Cross-domain defect guidance collaborative detection method and system for wind turbine generators

Abstract

The present application relates to the technical field of wind turbine operation and maintenance, and particularly relates to a wind turbine defect cross-domain guidance collaborative detection method and system, comprising: acquiring global scanning data of wind turbine blades and tower drums of non-stop wind turbines after a drone is calibrated by cross-domain space-time synchronization; acquiring near-domain reference data collected by a climbing drum robot at a preset reference position of the tower drum; taking the near-domain reference data as an anchor point, splicing the global scanning data to determine a defect suspected area and a spatial position; scheduling the climbing drum robot to a target site for spot detection, and acquiring multi-modal detection data; and combining real-time wind field conditions and unit operation parameters to perform interference compensation on the multi-modal data to obtain defect characteristic information. Through the present application, space-time unification and collaboration of the detection data of the drone and the climbing drum robot are realized, and the reliability of wind turbine defect identification and positioning is improved.

Description

Technical Field

[0001] This invention relates to the technical field of wind turbine operation and maintenance, and in particular to a cross-domain guided collaborative detection method and system for wind turbine defects. Background Technology

[0002] With the rapid development of the new energy industry, wind power has become a core component of the clean energy supply system, and the global installed capacity of wind power continues to expand. Wind turbines operate in complex outdoor environments for extended periods. Their towers and blades, as core load-bearing and energy-capturing components, are subjected to alternating loads, wind and sand erosion, and extreme weather conditions, making them prone to surface cracks, coating peeling, corrosion, and wear. If these defects are not identified and addressed in a timely manner, they will continue to escalate during operation, potentially leading to component failure, unplanned turbine shutdowns, and other significant safety and economic losses.

[0003] Currently, the mainstream technologies for surface defect detection in wind turbines include drone inspection and turbine-climbing robot inspection. Drone inspection offers advantages such as a wide operating range and high mobility, enabling large-scale macroscopic scanning of wind turbine towers and blades, and efficiently completing full-coverage initial screening of multiple turbines in a wind farm. Turbine-climbing robots can stably walk along the tower surface and are equipped with high-precision sensors to perform close-range, detailed inspection of the tower surface, offering significantly higher accuracy in identifying minute defects compared to long-distance inspection methods.

[0004] However, since wind turbines are in continuous operation and elastic deformation without stopping, and the external wind field environment is complex and changeable, the remote scanning of UAVs and the near-field acquisition of the climbing robot belong to different spatial reference systems and acquisition time sequences. When the detection data of the two are interacting in a coordinated manner, spatiotemporal data misalignment is very likely to occur, making it difficult to achieve accurate correlation and targeted guidance of defect locations. Summary of the Invention

[0005] This invention provides a cross-domain guided collaborative detection method and system for wind turbine defects, which can effectively solve the problems in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A cross-domain guided collaborative detection method for wind turbine defects includes: The full-domain scan data is obtained by performing a full-domain scan of the wind turbine blades and towers without stopping after the UAV is calibrated based on a preset cross-domain spatiotemporal synchronization benchmark. Acquire near-field reference data collected by the climbing robot at a preset reference position on the wind turbine tower; Using near-field benchmark data as anchor points, spatiotemporal feature stitching is performed on the full-field scan data to determine the suspected defect area and its spatial location; The robot is scheduled to move to the target site corresponding to the spatial location for fixed-point detection to obtain multimodal detection data; By combining real-time wind farm conditions and unit operating parameters, interference compensation is performed on the multimodal detection data to obtain compensated feature information, which is then used to characterize the defect features of the target site.

[0007] Furthermore, the method for establishing the pre-defined cross-domain spatiotemporal synchronization benchmark includes: After the climbing robot completes attitude locking at the preset reference position on the tower, the spatial coordinates of the preset reference position on the tower are collected to construct a spatial reference coordinate system; A global timing synchronization trigger signal is generated based on the spatial reference coordinate system, and the global timing synchronization trigger signal and the spatial reference coordinate system are sent to the UAV to complete the unification of spatial coordinates and timing alignment of data acquisition between the UAV and the climbing robot.

[0008] Furthermore, using near-field benchmark data as anchor points, spatiotemporal feature stitching is performed on the full-field scan data to determine suspected defect areas and their spatial locations, including: Using the spatial coordinates and temporal signals in the near-field reference data as indexes, the spatiotemporal features of the dynamic sequence images of the blades and the full-field images of the tower in the full-field scanning data are spatiotemporally stitched together to eliminate image misalignment caused by blade rotation and UAV flight attitude fluctuations, and to generate continuous feature maps of the entire surface of the blades and the tower. Pixel-level feature extraction is performed on continuous feature maps to identify and lock at least one suspected defect region among cracks, coating peeling, corrosion, and wear, and to extract the contour, grayscale, texture multi-dimensional feature vectors and their spatial locations of each suspected defect region.

[0009] Furthermore, by combining real-time acquired wind farm operating conditions and unit operating parameters, interference compensation is performed on the multi-modal detection data, including: The real-time wind farm operating conditions and unit operating parameters are matched with a pre-built operating condition structure coupling interference factor library to obtain the interference weight and compensation coefficient of the detection data in each dimension under the current operating conditions. Based on the compensation coefficient, interference decoupling and feature compensation are performed on a dimension-by-dimensional basis for multimodal detection data to eliminate feature shifts caused by non-defect factors.

[0010] Furthermore, controlling the climbing robot to complete attitude locking at a preset reference position on the tower includes: The robot is controlled to walk to the preset reference position on the tower via a magnetic segmented moving module, and then adaptively fits the conical curved surface of the tower to complete the attitude lock.

[0011] Furthermore, the UAV is calibrated based on a preset cross-domain spatiotemporal synchronization benchmark, including: The UAV uses the spatial reference coordinate system in the preset cross-domain spatiotemporal synchronization reference as the anchor point to complete the global calibration of its own flight attitude, so that the spatial coordinates collected by the UAV are mapped and bound to the spatial reference coordinate system.

[0012] Furthermore, the robot is directed to move to the target location corresponding to the spatial position for fixed-point detection, including: The tube-climbing robot plans the optimal walking path from its current position to the target site based on the SLAM algorithm and curvature adaptive control strategy, and autonomously moves to the target site to complete the detection posture lock.

[0013] Furthermore, based on the compensation coefficient, the multimodal detection data undergoes dimension-by-dimensional interference decoupling and feature compensation, including at least one of the following: Based on the characteristics of sound wave signals, compensate for frequency interference caused by wind noise; Based on visual image characteristics, compensate for distortions caused by changes in lighting and the curved surface of the tower; Based on the vibration spectrum characteristics, compensate for the spectral superposition interference caused by the vibration of the unit itself during operation; For tilt angle data, compensate for the tower's elastic deformation interference caused by wind load.

[0014] Furthermore, the near-field reference data includes the three-dimensional coordinates of the tower's preset reference position, the standard tilt angle, the reference vibration characteristics, and the standard visual image.

[0015] On the other hand, the present invention also provides a cross-domain guided collaborative detection system for wind turbine defects, comprising: The data acquisition module is used to acquire full-domain scan data obtained by the UAV after calibration based on a preset cross-domain spatiotemporal synchronization benchmark, which is used to perform full-domain scans on the blades and towers of wind turbine units that are running without stopping. The benchmark acquisition module is used to acquire near-field benchmark data collected by the climbing robot at a preset benchmark position on the wind turbine tower. The feature stitching module is used to stitch spatiotemporal features of the full-domain scan data with near-domain reference data as anchor points to determine the suspected defect area and its spatial location. The detection and scheduling module is used to schedule the climbing robot to move to the target site corresponding to the spatial position for fixed-point detection and obtain multimodal detection data; The interference compensation module is used to combine the real-time wind farm conditions and unit operating parameters to perform interference compensation on the multimodal detection data, and obtain the compensated feature information. The compensated feature information is used to characterize the defect features of the target site.

[0016] The technical solution of this invention can achieve the following technical effects: By pre-setting a cross-domain spatiotemporal synchronization benchmark, the spatial coordinates of the climbing robot and the UAV are unified and the acquisition time sequence is aligned. Spatiotemporal feature stitching with near-domain benchmark data as anchor points can eliminate image misalignment caused by blade rotation and UAV flight attitude fluctuations, and determine the suspected defect area and spatial location. Combining SLAM algorithm and curvature adaptive control strategy, the climbing robot can be scheduled and its attitude locked to the target site, ensuring the accuracy of close-range fine-grained detection. Through the matching of the working condition structure coupling interference factor library and the dimensional interference compensation, non-defect factors can be eliminated to obtain real and reliable defect feature information, comprehensively improving the accuracy and reliability of wind turbine defect identification and positioning.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the cross-domain guided collaborative detection method for wind turbine defects according to the present invention. Figure 2 This is a schematic diagram of the cross-domain guidance and collaborative detection system for wind turbine defects according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] like Figure 1 As shown, the cross-domain guided collaborative detection method for wind turbine defects of the present invention specifically includes the following steps: Step S1: Obtain the full-domain scan data obtained by the UAV performing a full-domain scan of the wind turbine blades and towers without stopping operation after calibration based on a preset cross-domain spatiotemporal synchronization benchmark. Step S2: Obtain near-field reference data collected by the climbing robot at the preset reference position of the wind turbine tower; Step S3: Using the near-field reference data as anchor points, spatiotemporal feature stitching is performed on the full-field scan data to determine the suspected defect area and its spatial location; Step S4: Schedule the climbing robot to move to the target site corresponding to the spatial location for fixed-point detection and obtain multimodal detection data; Step S5: Combine the real-time wind farm conditions and unit operating parameters to perform interference compensation on the multimodal detection data and obtain the compensated feature information. The compensated feature information is used to characterize the defect features of the target site.

[0023] In this embodiment, by combining cross-domain spatiotemporal synchronization benchmarks with near-domain benchmark data, not only is the spatial coordinate unification and acquisition timing alignment of the detection data from the UAV and the climbing robot achieved, but the near-domain benchmark data also serves as an anchor point to provide a reference for spatiotemporal feature stitching, eliminating image misalignment caused by blade rotation and UAV flight attitude fluctuations, and ensuring the accuracy of defect-suspected area localization. Through the targeted scheduling of the climbing robot based on SLAM algorithm and curvature adaptive control strategy, the located defect-suspected areas are transformed into target sites that can be detected at close range, realizing the connection between the UAV's full-domain macroscopic initial screening and the climbing robot's near-domain refined detection. Finally, by combining the interference compensation of wind farm conditions and unit operating parameters, non-defect interference in multimodal detection data is accurately stripped away, ensuring the authenticity of defect feature information. The combination of these three aspects makes the detection process a cross-domain collaborative closed-loop detection, which can improve the reliability and efficiency of wind turbine defect detection.

[0024] In a specific implementation, as one example, given that there is no unified spatiotemporal reference for the inspection of UAVs and tower-climbing robots, and considering the continuous operation of wind turbines and the complexity of the external wind field environment, direct data acquisition would lead to spatial misalignment and temporal offset during subsequent defect correlation. Therefore, before the UAV performs a full-domain scan, it is necessary to establish a cross-domain spatiotemporal synchronization reference, using the data collected by the tower-climbing robot at the physical reference position on the tower as the anchor point, and to perform spatial and temporal calibration of the UAV based on this reference. In this embodiment, after the tower-climbing robot completes attitude locking at the preset reference position on the tower, it constructs spatial reference coordinates, generates a temporal synchronization trigger signal, and sends it to the UAV, enabling the UAV to complete the unified calibration of its own flight attitude and acquisition timing before performing a full-domain scan. This results in full-domain scan data that can be traced back to the same reference in both space and time, as detailed below: Step S11: Obtain a preset cross-domain spatiotemporal synchronization benchmark; this cross-domain spatiotemporal synchronization benchmark is established based on the climbing robot, and the establishment method is as follows: First, the reference position of the tower is determined. Based on the preliminary detection and selection of the wind turbine tower structure and operating characteristics, the reference position must meet the following requirements: it should not be obstructed by the rotation of the blades, so that the UAV can clearly capture the spatial features of the position; and it should be located in the middle section of the tower, avoiding the area of ​​excessive wind load deformation at the top of the tower and the area of ​​vibration interference from the foundation at the bottom, so as to ensure that the collected spatial coordinates, tilt angle and vibration characteristics are stable and representative. The robot is then controlled to move to the preset reference position on the tower via the magnetic segmented moving module. The magnetic segmented moving module changes the magnetic force by adjusting the strength of the current inside the module, so that each segmented moving module is evenly attached to the surface of the tower. The robot moves in an alternating manner, keeping at least two segmented modules attached to the surface of the tower each time it moves, and then moving the other segmented modules. This allows the robot to move smoothly. After the climbing robot walks to the preset reference position, it activates the posture adaptive adjustment module. Through the onboard tilt sensor and distance sensor, it detects the degree of fit and posture angle between itself and the conical curved surface of the tower. It then adjusts the adsorption force and angle of each magnetic segment module to make the body of the climbing robot completely adaptively fit the conical curved surface of the tower, completes posture locking, and ensures that the climbing robot will not deviate in posture during the process of collecting reference data, thus ensuring the accuracy of the collected data. After the climbing robot completes attitude locking, it activates its onboard acquisition module to acquire the spatial coordinates of the reference position through the three-dimensional coordinate acquisition unit. The acquisition frequency is matched with the natural vibration frequency of the wind turbine to avoid coordinate deviation caused by the acquisition frequency being too low. The acquisition time is no less than one complete natural vibration cycle of the unit to ensure the integrity of the acquired data. After the acquisition is completed, the acquired three-dimensional coordinate data is denoised to remove abnormal data caused by wind field interference and equipment vibration, and retain stable coordinate data as the absolute spatial coordinates of the preset reference position of the tower. Based on the reference data collected by the climbing robot, a spatial reference coordinate system is constructed with the preset reference position as the origin, the direction of the tower axis as the Z-axis, the direction perpendicular to the tower surface and outward as the X-axis, and the direction parallel to the tangent of the tower surface as the Y-axis. After the spatial reference coordinate system is constructed, a global timing synchronization trigger signal is generated based on this coordinate system. The frequency of the timing trigger signal is consistent with the acquisition frequency of the three-dimensional coordinate acquisition unit of the climbing robot. The acquisition frequency of the three-dimensional coordinate acquisition unit is matched with the natural vibration frequency of the wind turbine, which can avoid resonance interference between the timing signal and the vibration of the turbine operation and ensure the stability of timing synchronization. The trigger signal and the spatial reference coordinate system are sent to the UAV to complete the construction of the cross-domain spatiotemporal synchronization reference. Step S12: After receiving the global timing synchronization trigger signal and the spatial reference coordinate system, the UAV uses the spatial reference coordinate system in the cross-domain spatiotemporal synchronization reference as the anchor point to start its own flight attitude calibration module, adjust its own flight attitude, and make the UAV's acquisition coordinate system consistent with the spatial reference coordinate system; the UAV aligns its own acquisition timing with the global timing synchronization trigger signal to ensure that the timing of the UAV's acquired data is completely synchronized with the timing of the climbing robot's acquired reference data, eliminating data deviation caused by timing misalignment; Step S13: After the UAV completes the calibration, it performs a full-domain scan of the wind turbine blades and towers that are running without stopping, and acquires full-domain scan data. During the scan, the UAV maintains the mapping and binding state with the spatial reference coordinate system in the preset cross-domain spatiotemporal synchronization reference, and receives the global timing synchronization trigger signal in real time to ensure that each set of scan data has the corresponding spatial coordinate information and timing information.

[0025] In this embodiment, by using the climbing robot as a spatiotemporal reference anchor point and leveraging its close proximity to the tower, a preset cross-domain spatiotemporal synchronization reference can be established. This enables the unification of the spatiotemporal reference between the UAV and the climbing robot during inspection, avoids spatial misalignment and temporal offset of the detection data, and allows the two to coordinate their acquisition processes. This results in the acquisition of effective full-domain scanning data, providing a data foundation for the accurate identification, targeted positioning, and multimodal data collaborative analysis of suspected defect areas.

[0026] In some embodiments of the present invention, near-field reference data serves as the anchor point for spatiotemporal feature stitching of full-field scanning data. Its data integrity and accuracy directly determine the accuracy of locating suspected defect areas. The three-dimensional coordinates in the near-field reference data provide a spatial positioning reference, ensuring the spatial correlation between full-field scanning data and near-field data. The standard tilt angle is the real-time tilt angle of the tower's preset reference position, reflecting the tower's reference deformation state and providing a reference for image misalignment compensation. The reference vibration characteristics are the inherent vibration frequency and amplitude of the tower's preset reference position, supplementing the timing synchronization basis and preventing timing offsets from affecting data collaboration. The standard visual image can serve as a visual stitching template, eliminating image distortion from UAV far-field scanning. This embodiment, through the collaborative work of an integrated acquisition module, enables the climbing robot to acquire near-field reference data containing three-dimensional coordinates, standard tilt angle, reference vibration characteristics, and a standard visual image at the tower's preset reference position. Specifically, the following operations are performed: Step S21: Confirm that the climbing robot has completed attitude locking at the preset reference position on the tower. Step S22: The acquisition module on the climbing robot includes a 3D coordinate acquisition unit, an tilt angle acquisition unit, a vibration acquisition unit, and a vision acquisition unit, all four units working synchronously. The 3D coordinates of the reference position have already been acquired during the cross-domain spatiotemporal synchronous reference establishment, and can be directly extracted as the 3D coordinates in the near-domain reference data. The tilt angle acquisition unit is a tilt angle sensor, with the same acquisition frequency and duration as the 3D coordinate acquisition. By averaging the acquired tilt angle data and removing abnormal tilt angle values ​​caused by instantaneous wind load impacts, the resulting average value is the standard tilt angle in the near-domain reference data. The vibration feature acquisition unit is a vibration sensor, with an acquisition frequency higher than the natural vibration frequency of the wind turbine, ensuring complete capture of all vibration signals. The data acquisition process covers multiple natural vibration cycles of the unit to ensure the representativeness of the collected vibration data. By filtering the vibration signals to remove interference signals such as wind noise and equipment vibration, characteristic parameters of the natural vibration frequency, amplitude, and spectrum are extracted and used as the benchmark vibration features in the near-field benchmark data. The visual acquisition unit is an industrial camera, with an acquisition resolution consistent with that of the UAV's full-field scanning images to ensure image stitching compatibility. Multiple images are acquired from different angles to ensure complete capture of the surface visual features of the benchmark location. By performing denoising, distortion correction, and grayscale normalization on the images, blurry and distorted images are removed, and clear and complete images are retained as the standard visual images in the near-field benchmark data. Step S23: Organize the collected near-field reference data in a unified manner; during the organization process, associate and bind the three-dimensional coordinates, standard tilt angle, reference vibration characteristics and standard visual image, and add corresponding time series labels to them. The time series labels are consistent with the global time series synchronization trigger signal.

[0027] In this embodiment, by taking the posture lock state of the climbing robot at the preset reference position of the tower as the basis, it can be ensured that the acquisition environment and equipment status meet the accuracy requirements. Through the coordinated acquisition and optimization processing of four types of parameters, the spatial, posture, temporal and visual reference requirements are fully covered, which can fundamentally avoid the full-domain scanning data splicing deviation caused by the lack of reference data or insufficient accuracy, and ensure the accurate coordination of the detection data of the UAV and the climbing robot.

[0028] In a specific implementation, as one example, to eliminate image misalignment caused by blade rotation without stopping and UAV flight attitude fluctuations, and to generate continuous and complete feature maps of the entire surface of the blades and tower, it is necessary to establish a spatiotemporal index association between the UAV's full-domain scan data and the near-domain reference data of the tower-climbing robot. However, traditional target detection networks have redundant parameters and high computational load, which cannot meet the real-time data processing needs of UAVs and tower-climbing robots. Furthermore, they lack the ability to extract pixel-level features of minute defects and have a single feature extraction dimension, which easily leads to missed or false detections of suspected defect areas. This embodiment uses near-domain reference data as an anchor point, combined with a spatiotemporal feature stitching algorithm and a lightweight improved target detection network, to achieve the stitching of full-domain scan data and the identification and localization of suspected defect areas. The specific implementation steps are as follows: Step S31: Extract the three-dimensional coordinates and time series signals from the near-field reference data as index parameters; Step S32: Based on the index parameters, perform spatiotemporal feature stitching on the full-domain scan data to eliminate image misalignment. The full-domain scan data includes dynamic sequence images of blades and full-domain images of the tower collected by the UAV. The dynamic sequence images of blades have inter-frame misalignment due to the continuous rotation of the blades, and the full-domain images of the tower have local image shifts due to fluctuations in the flight attitude of the UAV and elastic deformation of the tower. These deviations need to be eliminated by spatiotemporal feature stitching. First, using the three-dimensional coordinates in the near-field reference data as spatial indexes, the spatial coordinates of the tower full-field image and blade dynamic sequence image in the full-field scan data are mapped and aligned with the three-dimensional coordinates of the near-field reference data. Secondly, using the time-series signal of the near-field reference data as a time-series index, the acquisition time sequence of the full-field scan data is aligned frame by frame with the time-series signal of the near-field reference data to ensure the inter-frame temporal continuity of the dynamic sequence image of the blades; at the same time, combined with the standard tilt angle in the near-field reference data, the degree of elastic deformation of the tower under wind load is determined, and the offset area is returned to a reasonable position by adjusting the image pixel coordinates. Finally, for the dynamic sequence images of the blades, an inter-frame feature matching algorithm is used. Using the feature points of the standard visual image in the near-field reference data as a reference, feature matching is performed on each frame of the dynamic sequence images of the blades to eliminate image rotation and displacement deviations caused by blade rotation. For the full-domain images of the tower, a regional stitching algorithm is used. Tower images acquired by the UAV at different acquisition positions are aligned and stitched together in regions using the three-dimensional coordinates of the near-field reference data as anchor points to ensure that the stitched tower images are non-overlapping and non-broken. Finally, a continuous feature map of the entire surface of the blades and tower is generated, which can completely and clearly reflect the surface state of the blades and tower. Step S33: Construct a lightweight and improved object detection network. During the construction process, redundant convolutional layers in the traditional object detection network are removed, while the core feature extraction layer and classification layer are retained. Depthwise separable convolution is used to replace traditional convolution, reducing computation while preserving pixel-level features of the image. An attention mechanism is added to the feature extraction layer to focus on areas in the continuous feature map that may have defects, such as the area around the tower weld and the edge of the blade, which are prone to defects, thereby improving the efficiency and accuracy of defect feature extraction. The loss function is optimized in the classification layer, and differentiated classification weights are set for the feature differences of four types of defects: cracks, coating peeling, corrosion, and wear. To avoid misjudgments due to high similarity of different defect features, the network training process uses image samples of wind turbine blades and tower defects with different defect types, degrees, and lighting conditions. All training samples are manually labeled and confirmed to ensure that the trained network can accurately identify defects. Standard visual images from near-domain benchmark data are introduced as reference samples during training to enable the network to distinguish between normal surface features and defect features. After training, the network performance is verified by selecting defect samples and normal samples that were not used in training to ensure that the network's accuracy and recall for identifying the four types of defects meet the preset requirements.

[0029] Step S34: Input the continuous feature map of the entire surface of the blade and tower into the trained lightweight improved target detection network. The network first scans the continuous feature map pixel by pixel, extracting the grayscale, texture, and contour features of each pixel. Combined with the attention mechanism, it focuses on areas in the continuous feature map that are prone to defects, such as the blade edge and the area around the tower weld. The pixel-level features of these areas are enhanced, extracted, and analyzed. By comparing the normal surface features of the standard visual image in the near-field benchmark data, the network selects the set of pixels with abnormal features and initially identifies them as suspected defect areas. For the initially selected suspected defect areas, the network further analyzes the pixel-level features of the area through the classification layer to determine the defect type. At the same time, it accurately fits and locks the boundary of the suspected defect area, integrates the distribution range of all abnormal pixels in the area, forms the complete contour information of the suspected defect area, and clarifies the specific range of the suspected defect area. Step S35: Based on the locked contour of the suspected defect area, extract the multi-dimensional feature vector of the area; during the feature vector extraction process, contour fitting is performed on the contour features to extract the feature point coordinates of the contour and form a contour feature vector; gray-level statistical processing is performed on the gray-level features to extract feature parameters such as gray-level mean and gray-level variance to form a gray-level feature vector; texture analysis is performed on the texture features to extract feature parameters such as texture entropy and texture contrast to form a texture feature vector. The three types of feature vectors are integrated to serve as the multi-dimensional feature vector of the suspected defect area. Step S36: Map the pixel coordinates of the suspected defect area to the spatial reference coordinate system of the near-field reference data to obtain the spatial location coordinates of each suspected defect area.

[0030] In this embodiment, by using the spatial coordinates and temporal signals in the near-field reference data as the core index, combined with standard visual images and standard tilt angles for auxiliary calibration, spatiotemporal stitching of the full-field scanning data is achieved, fundamentally eliminating the image misalignment problem. By making targeted lightweight improvements to the target detection network, introducing an attention mechanism and differentiated classification weights, the recognition accuracy and real-time processing capability of minute defects are improved. At the same time, multi-dimensional feature vectors of contour, grayscale, and texture are extracted to provide comprehensive support for defect confirmation. Finally, the confirmed defect area is mapped to the spatial reference coordinate system of the near-field reference data to generate spatial location coordinates with engineering traceability for the navigation and fixed-point detection tasks of the climbing robot, thereby realizing a closed-loop detection system integrating full-field screening, precise positioning, and targeted re-inspection of minute defects on the blade and tower surfaces.

[0031] In a specific implementation, as one example, the application scenarios of wind turbines encompass multiple complex factors, including the special structure of the tower's conical curved surface, the dynamic interference of outdoor wind fields, and the irregular distribution of defects on the tower surface. During the process of scheduling the climbing robot to the target location, problems such as unreasonable path planning, posture deviation during movement, and inability to accurately reach the target location can easily occur. This embodiment utilizes a climbing robot based on SLAM algorithm and curvature adaptive control strategy to plan the optimal walking path, autonomously move to the target location, complete the detection posture lock, and collect multimodal detection data adapted to defect features to achieve fixed-point detection of the target location. The specific implementation steps are as follows: Step S41: Obtain the spatial coordinates of all target sites and the corresponding feature information of suspected defect areas, and perform preliminary sorting of multiple target sites; prioritize grouping target sites in the same tower height range into one group, further sorting within the group according to the continuous arrangement order of clockwise or counterclockwise along the circumference of the tower, and sorting between groups according to the order of progressively advancing from the tower reference position to the top or bottom of the tower, forming a target site detection sequence. Step S42: The climbing robot is in the reference position, and collects environmental data and its own posture data in real time and inputs them into the SLAM algorithm. The algorithm performs data fusion processing, removes abnormal data caused by wind field interference and sensor noise, and constructs a local environmental map of the tower based on the effective data. The map contains the tower surface contour information and obstacle location information. At the same time, the initial positioning of the climbing robot in the map is completed, providing position and environmental basis for path planning. Step S43: Based on the local environment map, extract the curvature parameters at different positions on the tower surface using the contour feature extraction algorithm to determine the curvature change pattern of the path-traversed area; combine the initial position of the climbing robot, the spatial coordinates of the target site, and the curvature change pattern to set path planning constraints, including: the path remains parallel to the tower axis to avoid path deviation leading to increased walking distance; the path avoids obvious obstacles on the tower surface to prevent them from interfering with the climbing robot's movement; the curvature change amplitude of the path-traversed area is controlled within the climbing robot's walking adaptation range to ensure stable posture during walking. Step S44: Based on the constraints and the target site detection sequence, plan the global initial optimal path and determine the overall walking route starting from the reference position, passing through each sorted target site in sequence, and finally returning to the reference position or the next area to be detected. Step S45: Decompose the global initial optimal path into several continuous moving units, control the climbing robot to start from the reference position, and gradually perform the movement operation through the magnetic segmented moving module; during the movement, the climbing robot collects its own posture and movement distance information in real time, the SLAM algorithm continuously updates the environmental map and robot localization based on the newly added environmental observation data, and identifies the position deviation caused by the tower curvature, obstacle distribution, and wind disturbance in real time; based on the updated environmental and local information, the global path is corrected in real time and replanned locally through the curvature adaptive control strategy, and the walking trajectory is dynamically adjusted so that the actual path always conforms to the curvature of the tower surface, avoids new or unforeseen obstacles, and corrects walking deviations, ensuring the real-time optimality and feasibility of the path; Step S46: During the movement, the adsorption force and movement speed of each magnetic segment module are adaptively adjusted according to the real-time planned path. When passing through areas with large curvature of the tower, the adsorption force is increased and the movement speed is reduced to ensure that the robot is in close contact with the surface of the tower and avoid posture deviation. When passing through areas with gentle curvature, the adsorption force is appropriately reduced and the movement speed is increased to improve movement efficiency. When the body contact gap is detected to be excessive or the tilt angle deviation is too large, the movement is stopped immediately and the posture is adjusted. After returning to normal, the robot continues to move. The robot moves stably along the real-time optimized path and arrives at the preset range around each target point in sequence. Step S47: After the climbing robot reaches the preset range around a single target site, it pauses its movement and enters the attitude locking stage; the adsorption angle and force of each magnetic segment module are adjusted to keep the body perpendicular to the surface of the suspected defect area, and the distance between the detection equipment and the detection area is controlled within the preset range to ensure complete coverage of the detection field of view; at the same time, the influence of tower elastic deformation caused by wind load on attitude is compensated by near-field reference data, and the body tilt angle is further fine-tuned to achieve stable attitude locking; Step S48: Under attitude lock, perform multimodal detection data acquisition for the current target site; the detection modes include visual detection, vibration detection, and acoustic detection, corresponding to defect surface features, internal structural features, and material damage features, respectively; during the acquisition process, configure the acquisition parameters of each sensor according to the features of the suspected defect area and the preliminary type determination results: for crack-type defects, increase the acquisition sensitivity of the acoustic sensor to focus on capturing the acoustic signal generated by the crack; for coating peeling, corrosion, and wear-type defects, increase the acquisition resolution of the industrial camera to focus on capturing surface visual features; for possible internal damage, activate the vibration sensor to acquire the vibration characteristics of the defect area; Step S49: After the current target site detection is completed, the attitude lock is released, the climbing robot resumes movement, and continues to repeat local path correction and replanning based on the real-time updated environment and positioning information, and sequentially completes the movement, attitude lock and data acquisition of the remaining target sites until all target site detections are completed.

[0032] In this embodiment, the SLAM algorithm can realize the simultaneous execution of map building, localization and path planning, and the curvature adaptive control strategy can adjust the walking parameters according to the real-time curvature of the tower. By combining the SLAM algorithm with the curvature adaptive control strategy, it is possible to realize the simultaneous execution of environmental perception, path planning and curvature adaptation. At the same time, multiple target sites are sorted in a coordinated manner to plan the optimal serial path, which is adapted to the conical curved surface structure of the tower, the dynamic environment and the needs of multi-target site detection.

[0033] In a specific implementation, as one example, during the multimodal detection data acquisition process, multiple non-defect interferences occur, such as wind field disturbances, unit operating vibrations, illumination changes, and tower elastic deformation. These interferences can directly cover or deviate from the inherent characteristics of defects, leading to distorted defect judgment. Existing interference compensation methods often use a single fixed threshold correction, which is prone to undercompensation or overcompensation, resulting in the masking or distortion of defect features. This embodiment pre-constructs a structural coupling interference factor library for operating conditions, matches the real-time collected wind field operating conditions and unit operating parameters to the corresponding interference entries, obtains the submodal interference weights and compensation coefficients, and then performs interference decoupling and feature compensation on the multimodal detection data dimension by dimension, realizing the removal of non-defect interferences and obtaining compensated feature information that can truly characterize the defect features of the target site. The specific implementation steps are as follows: Step S51: First, construct a database of structural coupling interference factors under different operating conditions through multiple sets of on-site calibration tests. The calibration objects cover different wind speeds, wind directions, unit speeds, power generation, and tower tilt angle deviations. Among them, wind field parameters directly cause wind noise and visual illumination disturbances, unit operating parameters correspond to tower-transmitted vibrations, and tower tilt angle reflects elastic deformation caused by wind loads. All three types of parameters are directly physically related to the detected interference and can be obtained through sensors. During the calibration process, four types of data—sound waves, visual data, vibration data, and tilt angle data—are collected in the defect-free area of ​​the tower. The characteristic intervals, contribution ratios, and correction directions of the interference under each operating condition are extracted to form standardized entries. Step S52: Obtain real-time wind speed and direction data through wind field sensing equipment, obtain real-time rotational speed and power generation data through the unit control cabinet interface, and obtain real-time tower tilt angle data through the tilt sensor carried by the tower climbing robot; Step S53: Using the combination of real-time operating parameters as the search condition, select the calibration entry with the smallest operating condition deviation in the interference factor library, and extract the multi-dimensional interference weight and compensation coefficient from the matching entry. The interference weight is used to characterize the degree of influence of the corresponding interference on a certain detection mode, and the compensation coefficient is used to determine the magnitude and direction of feature correction. Step S54: Perform interference decoupling and feature compensation on the multimodal detection data dimension by dimension based on the compensation coefficient; for acoustic signals, remove the interference components of the corresponding frequency band according to the compensation coefficient corresponding to wind noise, and retain the characteristic frequencies caused by the defect itself; for visual images, adjust the pixel coordinates and gray values ​​in reverse according to the illumination compensation coefficient and the surface distortion coefficient to restore the true contour and texture of the defect; for vibration spectrum, remove the fundamental frequency and harmonic superposition components according to the unit vibration compensation coefficient to highlight the inherent spectral characteristics of the defect area; for tilt angle data, correct the attitude offset according to the tower deformation compensation coefficient to restore the true relative position between the detection equipment and the defect area; each modal compensation process is executed independently. Step S55: Integrate the compensated features of each dimension to form compensated feature information that characterizes the defect features of the target site. It retains only the physical properties of the defect itself and does not contain non-defect interference caused by wind field, unit operation and tower deformation.

[0034] In this embodiment, by using the coupling correlation of operating conditions and structural features and the modal-by-modal dimensional compensation, each modal disturbance can be precisely corrected according to its corresponding source and degree of influence, and the direction and magnitude of compensation are based on evidence. Modal independent decoupling can avoid mutual influence between different disturbances and maintain the integrity of defect features. The operating condition and structural coupling disturbance factor library is built based on field calibration and is adapted to different wind farm environments and unit operating states. By eliminating feature shifts caused by non-defect factors, a real, stable and traceable feature basis can be provided for subsequent defect type determination, defect degree assessment and structural safety analysis, thereby improving the reliability of defect identification and determination results.

[0035] Based on the same inventive concept as the cross-domain guidance and collaborative detection method for wind turbine defects in the foregoing embodiments, this invention also provides a cross-domain guidance and collaborative detection system for wind turbine defects, such as... Figure 2 As shown, the system includes: The data acquisition module is used to acquire full-domain scan data obtained by the UAV after calibration based on a preset cross-domain spatiotemporal synchronization benchmark, which is used to perform full-domain scans on the blades and towers of wind turbine units that are running without stopping. The benchmark acquisition module is used to acquire near-field benchmark data collected by the climbing robot at a preset benchmark position on the wind turbine tower. The feature stitching module is used to stitch spatiotemporal features of the full-domain scan data with near-domain reference data as anchor points to determine the suspected defect area and its spatial location. The detection and scheduling module is used to schedule the climbing robot to move to the target site corresponding to the spatial position for fixed-point detection and obtain multimodal detection data; The interference compensation module is used to combine the real-time wind farm conditions and unit operating parameters to perform interference compensation on the multimodal detection data, and obtain the compensated feature information. The compensated feature information is used to characterize the defect features of the target site.

[0036] The system described above in this invention can effectively realize the cross-domain guided collaborative detection method for wind turbine defects, and the technical effects it can achieve are as described in the above embodiments, which will not be repeated here.

[0037] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A cross-domain guided collaborative detection method for wind turbine defects, characterized in that, include: The full-domain scan data is obtained by performing a full-domain scan of the wind turbine blades and towers without stopping after the UAV is calibrated based on a preset cross-domain spatiotemporal synchronization benchmark. Acquire near-field reference data collected by the climbing robot at a preset reference position on the wind turbine tower; Using the near-field reference data as anchor points, spatiotemporal feature stitching is performed on the full-field scan data to determine the suspected defect area and its spatial location; The climbing robot is directed to move to the target site corresponding to the spatial position to perform fixed-point detection and obtain multimodal detection data; By combining the real-time wind farm conditions and unit operating parameters, interference compensation is performed on the multimodal detection data to obtain compensated feature information, which is used to characterize the defect features of the target site.

2. The cross-domain guided collaborative detection method for wind turbine defects according to claim 1, characterized in that, The method for establishing the preset cross-domain spatiotemporal synchronization benchmark includes: After the climbing robot completes attitude locking at the preset reference position of the tower, it collects the spatial coordinates of the preset reference position of the tower and constructs a spatial reference coordinate system. A global timing synchronization trigger signal is generated based on the spatial reference coordinate system, and the global timing synchronization trigger signal and the spatial reference coordinate system are sent to the UAV to complete the unification of spatial coordinates and timing alignment of acquisition between the UAV and the climbing robot.

3. The cross-domain guided collaborative detection method for wind turbine defects according to claim 1, characterized in that, Using the near-field reference data as anchor points, spatiotemporal feature stitching is performed on the full-field scan data to determine the suspected defect area and its spatial location, including: Using the spatial coordinates and temporal signals in the near-field reference data as indexes, the spatiotemporal features of the blade dynamic sequence image and the tower full-field image in the full-field scan data are spatiotemporally stitched together to eliminate image misalignment caused by blade rotation and UAV flight attitude fluctuations, and to generate a continuous feature map of the entire surface of the blade and the tower. Pixel-level feature extraction is performed on the continuous feature map to identify and lock at least one suspected defect region among cracks, coating peeling, corrosion, and wear, and the contour, grayscale, texture multi-dimensional feature vectors and their spatial positions of each suspected defect region are extracted.

4. The cross-domain guided collaborative detection method for wind turbine defects according to claim 1, characterized in that, By combining real-time acquired wind farm operating conditions and unit operating parameters, interference compensation is performed on the multimodal detection data, including: The wind farm operating conditions and the unit operating parameters acquired in real time are matched with a pre-built operating condition structure coupling interference factor library to obtain the interference weight and compensation coefficient of the detection data in each dimension under the current operating condition. Based on the compensation coefficient, the multimodal detection data is subjected to interference decoupling and feature compensation in a dimensional manner to eliminate feature shifts caused by non-defect factors.

5. The cross-domain guided collaborative detection method for wind turbine defects according to claim 2, characterized in that, Controlling the climbing robot to lock its attitude at a preset reference position on the tower includes: The robot is controlled to walk to the preset reference position of the tower via a magnetic segmented moving module, and adaptively fits the conical curved surface of the tower to complete the attitude lock.

6. The cross-domain guided collaborative detection method for wind turbine defects according to claim 1 or 2, characterized in that, The UAV is calibrated based on the preset cross-domain spatiotemporal synchronization benchmark, including: The UAV uses the spatial reference coordinate system in the preset cross-domain spatiotemporal synchronization reference as an anchor point to complete the global calibration of its own flight attitude, so that the spatial coordinates collected by the UAV are mapped and bound to the spatial reference coordinate system.

7. The cross-domain guided collaborative detection method for wind turbine defects according to claim 1, characterized in that, The process of directing the climbing robot to move to the target site corresponding to the spatial location for fixed-point detection includes: The tube-climbing robot plans the optimal walking path from its current position to the target site based on the SLAM algorithm and curvature adaptive control strategy, and autonomously moves to the target site to complete the detection posture lock.

8. The cross-domain guided collaborative detection method for wind turbine defects according to claim 4, characterized in that, Based on the compensation coefficient, the multimodal detection data is subjected to dimensional interference decoupling and feature compensation, including at least one of the following: Based on the characteristics of sound wave signals, compensate for frequency interference caused by wind noise; Based on visual image characteristics, compensate for distortions caused by changes in lighting and the curved surface of the tower; Based on the vibration spectrum characteristics, compensate for the spectral superposition interference caused by the vibration of the unit itself during operation; For tilt angle data, compensate for the tower's elastic deformation interference caused by wind load.

9. The cross-domain guided collaborative detection method for wind turbine defects according to claim 1, characterized in that, The near-field reference data includes the three-dimensional coordinates, standard tilt angle, reference vibration characteristics, and standard visual image of the preset reference position of the tower.

10. A cross-domain guided collaborative detection system for wind turbine defects, characterized in that, include: The data acquisition module is used to acquire full-domain scan data obtained by the UAV after calibration based on a preset cross-domain spatiotemporal synchronization benchmark, which is used to perform full-domain scans on the blades and towers of wind turbine units that are running without stopping. The benchmark acquisition module is used to acquire near-field benchmark data collected by the climbing robot at a preset benchmark position on the wind turbine tower. The feature stitching module is used to stitch spatiotemporal features of the full-domain scan data with the near-domain reference data as anchor points to determine the suspected defect area and its spatial location. The detection and scheduling module is used to schedule the climbing robot to move to the target site corresponding to the spatial position for fixed-point detection and to obtain multimodal detection data. The interference compensation module is used to combine the real-time acquired wind farm conditions and unit operating parameters to perform interference compensation on the multimodal detection data and obtain compensated feature information. The compensated feature information is used to characterize the defect features of the target site.

Images