Method and system for detecting fan blade defects based on computer vision

Abstract

The invention discloses a method and a system for detecting fan blade defects based on computer vision. In computer image visual inspection of the fan blade, the shooting visual angle of a camera and the fan blade are of a three-dimensional structure with curvature, so that the shot image of the fan blade is distorted and deformed, and the defect identification precision is influenced. Therefore, by adopting a Mosaic data enhancement technology, the shot image of the fan blade is corrected according to the curvature of the fan blade, so that the analysis precision of subsequent blade defects is improved. In addition, an attention module CBAM is introduced, so that the detection efficiency of the visual image of the blade is improved; by improving the three-dimensional defect reconstruction method, the three-dimensional defect reconstruction precision of the fan blade is further improved; by subsequently introducing laser detection equipment and ultrasonic detection equipment, the accuracy of blade defect detection is further improved by adopting a fusion judgment method.

Description

Technical Field

[0001] The present invention belongs to the technical field of fan blade detection, and in particular relates to a method and system for fan blade defect detection based on computer vision. Background Art

[0002] Wind turbine blades (hereinafter referred to as wind turbines) are exposed to complex outdoor environments for long periods of time, facing erosion from multiple factors such as strong winds, rain, and ultraviolet rays. They are highly susceptible to defects such as wear, cracks, and corrosion, which can seriously impact wind turbine operation. Using drones to photograph wind turbine blades and then analyze the captured images for defects can detect these subtle issues early and allow for timely repairs, thus preventing blade defects from expanding under continuous mechanical stress and harsh environments. This in turn prevents sudden blade breakage, ensures stable wind turbine operation, reduces downtime caused by failures, and ensures the continuity and stability of wind power generation.

[0003] Currently, images of wind turbine blades are typically captured using drone-mounted cameras, followed by defect analysis models to identify the type of defects. However, the camera's viewing angle and the curvature of the blade itself can cause spatial distortion and deformation in the image. This can affect the type and size of defects in the blades during subsequent defect analysis. Existing technologies have not addressed this issue, and the accuracy of defect identification for wind turbine blades needs to be further improved. Summary of the Invention

[0004] In order to solve the above technical problems, the main purpose of the present invention is to provide a method and system for wind blade defect detection based on computer vision. By performing image correction on the photographed wind blades and using an improved defect recognition algorithm to perform defect analysis, the risk assessment accuracy of the wind blades can be improved.

[0005] In order to achieve the purpose of the present invention, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a method for detecting defects in wind turbine blades based on computer vision, comprising the following steps:

[0007] Step 1: Construct a set of wind turbine blade images: Use a camera to capture images of wind turbine blades, and then use Mosaic data enhancement technology to perform perspective transformation enhancement based on the surface curvature characteristics of the wind turbine blades.

[0008] Step 2: Build a defect detection model: Use the YOLOv11 architecture for lightweight deployment, design a dual-branch feature fusion network, and introduce the attention module (CBAM). Use a low-resolution camera to capture and inspect normal defect-free areas of the wind turbine blades. For suspected defect areas identified by the defect detection model with a confidence level greater than a preset value, switch to high-resolution camera shooting mode, capture high-resolution images of the local area, and then conduct further identification.

[0009] Step 3: 3D quantitative analysis of wind turbine blade defects: Combined with binocular vision point cloud data, the detected 2D defects are reconstructed in 3D to obtain quantitative parameters such as crack depth and damage area.

[0010] Furthermore, in step 1, the leaf image is processed using the Mosaic data enhancement technology, including the following steps:

[0011] Step 1.1, define the perspective transformation matrix:

[0012] where a 31 is the deformation compensation parameter, a 11 is the horizontal scaling parameter, a 22 is the vertical scaling parameter.

[0013] Step 1.2, dynamic adjustment of perspective transformation parameters: adjust the parameters of the perspective matrix M according to the curvature radius R of the blade surface to correct the image stretching caused by curvature and viewing angle.

[0014] Where f is the focal length of the camera, R is the local radius of curvature, and θ is the normal inclination angle of the blade surface.

[0015] Step 1.3: Calculate pixel fusion weights. Gaussian attenuation function is used to assign weights w to overlapping areas. i , suppressing mutations at splicing edges:

[0016] Where (x c ,y c ) is the coordinate of the center point of the subgraph in its local coordinate system, (x, y) is the coordinate in the global splicing coordinate system, and σ represents the standard deviation.

[0017] Step 1.4: perform stitching and fusion of the four images. The pixel fusion formula is: Among them I mosaic is the pixel value of the four sub-images through weighted fusion, w i is the pixel fusion weight; I irepresents the original pixel value of the i-th sub-image; T i (x,y) is the perspective transformation function that maps the global coordinates to the local coordinates of the i-th sub-image.

[0018] Furthermore, in step 2, the process of using the dual-branch feature fusion network to calculate the attention module CBAM is as follows: the input feature map F is divided into branch 1 and branch 2 after being output by the backbone network, and branch 1 performs global average pooling AvgPool on the feature map F to obtain F avg , and then perform nonlinear transformation through the multi-layer perceptron MLP to generate the channel weight vector W avg , W avg Perform channel splicing with the crack direction feature map OrientationMap to obtain the fusion feature W avg+dir , the OrientaionMap is the gradient direction histogram encoding extracted by HOG; branch 2 performs global maximum pooling MaxPool on the feature map F to obtain F max , and then generate the channel weight vector W through another multi-layer perceptron MLP processing max , W max Directly add it to OrientationMap to get the fusion feature W max+dir ;W of branch 1 avg+dir and W of branch 2 max+dir Splicing in the channel dimension and generating the channel attention weight map M through the Sigmoid activation function c (F).

[0019] Furthermore, in step 3, the three-dimensional reconstruction of the two-dimensional defect includes the following steps:

[0020] Step 3.1, pre-registration: Using the original model point set Q as the reference, perform pre-registration initialization transformation on the damage model point set P. The original model point set Q and the pre-registration damage model point set P0 can be expressed as P0 = {P i0 |P i0 ∈R 3 ,i=1,2···,n},Q0={q j0 |q j0 ∈R 3 ,j=1,2…,m};

[0021] Step 3.2, find the corresponding point: for any point p in the damage model point set P ik , find p ik To the point closest to the original model point set Q, record the point in the original model point set Q that is closest to p ik The closest point is q ik , forming the corresponding point set Q k ={q ik|q ik ∈R 3 ,i=1,2…,n}, the distance calculation formula is d ik =||p ik -q ik ||→min, k is the number of iterations.

[0022] Step 3.3: Dynamically screen the damaged area points and remove the corresponding point pairs that do not meet the following conditions (p i ,q i ):

[0023] Distance threshold screening: Check whether the distance between corresponding points is less than the set threshold d e , remove d ik >d e The point pair;

[0024] Curvature consistency screening: Calculate the local curvature C(p ik ) and C(q ik ), eliminating the curvature difference | C(p ik )-C(q ik )|>0.1 point pairs;

[0025] Normal vector angle screening: Calculate the normal vector angle θ of the corresponding point pair ik , remove θ ik Point pairs >15°;

[0026] Step 3.4, solve the transformation matrix: Based on the screened corresponding point set, minimize the curvature constraint energy function and solve the rotation transformation matrix R k and the translation transformation matrix T k , the curvature constraint energy function is expressed as E curv =∑||C(p i )-C(q i )|| 2 +λ∑||R·p i +Tq i || 2 →min, λ is the curvature constraint weight, and the rotation transformation matrix R k and the translation transformation matrix T k is the curvature constraint energy function E curv Minimize the rotation transformation matrix and translation transformation matrix after the kth iteration.

[0027] Step 3.5, update the relative position between models: Use the transformation matrix obtained in step (4) to rotate and translate the damage model point set P to obtain the new position P of the damage model k+1 =R k P k +Tk .

[0028] Step 3.6, iterative termination judgment: average distance between corresponding points Less than the given value ε, or reaching the maximum number of iterations N max , the iteration terminates; ε>0.

[0029] Furthermore, in step 3.3, for each point in the point cloud, the eigenvalues of the covariance matrix of the 50mm neighborhood of the point are calculated, and three eigenvalues λ1≥λ2≥λ3 are obtained. Then the local curvature of the point is expressed as

[0030] Furthermore, after completing the image recognition defects in step 3, step 4 can be executed for fusion verification, and step 4 includes: using laser detection equipment to re-verify the defect area identified by the image, and using ultrasonic detection equipment to cross-verify the hidden damage inside the defect area identified by the image.

[0031] Furthermore, in step 4, a decision fusion method is used to control the intervention timing of the laser detection equipment and the ultrasonic detection equipment, and whether to perform blade repair is determined based on the joint detection of image recognition, laser detection equipment and ultrasonic detection equipment.

[0032] Furthermore, the decision fusion method in step 4 includes the following steps:

[0033] Step 4.1: Assign basic probabilities. Set the confidence level for crack detection by the laser inspection equipment to m1, and the confidence level for internal damage detection by the ultrasonic inspection equipment to m2. The specific values of m1 and m2 are determined through calibration experiments or expert experience.

[0034] Step 4.2, formulate a conflict coefficient K to measure the degree of contradiction between the confidence of the laser detection equipment and the ultrasonic detection equipment, the conflict coefficient K = ∑ A∩B＝φ m1(A)m2(B), m1(A) and m2(B) represent the confidence of laser detection equipment and ultrasonic detection equipment respectively, ∑ A∩B＝φ m1(A)m2(B) represents all mutually exclusive hypothetical combinations of whether the laser detection equipment and the ultrasonic detection equipment can detect the defects of the wind turbine blades, and the product of the confidence scores of the two equipment is accumulated;

[0035] Step 4.3: Generate fusion confidence using Dempster synthesis rule m 12 (X) represents the joint confidence level of the fan blade defects identified by the fused laser detection equipment and ultrasonic detection equipment simultaneously; ∑ Y∩Z＝Xm1(Y)m2(Z) represents the sum of the product of the confidence level m1(Y) of the fan blade defects that can be detected by the laser detection equipment and the confidence level m2(Z) of the fan blade defects that can be detected by the ultrasonic detection equipment;

[0036] Step 4.4, decision logic setting: determine the threshold θ through ROC curve analysis or historical false positive and missed positive costs e , when m 12 (X)>θ e , it is determined to be a valid crack and the blade repair task is triggered.

[0037] In a second aspect, the present invention also provides a system for detecting defects in wind blades based on computer vision, which is characterized in that it includes a shooting device, a processor, a memory, and a computer program stored in the memory, wherein the shooting device is used to shoot images of wind blades, and the computer program can be executed by the processor to implement the above-mentioned method for detecting defects in wind blades based on visual images.

[0038] Furthermore, the shooting equipment also includes a ring fill light and a polarizing filter, which are used to achieve shadowless lighting on the blade surface and eliminate reflection interference.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The present invention uses Mosaic data enhancement technology to fuse the collected images, especially to adaptively adjust the captured blade images according to the structural characteristics and curvature of the wind turbine blades. The image after perspective transformation can be spatially corrected, eliminating the influence of the shooting angle and the spatial curvature structure of the wind turbine blades on the image, and improving the accuracy of subsequent defect location and size identification.

[0041] 2. The present invention uses the CBAM attention mechanism to control the image acquisition system. According to the control strategy, a low-resolution camera is used to shoot and detect in normal defect-free areas. For identified suspected defect areas, the camera switches to a high-resolution shooting mode to increase the pixels of the captured image. After high-precision shooting of the local area, further identification of the defect size and location is performed. The introduction of the CBAM attention mechanism, especially the introduction of HOG directional features in channel attention, customizes and improves the sensitivity of crack detection. On the one hand, it allows the low-resolution camera to focus on finding blade defects based on the set defect features, speeding up the defect identification speed of the entire blade. On the other hand, using a low-resolution camera for shooting first also saves device storage and computing resources. When performing defect analysis and calculation of the blade image, it can pay more attention to the key feature information generated by the defect, thereby improving computing efficiency.

[0042] 3. Traditionally, damage boundaries have been determined by calculating the curvature of each point in point cloud data and constructing the damage boundary using abrupt changes in curvature. Compared to traditional methods, this invention, taking into account the large curvature of wind turbine blades, further improves the accuracy of 3D defect reconstruction of wind turbine blades by improving curvature constraints and damage point screening mechanisms. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The disclosure of the present invention will become more easily understood with reference to the accompanying drawings. Those skilled in the art will readily appreciate that these drawings are merely used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0044] Figure 1 The figure is a flow chart of fan blade defect identification according to an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the calculation flow of the CBAM module channel attention using dual-branch feature fusion according to an embodiment of the present invention.

[0046] Figure 3 Schematic diagram of a method for three-dimensional reconstruction of a two-dimensional defect according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of this application.

[0048] The separating agent and the method of use provided by the present invention are described in detail below with reference to the embodiments, but they should not be construed as limiting the scope of protection of the present invention.

[0049] This embodiment provides a method for detecting defects in wind turbine blades based on computer vision. This method can correct the effects of blade spatial curvature and camera angle on the size and location of defects on wind turbine blades, and improve the accuracy of three-dimensional reconstruction of defects.

[0050] Reference Figure 1 , is a schematic diagram of the flow chart of fan blade defect identification according to this embodiment, which includes the following steps:

[0051] Step 1: Construct a set of wind turbine blade images: Use a camera to capture images of wind turbine blades, and then use Mosaic data enhancement technology to perform perspective transformation enhancement based on the surface curvature characteristics of the wind turbine blades.

[0052] In step 1 of this embodiment, the blade image is processed using Mosaic data enhancement technology based on the large curvature of the wind turbine blade, including the following steps:

[0053] Step 1.1, define the perspective transformation matrix:

[0054] where a 31 is the deformation compensation parameter, a 11 is the horizontal scaling parameter, a 22 is the vertical scaling parameter.

[0055] Step 1.2: Dynamically adjust the parameters of the perspective matrix M according to the blade surface curvature radius R to correct the image stretching caused by curvature and viewing angle. The specific adjustment parameters are:

[0056] Where f is the focal length of the camera, R is the local curvature radius of the wind turbine blade, and θ is the normal inclination angle of the blade surface. 11 and a 22 The image stretching caused by curvature and viewing angle is corrected by the proportional relationship between the camera focal length and the curvature radius of the wind turbine blade. 31 Dynamic adjustments are made to the curvature characteristics of different parts of the fan blades to achieve the purpose of deformation compensation.

[0057] Step 1.3: Calculate pixel fusion weights. Gaussian attenuation function is used to assign weights w to overlapping areas. i , suppressing mutations at splicing edges:

[0058] Where (x c ,y c ) are the coordinates of the center point of the subgraph in its local coordinate system; (x, y) are the coordinates in the global stitching coordinate system, which have been normalized to the range [0, 1]; σ represents the standard deviation.

[0059] Step 1.4: perform stitching and fusion of the four images. The pixel fusion formula is:

[0060] Among them I mosaic is the pixel value of the four sub-images through weighted fusion, w i is the pixel fusion weight; I i represents the original pixel value of the i-th sub-image; T i (x,y) is the perspective transformation function that maps the global coordinates to the local coordinates of the i-th sub-image.

[0061] Step 2: Build a defect detection model: Use the YOLOv11 architecture for lightweight deployment, design a dual-branch feature fusion network, and introduce the attention module (CBAM). Use a low-resolution 640×640 camera to shoot and inspect wind blades in normal defect-free areas. For suspected defect areas with a confidence level greater than 0.7 identified by the defect detection model, switch to the high-resolution 1280×1280 camera shooting mode, perform high-resolution shooting of the local area, and then perform further identification.

[0062] Figure 2 The calculation process of the CBAM module channel attention using dual-branch feature fusion is shown. Specifically in this embodiment, the process of using the dual-branch feature fusion network to calculate the attention module CBAM is as follows: the input feature map F is divided into branch 1 and branch 2 after being output by the backbone network. Branch 1 performs global average pooling AvgPool on the feature map F to obtain F avg , and then perform nonlinear transformation through the multi-layer perceptron MLP to generate the channel weight vector W avg , W avg Perform channel splicing with the crack direction feature map OrientationMap to obtain the fusion feature W avg+dir , the OrientaionMap is the gradient direction histogram encoding extracted by HOG; branch 2 performs global maximum pooling MaxPool on the feature map F to obtain F max , and then generate the channel weight vector W through another multi-layer perceptron MLP processing max , W max Directly add it to OrientationMap to get the fusion feature W max+dir ;W of branch 1 avg+dir and W of branch 2 max+dir Splicing in the channel dimension and generating the channel attention weight map M through the Sigmoid activation function c (F).

[0063] A dual-branch feature fusion network performs attention module (CBAM) calculations. Branch 1 performs average pooling to characterize the global average response of the channel, while branch 2 performs maximum pooling to enhance the local significant response of the channel. This, combined with crack direction features, improves the detection capability of discrete defects (such as lightning strike points). Specifically, the CBAM attention mechanism extracts crack direction features from the directional gradient histogram (HOG), allowing the camera's visual image detection to focus more on identifying the wind turbine crack itself, helping to improve crack detection sensitivity and efficiency.

[0064] Step 3: 3D quantitative analysis of wind turbine blade defects: Combined with binocular vision point cloud data, the detected 2D defects are reconstructed in 3D to obtain quantitative parameters such as crack depth and damage area.

[0065] Reference Figure 3 The method for performing three-dimensional reconstruction of a two-dimensional defect according to this embodiment is schematically shown, and specifically includes the following steps:

[0066] Step 3.1, pre-registration: Using the original model point set Q as the reference, perform pre-registration initialization transformation on the damage model point set P. The original model point set Q and the pre-registration damage model point set P0 can be expressed as P0 = {P i0 |P i0 ∈R 3 ,i=1,2···,n},Q0={q j0 |q j0 ∈R 3 ,j=1,2…,m};

[0067] Step 3.2, find the corresponding point: for any point p in the damage model point set P ik , find p ik To the point closest to the original model point set Q, record the point in the original model point set Q that is closest to p ik The closest point is q ik , forming the corresponding point set Q k =q ik |q ik ∈R 3 ,i=1,2…,n}, the distance calculation formula is d ik =||p ik -q ik ||→min, k is the number of iterations.

[0068] Step 3.3: Dynamically screen the damaged area points and remove the corresponding point pairs that do not meet the following conditions (p i ,q i ):

[0069] ① Distance threshold screening: Check whether the distance between corresponding points is less than the set threshold d e , remove d ik >d e The point is right.

[0070] ② Curvature consistency screening: Calculate the local curvature C(p ik ) and C(q ik ), eliminating the curvature difference | C(p ik )-C(q ik )|>0.1; for each point in the point cloud, calculate the eigenvalue of the covariance matrix of the 50mm neighborhood of the point, and get three eigenvalues λ1≥λ2≥λ3, then the local curvature of the point is defined as

[0071] ③ Normal vector angle screening: Calculate the normal vector angle θ of the corresponding point pair ik , remove θ ik >15° point pairs.

[0072] Step 3.4, solve the transformation matrix: Based on the screened corresponding point set, minimize the curvature constraint energy function and solve the rotation transformation matrix R k and the translation transformation matrix T k , the curvature constraint energy function is expressed as E curv =∑||C(p i )-C(q i )|| 2 +λ∑||R·p i +Tq i || 2 →min, the curvature constraint weight λ is preferably 0.3, and the rotation transformation matrix R k and the translation transformation matrix T k is the curvature constraint energy function E curv Minimize the rotation transformation matrix and translation transformation matrix after the kth iteration.

[0073] Step 3.5, update the relative position between models: Use the transformation matrix obtained in step (4) to rotate and translate the damage model point set P to obtain the new position P of the damage model k+1 =P k P k +T k .

[0074] Step 3.6, iterative termination judgment: average distance between corresponding points Less than the given value ε, or reaching the maximum number of iterations N max , the iteration terminates; ε>0.

[0075] Compared with the traditional 3D reconstruction method, the 3D reconstruction method adopted in this embodiment has the following improvements:

[0076] (1) Define the local curvature as It can effectively characterize the local concave-convex characteristics of the surface, making the three-dimensional reconstruction method more suitable for the characteristics of large curvature of the blade surface.

[0077] (2)E curv is the total target value of the curvature constraint energy function, and point cloud alignment and curvature consistency are achieved by minimizing this value. (∑||C(p i )-C(q i )|| 2 ) represents curvature constraint consistency, ||R·p i +Tq i|| 2 Represents geometric constraints. In the process of point cloud registration, in addition to the geometric distance constraint, the curvature constraint consistency (∑||C(p i )-C(q i )|| 2 ), using curvature differences as part of the energy function to suppress mismatches caused by surface deformation or noise. At the same time, a balance parameter λ is set in the curvature constraint energy function to balance the two constraints, achieving balanced constraints and smooth matching.

[0078] (3) Aiming at the complex curved surface of wind turbine blades, multiple damage point screening mechanisms are established to eliminate damage points that do not meet the conditions, thereby improving the matching accuracy of damage points and thus improving the overall accuracy of three-dimensional reconstruction.

[0079] In step 4, to further improve the accuracy of blade defect detection, this embodiment also uses laser and ultrasonic testing equipment to cross-validate the image detection results. A decision fusion method is used to control the timing of laser and ultrasonic intervention. Based on the combined detection of image recognition, laser, and ultrasonic testing, a decision is made whether to perform blade repair.

[0080] Specifically, the decision fusion method in step 4 includes the following steps:

[0081] Step 4.1: Assign basic probabilities. Set the confidence level for crack detection by the laser inspection equipment to m1, and the confidence level for internal damage detection by the ultrasonic inspection equipment to m2. The specific values of m1 and m2 are determined through calibration experiments or expert experience.

[0082] Step 4.2, formulate a conflict coefficient K to measure the degree of contradiction between the confidence of the laser detection equipment and the ultrasonic detection equipment, the conflict coefficient K = ∑ A∩B＝φ m1(A)m2(B), m1(A) and m2(B) represent the confidence of laser detection equipment and ultrasonic detection equipment respectively, ∑ A∩B＝φ m1(A)m2(B) represents all mutually exclusive hypothetical combinations of whether the laser detection equipment and the ultrasonic detection equipment can detect defects in the wind turbine blades, and the product of the confidence scores of the two devices is accumulated.

[0083] Step 4.3: Generate fusion confidence using Dempster synthesis rule m 12 (X) represents the joint confidence level of the fan blade defects identified by the fused laser detection equipment and ultrasonic detection equipment simultaneously; ∑ Y∩Z＝Xm1(Y)m2(Z) represents the sum of the products of the confidence level m1(Y) that the laser detection equipment can detect defects in the fan blades and the confidence level m2(Z) that the ultrasonic detection equipment can detect defects in the fan blades. 1-K is used for normalization to eliminate conflicting influences.

[0084] Step 4.4, decision logic setting: determine the threshold θ through ROC curve analysis or historical false positive and missed positive costs e , when m 12 (X)>θ e , it is determined to be a valid crack and the blade repair task is triggered.

[0085] By integrating the comprehensive judgment of image detection system, laser detection equipment and ultrasonic detection equipment, the accuracy of wind turbine blade defect detection can be further improved.

[0086] In addition, this embodiment also provides a system for detecting defects in wind blades based on computer vision, which is characterized by including a shooting device, a processor, a memory, and a computer program stored in the memory. The shooting device is used to capture images of wind blades, and the computer program can be executed by the processor to implement the above-mentioned method for detecting defects in wind blades based on visual images.

[0087] Furthermore, the shooting equipment also includes a ring fill light and a polarizing filter, which are used to achieve shadowless lighting on the blade surface and eliminate reflection interference.

[0088] The above is a further detailed description of the present invention in conjunction with specific / preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. Those skilled in the art of the art to which the present invention belongs may make several substitutions or modifications to the described embodiments without departing from the scope of the present invention, and such substitutions or modifications should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for detecting defects in fan blades based on computer vision, characterized in that: The following steps are involved: Step 1: Construct a set of wind turbine blade images: Use a camera to capture images of wind turbine blades, and then use Mosaic data enhancement technology to perform perspective transformation enhancement based on the surface curvature characteristics of the wind turbine blades. Step 2: Build a defect detection model: Use the YOLOv11 architecture for lightweight deployment, design a dual-branch feature fusion network, and introduce the attention module (CBAM). Use a low-resolution camera to capture and inspect normal defect-free areas of the wind turbine blades. For suspected defect areas identified by the defect detection model with a confidence level greater than a preset value, switch to high-resolution camera shooting mode, capture high-resolution images of the local area, and then conduct further identification. Step 3: 3D quantitative analysis of wind turbine blade defects: Combined with binocular vision point cloud data, the detected 2D defects are reconstructed in 3D to obtain quantitative parameters such as crack depth and damage area.

2. The method for detecting fan blade defects based on computer vision according to claim 1, characterized in that: In step 1, the leaf image is processed using the Mosaic data enhancement technology, including the following steps: Step 1.1, define the perspective transformation matrix: where a 31 is the deformation compensation parameter, a 11 is the horizontal scaling parameter, a 22 is the vertical scaling parameter; Step 1.2, dynamic adjustment of perspective transformation parameters: adjust the parameters of the perspective matrix M according to the curvature radius R of the blade surface to correct the image stretching caused by curvature and viewing angle. Where f is the focal length of the camera, R is the local radius of curvature, and θ is the normal inclination angle of the blade surface; Step 1.3: Calculate pixel fusion weights. Gaussian attenuation function is used to assign weights to overlapping areas to suppress sudden changes in the splicing edges. Step 1.4: stitch and fuse the four images.

3. The method for detecting fan blade defects based on computer vision according to claim 1, characterized in that: In step 2, the process of using the dual-branch feature fusion network to calculate the attention module CBAM is as follows: the input feature map F is divided into branch 1 and branch 2 after being output by the backbone network, and branch 1 performs global average pooling AvgPool on the feature map F to obtain F avg , and then perform nonlinear transformation through the multi-layer perceptron MLP to generate the channel weight vector W avg , W avg Perform channel splicing with the crack direction feature map OrientationMap to obtain the fusion feature W avg+dir , the OrientaionMap is a gradient direction histogram encoding extracted by HOG; Branch 2 performs global maximum pooling MaxPool on the feature map F to obtain F max , and then generate the channel weight vector W through another multi-layer perceptron MLP processing max , W max Directly add it to OrientationMap to get the fusion feature W max+dir ;W of branch 1 avg+dir and W of branch 2 max+dir Splicing in the channel dimension and generating the channel attention weight map M through the Sigmoid activation function c (F).

4. The method for detecting fan blade defects based on computer vision according to claim 1, characterized in that: In step 3, the three-dimensional reconstruction of the two-dimensional defect includes the following steps: Step 3.1, pre-registration: Using the original model point set Q as the reference, perform pre-registration initialization transformation on the damage model point set P. The original model point set Q and the pre-registration damage model point set P0 can be expressed as P0 = {P i0 |P i0 ∈R 3 ,i=1,2…,n},Q0={q j0 |q j0 ∈R 3 ,j=1,2…,m}; Step 3.2, find the corresponding point: for any point p in the damage model point set P ik , find p ik To the point closest to the original model point set Q, record the point in the original model point set Q that is closest to p ik The closest point is q ik , forming the corresponding point set Q k ={q ik |q ik ∈R 3 ,i=1,2…,n}; Step 3.3: Dynamically screen the damaged area points and remove the corresponding point pairs that do not meet the following conditions (p i ,q i ): Distance threshold screening: Check whether the distance between corresponding points is less than the set threshold d e , remove d ik >d e The point pair; Curvature consistency screening: Calculate the local curvature C(p ik ) and C(q ik ), eliminating the curvature difference | C(p ik )-C(q ik )|>0.1 point pairs; Normal vector angle screening: Calculate the normal vector angle θ of the corresponding point pair ik , remove θ ik Point pairs >15°; Step 3.4, solve the transformation matrix: Based on the screened corresponding point set, minimize the curvature constraint energy function and solve the rotation transformation matrix R k and the translation transformation matrix T k , the curvature constraint energy function is expressed as E curv =∑||C(p i )-C(q i )|| 2 +λ∑||R·p i +Tq i || 2 →min, λ is the curvature constraint weight, and the rotation transformation matrix R k and the translation transformation matrix T k is the curvature constraint energy function F curv Minimize the rotation transformation matrix and translation transformation matrix after the kth iteration; Step 3.5, update the relative position between models: Use the transformation matrix obtained in step (4) to rotate and translate the damage model point set P to obtain the new position P of the damage model k+1 =R k P k +T k ; Step 3.6, iterative termination judgment: average distance between corresponding points Less than the given value ε, or reaching the maximum number of iterations N max , the iteration terminates; 5. The method for detecting fan blade defects based on computer vision according to claim 4, characterized in that: In step 3.3, for each point in the point cloud, the eigenvalues of the covariance matrix of the 50mm neighborhood of the point are calculated, and three eigenvalues λ1≥λ2≥λ3 are obtained. Then the local curvature of the point is expressed as 6. The method for detecting defects in wind turbine blades based on computer vision according to any one of claims 1 to 5, characterized in that: After completing the image recognition defects in step 3, step 4 can also be executed for fusion verification. Step 4 includes: using laser detection equipment to re-verify the defect area identified by the image, and using ultrasonic detection equipment to cross-verify the hidden damage inside the defect area identified by the image.

7. The method for detecting fan blade defects based on computer vision according to any one of claim 6, characterized in that: In step 4, a decision fusion method is used to control the intervention timing of the laser detection equipment and the ultrasonic detection equipment, and whether to perform blade repair is determined based on the joint detection of image recognition, laser detection equipment and ultrasonic detection equipment.

8. The method for detecting fan blade defects based on computer vision according to claim 7, characterized in that: The decision fusion method described in step 4 includes the following steps: Step 4.1: Assign basic probabilities. Set the confidence level for crack detection by the laser inspection equipment to m1, and the confidence level for internal damage detection by the ultrasonic inspection equipment to m2. The specific values of m1 and m2 are determined through calibration experiments or expert experience. Step 4.2, formulate a conflict coefficient K to measure the degree of contradiction between the confidence of the laser detection equipment and the ultrasonic detection equipment, the conflict coefficient K = ∑ A∩B＝φ m1(A)m2(B), m1(A) and m2(B) represent the confidence of laser detection equipment and ultrasonic detection equipment respectively, ∑ A∩B＝φ m1(A)m2(B) represents all mutually exclusive hypothetical combinations of whether the laser detection equipment and the ultrasonic detection equipment can detect the defects of the wind turbine blades, and the product of the confidence scores of the two equipment is accumulated; Step 4.3: Use Dempster synthesis rule to generate fusion confidence m 12 (X); Step 4.4, decision logic setting: determine the threshold θ through ROC curve analysis or historical false positive and missed positive costs e , when m 12 (X)>θ e , it is determined to be a valid crack and the blade repair task is triggered.

9. A system for detecting defects in fan blades based on computer vision, characterized in that: The method comprises a photographing device, a processor, a memory and a computer program stored in the memory, wherein the photographing device is used to photograph images of wind turbine blades, and the computer program can be executed by the processor to implement a method for detecting defects in wind turbine blades based on visual images as described in any one of claims 1 to 8.

10. The system for detecting defects in wind turbine blades based on computer vision according to claim 9, characterized in that: The shooting equipment also includes a ring fill light and a polarizing filter.

Images