows that have been removed from real scour pits.

[0046] Furthermore, the step of performing three-dimensional reconstruction of the grayscale image of the wind turbine pile foundation scour pit to obtain the point cloud data of the pile foundation scour pit specifically includes:

[0047] Let the grayscale value of a pixel (x'', y'') in the grayscale image of the scour pit of the wind turbine pile foundation be I(x'', y''), and normalize it as follows:

[0048]

[0049] In the formula, I(x'', y'') and I'(x'', y'') are the gray values ​​of pixel (x'', y'') before and after normalization, respectively; I min I max These are the minimum and maximum grayscale values ​​of the grayscale map of the scour pits in wind turbine pile foundations, respectively.

[0050] Construct the following Gaussian kernel G for I'(x'', y''). σ (ξ, η):

[0051]

[0052] In the formula, σ is the standard deviation, and ξ and η are the horizontal and vertical offsets of the Gaussian kernel center, respectively;

[0053] Based on I'(x'', y'') and G σ The 2D convolution of (ξ, η) yields the initial depth z0 of pixel (x'', y''):

[0054]

[0055] In the formula, (x''−ξ, y''−η) represents the pixel point after the current pixel (x'', y'') is offset by the Gaussian kernel center (ξ, η);

[0056] Repeat the above normalization, Gaussian kernel construction, and 2D convolution steps to calculate the initial depth z0 of each pixel;

[0057] Assume the error function corresponding to pixel (x'', y'') in the grayscale image of the wind turbine pile foundation scour pit is expressed as:

[0058]

[0059] In the formula, s1 and s2 are the illumination parameters in the x-axis direction and y-axis direction, respectively, ∂z / ∂x'' is the slope of the terrain in the x-direction, and ∂z / ∂y'' is the slope of the terrain in the y-direction;

[0060] Based on the initial depth of each pixel and the corresponding error function, the initial depth of each pixel in the grayscale image of wind turbine pile foundation scour pits is solved iteratively using a combination of gradient descent and constraints to achieve a three-dimensional reconstruction of the effective grayscale image of wind turbine pile foundation scour pits.

[0061]

[0062] In the formula, z k-1 (x'', y'') and z k (x'', y'') are the depth estimates of pixel (x'', y'') after the (k-1)th and kth iterations, respectively; when k=1, z 0 (x'', y'')=z 0 λ is the gradient descent step size; μ is the constraint weight; E k-1 (x'', y'') is the error function of pixel (x'', y'') at the (k-1)th iteration, and is calculated as follows:

[0063]

[0064] The parameters λ and μ are obtained as follows:

[0065]

[0066]

[0067] In the formula, △I represents the brightness range of the image;

[0068] The iterative convergence condition for 3D reconstruction is the error E. k-1 The absolute value of (x'', y'') is less than the preset tolerance threshold e. a Or the number of iterations k reaches the preset maximum value k max ,Right now:

[0069]

[0070] Let p = (x'', y'', z'') be the point cloud of the scour pit of the wind turbine pile foundation after k iterations. k If (x'', y'') is represented, then the cluster of scour pits on the wind turbine pile foundation is denoted as P.

[0071] Furthermore, the acquisition of multiple individual scour pit point cloud data from the pile foundation scour pit point cloud data specifically includes:

[0072] Assuming there are n points in the set P of scour pits on wind turbine pile foundations, ∀ i∈[1, …, n], search for the distance p from the point cloud in the set P of scour pits on wind turbine pile foundations. i Point cloud p ∈ P less than r j (j∈[1, …, n]), and form p i The neighborhood point set N(p) i ),Right now:

[0073]

[0074] Based on the neighborhood point set N(p) i) Calculate point cloud p i Covariance matrix Cov i :

[0075]

[0076] In the formula, |·| represents the number of point clouds in the set; It is p i The centroid of the neighborhood point set;

[0077] For the covariance matrix Cov i Perform eigenvalue decomposition to obtain three eigenvalues ​​λ1, λ2, and λ3; the eigenvector n corresponding to the smallest eigenvalue. i =(n x , n y , n z ) T That is, point cloud p i The normal vector;

[0078] Calculate the point cloud p based on the above three eigenvalues. i curvature C i :

[0079]

[0080] After obtaining the curvature of all point clouds, remove point clouds with curvature greater than a set threshold to obtain the filtered point cloud set P. f ;

[0081] After retaining the filtered data, point cloud P f All normal vectors (n) x , n y , n z Point clouds whose components in the x, y, and z directions are greater than the corresponding component thresholds Tx, Ty, and Tz are used to obtain the effective scour pit point cloud set P. i ;

[0082] Depth detection is performed along the z-coordinate direction of the point cloud to remove the effective scour pit point set P. i For point clouds exceeding a depth threshold, only point clouds within a certain threshold are retained to form the bottom point cloud set P. b ;

[0083] The underlying point cloud P is defined according to the following formula. b Clustering the scour pit point cloud in the data yields multiple individual bottom-level point clouds:

[0084] ∀ I, j∈[1, …, s], i≠j, p i ∈ P b p j ∈ P b Find pi Neighborhood set N'(p) i ):

[0085]

[0086] In the formula, ε is the distance threshold between adjacent points;

[0087] Find the underlying point cloud P b All point clouds have neighborhood sets, and the point clouds with more than a preset minimum number of points in their neighborhood sets are defined as core point clouds. Assume there are γ core point clouds in total, and the set formed is P. γ Then the core point cloud satisfies:

[0088]

[0089] In the formula, sum(·) represents the point cloud count;

[0090] core point p o As the initial point of cluster A, then in set P γ p o neighborhood N ∈ (p o And in this neighborhood, find any core point p that is not in cluster A. o Then add it to cluster A, i.e., A←A ∪ p o Finally, perform the same operation until no more core points can be added, completing the final clustering of cluster A; after completing the clustering of cluster A, from set P... γ Then, select non-cluster A core points and perform clustering operations according to cluster A to form clusters B, C, ..., thus forming a bottom-level point cloud dataset P containing several clusters. c ;

[0091] The number of layers n for the stacking expansion of the point cloud in each cluster is determined based on the z-coordinate depth value of the point cloud in each cluster.

[0092]

[0093] In the formula, △z is the depth increment of each point cloud layer; This represents the average depth of the entire point cloud.

[0094] Suppose the point cloud in cluster A is p l =(x l , y l , z l For all point clouds in cluster A, create a new point cloud with n layers; for each layer, increase the z-coordinate depth value by k ⋅ Δz to obtain a k-layer point cloud dataset A. k :

[0095]

[0096] In the formula, k is the current layer number, k=1,2,…,n;

[0097] Merging each layer of point cloud data ultimately forms the target point cloud dataset A. f :

[0098]

[0099] After completing the stacked expansion of cluster A, the expansion steps are repeated for all clusters to finally obtain the initial scour pit point cloud dataset P after morphological restoration. r ;

[0100] The initial scour pit point cloud dataset P after morphological restoration r The point cloud of each cluster is proportionally corrected to obtain the corrected set of scour pit points P. f The proportional correction parameter s is obtained by the following formula:

[0101]

[0102] In the formula, z b This represents the actual depth of the seabed pile foundation.

[0103] Furthermore, the volume calculation of each individual scour pit point cloud data to achieve scour detection specifically includes:

[0104] Let P be the corrected set of scour pits. f The generated tetrahedral set is T = (T1, T2, …, T i ), each tetrahedron T i It consists of 4 points, namely T i =(p i1 , p i2 , p i3 , p i4 Let the coordinates of the four vertices of the tetrahedron be p, ... i1 =(x i1 ,y i1 , z i1 ), p i2 =(x i2 , y i2 , z i2 ), p i3 =(x i3 , y i3 , z i3 ), p i4 =(x i4 , y i4 , z i4 Find the circumcenter c of the four vertices:

[0105]

[0106] According to the circumcenter c=(x ic , y ic , z ic Calculate the circumsphere radius R of each tetrahedron:

[0107]

[0108] Retain the tetrahedron that satisfies circumsphere radius R ≤ α;

[0109] Introducing weighting factor w i The volume contribution of each retained tetrahedron is corrected, and the total scour pit volume is finally obtained by accumulating the volumes of all retained tetrahedrons:

[0110]

[0111] In the formula, w i L is the shortest side of tetrahedron i i min With the longest side L i max The ratio of .

[0112] Beneficial Effects: This invention reads side-scan sonar images of wind turbine pile foundations to be scoured; identifies and extracts scour pits from these images to obtain effective grayscale images of wind turbine pile foundation scour pits; uses the parameter-optimized SFS method to perform 3D reconstruction of the effective grayscale images of wind turbine pile foundation scour pits to obtain pile foundation scour pit point cloud data; filters the scour pit point cloud data based on curvature and normal vectors, and obtains multiple individual scour pit point cloud data through z-coordinate depth descent detection, clustering, and morphological restoration; and uses the weighted AlphaShape algorithm to accurately calculate the scour volume of each scour pit point cloud. Through the above technical means, this invention has the following advantages:

[0113] (1) The method for identifying and extracting scour pits in pile foundation images designed in this invention is based on the characteristics of the terrain after pile foundation scour, such as the large continuous black shadows on the side scan image. Based on connected component analysis and regional feature extraction, features such as size, centroid and boundary are introduced to identify the scour pit area, thus forming the identification and extraction of scour pit area in pile foundation images. This removes the influence of the shadow generated by the pile foundation on subsequent processing, and improves the accuracy of side scan sonar pile foundation scour detection by targeted detection in the later stage.

[0114] (2) The three-dimensional reconstruction of pile foundation scour pit images designed in this invention introduces an iterative process combining gradient descent and constraints based on the SFS method. The termination of the iteration is determined by the absolute value of the error being less than the tolerance threshold or the number of iterations reaching the preset maximum value. The gradient descent iteration has the characteristics of convergence stability and insensitivity to noise, and it does not depend heavily on external information. Therefore, it can stably obtain the three-dimensional reconstruction results. The design of the initial extraction of scour pit point cloud is based on curvature and normal vector to obtain an initial effective scour pit point cloud set. The design of the extraction and morphology restoration of the bottom point cloud of scour pit adopts the density clustering method on the basis of point cloud z-coordinate depth descent detection and morphology restoration, which removes the influence of environmental noise and terrain point cloud noise, thereby accurately extracting each scour pit point cloud, reducing the error of volume calculation, and thus improving the calculation accuracy of scour pit volume.

[0115] (3) The pile foundation scour volume calculation method designed in this invention is based on clustering segmentation and morphological recovery to obtain individual scour pit point clouds, and uses the objective function designed based on volume calculation to process them. By adding weight factors to correct the volume of each tetrahedron, a high-precision scour pit volume is obtained, thereby improving the accuracy of side-scan sonar pile foundation scour detection.

[0116] In summary, the method of this invention achieves precise quantification of scour volume detection, providing an advanced and effective technical means for the management and maintenance of offshore wind power pile foundations. Attached Figure Description

[0117] Figure 1 This is a flowchart of the wind turbine pile foundation scour detection method based on side-scan sonar proposed in this invention.

[0118] Figure 2 This is a grayscale image of a side-scan sonar for a wind turbine pile foundation;

[0119] Figure 3 This is a binary image of a side-scan sonar on a wind turbine pile foundation;

[0120] Figure 4 The connected components in the final retained binary graph of the wind turbine pile foundation side-scan sonar;

[0121] Figure 5 It consists of five independent scour pit areas;

[0122] Figure 6 Grayscale diagram of effective wind turbine pile foundation scour pits;

[0123] Figure 7 Point cloud of scour pits for reconstructed wind turbine pile foundations; Figure 7 (a) in the text represents a three-dimensional view. Figure 7 (b) in the diagram represents the top view;

[0124] Figure 8 A schematic diagram for screening point clouds of scour pits caused by wind turbine pile drivers. Figure 8 In the diagram, (a) represents point cloud filtering based on curvature. Figure 8 (b) in the figure represents point cloud filtering based on normal vectors;

[0125] Figure 9 Point clouds for the underlying scour pits;

[0126] Figure 10 The corrected point cloud of scour pits. Figure 10 (a) in the image represents a low-level scour pit point cloud. Figure 10 (b) in the image represents the corrected top-view point cloud. Figure 10 (c) in the diagram represents the corrected side view point cloud.

[0127] Figure 11 This is a schematic diagram of the four vertices of the point cloud and the circumcenter.

[0128] Figure 12 This is a grayscale image of a side-scan sonar simulation environment containing a single scour pit. Figure 12 (a) in the image represents a grayscale image containing cylindrical scour pits. Figure 12 (b) in the image represents a grayscale image containing square scour pits;

[0129] Figure 13 This is a 3D reconstruction of a side-scan sonar simulation environment containing a single scour pit. Figure 13 (a) in the figure represents a 3D reconstruction of a cylindrical scour pit. Figure 13 (b) in the figure represents a three-dimensional reconstruction of the square scour pit;

[0130] Figure 14 The images show the extracted point clouds of two sets of single scour pits. Figure 14 In the diagram, (a) represents the point cloud of a cylindrical scour pit. Figure 14 (b) in the figure represents a square scour pit point cloud;

[0131] Figure 15 This is a grayscale image of a side-scan sonar simulation environment containing double scour pits. Figure 15 In the image, (a) represents a grayscale image containing two cylindrical scour pits. Figure 15 (b) in the text indicates a scour pit containing two square scour pits;

[0132] Figure 16 This is a 3D reconstruction of a side-scan sonar simulation environment containing double scour pits. Figure 16 (a) in the figure represents a 3D reconstruction of a double-cylindrical scour pit. Figure 16 (b) in the figure represents a three-dimensional reconstruction of the quadrangular scour pit;

[0133] Figure 17 The images show the extracted point cloud results for four groups of single scour pits. Figure 17 In the diagram, (a) represents the point cloud of a large cylindrical scour pit. Figure 17 (b) in the image represents the point cloud of a small cylindrical scour pit. Figure 17 (c) in the figure represents the point cloud of a large rectangular scour pit. Figure 17 In the diagram, (d) represents the point cloud of a small rectangular scour pit;

[0134] Figure 18 These are four sets of side-scan sonar images of offshore wind turbine pile foundations. Figure 18 (a) in the image represents the image of pile #1. Figure 18 (b) in the image represents the image of pile #2. Figure 18 (c) in the image represents the image of pile #3. Figure 18 (d) in the image represents the image of pile #4.

[0135] Figure 19 The data consists of point cloud data reconstructed from four sets of images in 3D. Figure 19 (a) in the text represents the point cloud data of pile foundation #1. Figure 19 (b) in the figure represents the point cloud data of pile foundation #2. Figure 19 (c) in the text represents the point cloud data of pile foundation #3. Figure 19 (d) in the text represents the point cloud data of pile foundation #4;

[0136] Figure 20 These are point cloud data after morphological restoration of four sets of images. Figure 20 In the diagram, (A) represents the point cloud data of pile foundation #1, which includes four scour pits: (a), (b), (c), and (d). Figure 20 (B) in the figure represents the point cloud data of pile foundation #2, which includes four scour pits: (a), (b), (c), and (d). Figure 20 (C) in the figure represents the point cloud data of pile foundation #3, which includes four scour pits: (a), (b), (c), and (d). Figure 20 (D) in the figure represents the point cloud data of pile foundation No. 4, which includes three scour pits (a), (b), and (c). Detailed Implementation

[0137] The technical solution of the present invention will now be further described in conjunction with the accompanying drawings and embodiments.

[0138] like Figure 1 As shown in the figure, this invention proposes a method for detecting scour of wind turbine pile foundations based on side-scan sonar to obtain high-precision scour pit volume data. The specific steps are as follows:

[0139] Step 1: Read the side-scan sonar image of the wind turbine pile foundation to be scoured;

[0140] Step 2: Binarize the side-scan sonar image of the wind turbine pile foundation to obtain a binary image containing the scour pit area. This includes the following operations:

[0141] Step 2-1: Convert the side-scan sonar image of the wind turbine pile foundation from RGB format to grayscale image using the following grayscale conversion operation, such as... Figure 2 As shown:

[0142]

[0143] In the formula, I R I G I B These are the red, green, and blue pixel values ​​of a certain pixel in the side-scan sonar image of the wind turbine pile foundation, and I is the grayscale value of that pixel after conversion.

[0144] Step 2-2: Further convert the grayscale image into a binary image containing the scour pit region through binarization, such as... Figure 3 As shown, this is done to distinguish the pile foundation area from the background area in the image, so as to extract the scour pit area.

[0145] Step 3: Extract scour pits based on connected components and region features. Specific operations include:

[0146] Step 3-1: Identify multiple connected components in the binary graph containing the scour pit area;

[0147] Suppose there exists a path (p1, p2, ..., p) in a binary image consisting of n points with a pixel value of 255. n If the distance between adjacent points is less than a certain value, then points p1 and p2 are defined. n It is connected, that is:

[0148]

[0149] In the formula, ‖·‖ represents the Euclidean distance, and ϵ is the neighborhood distance threshold;

[0150] Step 3-2: After obtaining multiple connected domains, the pile foundation area is extracted by sequentially filtering based on the size, centroid, and boundary features of each connected domain.

[0151] Since the size of each connected component depends on the number of pixels with a grayscale value of 255, the size of the connected components can be selected based on the sum of the grayscale values ​​of the pixels in each connected component, and will be greater than a grayscale value threshold. Connected components are preserved to avoid extracting excessively small noise points. Let the sum of the gray values ​​of all pixels in the i-th connected component be... If it satisfies the following formula, the flow domain is retained to avoid excessively small noise being extracted;

[0152]

[0153] Since the central region of the side-scan sonar image is the seabed line, and the pile foundation area is not near this region, centroid filtering is performed on the retained areas after size filtering in order to preserve the pile foundation areas far from the seabed line.

[0154] Let W be the width of the side-scan sonar image of the wind turbine pile foundation, then the x-coordinate of the middle column of the image is c:

[0155]

[0156] Let c be the x-coordinate of the centroid of the region retained after size filtering. x If the region is far from the x-coordinate c, then the region is further preserved using the following formula:

[0157]

[0158] In the formula, δ represents the centroid distance threshold;

[0159] Since the retained region may contain elongated and irregularly shaped non-target regions, for all connected components retained after centroid filtering, the maximum and minimum coordinates (x, y, y) of each connected component are used again. max , y max ), (x min , y min Bounding box filtering is performed, that is, connected components that meet the following conditions are ultimately retained, such as... Figure 4 As shown:

[0160]

[0161] In the formula, w is the width of the bounding box; h is the length of the bounding box; w min h is the minimum width threshold. min Minimum length threshold;

[0162] Suppose that m connected components are retained after bounding box filtering, then the set of these connected components is denoted as R(m).

[0163] Step 4: Expand and merge regions while preserving connected domains to obtain a grayscale map of effective wind turbine pile foundation scour pits. This includes the following operations:

[0164] Step 4-1: Due to noise, uneven lighting, and imperfect thresholding in the image, the retained connected component set R(m) may only contain a portion of the scour pit region. To ensure complete extraction of the scour pit region, a neighborhood search method is used for region expansion.

[0165] For each connected region in R(m), search for neighboring pixels with a value of 0 around each pixel in the region, but whose distance from each other does not exceed a threshold, and perform the following k-iteration expansion:

[0166]

[0167] In the formula, (x', y') and (x, y) represent the pixel coordinates before and after expansion, respectively; R k (m) 、 R k-1 (m) represents the set of m connected components after k and k-1 iterations, respectively; ω is the expansion threshold.

[0168] The iteration should stop when the following conditions are met:

[0169]

[0170] In the formula, d max This represents the maximum extended distance.

[0171] Iterative expansion of all connected regions in the connected region set R(m) yields all complete scour pit regions and forms a new connected region set R'(m).

[0172] Step 4-2: After the connected domains are expanded, in order to ensure that the connected or overlapping domains can be effectively merged, the regions are merged.

[0173] Calculate the minimum boundary distance d(R'(i), R'(j)) between any two connected components R'(i) and R'(j) in the set of connected components R'(m):

[0174]

[0175] In the formula, ∂R'(i) and ∂R'(j) are the boundary point sets of the connected domains R'(i) and R'(j), respectively.

[0176] Determine whether R'(i) and R'(j) satisfy the merging condition, i.e., whether d(R'(i),R'(j)) is less than or equal to the maximum allowed merging distance threshold l; if so, merge R'(i) and R'(j) according to the following formula to obtain a new connected component R'(ij); otherwise, R'(i) and R'(j) belong to two separate connected components.

[0177]

[0178] By repeating the process of merging regions, all the individual connected domains can be formed, thus creating each independent scour pit region.

[0179] like Figure 5 As shown, for Figure 4 After regional expansion and merging, five independent scour pit regions were obtained.

[0180] Step 4-3: Due to the presence of pile shadows that are not actual scour pits in the expanded and merged scour pit area, in order to obtain the effective scour pit area, these shadows need to be identified and effectively removed. Simultaneously, considering that the 3D reconstruction of the wind turbine pile foundation side-scan sonar image requires grayscale information, the binary image with the removed pile shadows is merged with the grayscale image of the pile foundation scour pit to obtain the final effective grayscale image of the wind turbine pile foundation scour pit, as shown below. Figure 6 As shown.

[0181] Step 5: Using the parameter-optimized SFS method, a 3D reconstruction of the grayscale image of the effective wind turbine pile foundation scour pit is performed to obtain the pile foundation scour pit point cloud data. Since the accurate calculation of the volume of the wind turbine pile foundation scour pit is based on 3D point cloud data, it is necessary to perform a 3D reconstruction of the acquired grayscale image of the effective wind turbine pile foundation scour pit to obtain the pile foundation point cloud data. Because the Shadow Recovering Shape Method (SFS method) can deduce the 3D shape features of an object's surface based on the brightness variations of the object's surface in a 2D image, combined with an optical reflection model, it can be used to perform a 3D reconstruction of the grayscale image of the effective wind turbine pile foundation scour pit. However, in the conventional SFS method, the initial depth value of the pile foundation image at the z-coordinate is randomly selected, which can easily reduce the convergence speed during reconstruction iterations. Therefore, the initial depth value of the z-coordinate is generated by Gaussian kernel convolution, and a brightness-driven adaptive parameter is constructed to optimize the SFS method.

[0182] The specific operations include:

[0183] Step 5-1: Obtaining the initial depth value of the z-coordinate of the pile foundation image in the SFS method.

[0184] Let the grayscale value of a pixel (x'', y'') in the grayscale image of the effective wind turbine pile foundation scour pit be I(x'', y''), and normalize it as follows:

[0185]

[0186] In the formula, I(x'', y'') and I'(x'', y'') are the gray values ​​of pixel (x'', y'') before and after normalization, respectively; I min I max These are the minimum and maximum grayscale values ​​of the effective wind turbine pile foundation scour pit grayscale map, respectively.

[0187] Construct the following Gaussian kernel G for I'(x'', y''). σ (ξ, η):

[0188]

[0189] In the formula, σ is the standard deviation, which determines the smoothness; ξ and η are the horizontal and vertical offsets of the Gaussian kernel center, respectively.

[0190] Based on I'(x'', y'') and G σ The 2D convolution of (ξ, η) yields the initial depth z0 of pixel (x'', y''):

[0191]

[0192] In the formula, (x''−ξ, y''−η) represents the pixel point after the current pixel (x'', y'') of the image is offset by the Gaussian kernel center (ξ, η).

[0193] Repeat the above normalization, Gaussian kernel construction, and 2D convolution steps to calculate the initial depth z0 of each pixel.

[0194] The initial depth provides a reasonable starting point for subsequent 3D reconstruction iterations, reducing the number of iterations. Furthermore, this initial depth is closer to the initial solution of the actual terrain, thus helping to improve the grayscale of the effective wind turbine pile foundation scour pit area. Figure 3 Accuracy in dimensional reconstruction and volume calculation.

[0195] Step 5-2: 3D reconstruction of grayscale map of effective wind turbine pile foundation scour pits.

[0196] Since the brightness of measured side-scan sonar images of wind turbine pile foundations is usually affected by both lighting conditions and seabed surface shape, it is necessary to establish the relationship between image brightness and seabed surface shape in order to infer the precise depth of seabed topography from image brightness information and thus accurately reconstruct the topography surface.

[0197] Assuming the grayscale image of the effective wind turbine pile foundation scour pit is I(x'', y''), the error function corresponding to pixel (x'', y'') is expressed as:

[0198]

[0199] In the formula, s1 and s2 are the illumination parameters in the x-axis and y-axis directions, respectively, ∂z / ∂x'' is the slope of the terrain in the x-direction, and ∂z / ∂y'' is the slope of the terrain in the y-direction.

[0200] Based on the initial depth of each pixel and the corresponding error function, the initial depth of each pixel in the grayscale image of the effective wind turbine pile foundation scour pit is solved iteratively using a combination of gradient descent and constraints to achieve the 3D reconstruction of the grayscale image of the effective wind turbine pile foundation scour pit:

[0201]

[0202] In the formula, z k-1 (x'', y'') and z k(x'', y'') are the depth estimates of pixel (x'', y'') after the (k-1)th and kth iterations, respectively. When k=1, z 0 (x'', y'')=z 0 Λ is the gradient descent step size, i.e., the magnitude of depth iteration adjustment; μ is the constraint weight, used to adjust the balance between the current depth and the initial depth; E k-1 (x'', y'') is the error function of pixel (x'', y'') at the (k-1)th iteration, and is calculated as follows:

[0203]

[0204] The parameters λ and μ are obtained as follows:

[0205]

[0206]

[0207] In the formula, △I represents the brightness range of the image, which is the difference between the maximum and minimum pixel grayscale values.

[0208] The iterative convergence condition for 3D reconstruction is the error E. k-1 The absolute value of (x'', y'') is less than the preset tolerance threshold e. a Or the number of iterations k reaches the preset maximum value k max ,Right now:

[0209]

[0210] Let p = (x'', y'', z'') be the point cloud of the scour pit of the wind turbine pile foundation after k iterations. k If (x'', y'') is represented, then the cluster of scour pits on the wind turbine pile foundation is denoted as P.

[0211] like Figure 7 The image shows the reconstructed point cloud of scour pits in the wind turbine foundation. The color represents the depth of the point cloud, with darker colors indicating greater depth.

[0212] Step 6: Filter the scour pit point cloud data based on curvature and normal vectors, and obtain multiple individual scour pit point cloud datasets through z-coordinate depth descent detection, clustering, and morphological restoration. Specifically, this includes:

[0213] Step 6-1: Filter the 3D point cloud data based on curvature and normal vector.

[0214] To efficiently extract the scour point cloud from the wind turbine pile foundation scour point cloud set P, it is necessary to filter out the non-depression point cloud. For example... Figure 8The figure shows a schematic diagram of point cloud filtering for scour pits in wind turbine pile drivers. As can be seen from the figure, there is a significant difference in curvature and normal vector between the pile foundation point clouds in concave and non-concave areas. Therefore, point clouds in concave areas can be filtered first based on curvature; then, point clouds in scour pits can be effectively and accurately extracted based on normal vectors. The calculation of curvature also requires the normal vector of the point cloud, so it is extracted first.

[0215] Assuming there are n points in the wind turbine pile foundation point cloud P, ∀ i∈[1, …,n], search for the distance p from the point cloud in the wind turbine pile foundation scour pit point cloud P. i Point cloud p ∈ P less than r j (j∈[1, …,n]), and form p i The neighborhood point set N(p) i ),Right now:

[0216]

[0217] Based on the neighborhood point set N(p) i ) Calculate point cloud p i Covariance matrix Cov i :

[0218]

[0219] In the formula, |·| is the cardinality of the set, that is, the number of point clouds in the set; It is p i The centroid of the neighborhood point set.

[0220] For the covariance matrix Cov i Perform eigenvalue decomposition to obtain three eigenvalues ​​λ1, λ2, and λ3; the eigenvector n corresponding to the smallest eigenvalue. i =(n x , n y , n z ) T That is, point cloud p i The normal vector.

[0221] Calculate the point cloud p based on the above three eigenvalues. i curvature C i :

[0222]

[0223] Following the above process, after obtaining the curvature of all point clouds, point cloud filtering for concave regions can be performed, i.e., removing point clouds with curvature greater than a set threshold, thus forming a filtered point cloud set P. f The effective and accurate extraction of point clouds of pile foundation scour pits is mainly achieved by retaining P f All normal vectors (n) x, n y , n z Point clouds whose components in the x, y, and z directions are greater than the corresponding component thresholds Tx, Ty, and Tz are used to obtain the initial effective scour pit point cloud set P. i .

[0224] Step 6-2: Obtain multiple individual scour pit point cloud data through z-coordinate depth descent detection, clustering, and morphological restoration.

[0225] Due to the initial effective scour pits clustered after screening, P i The dataset contains environmental noise and terrain point cloud noise. To facilitate subsequent volume calculation of the scour pit point cloud, the underlying point cloud is preserved, and layering and expansion are performed in the vertical direction to reconstruct the points required for volume calculation, thus forming the target scour pit point cloud. The method for preserving the underlying point cloud is to perform depth descent detection along the z-coordinate direction of the point cloud, i.e., removing the initial effective scour pit point cloud set P. i For point clouds exceeding a depth threshold, only point clouds within a certain threshold are retained, forming a bottom-level point cloud set P. b Assume P b There are s point clouds in total, ∀ I ∈ [1,…, s], p i ∈ P b ,but:

[0226]

[0227] In the formula, z i For point cloud p i The z-coordinate depth value, z max For P i The z-coordinate depth value of the highest point cloud in the middle; △z is the depth threshold.

[0228] like Figure 9 As shown, this is the obtained point cloud of the underlying scour pits.

[0229] Because interference issues can occur when calculating the volume of scour pit point clouds, density clustering is used to cluster the bottom-level scour pit point clouds as follows, thereby obtaining multiple individual bottom-level point clouds.

[0230] ∀ I, j∈[1, …, s], i≠j, p i ∈ P b p j ∈ P b Find p i Neighborhood set N'(p) i ):

[0231]

[0232] In the formula, ε is the distance threshold between adjacent points.

[0233] Find the underlying point cloud P b The set of all point clouds is defined as a neighborhood set, and the point cloud with more than a preset minimum number of points in its neighborhood set is defined as a core point cloud. Assume there are γ core point clouds, and the set formed is P. γ Then the core point cloud satisfies:

[0234]

[0235] In the formula, sum(·) represents the point cloud count.

[0236] Cluster the core points to obtain individual bottom-level point clouds. ∀ o∈[1, …, γ], first cluster the core points p o As the initial point of cluster A, i.e., A = {p o}; then in set P γ p o neighborhood N ∈ (p o And in this neighborhood, find any core point p that is not in cluster A. o Then add it to cluster A, i.e., A←A ∪ p o Finally, perform the same operation, repeating the process for the newly added core points in cluster A in set P. γ The process involves constructing neighborhoods within the new neighborhoods and adding non-cluster A core points until no more core points can be added, thus completing the final clustering of cluster A. After clustering cluster A is complete, the clustering is performed from set P. γ Then, select non-cluster A core points and perform clustering operations on cluster A to form clusters B, C, ..., thus forming a bottom-level point cloud dataset P containing several clusters. c .

[0237] The number of layers n of the point cloud stacking expansion for each cluster is determined based on the z-coordinate depth value of the point cloud in each cluster.

[0238]

[0239] In the formula, △z is the depth increment of each point cloud layer; This represents the average depth of the entire point cloud.

[0240] Suppose the point cloud in cluster A is p l =(x l , y l , z l For all point clouds in cluster A, create a new point cloud with n layers; for each layer, increase the z-coordinate depth value by k ⋅ Δz to obtain a k-layer point cloud dataset A. k :

[0241]

[0242] In the formula, k is the current layer number, k=1,2,…,n;

[0243] Merging each layer of point cloud data ultimately forms the target point cloud dataset A. f :

[0244]

[0245] After completing the stacked expansion of cluster A, the expansion steps are repeated for all clusters to finally obtain the initial scour pit point cloud dataset P after morphological restoration. r ;

[0246] Counter-shrinkage pit set P r The point cloud of each cluster is proportionally corrected to obtain the corrected set of scour pit points P. f The proportional correction parameter s is obtained by the following formula:

[0247]

[0248] In the formula, z b This represents the actual depth of the seabed pile foundation.

[0249] like Figure 10 As shown, this is the final corrected point cloud of the scour pits.

[0250] Step 7: Use the weighted AlphaShape algorithm to perform accurate volume calculation for each scour pit point cloud.

[0251] To accurately assess the scour of wind turbine foundation piles, quantitative analysis can be performed by calculating the volume of the scour pit point cloud. The AlphaShape algorithm uses a shape control parameter α to select the tetrahedrons needed for volume calculation from all tetrahedrons formed by the boundary point cloud, thus calculating the target volume. This algorithm has significant advantages in generating closed surface models and can measure the volume of closed surface models with high accuracy; therefore, it can be used to calculate the volume of the scour pit point cloud.

[0252] like Figure 11 As shown, this algorithm generates multiple tetrahedra by using the four adjacent point clouds at the boundary of the scour crater point cloud to encapsulate the scour crater point cloud, and obtains the scour crater volume by calculating the volume of all tetrahedra. Although this algorithm can control the smoothness and detail of the tetrahedra by adjusting the shape control parameter α, it cannot individually adjust each tetrahedron of the model, i.e., it cannot correct the volume contribution of each tetrahedron, which will affect the accuracy of scour detection.

[0253] Let point cloud P f The generated tetrahedral set is T = (T1, T2, …, Ti ), each tetrahedron T i It consists of 4 points, namely T i =(p i1 , p i2 , p i3 , p i4 Let the coordinates of the four vertices of the tetrahedron be p, ... i1 =(x i1 , y i1 , z i1 ), p i2 =(x i2 , y i2 , z i2 ), p i3 =(x i3 , y i3 , z i3 ), p i4 =(x i4 , y i4 , z i4 From this, we can find the circumcenter c of the four vertices:

[0254]

[0255] Based on the calculated circumcenter c=(x ic , y ic , z ic Calculate the circumsphere radius R of each tetrahedron:

[0256]

[0257] Since the circumsphere radius R of a tetrahedron is closely related to the spatial distribution between its four vertices, a smaller R indicates that the four points are closer together, i.e., the local point cloud is denser. The geometry of the scour pit boundary is composed of a dense point cloud, so its corresponding tetrahedron circumsphere radius is also small; conversely, tetrahedrons with larger circumsphere radii are often composed of sparse points or noise points, making it difficult to represent the scour pit boundary structure. Therefore, tetrahedrons are selected based on R, i.e., tetrahedrons satisfying R ≤ α are retained, and all the tetrahedrons retained at the end form the scour pit model.

[0258] Because point cloud data can be affected by noise and uneven sampling, some tetrahedrons may exhibit elongated or distorted shapes. Therefore, a weighting factor w is introduced. i The volume contribution of each tetrahedron is corrected, and the total scour pit volume is finally obtained by accumulating the volumes of all tetrahedrons:

[0259]

[0260] In the formula w iL is the shortest side of tetrahedron i i min With the longest side L i max The ratio:

[0261] .

[0262] This embodiment focuses on the side-scan sonar images of wind turbine pile foundations acquired by side-scan sonar. Based on the obtained grayscale images of effective wind turbine pile foundation scour pits, the parameter-optimized SFS method is used to realize the three-dimensional reconstruction of the effective wind turbine pile foundation scour pit grayscale images. The weight-corrected AlphaShape algorithm is used to accurately calculate the scour volume of each scour pit point cloud, thereby improving the efficiency and accuracy of side-scan sonar pile foundation scour detection.

[0263] To verify the effectiveness of the side-scan sonar pile foundation scour detection method proposed in this embodiment, a side-scan sonar simulation environment with a single scour pit, a side-scan sonar simulation environment with a double scour pit, and real wind turbine pile foundation side-scan sonar images were used for scour detection verification.

[0264] Figure 12 This is a grayscale image of a side-scan sonar simulation environment containing cylindrical and square single scour pits, obtained by modeling in SolidWorks software, point cloudification in CloudCompare software, adding light sources and adjusting lighting directions. Figure 13 This is a 3D reconstruction of a side-scan sonar simulation environment containing a single scour pit. Figure 14 These are two sets of single scour pit point clouds extracted through dual extraction of curvature and normal vectors, preservation of the underlying point cloud, and subsequent point cloud morphology restoration. Figure 14 As can be seen, the scour pits in each group can be completely extracted and their shapes restored. The final volume detection results are shown in Table 1. The table shows that the wind turbine pile foundation scour detection method of this patent has very good detection accuracy when processing single scour pit point cloud data, with an average error of only 0.34%. This indicates that the method can meet high standards in terms of accuracy.

[0265] Table 1

[0266]

[0267] Figure 15 The grayscale images of the side-scan sonar simulation environments, containing two cylindrical and two square scour pits, were obtained after modeling in SolidWorks software, point cloudification in CloudCompare software, and processing such as adding light sources and lighting directions. Figure 16 This is a 3D reconstruction of a side-scan sonar simulation environment containing double scour pits. Figure 17These are four sets of single-scour crater point clouds extracted through dual extraction of curvature and normal vectors, preservation of the underlying point cloud, and subsequent point cloud morphology restoration. Figure 17 As can be seen, each group can be completely extracted and its shape restored into an individual scour pit point cloud. The final volume detection results are shown in Table 2. As can be seen from the table, the maximum volume detection error of the four independent scour pits in the two simulation environments is 5.80%, and the average detection error is 4.775%. This shows that the wind power pile foundation scour detection method of this patent can still effectively and accurately perform volume calculations when facing a dual scour pit model environment.

[0268] Table 2

[0269]

[0270] For the two types of scour pit side-scan sonar simulation environments mentioned above, the wind turbine pile foundation scour detection method designed in this embodiment can identify scour pit areas based on the characteristics of the terrain after pile foundation scour, such as large continuous black shadows on the side-scan image. It introduces features such as size, centroid, and boundary to identify scour pit areas based on connected component analysis and regional feature extraction. This forms the scour pit area identification and extraction of wind turbine pile foundation test sonar images, removes the influence of shadows generated by the pile foundation on subsequent processing, and thus improves the accuracy of side-scan sonar pile foundation scour detection. The designed 3D reconstruction of wind turbine pile foundation scour pit images is based on the SFS method, using an iterative approach combining gradient descent and constraints. It determines whether the iteration terminates when the absolute value of the error is less than a tolerance threshold or the number of iterations reaches a preset maximum value, thus obtaining 3D point cloud data. A point cloud filtering method based on curvature and normal vectors is designed to obtain an initial set of valid scour pit point clouds. The designed extraction and morphological restoration of the bottom-layer point cloud of scour pits incorporates density clustering based on point cloud z-coordinate depth descent detection and morphological restoration, removing the influence of environmental noise and terrain noise, thereby accurately extracting the point cloud of each scour pit and reducing the error in volume calculation. The designed wind turbine pile foundation scour volume calculation method, based on the individual scour pit point clouds obtained from clustering segmentation and morphological restoration, corrects the point cloud data through proportional correction parameters and adds weighting factors to correct the volume contribution of each tetrahedron, thereby improving the accuracy of side-scan sonar detection of wind turbine pile foundation scour.

[0271] Figure 18 and Figure 19 Four sets of side-scan sonar images of offshore wind turbine pile foundations and their three-dimensional reconstructed point cloud maps are presented. Figure 20 The figure shows the point cloud of the scour pit for each pile foundation. As can be clearly seen from the figure, after the extraction and 3D reconstruction of the scour pit area in the side-scan sonar image of the actual offshore wind power pile foundation, a relatively complete scour pit point cloud can still be obtained through point cloud extraction and morphology restoration, and its accurate scour pit volume can still be calculated, as shown in Table 3.

[0272] Table 3

[0273]

[0274] Therefore, the wind turbine pile foundation scour detection method based on side-scan sonar in this embodiment realizes reliable scour volume detection of wind turbine pile foundation scour pits based on accurate reconstruction of the wind turbine pile foundation, providing a guarantee for efficient and high-precision operation and maintenance of offshore wind turbine pile foundations.

Claims

1. A method for detecting scour of wind turbine pile foundations based on side-scan sonar, characterized in that: Includes the following steps: Read the side-scan sonar images of the wind turbine pile foundation to be scoured; Image processing was performed on the side-scan sonar images of the wind turbine pile foundation to obtain a grayscale image of the scour pits of the wind turbine pile foundation. The grayscale image of the wind turbine pile foundation scour pit is reconstructed in three dimensions to obtain the point cloud data of the pile foundation scour pit; Multiple individual scour pit point cloud data are obtained from the pile foundation scour pit point cloud data; Volume calculation is performed on the point cloud data of each individual scour pit to achieve scour detection.

2. The method for detecting scour of wind turbine pile foundations based on side-scan sonar according to claim 1, characterized in that: The image processing of the side-scan sonar images of the wind turbine pile foundation to obtain a grayscale image of the wind turbine pile foundation scour pits specifically includes: The side-scan sonar images of wind turbine pile foundations are converted to grayscale to obtain grayscale images; The grayscale image is binarized to obtain a binary image containing the scour pit region; Connectivity component identification is performed on the binary graph containing scour pits. Based on the size, centroid, and boundary features of each connected component, the identified connected components are filtered to obtain the retained connected components. After performing neighborhood expansion on all retained connected domains, merging the regions of the retained connected domains after neighborhood expansion, and removing the pile shadows of non-real scour pits, a binary map with the pile shadows of non-real scour pits removed is obtained. The binary image of the pile foundation with the non-real scour pits removed is merged with the grayscale image to obtain the grayscale image of the wind turbine pile foundation scour pits.

3. The method for detecting scour of wind turbine pile foundations based on side-scan sonar according to claim 2, characterized in that: The process of identifying connected components in a binary graph containing scour pits involves filtering the identified connected components based on their size, centroid, and boundary features to obtain the retained connected components. Specific operations include: Suppose that in the binary image containing the scour pit region there exists a path (p1, p2, ..., p) consisting of n points with a pixel value of 255. n If the distance between adjacent points is less than the neighborhood distance threshold, then points p1 and p2 are defined. n It is connected, represented as: ； In the formula, ‖·‖ represents the Euclidean distance, and ϵ is the neighborhood distance threshold; After obtaining multiple connected components, let the sum of the gray values ​​of all pixels in the i-th connected component be . If it satisfies the following formula, then the connected component is retained; ； Let W be the width of the side-scan sonar image of the wind turbine pile foundation, then the x-coordinate of the middle column of the side-scan sonar image of the wind turbine pile foundation is c: ； Let c be the x-coordinate of the centroid of the connected components retained after filtering by the sum of gray values. x Connected regions that satisfy the following equation are preserved: ； In the formula, δ represents the centroid distance threshold; For all connected components retained after centroid filtering, based on the maximum coordinate (x) of each connected component... max , y max Minimum coordinates (x) min , y min Perform bounding box filtering and ultimately retain connected components that meet the following conditions: ； In the formula, w is the width of the bounding box, h is the length of the bounding box, and w min h is the minimum width threshold. min This is the minimum length threshold.

4. The method for detecting scour of wind turbine pile foundations based on side-scan sonar according to claim 3, characterized in that: The process of expanding the neighborhood of all retained connected components, merging the regions of the expanded retained connected components, and removing the pile shadows of non-real scour pits to obtain a binary image with the pile shadows of non-real scour pits removed specifically includes: Suppose that m connected components are preserved, forming a set of connected components R(m); For each connected component, search for neighboring pixels with a value of 0 around each pixel in the connected component, but whose distance from each other does not exceed the expansion threshold, and perform the following k iterations of expansion: ； In the formula, (x', y') and (x, y) represent the pixel coordinates before and after expansion, respectively; R k (m) 、 R k-1 (m) represent the set of m connected components after k and k-1 iterations, respectively, and ω is the expansion threshold; The iterative expansion stops and a new set of connected components R'(m) is formed when the following conditions are met: ； In the formula, d max Indicates the maximum expansion distance; Calculate the minimum boundary distance d(R'(i), R'(j)) between any two connected components R'(i) and R'(j) in the new set of connected components R'(m): ； In the formula, ∂R'(i) and ∂R'(j) are the boundary point sets of the connected domains R'(i) and R'(j), respectively; Determine whether d(R'(i),R'(j)) is less than or equal to the maximum allowed merging distance threshold l; if so, it means that R'(i) and R'(j) meet the merging condition, and merge R'(i) and R'(j) according to the following formula to obtain a new connected component R'(ij); otherwise, R'(i) and R'(j) belong to two separate connected components. ； Identify and remove pile shadows that are not real scour pits from all the individual connected domains obtained, and obtain a binary map of pile shadows that have been removed from real scour pits.

5. The method for detecting scour of wind turbine pile foundations based on side-scan sonar according to claim 1, characterized in that: The process of performing three-dimensional reconstruction of the grayscale image of the wind turbine pile foundation scour pit to obtain point cloud data of the pile foundation scour pit specifically includes: Let the grayscale value of a pixel (x'', y'') in the grayscale image of the scour pit of the wind turbine pile foundation be I(x'', y''), and normalize it as follows: ； In the formula, I(x'', y'') and I'(x'', y'') are the gray values ​​of pixel (x'', y'') before and after normalization, respectively; I min I max These are the minimum and maximum grayscale values ​​of the grayscale map of the scour pits in wind turbine pile foundations, respectively. Construct the following Gaussian kernel G for I'(x'', y''). σ (ξ, η): ； In the formula, σ is the standard deviation, and ξ and η are the horizontal and vertical offsets of the Gaussian kernel center, respectively; Based on I'(x'', y'') and G σ The 2D convolution of (ξ, η) yields the initial depth z0 of pixel (x'', y''): ； In the formula, (x''−ξ, y''−η) represents the pixel point after the current pixel (x'', y'') is offset by the Gaussian kernel center (ξ, η); Repeat the above normalization, Gaussian kernel construction, and 2D convolution steps to calculate the initial depth z0 of each pixel; Assume the error function corresponding to pixel (x'', y'') in the grayscale image of the wind turbine pile foundation scour pit is expressed as: ； In the formula, s1 and s2 are the illumination parameters in the x-axis direction and y-axis direction, respectively, ∂z / ∂x'' is the slope of the terrain in the x-direction, and ∂z / ∂y'' is the slope of the terrain in the y-direction; Based on the initial depth of each pixel and the corresponding error function, the initial depth of each pixel in the grayscale image of wind turbine pile foundation scour pits is solved iteratively using a combination of gradient descent and constraints to achieve a three-dimensional reconstruction of the effective grayscale image of wind turbine pile foundation scour pits. ； In the formula, z k-1 (x'', y'') and z k (x'', y'') are the depth estimates of pixel (x'', y'') after the (k-1)th and kth iterations, respectively; when k=1, z 0 (x'', y'')=z 0 λ is the gradient descent step size; μ is the constraint weight; E k-1 (x'', y'') is the error function of pixel (x'', y'') at the (k-1)th iteration, and is calculated as follows: ； The parameters λ and μ are obtained as follows: ； ； In the formula, △I represents the brightness range of the image; The iterative convergence condition for 3D reconstruction is the error E. k-1 The absolute value of (x'', y'') is less than the preset tolerance threshold e. a Or the number of iterations k reaches the preset maximum value k max ,Right now: ； Let p = (x'', y'', z'') be the point cloud of a wind turbine pile foundation scour pit after k iterations. k If (x'', y'') is represented, then the cluster of scour pits on the wind turbine pile foundation is denoted as P.

6. The method for detecting scour of wind turbine pile foundations based on side-scan sonar according to claim 1, characterized in that: The acquisition of multiple individual scour pit point cloud data from the pile foundation scour pit point cloud data specifically includes: Assuming there are n points in the set P of scour pits on wind turbine pile foundations, ∀ i∈[1, …, n], search for the distance p from the point cloud in the set P of scour pits on wind turbine pile foundations. i Point cloud p ∈ P less than r j (j∈[1, …, n]), and form p i The neighborhood point set N(p) i ),Right now: ； Based on the neighborhood point set N(p) i ) Calculate point cloud p i Covariance matrix Cov i : ； In the formula, |·| represents the number of point clouds in the set; It is p i The centroid of the neighborhood point set; For the covariance matrix Cov i Perform eigenvalue decomposition to obtain three eigenvalues ​​λ1, λ2, and λ3; the eigenvector n corresponding to the smallest eigenvalue. i =(n x , n y , n z ) T That is, point cloud p i The normal vector; Calculate the point cloud p based on the above three eigenvalues. i curvature C i : ； After obtaining the curvature of all point clouds, remove point clouds with curvature greater than a set threshold to obtain the filtered point cloud set P. f ; After retaining the filtered data, point cloud P f All normal vectors (n) x , n y , n z Point clouds whose components in the x, y, and z directions are greater than the corresponding component thresholds Tx, Ty, and Tz are used to obtain the effective scour pit point cloud set P. i ; Depth detection is performed along the z-coordinate direction of the point cloud to remove the effective scour pit point set P. i For point clouds exceeding a depth threshold, only point clouds within a certain threshold are retained to form the bottom point cloud set P. b ; The underlying point cloud P is defined according to the following formula. b Clustering the scour pit point cloud in the data yields multiple individual bottom-level point clouds: ∀ I, j∈[1, …, s], i≠j, p i ∈ P b p j ∈ P b Find p i Neighborhood set N'(p) i ): ； In the formula, ε is the distance threshold between adjacent points; Find the underlying point cloud P b All point clouds have neighborhood sets, and the point clouds with more than a preset minimum number of points in their neighborhood sets are defined as core point clouds. Assume there are γ core point clouds in total, and the set formed is P. γ Then the core point cloud satisfies: ； In the formula, sum(·) represents the point cloud count; core point p o As the initial point of cluster A, then in set P γ p o neighborhood N ∈ (p o And in this neighborhood, find any core point p that is not in cluster A. o Then add it to cluster A, i.e., A←A ∪ p o Finally, perform the same operation until no more core points can be added, completing the final clustering of cluster A; after completing the clustering of cluster A, from set P... γ Then, select non-cluster A core points and perform clustering operations according to cluster A to form clusters B, C, ..., thus forming a bottom-level point cloud dataset P containing several clusters. c ; The number of layers n for the stacking expansion of the point cloud in each cluster is determined based on the z-coordinate depth value of the point cloud in each cluster. ； In the formula, △z is the depth increment of each point cloud layer; This represents the average depth of the entire point cloud. Suppose the point cloud in cluster A is p l =(x l , y l , z l For all point clouds in cluster A, create a new point cloud with n layers; for each layer, increase the z-coordinate depth value by k ⋅ Δz to obtain a k-layer point cloud dataset A. k : ； In the formula, k is the current layer number, k=1,2,…,n; Merging each layer of point cloud data ultimately forms the target point cloud dataset A. f : ； After completing the stacked expansion of cluster A, the expansion steps are repeated for all clusters to finally obtain the initial scour pit point cloud dataset P after morphological restoration. r ; The initial scour pit point cloud dataset P after morphological restoration r The point cloud of each cluster is proportionally corrected to obtain the corrected set of scour pit points P. f The proportional correction parameter s is obtained by the following formula: ； In the formula, z b This represents the actual depth of the seabed pile foundation.

7. The method for detecting scour of wind turbine pile foundations based on side-scan sonar according to claim 1, characterized in that: The aforementioned volume calculation of each individual scour pit point cloud data to achieve scour detection specifically includes: Let P be the corrected set of scour pits. f The generated tetrahedral set is T = (T1, T2, …, T i ), each tetrahedron T i It consists of 4 points, namely T i =(p i1 , p i2 , p i3 , p i4 Let the coordinates of the four vertices of the tetrahedron be p, ... i1 =(x i1 , y i1 ,z i1 ), p i2 =(x i2 , y i2 , z i2 ), p i3 =(x i3 , y i3 , z i3 ), p i4 =(x i4 , y i4 , z i4 Find the circumcenter c of the four vertices: ； According to the circumcenter c=(x ic , y ic , z ic Calculate the circumsphere radius R of each tetrahedron: ； Retain the tetrahedron that satisfies circumsphere radius R ≤ α; Introducing weighting factor w i The volume contribution of each retained tetrahedron is corrected, and the total scour pit volume is finally obtained by accumulating the volumes of all retained tetrahedrons: ； In the formula, w i L is the shortest side of tetrahedron i i min With the longest side L i max The ratio of .

Images