A visual detection, diagnosis and early warning method for wind power tower cylinder curved surface defects

Abstract

The application discloses a kind of visual detection and diagnosis and early warning method for wind power tower cylinder curved surface defect, comprising the following steps: collecting data and obtaining tower cylinder curved surface detection image and tower cylinder curved surface coordinate data after pretreatment;Curved surface unfolding correction is carried out to tower cylinder curved surface detection image, and tower cylinder curved surface unfolding image and unfolding coordinate corresponding relationship are obtained;Curved surface interference area is identified and processed to obtain tower cylinder curved surface standard detection image;Defect candidate region is extracted to construct defect candidate region set;Defect morphology evolution input data is constructed;According to the curved surface constraint Wulff defect morphology evolution network, defect stable morphology contour data is obtained;Defect diagnosis feature data is constructed;Historical defect diagnosis feature data is called, and defect risk level, defect diagnosis result and defect early warning result are generated;Wind power tower cylinder curved surface defect detection and diagnosis and early warning report are generated.The application utilizes curved surface vision and Wulff morphology evolution technology, realizes tower cylinder defect accurate detection and reliable early warning.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence defect recognition technology, and in particular to a visual detection, diagnosis and early warning method for defects on the curved surface of wind turbine towers. Background Technology

[0002] Wind turbine towers operate outdoors for extended periods, making their surfaces susceptible to damage from wind, sand, rain, salt spray, temperature variations, and mechanical stress. This can lead to surface defects such as cracks, corrosion, coating peeling, dents, and weld abnormalities. Current inspection methods typically involve manual inspections, manual verification of drone images, or the use of standard image recognition models to detect defects in tower surface images to determine their location and type.

[0003] However, the surface of wind turbine towers is a large-sized cylindrical curved surface, and the acquired images are prone to surface stretching, local reflections, shadow gradients, weld textures, and coating color differences. Existing methods often directly judge defects based on image edges, colors, or segmentation results, making it difficult to perform stable morphological correction of candidate contours under surface constraints. This can easily lead to broken boundaries of real defects, misidentification of false defects, and incomplete defect contours, thus affecting the reliability of subsequent defect diagnosis and risk warning.

[0004] Therefore, how to provide a visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a visual detection, diagnosis, and early warning method for surface defects in wind turbine towers. This invention utilizes surface vision and Wulff morphology evolution technology to achieve accurate detection and reliable early warning of tower defects.

[0006] A visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to an embodiment of the present invention includes the following steps: Collect visual images of the curved surface area of ​​the wind turbine tower and auxiliary data for tower detection, and preprocess them to obtain the tower surface detection image and tower surface coordinate data. Based on the tower inspection auxiliary data and tower surface coordinate data, the tower surface inspection image is corrected by surface unfolding to obtain the correspondence between the unfolded tower surface image and the unfolded coordinates. Based on the unfolded image of the tower surface, the interference region of the curved surface is identified, and interference suppression processing is performed to obtain the standard detection image of the tower surface; Based on the standard inspection image of the tower surface, candidate defect regions are extracted, and a set of candidate defect regions is constructed. Based on the defect candidate region set, tower surface coordinate data, unfolded coordinate correspondence, and surface interference region, input data for defect morphology evolution is constructed. The defect morphology evolution input data is input into the surface-constrained Wulff defect morphology evolution network, and the morphology evolution correction under surface constraints is performed on the defect candidate region set to obtain stable defect morphology contour data. Based on the defect stability morphology contour data, the correspondence between unfolded coordinates and the tower surface coordinate data, defect diagnostic feature data are constructed. Based on defect diagnostic feature data, retrieve historical defect diagnostic feature data to generate defect risk level, defect diagnosis result and defect warning result; Generate reports on the detection, diagnosis, and early warning of defects on the curved surface of wind turbine towers.

[0007] Optionally, the preprocessing includes preprocessing the visual image of the tower surface to obtain a tower surface detection image, and establishing tower surface coordinate data based on tower detection auxiliary data.

[0008] Optionally, the generation of the correspondence between the unfolded image of the tower surface and the unfolded coordinates includes: Based on the shooting posture data and tower structure parameters in the tower detection auxiliary data, surface projection matching is performed on the image position in the tower surface detection image to obtain the surface projection relationship between the image position and the tower surface coordinate data. Based on the surface projection relationship, the image position in the tower surface detection image is converted into the corresponding height coordinates and circumferential coordinates, and arranged on the unfolding plane according to the height coordinates and circumferential coordinates to obtain the initial surface unfolding image and the initial unfolding coordinate relationship; Based on the initial unfolded coordinate relationship, the pixel values ​​between adjacent unfolded positions in the initial unfolded surface image are interpolated, and the unfolded position offset in the initial unfolded coordinate relationship is corrected to obtain the unfolded surface image of the tower and the corrected unfolded coordinate relationship. Based on the image surface coordinate set and the corrected unfolded coordinate relationship, the unfolded coordinate correspondence between the image position in the tower surface detection image and the unfolded position in the unfolded tower surface image is generated.

[0009] Optionally, the generation of the standard inspection image of the tower surface includes: Extract surface interference feature data from the unfolded image of the tower surface, identify the location of surface interference in the unfolded image of the tower surface based on the surface interference feature data, and construct the surface interference region. Based on the location distribution of the surface interference region in the unfolded image of the tower surface and the surface interference feature data, the interference suppression parameters corresponding to the surface interference region are generated. The tower surface unfolded image is subjected to interference suppression processing according to the interference suppression parameters to obtain the interference suppression image. The standard inspection image of the tower surface is then generated based on the interference suppression image.

[0010] Optionally, the generation of the defect candidate region set includes: Based on the grayscale changes, color changes, texture changes, and edge changes in the standard inspection image of the tower surface, defect candidate feature data are extracted; Based on the defect candidate feature data, the defect candidate locations in the standard inspection image of the tower surface are determined, and an initial defect candidate region is generated; The initial defect candidate regions are merged to obtain merged defect candidate regions, and the merged defect candidate regions are then filtered to obtain defect candidate regions. Based on the location of the defect candidate regions in the standard inspection image of the tower surface, a set of defect candidate regions is constructed.

[0011] Optionally, the construction of the defect morphology evolution input data includes: Based on the set of candidate defect regions, extract boundary points on the region boundary to form a boundary point sequence and generate candidate contour data; Based on the coordinate correspondence, the boundary point sequence in the candidate contour data is mapped to the tower surface coordinate data to obtain the boundary point surface coordinate sequence. The surface spacing, circumferential change and height change between adjacent boundary points are calculated and a surface contour constraint sequence is formed. The local bending direction and continuous direction of the candidate contour data on the wind turbine tower surface are determined based on the surface contour constraint sequence. The boundary point surface coordinate sequence, surface contour constraint sequence, local bending direction and continuous direction are then combined in the order of the boundary point sequence to obtain the surface constraint data. Based on the positional overlap between the surface interference region and the boundary point sequence in the candidate contour data, interference overlap boundary points are determined. Based on the proximity distance between the surface interference region and the boundary point sequence in the candidate contour data, interference neighboring boundary points are determined. Interference overlap boundary points, interference neighboring boundary points, and boundary point sequences are marked accordingly to obtain the boundary point interference marking sequence. Interference suppression weight data is generated according to the arrangement order of the boundary point sequence. The contour extension direction between adjacent boundary points is calculated based on the candidate contour data, the boundary response intensity is calculated based on the boundary point sequence, and the contour extension direction is corrected based on the surface constraint data to obtain the surface-corrected contour direction. The boundary response intensity is weighted according to the interference suppression weight data to obtain the suppressed boundary response intensity. The surface correction contour direction and the suppressed boundary response intensity are then combined in the order of the boundary point sequence to obtain the boundary direction response data. Candidate contour data, surface constraint data, interference suppression weight data, and boundary direction response data are combined by aligning points according to the same defect candidate region to construct input data for defect morphology evolution.

[0012] Optionally, the generation of the defect stable topography contour data includes: The defect morphology evolution input data is input into the surface-constrained Wulff defect morphology evolution network, which includes an input alignment layer, an interface energy construction layer, a surface constraint modulation layer, an interference weight coupling layer, a Wulff morphology evolution layer, and a stable contour output layer. In the input alignment layer, candidate contour data, surface constraint data, interference suppression weight data and boundary direction response data are read according to the same defect candidate region. The surface constraint data, interference suppression weight data and boundary direction response data are aligned according to the boundary point sequence in the candidate contour data to obtain the contour evolution feature sequence. In the interface energy construction layer, based on the boundary direction response data in the contour evolution feature sequence, the surface correction contour direction and the suppressed boundary response intensity corresponding to each boundary point are extracted, and the directional interface energy corresponding to each boundary point is calculated based on the surface correction contour direction and the suppressed boundary response intensity. The candidate contour interface energy sequence is generated according to the arrangement order of the boundary point sequence. In the surface constraint modulation layer, based on the surface constraint data in the contour evolution feature sequence, the surface contour constraint sequence, local bending direction and surface continuity direction are extracted. Then, the candidate contour interface energy sequence is subjected to surface direction constraint and surface continuity constraint based on the surface contour constraint sequence, local bending direction and surface continuity direction to obtain the surface constraint interface energy sequence. In the interference weight coupling layer, the interference suppression weight data in the contour evolution feature sequence is used to perform point-by-point weight coupling on the surface constraint interface energy sequence to obtain the interference coupling interface energy sequence. In the Wulff morphology evolution layer, the interface energy convergence direction corresponding to each boundary point in the boundary point sequence is determined according to the interference coupling interface energy sequence. Within the surface position range limited by the surface constraint data, the position of the boundary point sequence in the candidate contour data is updated according to the interface energy convergence direction to obtain the evolution contour data. In the stable profile output layer, the surface continuity of the evolving profile data is checked based on the surface constraint data, and the connection relationship of the evolving boundary points is determined according to the profile arrangement order in the evolving profile data. The boundary closure is checked based on the connection relationship of the evolving boundary points, and the position of the boundary points that fail the surface continuity check or boundary closure check is corrected to obtain the stable morphological profile data of the defect.

[0013] Optionally, the construction of the defect diagnostic feature data includes: Contour sampling is performed on the stable morphology contour data of the defect to obtain stable boundary points and determine the connection relationship of stable boundary points; Based on the stable boundary points and their connection relationships, the stable profile closure line is determined, and the defect profile area, defect profile length, defect profile closure degree, and defect profile continuity are calculated to obtain the basic data of the defect morphology. The connection direction between adjacent stable boundary points is determined based on the connection relationship between stable boundary points, and the distribution density of stable boundary points is calculated based on the distribution position of stable boundary points on the closed line of the stable profile. Based on the connection direction and the distribution density of stable boundary points, the main extension direction of the defect contour, the number of defect contour branches, and the abrupt change position of the defect contour boundary are extracted and combined to obtain the defect morphology features. Based on the coordinate correspondence, the stable boundary points are mapped to the tower surface coordinate data to obtain the stable boundary point surface coordinate sequence; Based on the stable boundary point surface coordinate sequence, the height distribution range, circumferential distribution range, and contour center surface coordinates of the defect on the wind turbine tower surface are determined, and the height distribution range, circumferential distribution range, and contour center surface coordinates constitute the defect location characteristics. The structural proximity features are constructed based on the stable boundary point surface coordinate sequence and the tower surface coordinate data. Defect morphology features, defect location features, and structural proximity features are combined according to the same stable morphology contour data of the defect to construct defect diagnostic feature data.

[0014] Optionally, the generation of the defect risk level, defect diagnosis result, and defect early warning result includes: Based on the defect location features in the defect diagnosis feature data, determine the location of the wind turbine tower surface corresponding to the defect, and retrieve historical defect diagnosis feature data according to the location of the wind turbine tower surface. Based on the defect diagnosis feature data and historical defect diagnosis feature data, the defect morphology features, defect location features and structural proximity features corresponding to the same wind turbine tower surface location are compared to obtain defect feature change data. Based on the defect feature change data, extract the defect contour area change, defect contour length change, defect contour main extension direction change, contour center surface coordinate change, and surface proximity distance change, and construct defect evolution judgment data. The defect type is determined based on the defect morphology features in the defect diagnosis feature data, the defect expansion state is determined based on the defect evolution judgment data, and the structural proximity state is determined based on the structural proximity features and defect evolution judgment data in the defect diagnosis feature data. The defect diagnosis result is composed of the defect type, defect expansion state, and structural proximity state. Based on defect evolution judgment data and defect diagnosis results, defects are classified into risk levels, defect risk grades are generated, and defect early warning results are generated.

[0015] Optionally, the generation of the wind turbine tower surface defect detection, diagnosis, and early warning report includes: Generate defect diagnosis records based on defect diagnosis results; Based on the defect risk level, the defect diagnosis record is marked with risk to obtain the defect risk record; Based on the defect warning results, the defect risk records are marked with warnings to obtain defect warning records; Based on the defect warning records, generate a wind turbine tower surface defect detection, diagnosis and warning report according to the location of the wind turbine tower surface.

[0016] The beneficial effects of this invention are: This application acquires visual images of the wind turbine tower surface and auxiliary tower inspection data by collecting images of the curved surface area of ​​the wind turbine tower. It then combines the shooting posture data and tower structural parameters to establish coordinate data of the curved surface of the tower. This allows subsequent inspection to move beyond ordinary planar image recognition and instead process the actual spatial position of the cylindrical curved surface of the wind turbine tower. By generating a corresponding relationship between the unfolded image of the tower surface and the unfolded coordinates through surface unfolding correction, the edge stretching, local deformation, and positional deviation caused by cylindrical surface imaging can be reduced. This enables the defect candidate area to establish a correspondence with the height and circumferential position on the curved surface of the tower, thereby improving the accuracy of defect location and the reliability of subsequent diagnostic data.

[0017] This application further identifies surface interference regions based on the image unfolding of the tower surface and performs interference suppression processing on the surface interference regions. It can weaken local reflections, shadow gradients, weld textures and coating color differences that are common in outdoor inspection of wind turbine towers, and reduce the situation where the above interferences are misidentified as cracks, corrosion or coating peeling. By generating candidate contour data, surface constraint data, interference suppression weight data and boundary direction response data based on the defect candidate region set, the subsequent processing of the defect contour is simultaneously restricted by the surface geometric constraints, interference intensity constraints and boundary direction constraints, avoiding false detection and false negative detection caused by relying solely on grayscale, color or edge response.

[0018] This application employs a surface-constrained Wulff defect morphology evolution network to perform morphology evolution correction on candidate contours in the defect candidate region set under surface constraints. This transforms ordinary defect candidate contours into stable defect morphology contour data. This processing method can stabilize and correct candidate contours that are broken or discontinuous due to the influence of reflection, shadow, and weld texture, resulting in a final defect contour with better continuity, closure, and surface position consistency. Since the stable defect morphology contour data is further used to extract defect morphology features, defect location features, and structural proximity features, it provides a more stable data foundation for defect type determination, defect expansion state determination, and structural proximity state determination.

[0019] This application also incorporates historical defect diagnosis feature data from the same wind turbine tower surface location to generate defect risk levels, defect diagnosis results, and defect early warning results. This ensures that the detection results are not limited to single image recognition but reflect changes in defect contour area, defect contour length, main extension direction of the defect contour, contour center surface coordinates, and adjacent surface distances. This method can identify whether defects are expanding, whether they are moving towards welds, flange connections, or opening edges, and generate corresponding early warning results and detection diagnosis reports, thereby improving the accuracy of wind turbine tower surface defect detection, the specificity of diagnosis, and the timeliness of maintenance early warnings. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers proposed in this invention. Figure 2 This is a schematic diagram of the process for identifying and suppressing interference in a visual detection, diagnosis and early warning method for curved surface defects of wind turbine towers proposed in this invention. Figure 3 This is a schematic diagram illustrating the process of constructing defect diagnosis feature data and generating defect risk warning results in a visual detection, diagnosis, and early warning method for curved surface defects of wind turbine towers proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers includes the following steps: Visual images of the curved surface of wind turbine towers and auxiliary data for tower inspection are collected and preprocessed to obtain curved surface inspection images and coordinate data of the towers. A visual inspection, diagnosis and early warning method for defects on the curved surface of wind turbine towers is proposed. The auxiliary data for tower inspection includes shooting posture data and tower structural parameters. Based on the tower inspection auxiliary data and tower surface coordinate data, the tower surface inspection image is corrected by surface unfolding to obtain the correspondence between the unfolded tower surface image and the unfolded coordinates. Based on the unfolded image of the tower surface, the interference region of the surface is identified, and the interference suppression processing of the unfolded image of the tower surface is performed according to the interference region to obtain the standard detection image of the tower surface. Based on the standard inspection image of the tower surface, candidate defect regions are extracted, and a set of candidate defect regions is constructed. Based on the defect candidate region set, tower surface coordinate data, unfolded coordinate correspondence, and surface interference region, candidate contour data, surface constraint data, interference suppression weight data, and boundary direction response data are generated, and defect morphology evolution input data are constructed. The defect morphology evolution input data is input into the surface-constrained Wulff defect morphology evolution network, and the morphology evolution correction of the candidate contours in the defect candidate region set under surface constraints is performed to obtain stable defect morphology contour data. Based on the defect stability profile data, the correspondence between unfolded coordinates and the tower surface coordinate data, defect morphology features, defect location features and structural proximity features are extracted, and defect diagnostic feature data are constructed. Based on the defect diagnosis feature data, historical defect diagnosis feature data of the same wind turbine tower surface location are retrieved to generate defect risk level, defect diagnosis result and defect warning result; Based on the defect risk level, defect diagnosis results, and defect early warning results, a report on the detection, diagnosis, and early warning of defects on the curved surface of wind turbine towers is generated.

[0023] In this embodiment, the preprocessing includes preprocessing the visual image of the tower surface to obtain the tower surface detection image, and establishing tower surface coordinate data based on the tower detection auxiliary data. A visual detection, diagnosis and early warning method for wind power tower surface defects is provided. The tower detection auxiliary data includes shooting posture data and tower structural parameters.

[0024] In this embodiment, the generation of the correspondence between the unfolded image of the tower surface and the unfolded coordinates includes: Based on the shooting posture data and tower structure parameters in the tower detection auxiliary data, surface projection matching is performed on the image position in the tower surface detection image to obtain the surface projection relationship between the image position and the tower surface coordinate data. During surface projection matching, the camera optical axis direction and imaging plane position are determined using shooting posture data, and the tower cylindrical surface equation is determined using tower structural parameters. The image position in the tower surface detection image is back-projected along the camera optical axis onto the tower cylindrical surface. When the back-projection line intersects with the tower cylindrical surface, the tower surface coordinates corresponding to the intersection point are used as the matching result of the image position, thus forming a surface projection relationship. Based on the surface projection relationship, the image position in the tower surface detection image is converted into the corresponding height coordinates and circumferential coordinates to obtain the image surface coordinate set; Based on the image surface coordinate set, the image positions in the tower surface detection image are arranged on the unfolding plane according to the height coordinate and circumferential coordinate, so as to obtain the initial unfolded image and the initial unfolded coordinate relationship; Based on the initial unfolded coordinate relationship, the pixel values ​​between adjacent unfolded positions in the initial unfolded surface image are interpolated, and the unfolded position offset in the initial unfolded coordinate relationship is corrected to obtain the unfolded surface image of the tower and the corrected unfolded coordinate relationship. Based on the image surface coordinate set and the corrected unfolded coordinate relationship, the unfolded coordinate correspondence between the image position in the tower surface detection image and the unfolded position in the unfolded tower surface image is generated.

[0025] In this embodiment, the generation of the standard inspection image of the tower surface includes: Based on the brightness distribution, color distribution, texture continuity, and edge response in the unfolded image of the tower surface, surface interference feature data is extracted; Identify the location of surface interference in the unfolded image of the tower surface based on surface interference feature data, and construct the surface interference region based on the location of surface interference. When identifying the location of surface interference, the location with a sudden increase in brightness and uninterrupted texture continuity is identified as the location of reflective surface interference; the location with a continuous decrease in brightness and a unidirectional gradual change in edge response is identified as the location of shadow surface interference; the location with a continuous linear extension that is consistent with the circumferential or height direction of the tower is identified as the location of weld texture surface interference; and the location with a sudden change in color distribution but no closed boundary formed by edge response is identified as the location of coating color difference surface interference. The above locations are then combined to form the surface interference area. Based on the location distribution of the surface interference region in the unfolded image of the tower surface and the surface interference feature data, the interference suppression parameters corresponding to the surface interference region are generated. Based on the interference suppression parameters, interference suppression processing is performed on the surface interference region in the tower surface unfolding image to obtain the interference suppression image; A standard detection image of the tower surface is generated based on the interference suppression image.

[0026] In this embodiment, the generation of the defect candidate region set includes: Based on the grayscale changes, color changes, texture changes, and edge changes in the standard inspection image of the tower surface, defect candidate feature data are extracted; Based on the defect candidate feature data, the defect candidate locations in the standard inspection image of the tower surface are determined, and an initial defect candidate region is generated based on the defect candidate locations; Based on the positional continuity between the initial defect candidate regions, the initial defect candidate regions are merged to obtain the merged defect candidate regions; Based on the area and boundary integrity of the merged defect candidate regions, the merged defect candidate regions are screened to obtain defect candidate regions; When screening candidate regions for merging defects, the number of pixels contained in the candidate region for merging defects is calculated as the region area, and the ratio of the length of continuous boundary segments to the total length of the boundary in the candidate region for merging defects is calculated as the boundary integrity. When the region area is zero, the boundary integrity is zero, or the candidate region for merging defects does not form a connectable boundary, the candidate region for merging defects is removed, and the remaining candidate regions for merging defects are retained as candidate regions for defects. Based on the location of the defect candidate regions in the standard inspection image of the tower surface, a set of defect candidate regions is constructed.

[0027] In this embodiment, the construction of the defect morphology evolution input data includes: Based on the region boundaries of each defect candidate region in the defect candidate region set in the standard inspection image of the tower surface, the boundary points on the region boundaries are extracted, and a boundary point sequence is formed according to the adjacent connection order of the boundary points on the region boundaries. Candidate contour data is then generated based on the boundary point sequence. Based on the coordinate correspondence, the boundary point sequence in the candidate contour data is mapped to the tower surface coordinate data to obtain the boundary point surface coordinate sequence. The surface spacing, circumferential variation and height variation between adjacent boundary points are calculated based on the boundary point surface coordinate sequence, and the surface spacing, circumferential variation and height variation constitute the surface contour constraint sequence. When calculating the surface spacing, the surface distance is calculated based on the height coordinate difference and circumferential coordinate difference between adjacent boundary points in the boundary point surface coordinate sequence. The circumferential coordinate difference is converted into arc length according to the tower radius. The circumferential change is the circumferential coordinate difference between adjacent boundary points, and the height change is the height coordinate difference between adjacent boundary points. The surface spacing is determined by the height change and the circumferential arc length. The local bending direction and continuous direction of the candidate contour data on the wind turbine tower surface are determined based on the surface contour constraint sequence. The boundary point surface coordinate sequence, surface contour constraint sequence, local bending direction and continuous direction are then combined in the order of the boundary point sequence to obtain the surface constraint data. When determining the local bending direction, the circumferential variation of multiple consecutive adjacent boundary points in the surface profile constraint sequence is accumulated to obtain the bending trend of the candidate profile along the circumference of the tower. When determining the continuous direction of the surface, the surface spacing and height variation of multiple consecutive adjacent boundary points are sorted, and the direction with continuous surface spacing and consistent height variation direction is retained as the continuous direction of the surface. Based on the positional overlap between the surface interference region and the boundary point sequence in the candidate contour data, interference overlap boundary points are determined. Based on the proximity distance between the surface interference region and the boundary point sequence in the candidate contour data, interference neighboring boundary points are determined. Interference overlap boundary points, interference neighboring boundary points, and boundary point sequences are then marked accordingly to obtain the boundary point interference marking sequence. Based on the boundary point interference labeling sequence, the first interference suppression weight is assigned to the interference coincident boundary points, the second interference suppression weight is assigned to the interference neighboring boundary points, and the third interference suppression weight is assigned to the boundary points in the boundary point sequence that are not labeled as interference coincident boundary points or interference neighboring boundary points. The interference suppression weight data is generated according to the arrangement order of the boundary point sequence. Among them, the first interference suppression weight, the second interference suppression weight, and the third interference suppression weight are all values ​​that are greater than zero and not greater than one. The first interference suppression weight is less than the second interference suppression weight, the second interference suppression weight is less than the third interference suppression weight, the first interference suppression weight corresponding to the interference coincident boundary point takes the minimum value, the second interference suppression weight corresponding to the interference adjacent boundary point takes the middle value, and the third interference suppression weight corresponding to the unmarked boundary point takes the maximum value, so that the traction effect of the boundary point in the surface interference area in the subsequent morphological evolution is weakened. The contour extension direction between adjacent boundary points is calculated based on the boundary point sequence in the candidate contour data. The boundary response intensity is calculated based on the pixel changes of the boundary point sequence in the standard detection image of the tower surface. The contour extension direction is corrected based on the local bending direction and the continuous direction of the surface in the surface constraint data to obtain the corrected contour direction. When calculating the boundary response intensity, the average pixel value, average color value, and average texture response value on both sides of the boundary are extracted with the boundary point as the center, and the weighted result of the difference between the two sides is calculated as the boundary response intensity. When correcting the surface direction, the contour extension direction is projected onto the direction range defined by the local bending direction and the continuous direction of the surface to obtain the surface correction contour direction. The boundary response intensity is weighted according to the interference suppression weight data to obtain the suppressed boundary response intensity. The surface correction contour direction and the suppressed boundary response intensity are then combined in the order of the boundary point sequence to obtain the boundary direction response data. Candidate contour data, surface constraint data, interference suppression weight data, and boundary direction response data are aligned and combined according to the same defect candidate region. This ensures that the boundary point sequence in the candidate contour data, the boundary point surface coordinate sequence in the surface constraint data, the surface contour constraint sequence, the interference suppression weight data, and the boundary direction response data maintain a point correspondence, thus constructing the defect morphology evolution input data.

[0028] In this embodiment, the generation of stable defect topography contour data includes: The defect morphology evolution input data is input into the surface-constrained Wulff defect morphology evolution network. A visual detection, diagnosis and early warning method for surface defects of wind turbine towers is proposed. The surface-constrained Wulff defect morphology evolution network includes an input alignment layer, an interface energy construction layer, a surface constraint modulation layer, an interference weight coupling layer, a Wulff morphology evolution layer and a stable contour output layer. In the input alignment layer, candidate contour data, surface constraint data, interference suppression weight data and boundary direction response data are read according to the same defect candidate region. The surface constraint data, interference suppression weight data and boundary direction response data are aligned according to the boundary point sequence in the candidate contour data to obtain the contour evolution feature sequence. In the interface energy construction layer, based on the boundary direction response data in the contour evolution feature sequence, the surface correction contour direction and the suppressed boundary response intensity corresponding to each boundary point are extracted, and the directional interface energy corresponding to each boundary point is calculated based on the surface correction contour direction and the suppressed boundary response intensity. The candidate contour interface energy sequence is generated according to the arrangement order of the boundary point sequence. In the surface constraint modulation layer, based on the surface constraint data in the contour evolution feature sequence, the surface contour constraint sequence, local bending direction and surface continuity direction are extracted. Then, the candidate contour interface energy sequence is subjected to surface direction constraint and surface continuity constraint based on the surface contour constraint sequence, local bending direction and surface continuity direction to obtain the surface constraint interface energy sequence. When constraining the continuity of a surface, the location where the surface spacing between adjacent boundary points suddenly increases or the direction of the change in height or circumferential change changes abruptly is taken as the location of the deviation from the continuity of the surface. The interface energy corresponding to the deviation from the continuity of the surface is increased to obtain the interface energy sequence of the surface constraint. In the interference weight coupling layer, based on the interference suppression weight data in the contour evolution feature sequence, the surface constraint interface energy sequence is coupled point by point. The interface energy corresponding to each boundary point in the surface constraint interface energy sequence is multiplied with the interference suppression weight corresponding to the same boundary point in the interference suppression weight data to obtain the interference coupling interface energy sequence. In the Wulff morphology evolution layer, the interface energy convergence direction corresponding to each boundary point in the boundary point sequence is determined according to the interference coupling interface energy sequence. Within the surface position range limited by the surface constraint data, the position of the boundary point sequence in the candidate contour data is updated according to the interface energy convergence direction to obtain the evolution contour data. When determining the interface energy convergence direction, compare the changes in the interference coupling interface energy of the same boundary point along multiple adjacent candidate directions within the surface position range defined by the surface constraint data, and select the direction that reduces the interference coupling interface energy as the interface energy convergence direction. When updating the position, move the boundary point to the adjacent surface position along the interface energy convergence direction, and keep the updated boundary point still corresponding to the same defect candidate region. In the stable profile output layer, the surface continuity of the evolving profile data is checked based on the surface constraint data, and the connection relationship of the evolving boundary points is determined according to the profile arrangement order in the evolving profile data. The boundary closure is checked based on the connection relationship of the evolving boundary points, and the position of the boundary points that fail the surface continuity check or boundary closure check is corrected to obtain the defect stable morphology profile data. During surface continuity verification, the surface spacing, circumferential variation, and height variation between adjacent boundary points in the evolution profile data are checked to ensure continuous change. During boundary closure verification, the connection relationship between the evolution boundary points is used to determine whether the first and last boundary points of the evolution profile data form a closed connection. For unclosed positions, the position is corrected according to the interface energy convergence direction of adjacent boundary points and the surface continuity direction.

[0029] In this embodiment, the construction of defect diagnostic feature data includes: Contour sampling is performed on the stable morphology contour data of the defect to obtain stable boundary points, and the connection relationship of stable boundary points is determined according to the arrangement order of stable boundary points in the stable morphology contour data of the defect. During contour sampling, contour points are extracted along the contour arrangement order of the defect stable morphology contour data at the same sampling interval to obtain stable boundary points. When there is a break between adjacent stable boundary points, connection points are added according to the contour arrangement order in the defect stable morphology contour data, and stable boundary point connection relationship is formed by stable boundary points and connection points. The stable profile closure line is determined based on the stable boundary points and their connection relationships. The defect profile area, defect profile length, defect profile closure degree, and defect profile continuity are then calculated based on the stable profile closure line to obtain the basic data of the defect morphology. The connection direction between adjacent stable boundary points is determined based on the connection relationship of stable boundary points. The distribution density of stable boundary points is calculated based on their distribution position on the closed line of the stable profile. The main extension direction of the defect profile, the number of branches of the defect profile, and the abrupt change position of the defect profile boundary are extracted based on the connection direction and the distribution density of stable boundary points. Among them, the main extension direction of the defect profile is determined according to the direction in which the continuous connection direction appears most frequently in the connection relationship of the stable boundary points; the number of branches of the defect profile is determined according to the number of bifurcations of the connection direction on the closed line of the stable profile; the abrupt change position of the defect profile boundary is determined according to the position where the change amplitude of the adjacent connection direction exceeds the direction change threshold; the direction change threshold is determined by the average value of the change amplitude of all adjacent connection directions on the closed line of the stable profile. The defect morphology features are obtained by combining the basic data of the defect morphology, the main extension direction of the defect contour, the number of branches of the defect contour, and the abrupt change position of the defect contour boundary according to the same stable morphology contour data of the defect. Based on the coordinate correspondence, the stable boundary points are mapped to the tower surface coordinate data to obtain the stable boundary point surface coordinate sequence; Based on the stable boundary point surface coordinate sequence, the height distribution range, circumferential distribution range, and contour center surface coordinates of the defect on the wind turbine tower surface are determined, and the height distribution range, circumferential distribution range, and contour center surface coordinates constitute the defect location characteristics. Based on the stable boundary point surface coordinate sequence and tower surface coordinate data, the structural proximity position of the defect stable morphology profile data in the tower surface coordinate data is determined, and the surface proximity distance, proximity direction and proximity overlap range between the defect stable morphology profile data and the structural proximity position are calculated. The surface proximity distance, proximity direction and proximity overlap range constitute the structural proximity feature. The tower structure location is determined based on the tower surface coordinate data. The tower structure location includes the weld location, flange connection location, and opening edge location. The adjacent location of the structure is determined based on the surface distance between the stable boundary point surface coordinate sequence and the tower structure location. The minimum surface distance from the stable boundary point surface coordinate sequence to the adjacent location of the structure is calculated as the surface proximity distance. The direction from the surface coordinate of the profile center to the adjacent location of the structure is calculated as the proximity direction. Defect morphology features, defect location features, and structural proximity features are combined according to the same stable morphology contour data of the defect to construct defect diagnostic feature data.

[0030] In this embodiment, the generation of defect risk level, defect diagnosis result, and defect early warning result includes: Based on the defect location features in the defect diagnosis feature data, determine the location of the wind turbine tower surface corresponding to the defect, and retrieve historical defect diagnosis feature data according to the location of the wind turbine tower surface. Based on the defect diagnosis feature data and historical defect diagnosis feature data, the defect morphology features, defect location features and structural proximity features corresponding to the same wind turbine tower surface location are compared to obtain defect feature change data. Based on the defect feature change data, extract the defect contour area change, defect contour length change, defect contour main extension direction change, contour center surface coordinate change, and surface proximity distance change, and construct defect evolution judgment data. The change in defect contour area is obtained by subtracting the historical defect contour area from the current defect contour area; the change in defect contour length is obtained by subtracting the historical defect contour length from the current defect contour length; the change in the main extension direction of the defect contour is obtained by the angle between the main extension direction of the current defect contour and the main extension direction of the historical defect contour; the change in the coordinates of the contour center surface is obtained by the surface distance between the coordinates of the current contour center surface and the coordinates of the historical contour center surface; and the change in the surface proximity distance is obtained by subtracting the historical surface proximity distance from the current surface proximity distance. The defect type is determined based on the defect morphology features in the defect diagnosis feature data, the defect expansion state is determined based on the defect evolution judgment data, and the structural proximity state is determined based on the structural proximity features and defect evolution judgment data in the defect diagnosis feature data. The defect diagnosis result is composed of the defect type, defect expansion state, and structural proximity state. Based on defect evolution judgment data and defect diagnosis results, defects are classified into risk levels to generate defect risk grades. When classifying risks, a high-level defect risk level is generated when both the change in defect contour area and the change in defect contour length are positive, and the change in the distance between adjacent surfaces is negative; a medium-level defect risk level is generated when either the change in defect contour area or the change in defect contour length is positive, and the change in the distance between adjacent surfaces is not negative; and a low-level defect risk level is generated when the changes in defect contour area, the change in defect contour length, and the change in the coordinates of the contour center surface are all zero. Based on the defect risk level and defect diagnosis results, generate defect warning results; When generating defect warning results, high-level defect risk levels correspond to immediate review of warning results, medium-level defect risk levels correspond to follow-up inspection of warning results, and low-level defect risk levels correspond to routine record of warning results. When the structural proximity status in the defect diagnosis results shows that the defect is moving closer to the structural proximity, the defect warning result is upgraded by one level.

[0031] In this embodiment, the detection, diagnosis, and early warning report generation of surface defects in wind turbine towers includes: Based on the defect diagnosis results, a defect diagnosis record is generated. A visual detection, diagnosis and early warning method for defects on the curved surface of wind turbine towers is proposed. The defect diagnosis record includes a defect type record, a defect location record and a defect status record. Based on the defect risk level, the defect diagnosis records are marked with risk to obtain the defect risk records; Based on the defect warning results, the defect risk records are marked with warnings to obtain defect warning records; Based on the defect warning records, generate a wind turbine tower surface defect detection, diagnosis and warning report according to the location of the wind turbine tower surface.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to the tower inspection scenario of a coastal wind farm. This wind farm is constantly exposed to sea breezes, salt spray, rain, strong sunlight, and temperature variations. The tower surface is prone to cracks, corrosion, coating peeling, flaking near welds, and localized depressions. Because the tower is a large-sized cylindrical curved surface, images captured by drones are prone to edge stretching, localized reflections, shadow gradients, weld textures, and coating color differences. These interferences can cause the boundaries of actual defects to break, and can also lead to reflections, rain streaks, or weld textures being mistaken for defects, thus affecting subsequent maintenance judgments.

[0033] In practical applications, inspection personnel use drones to conduct a circumferential survey along the tower's height, acquiring visual images of the tower surface. Simultaneously, they record the shooting posture data and tower structural parameters. The system preprocesses these visual images to obtain tower surface inspection images and establishes tower surface coordinate data based on the shooting posture data and tower structural parameters, ensuring that the positions in the images correspond to the actual positions on the tower surface. Subsequently, the system performs surface unfolding correction on the tower surface inspection images, obtaining the correspondence between the unfolded tower surface image and the unfolded coordinates. This reduces image stretching and positional deviations caused by the cylindrical surface, providing a foundation for subsequent defect localization.

[0034] After obtaining the unfolded image of the tower surface, the system identifies surface interference regions based on brightness distribution, color distribution, texture continuity, and edge response. For locations with a sudden increase in brightness but uninterrupted texture, the system identifies them as reflective surface interference locations; for locations with a continuous decrease in brightness and gradual edge changes, the system identifies them as shadow-type surface interference locations; for locations extending continuously along the circumferential or height direction of the tower, the system identifies them as weld texture-type surface interference locations; and for locations with abrupt color changes but no closed boundary, the system identifies them as coating color difference-type surface interference locations. The system generates interference suppression parameters based on the surface interference regions and performs interference suppression processing on the unfolded tower surface image to obtain a standard inspection image of the tower surface, making the true defect boundaries clearer and weakening the influence of false boundaries.

[0035] Next, the system extracts candidate defect regions based on the standard inspection image of the tower surface and constructs a set of candidate defect regions. For each candidate defect region, the system extracts boundary points on the region boundary to form a boundary point sequence and generates candidate contour data. Then, based on the coordinate correspondence and the tower surface coordinate data, the system maps the candidate contours to the tower surface position to obtain surface constraint data. At the same time, it generates interference suppression weight data by combining the surface interference region and generates boundary direction response data based on the pixel changes of the boundary point sequence in the standard inspection image of the tower surface. The above data are combined according to the same candidate defect region to form the input data for defect morphology evolution.

[0036] The system inputs the defect morphology evolution data into a surface-constrained Wulff defect morphology evolution network. This network corrects the candidate contours under surface constraints based on candidate contour data, surface constraint data, interference suppression weights, and boundary direction response data. For boundaries fractured due to reflection or shadows, the system corrects them using surface continuity constraints. For pseudo-boundaries near weld textures or coating color differences, the system reduces their traction effect using interference suppression weights. For true defect boundaries, the system performs contour stabilization updates based on the interface energy convergence direction, ultimately obtaining stable defect morphology contour data. This stable defect morphology contour data more completely reflects the actual defect boundary, providing a more reliable basis for subsequent diagnosis.

[0037] During the diagnostic phase, the system extracts defect morphology features based on stable defect contour data, extracts defect location features based on the correspondence between unfolded coordinates and tower surface coordinate data, and extracts structural proximity features based on the surface proximity relationships between the defect and welds, flange connections, and opening edges. This constructs defect diagnostic feature data. The system then retrieves historical defect diagnostic feature data for the same wind turbine tower surface location to determine whether the defect contour is expanding, whether its location is shifting, or whether it is approaching a structurally sensitive location. It then generates a defect risk level, defect diagnostic results, and defect warning results. When a defect shows continuous expansion, its contour moves closer to a structurally adjacent location, or its boundary abruptly intensifies, the system generates a corresponding warning.

[0038] Verification through on-site inspections, manual checks, and maintenance records shows that this invention can effectively reduce false detections caused by surface stretching, localized reflections, shading gradients, weld textures, and coating color differences, improving the continuity and integrity of defect contours. Based on the wind turbine tower surface defect detection, diagnosis, and early warning reports output by the system, inspection personnel can quickly determine the tower location of the defect, the defect morphology, the risk level, and subsequent handling methods, thereby improving the efficiency of wind turbine tower inspections, the accuracy of defect diagnosis, and the reliability of risk warnings.

[0039] Table 1 Comparison of Detection, Diagnosis and Early Warning Performance of Curved Surface Defects in Wind Turbine Towers

[0040] As shown in Table 1, the method of the present invention achieves a defect detection rate of 90.7%, which is an improvement over the 82.6% of manual inspection combined with ordinary image annotation, the 88.9% of ordinary deep learning image segmentation, and the 86.3% of surface unfolding correction combined with conventional edge detection. The reason for this is that the present invention does not directly output defect results based on the grayscale, color, or edge response in the tower surface image. Instead, it first performs surface unfolding correction on the tower surface detection image, then identifies the surface interference area and performs interference suppression processing. This weakens the influence of reflection, shadow, weld texture, and coating color difference on the defect boundary. The present invention can retain the boundary response of the real defect area, thereby improving the defect detection rate.

[0041] Regarding the false defect false detection rate, the method of this invention is 5.9%, which is lower than the 18.7% of manual inspection combined with ordinary image annotation method, the 12.4% of ordinary deep learning image segmentation method, and the 14.8% of surface unfolding correction combined with conventional edge detection method. This result shows that the present invention generates interference suppression weight data through surface interference regions and weakens the traction effect of interference coincident boundary points and interference adjacent boundary points in the surface-constrained Wulff defect morphology evolution network, making it difficult for false defect boundaries to enter the final stable defect morphology contour data, thus reducing the false detection rate.

[0042] Regarding the defect contour integrity rate, the method of this invention achieves 88.8%, which is an improvement compared to 76.4% for manual inspection combined with ordinary image annotation, 83.1% for ordinary deep learning image segmentation, and 81.5% for surface unfolding correction combined with conventional edge detection. This metric directly reflects the core advantages of this invention. The surface of a wind turbine tower is a cylindrical curved surface, and defect boundaries in images are easily affected by surface stretching, local occlusion, and changes in illumination, resulting in discontinuities. Ordinary methods typically only output a rough defect area, making it difficult to guarantee contour closure and continuity. This invention further transforms candidate defect regions into candidate contour data, surface constraint data, interference suppression weight data, and boundary direction response data, and inputs these into the surface-constrained Wulff defect morphology evolution network for morphology evolution correction. This allows the candidate contours to be updated along the continuous direction of the tower surface, ultimately forming more complete and stable defect morphology contour data.

[0043] Regarding the average positioning error, the method of this invention achieves 6.4 cm, which is superior to 18.5 cm for manual inspection combined with ordinary image annotation, 13.2 cm for ordinary deep learning image segmentation, and 11.7 cm for surface unfolding correction combined with conventional edge detection. This advantage stems from the continuous use of the correspondence between the tower surface coordinate data and the unfolded coordinates. This invention does not simply determine the defect location in a two-dimensional image, but rather maps the position in the visual image of the tower surface to the tower surface coordinate data, allowing the defect location to be expressed through the height distribution range, circumferential distribution range, and contour center surface coordinates. This enables inspectors to more quickly locate the actual surface position of the tower when verifying defects, reducing positioning deviations caused by image unfolding errors, changes in shooting angle, and manual estimation.

[0044] In terms of early warning accuracy, the method of this invention achieves 90.1%, which is higher than the 79.2% of manual inspection combined with ordinary image annotation, the 84.6% of ordinary deep learning image segmentation, and the 82.8% of surface unfolding correction combined with conventional edge detection. This is because the early warning of this invention is not a single image judgment, but is based on comparing defect diagnostic feature data with historical defect diagnostic feature data, and further extracting the changes in defect contour area, defect contour length, main extension direction of defect contour, center surface coordinates, and adjacent surface distances. When a defect shows an expanding trend, or moves closer to the weld, flange connection, or opening edge, the system can generate a corresponding defect risk level and defect early warning result. Therefore, this invention can more accurately distinguish between stable defects and defects with development risk.

[0045] Regarding the average processing time per tower, the method of this invention is 15.3 minutes, shorter than the 46.8 minutes of manual inspection combined with ordinary image annotation, the 24.5 minutes of ordinary deep learning image segmentation, and the 21.6 minutes of surface unfolding correction combined with conventional edge detection. Although this invention adds surface-constrained Wulff defect morphology evolution network processing, the overall processing efficiency is still superior because the front-end reduces invalid regions involved in the calculation through surface interference region identification, interference suppression processing, and defect candidate region set construction. Subsequently, morphology evolution correction is only performed on candidate contours. In particular, in the manual review stage, the defect stable morphology contour data, defect risk level, defect diagnosis results, and defect early warning results output by this invention can directly form a wind turbine tower surface defect detection, diagnosis, and early warning report, reducing the time spent on manual searching, manual comparison, and repeated confirmation.

[0046] As can be seen from the data in Table 1, the present invention outperforms the comparative methods in terms of defect detection rate, false defect detection rate, defect contour integrity rate, average positioning error, early warning accuracy, and average processing time per tower. The fundamental reason for this performance improvement lies in the fact that the present invention does not simply identify surface defects of wind turbine towers as ordinary two-dimensional image defects. Instead, it establishes a complete processing chain around the cylindrical surface of the tower, encompassing surface unfolding correction, surface interference suppression, defect candidate contour construction, surface constraint Wulff morphology evolution correction, and historical diagnostic feature comparison and early warning. This makes the final output results more stable in terms of contour integrity, position reliability, and early warning effectiveness.

[0047] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers, characterized in that, Includes the following steps: Collect visual images of the curved surface area of ​​the wind turbine tower and auxiliary data for tower detection, and preprocess them to obtain the tower surface detection image and tower surface coordinate data. Based on the tower inspection auxiliary data and tower surface coordinate data, the tower surface inspection image is corrected by surface unfolding to obtain the correspondence between the unfolded tower surface image and the unfolded coordinates. Based on the unfolded image of the tower surface, the interference region of the curved surface is identified, and interference suppression processing is performed to obtain the standard detection image of the tower surface; Based on the standard inspection image of the tower surface, candidate defect regions are extracted, and a set of candidate defect regions is constructed. Based on the defect candidate region set, tower surface coordinate data, unfolded coordinate correspondence, and surface interference region, input data for defect morphology evolution is constructed. The defect morphology evolution input data is input into the surface-constrained Wulff defect morphology evolution network, and the morphology evolution correction under surface constraints is performed on the defect candidate region set to obtain stable defect morphology contour data. Based on the defect stability morphology contour data, the correspondence between unfolded coordinates and the tower surface coordinate data, defect diagnostic feature data are constructed. Based on defect diagnostic feature data, retrieve historical defect diagnostic feature data to generate defect risk level, defect diagnosis result and defect warning result; Generate reports on the detection, diagnosis, and early warning of defects on the curved surface of wind turbine towers.

2. The visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to claim 1, characterized in that, The preprocessing includes preprocessing the visual image of the tower surface to obtain the tower surface detection image, and establishing tower surface coordinate data based on the tower detection auxiliary data.

3. The visual detection, diagnosis, and early warning method for curved surface defects in wind turbine towers according to claim 1, characterized in that, The generation of the correspondence between the unfolded image of the tower surface and the unfolded coordinates includes: Based on the shooting posture data and tower structure parameters in the tower detection auxiliary data, surface projection matching is performed on the image position in the tower surface detection image to obtain the surface projection relationship between the image position and the tower surface coordinate data. Based on the surface projection relationship, the image position in the tower surface detection image is converted into the corresponding height coordinates and circumferential coordinates, and arranged on the unfolding plane according to the height coordinates and circumferential coordinates to obtain the initial surface unfolding image and the initial unfolding coordinate relationship; Based on the initial unfolded coordinate relationship, the pixel values ​​between adjacent unfolded positions in the initial unfolded surface image are interpolated, and the unfolded position offset in the initial unfolded coordinate relationship is corrected to obtain the unfolded surface image of the tower and the corrected unfolded coordinate relationship. Based on the image surface coordinate set and the corrected unfolded coordinate relationship, the unfolded coordinate correspondence between the image position in the tower surface detection image and the unfolded position in the unfolded tower surface image is generated.

4. The visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to claim 1, characterized in that, The generation of the standard inspection image of the tower surface includes: Extract surface interference feature data from the unfolded image of the tower surface, identify the location of surface interference in the unfolded image of the tower surface based on the surface interference feature data, and construct the surface interference region. Based on the location distribution of the surface interference region in the unfolded image of the tower surface and the surface interference feature data, the interference suppression parameters corresponding to the surface interference region are generated. The tower surface unfolded image is subjected to interference suppression processing according to the interference suppression parameters to obtain the interference suppression image. The standard inspection image of the tower surface is then generated based on the interference suppression image.

5. The visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to claim 1, characterized in that, The generation of the defect candidate region set includes: Based on the grayscale changes, color changes, texture changes, and edge changes in the standard inspection image of the tower surface, defect candidate feature data are extracted; Based on the defect candidate feature data, the defect candidate locations in the standard inspection image of the tower surface are determined, and an initial defect candidate region is generated; The initial defect candidate regions are merged to obtain merged defect candidate regions, and the merged defect candidate regions are then filtered to obtain defect candidate regions. Based on the location of the defect candidate regions in the standard inspection image of the tower surface, a set of defect candidate regions is constructed.

6. The visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to claim 1, characterized in that, The construction of the defect morphology evolution input data includes: Based on the set of candidate defect regions, extract boundary points on the region boundary to form a boundary point sequence and generate candidate contour data; Based on the coordinate correspondence, the boundary point sequence in the candidate contour data is mapped to the tower surface coordinate data to obtain the boundary point surface coordinate sequence. The surface spacing, circumferential change and height change between adjacent boundary points are calculated and a surface contour constraint sequence is formed. The local bending direction and continuous direction of the candidate contour data on the wind turbine tower surface are determined based on the surface contour constraint sequence. The boundary point surface coordinate sequence, surface contour constraint sequence, local bending direction and continuous direction are then combined in the order of the boundary point sequence to obtain the surface constraint data. Based on the positional overlap between the surface interference region and the boundary point sequence in the candidate contour data, interference overlap boundary points are determined. Based on the proximity distance between the surface interference region and the boundary point sequence in the candidate contour data, interference neighboring boundary points are determined. Interference overlap boundary points, interference neighboring boundary points, and boundary point sequences are marked accordingly to obtain the boundary point interference marking sequence. Interference suppression weight data is generated according to the arrangement order of the boundary point sequence. The contour extension direction between adjacent boundary points is calculated based on the candidate contour data, the boundary response intensity is calculated based on the boundary point sequence, and the contour extension direction is corrected based on the surface constraint data to obtain the surface-corrected contour direction. The boundary response intensity is weighted according to the interference suppression weight data to obtain the suppressed boundary response intensity. The surface correction contour direction and the suppressed boundary response intensity are then combined in the order of the boundary point sequence to obtain the boundary direction response data. Candidate contour data, surface constraint data, interference suppression weight data, and boundary direction response data are combined by aligning points according to the same defect candidate region to construct input data for defect morphology evolution.

7. The visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to claim 1, characterized in that, The generation of the defect stable topography contour data includes: The defect morphology evolution input data is input into the surface-constrained Wulff defect morphology evolution network, which includes an input alignment layer, an interface energy construction layer, a surface constraint modulation layer, an interference weight coupling layer, a Wulff morphology evolution layer, and a stable contour output layer. In the input alignment layer, candidate contour data, surface constraint data, interference suppression weight data and boundary direction response data are read according to the same defect candidate region. The surface constraint data, interference suppression weight data and boundary direction response data are aligned according to the boundary point sequence in the candidate contour data to obtain the contour evolution feature sequence. In the interface energy construction layer, based on the boundary direction response data in the contour evolution feature sequence, the surface correction contour direction and the suppressed boundary response intensity corresponding to each boundary point are extracted, and the directional interface energy corresponding to each boundary point is calculated based on the surface correction contour direction and the suppressed boundary response intensity. The candidate contour interface energy sequence is generated according to the arrangement order of the boundary point sequence. In the surface constraint modulation layer, based on the surface constraint data in the contour evolution feature sequence, the surface contour constraint sequence, local bending direction and surface continuity direction are extracted. Then, the candidate contour interface energy sequence is subjected to surface direction constraint and surface continuity constraint based on the surface contour constraint sequence, local bending direction and surface continuity direction to obtain the surface constraint interface energy sequence. In the interference weight coupling layer, the interference suppression weight data in the contour evolution feature sequence is used to perform point-by-point weight coupling on the surface constraint interface energy sequence to obtain the interference coupling interface energy sequence. In the Wulff morphology evolution layer, the interface energy convergence direction corresponding to each boundary point in the boundary point sequence is determined according to the interference coupling interface energy sequence. Within the surface position range limited by the surface constraint data, the position of the boundary point sequence in the candidate contour data is updated according to the interface energy convergence direction to obtain the evolution contour data. In the stable profile output layer, the surface continuity of the evolving profile data is checked based on the surface constraint data, and the connection relationship of the evolving boundary points is determined according to the profile arrangement order in the evolving profile data. The boundary closure is checked based on the connection relationship of the evolving boundary points, and the position of the boundary points that fail the surface continuity check or boundary closure check is corrected to obtain the stable morphological profile data of the defect.

8. A visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to claim 1, characterized in that, The construction of the defect diagnostic feature data includes: Contour sampling is performed on the stable morphology contour data of the defect to obtain stable boundary points and determine the connection relationship of stable boundary points; Based on the stable boundary points and their connection relationships, the stable profile closure line is determined, and the defect profile area, defect profile length, defect profile closure degree, and defect profile continuity are calculated to obtain the basic data of the defect morphology. The connection direction between adjacent stable boundary points is determined based on the connection relationship between stable boundary points, and the distribution density of stable boundary points is calculated based on the distribution position of stable boundary points on the closed line of the stable profile. Based on the connection direction and the distribution density of stable boundary points, the main extension direction of the defect contour, the number of defect contour branches, and the abrupt change position of the defect contour boundary are extracted and combined to obtain the defect morphology features. Based on the coordinate correspondence, the stable boundary points are mapped to the tower surface coordinate data to obtain the stable boundary point surface coordinate sequence; Based on the stable boundary point surface coordinate sequence, the height distribution range, circumferential distribution range, and contour center surface coordinates of the defect on the wind turbine tower surface are determined, and the height distribution range, circumferential distribution range, and contour center surface coordinates constitute the defect location characteristics. The structural proximity features are constructed based on the stable boundary point surface coordinate sequence and the tower surface coordinate data. Defect morphology features, defect location features, and structural proximity features are combined according to the same stable morphology contour data of the defect to construct defect diagnostic feature data.

9. A visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to claim 1, characterized in that, The generation of the defect risk level, defect diagnosis result, and defect early warning result includes: Based on the defect location features in the defect diagnosis feature data, determine the location of the wind turbine tower surface corresponding to the defect, and retrieve historical defect diagnosis feature data according to the location of the wind turbine tower surface. Based on the defect diagnosis feature data and historical defect diagnosis feature data, the defect morphology features, defect location features and structural proximity features corresponding to the same wind turbine tower surface location are compared to obtain defect feature change data. Based on the defect feature change data, extract the defect contour area change, defect contour length change, defect contour main extension direction change, contour center surface coordinate change, and surface proximity distance change, and construct defect evolution judgment data. The defect type is determined based on the defect morphology features in the defect diagnosis feature data, the defect expansion state is determined based on the defect evolution judgment data, and the structural proximity state is determined based on the structural proximity features and defect evolution judgment data in the defect diagnosis feature data. The defect diagnosis result is composed of the defect type, defect expansion state, and structural proximity state. Based on defect evolution judgment data and defect diagnosis results, defects are classified into risk levels, defect risk grades are generated, and defect early warning results are generated.

10. A visual detection, diagnosis, and early warning method for defects on the curved surface of wind turbine towers according to claim 1, characterized in that, The generation of the wind turbine tower surface defect detection, diagnosis, and early warning report includes: Generate defect diagnosis records based on defect diagnosis results; Based on the defect risk level, the defect diagnosis record is marked with risk to obtain the defect risk record; Based on the defect warning results, the defect risk records are marked with warnings to obtain defect warning records; Based on the defect warning records, generate a wind turbine tower surface defect detection, diagnosis and warning report according to the location of the wind turbine tower surface.

Images