Fan blade unmanned aerial vehicle inspection defect intelligent identification system

Abstract

The application discloses a fan blade unmanned aerial vehicle inspection defect intelligent identification system, comprising: a route planning and autonomous navigation module, a multi-source data acquisition module, a data preprocessing and feature enhancement module, a multi-modal fusion detection module, a defect comprehensive analysis decision module and a data interaction and visualization module; the defect comprehensive analysis decision module classifies and grades defects according to received comprehensive defect detection results, and calculates the minimum included angle between the blade and the tower cylinder according to blade spatial position data to evaluate the tower scanning risk. The application generates a priority-ordered operation and maintenance decision suggestion based on a maintenance urgency index through a built-in blade structure importance zoning model in the defect comprehensive analysis decision module and in combination with the tower scanning risk evaluation result; complete decision basis from defect discovery to maintenance decision is provided for operation and maintenance personnel, and the scientificity and efficiency of operation and maintenance management are significantly improved.

Description

Technical Field

[0001] This application belongs to the field of wind power equipment operation and maintenance technology, specifically involving an intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection. Background Technology

[0002] With the deepening of the global energy transition, the wind power industry has continued its rapid development, and my country's installed wind power capacity has ranked first in the world for many consecutive years. As the core component for capturing wind energy, wind turbine blades operate in harsh environments such as high salt spray, high humidity, strong ultraviolet radiation, and alternating loads, making them highly susceptible to various surface and internal defects such as cracks, corrosion, flaking, pinholes, and lightning strike damage. If blade damage is not detected and addressed in a timely manner, it can lead to serious accidents such as blade breakage, causing enormous economic losses and safety risks.

[0003] Currently, defect detection of wind turbine blades mainly relies on traditional manual inspection methods, including telescope observation, rope climbing observation, and inspection using high-altitude work platforms. These traditional methods have many inherent drawbacks: First, they are inefficient, with the inspection of a single wind turbine often taking several hours or even longer; second, they pose high safety risks, as inspectors must work at high altitudes or in adverse weather conditions, making personal safety difficult to guarantee; third, the inspection is highly subjective, with results heavily dependent on the inspectors' experience and visual judgment, leading to a high rate of missed detections; and fourth, traditional inspections usually require the wind turbine to be shut down, directly resulting in power generation losses.

[0004] In recent years, the rapid development of drone technology has provided a new technological approach for wind turbine blade inspection. Drones equipped with high-definition cameras and other sensors take aerial photographs of the blades, and combined with image processing and machine learning algorithms for defect identification, has become an important means of non-destructive testing of blades. However, existing drone-based blade inspection systems still have the following technical shortcomings:

[0005] Firstly, most existing inspection route planning methods require the wind turbine to be shut down or to obtain precise real-time blade attitude information in advance, making it difficult to achieve high-quality data acquisition under normal wind turbine operation (blades continuously rotating). Although some systems have achieved non-stop inspection, their routes are fixed and lack the ability to adaptively optimize based on historical inspection results.

[0006] Secondly, existing systems mostly use a single sensor (such as a visible light camera or an infrared thermal imager) for data acquisition. Single-modal detection methods have significant limitations: visible light images are insufficient for identifying internal blade defects and early latent damage; infrared thermal imaging lacks sufficient resolution for fine surface cracks; and lidar point cloud data lacks texture information, making it difficult to directly determine the defect type. The detection accuracy and reliability of single-modal detection methods are limited.

[0007] Third, existing intelligent detection algorithms face problems such as weak environmental adaptability and insufficient performance in detecting multiple types of defects in drone aerial photography scenarios. Images acquired by drones are greatly affected by factors such as changes in lighting, shooting angle, and blurring due to leaf motion. In addition, the inter-class imbalance of defect samples (scarcity of samples of minute defects) makes it difficult for existing detection models to simultaneously ensure the recognition accuracy of various types of defects.

[0008] Fourth, existing systems are mostly simple pipelines of "collection-identification", lacking a complete data closed loop from data collection, multimodal fusion detection to defect classification assessment, operation and maintenance decision generation and route iteration optimization, making it difficult to achieve continuous self-improvement of inspection strategies. Summary of the Invention

[0009] This application provides an intelligent defect identification system for unmanned aerial vehicle (UAV) inspection of wind turbine blades, which aims to solve the problems in the prior art such as the need to stop the UAV during inspection, lack of adaptive optimization of flight routes, limited accuracy of single-mode detection, and lack of complete data closed loop.

[0010] A wind turbine blade unmanned aerial vehicle (UAV) inspection defect intelligent identification system includes: a route planning and autonomous navigation module, a multi-source data acquisition module, a data preprocessing and feature enhancement module, a multimodal fusion detection module, a defect comprehensive analysis and decision-making module, and a data interaction and visualization module;

[0011] The route planning and autonomous navigation module generates a hemispherical inspection route with the wind turbine hub center as the center and a preset safety distance as the radius based on the input wind turbine structural parameters, and outputs the planned route data to the multi-source data acquisition module.

[0012] The multi-source data acquisition module is mounted on the UAV platform. Based on the received planned flight path data, it synchronously controls the execution of visible light image acquisition, infrared thermal imaging acquisition, and lidar point cloud acquisition, and outputs the acquired multi-source data to the data preprocessing and feature enhancement module.

[0013] The data preprocessing and feature enhancement module performs noise reduction, contrast enhancement, multi-scale detail enhancement and normalization on the received multi-source data, and outputs the processed visible light image data, infrared thermal imaging data and lidar point cloud data to the multimodal fusion detection module.

[0014] The multimodal fusion detection module performs defect target detection on visible light image data, thermal anomaly detection on infrared thermal imaging data, and geometric deformation detection on lidar point cloud data. Then, it performs spatiotemporal alignment and confidence weighted fusion on the detection results of the three modes through multimodal feature fusion based on attention mechanism. Based on the fused features, it comprehensively determines the existence, type, severity, and spatial location of the defect, and outputs the comprehensive defect detection result to the defect comprehensive analysis and decision module.

[0015] The defect comprehensive analysis and decision-making module classifies and grades defects based on the received comprehensive defect detection results, calculates the minimum angle between the blade and the tower based on the blade spatial position data to assess the tower sweep risk, and then generates operation and maintenance decision suggestions by combining the defect classification results and the tower sweep risk assessment results, and outputs the operation and maintenance decision data to the data interaction and visualization module.

[0016] The data interaction and visualization module stores, visualizes in three dimensions, and generates reports on the received operation and maintenance decision data. It also feeds back the defect distribution data, detection accuracy data, and route execution data of this inspection to the route planning and autonomous navigation module for iterative optimization of subsequent inspection routes.

[0017] Optionally, the route planning and autonomous navigation module includes: a wind turbine parameter input submodule, a hemispherical inspection route generation submodule, a dynamic obstacle avoidance submodule, and a route optimization submodule;

[0018] The wind turbine parameter input submodule receives and stores the wind turbine's structural parameters and geographic information parameters, and transmits them to the hemispherical inspection route generation submodule.

[0019] The hemispherical inspection route generation submodule generates wind turbine structural parameters based on the wind turbine parameter input submodule, establishes a local coordinate system of the wind turbine with the spatial coordinates of the hub center as the origin, evenly distributes several track points on the hemispherical surface, and converts the track point sequence into a route file for output to the UAV flight control system; the submodule also sets the exposure timing parameters for image acquisition based on the rated speed range of the blades, so that the camera shutter trigger time matches the time window when the blades pass through the preset acquisition area;

[0020] The dynamic obstacle avoidance submodule receives multi-source perception data from the UAV's onboard obstacle avoidance sensor in real time. It has a built-in local path replanning unit. When it is determined that there is an obstacle within the preset safe distance of the current planned route, it triggers a local route correction process. After the UAV bypasses the obstacle and returns to the original route, it resumes the inspection task along the original planned route.

[0021] The route optimization submodule receives historical inspection data and defect distribution information from the data interaction and visualization module. It performs statistical analysis on the defect types, spatial locations, and severity levels detected in the historical inspections to generate a heat map of the defect probability distribution on the blade surface. Based on this heat map, it dynamically adjusts the spatial distribution density of the track points in the route planning of subsequent inspections. The optimized route parameters are then updated in the route template library of the hemispherical inspection route generation submodule.

[0022] Optionally, the multi-source data acquisition module is deployed on an unmanned aerial vehicle (UAV) platform and includes: a visible light image acquisition submodule, an infrared thermal imaging acquisition submodule, a lidar point cloud acquisition submodule, an acquisition control submodule, and a data caching submodule;

[0023] The data acquisition and control submodule receives the planned route data output by the route planning and autonomous navigation module. Based on the preset acquisition points, acquisition times, and acquisition parameters in the route, it synchronously controls the visible light image acquisition submodule, infrared thermal imaging acquisition submodule, and lidar point cloud acquisition submodule to perform data acquisition. This submodule has a built-in high-precision clock synchronization unit, which uses the second pulse signal of the Global Navigation Satellite System as a unified clock reference to synchronize the data acquisition time of each sensor. When the UAV arrives at the preset acquisition point, it simultaneously sends trigger pulses to the three sensors.

[0024] The data caching submodule aligns and indexes visible light image frames, infrared thermal imaging frames, and lidar point cloud frames according to a unified time axis, based on the timestamp and spatial pose information provided by the acquisition control submodule. This forms a multimodal data record with the acquisition time as the primary key. Each record contains visible light image data, infrared thermal image data, 3D point cloud data, and corresponding UAV pose metadata for the same spatial region of the blade at the same time. After the inspection task is completed, this submodule outputs the complete multimodal data package to the data preprocessing and feature enhancement module.

[0025] Optionally, the data preprocessing and feature enhancement module includes: an image denoising submodule, a contrast enhancement submodule, a multi-scale detail enhancement submodule, and a data normalization submodule;

[0026] The image denoising submodule receives visible light image data output by the multi-source data acquisition module, uses a non-local mean algorithm to remove background noise in the image, and uses the non-local self-similarity redundancy information in the image to perform a weighted average estimation of each pixel.

[0027] The contrast enhancement submodule uses a limited contrast adaptive histogram equalization algorithm on the denoised image to divide the image into several rectangular regions, perform histogram equalization on each region, and then perform smooth fusion.

[0028] The multi-scale detail enhancement submodule constructs a multi-scale feature pyramid for the contrast-enhanced image, and splices and fuses the detail feature maps at each scale with the original feature map to generate an enhanced feature map containing multi-scale detail information.

[0029] The data normalization submodule performs unified normalization processing on the enhanced visible light image, infrared thermal imaging data, and lidar point cloud data, converting the data of the three modalities into a unified tensor format before outputting them to the multimodal fusion detection module.

[0030] Optionally, the multimodal fusion detection module includes: a visible light target detection submodule, an infrared anomaly detection submodule, a point cloud geometric analysis submodule, a multimodal feature fusion submodule, and a defect comprehensive identification submodule;

[0031] The visible light target detection submodule uses an improved YOLO detection network to detect defect targets in the preprocessed visible light image. It introduces receptive field attention convolution to replace some traditional convolution modules in the backbone network, dynamically adjusts the weights of different receptive fields, and uses an improved loss function with a modulation factor in the detection head to alleviate the problem of imbalance between positive and negative samples.

[0032] The infrared anomaly detection submodule performs a two-level anomaly determination process on the preprocessed infrared thermal imaging data: the first level is global adaptive threshold detection, which statistically analyzes the temperature distribution of the entire infrared image and initially marks the candidate thermal anomaly region with a global anomaly temperature threshold; the second level is local contrast verification, which calculates the difference between the average temperature of the candidate region and the average temperature of the surrounding normal region, and confirms the region as a valid thermal anomaly when the temperature difference exceeds a preset threshold.

[0033] The point cloud geometric analysis submodule performs three-dimensional reconstruction and geometric analysis on the pre-processed lidar point cloud data, registers and compares the blade point cloud data obtained from the current inspection with historical point cloud data or blade design models, and calculates the spatial distance deviation between corresponding points to generate a deformation deviation distribution map.

[0034] The multimodal feature fusion submodule receives the output results of the visible light target detection submodule, the infrared anomaly detection submodule, and the point cloud geometric analysis submodule. Based on the timestamp and spatial pose information when the UAV collects data, it converts the detection results of the three modes into a unified local coordinate system of the wind turbine to establish a spatial correspondence. Then, it uses a cross-modal attention mechanism to adaptively weight and sum the features of each mode to obtain the unified feature tensor after fusion.

[0035] The defect comprehensive identification submodule, based on the fused feature tensor output by the multimodal feature fusion submodule, simultaneously performs binary classification of defect existence, multi-class classification of defect type, and regression of defect severity level through the built-in multi-task classification and regression network, and outputs a comprehensive defect detection result.

[0036] Optionally, the defect comprehensive analysis and decision-making module includes: a defect classification and grading submodule, a tower sweep risk assessment submodule, and an operation and maintenance decision generation submodule;

[0037] The defect classification and grading submodule receives the comprehensive defect detection results output by the multimodal fusion detection module, classifies the defects according to the preset defect classification criteria, and classifies the severity level of the defects according to the size, depth, and location parameters of the defects combined with the blade structure importance partitioning model.

[0038] The tower sweep risk assessment submodule calculates the minimum angle between the blade and the tower based on the blade spatial position data and the blade geometric information provided by the point cloud geometric analysis submodule. The preset safe angle threshold is used as the judgment criterion. When the minimum angle is greater than the safe threshold, it is determined that there is no tower sweep risk. When the minimum angle is between the safe threshold and the warning threshold, a first-level tower sweep risk warning signal is generated. When the minimum angle is less than the warning threshold, a second-level tower sweep risk warning signal is generated.

[0039] The operation and maintenance decision generation submodule calculates the maintenance urgency index of each defect based on the defect classification and grading submodule and the risk assessment results of the tower sweeping risk assessment submodule. Based on the risk priority maintenance decision model, the submodule sorts all detected defects by priority and generates corresponding operation and maintenance decision suggestions.

[0040] Optionally, the data interaction and visualization module includes: a data storage submodule, a 3D visualization submodule, an inspection report generation submodule, and a data feedback submodule;

[0041] The data storage submodule stores the comprehensive diagnostic report, defect detection results, and operation and maintenance decision data output by the defect comprehensive analysis and decision module into the database, establishes an independent defect file database table for each blade, and records all defect information detected in each inspection and subsequent maintenance records in chronological order to form a complete traceability chain.

[0042] The three-dimensional visualization submodule is based on the three-dimensional model of the wind turbine. It spatially positions and marks the defect detection results on the three-dimensional model, presents the type, location and severity of the defects in a visual way, and supports time-series comparison view to overlay and display defect data from multiple inspections on the same three-dimensional model.

[0043] The inspection report generation submodule automatically generates a graphic inspection report based on defect detection results and operation and maintenance decision data;

[0044] The data feedback submodule feeds back the defect distribution statistics, detection accuracy evaluation data, and route execution data of this inspection to the route optimization submodule of the route planning and autonomous navigation module. The route optimization submodule then performs trackpoint density redistribution, gimbal parameter adjustment, and route smoothing optimization based on the feedback data, thus realizing a data closed loop of collection-detection-decision-optimization.

[0045] Optionally, the visible light image acquisition submodule is equipped with a high-resolution three-axis stabilized gimbal camera, integrating a wide-angle camera and a zoom camera dual-optical-path system. The wide-angle camera is used for wide-area shooting of the entire blade, and the zoom camera has an optical zoom capability of no less than 20x for long-distance close-up acquisition of local areas of the blade. The gimbal system has a three-axis mechanical stabilization function.

[0046] The infrared thermal imaging acquisition submodule is equipped with an uncooled infrared thermal imager, which is used to simultaneously acquire thermal imaging data of the blade surface and capture temperature anomaly areas caused by defects.

[0047] The lidar point cloud acquisition submodule is equipped with a lidar scanner, which is used to acquire three-dimensional point cloud data of the wind turbine blades and the surrounding environment, and obtain the geometric contour information and spatial attitude information of the blades.

[0048] Optionally, the multi-scale detail enhancement submodule performs multi-level average pooling operations on the input image to construct a feature pyramid representation with four scale levels from fine to coarse. For each pooling level, the feature map of the previous level is upsampled to the original scale and then subjected to element-wise difference operation with the feature map of the original scale to obtain the detail feature map at that scale. After completing the difference calculation at each scale, the detail feature maps at all scales are concatenated and fused with the original feature map through the channel dimension to generate an enhanced feature map containing multi-scale detail information.

[0049] Optionally, the multimodal feature fusion submodule constructs a three-input feature fusion network, extracting feature vectors from visible light detection results, infrared detection results, and point cloud geometric analysis results through three parallel feature encoding branches; a cross-modal attention mechanism is introduced, using features from the visible light modality as queries and features from the infrared and point cloud modalities as keys and values, respectively. The correlation weights between the features of each modality are calculated through scaled dot product attention, and the features of each modality are adaptively weighted and summed according to the attention weights to obtain the fused unified feature tensor.

[0050] Compared with the prior art, this application has at least the following beneficial effects:

[0051] This application utilizes a hemispherical inspection route centered on the wind turbine hub, coupled with exposure timing parameters set according to the blade's rated rotational speed range. This allows the UAV to complete high-quality data acquisition while the wind turbine is running normally and the blades are continuously rotating, without requiring the turbine to be shut down or yawed. Compared to existing inspection methods that require shutdown, this application significantly reduces power generation losses caused by inspections, improving the economic efficiency of wind farm operation.

[0052] This application simultaneously acquires data from three modalities—visible light images, infrared thermal imaging, and lidar point clouds—through a multi-source data acquisition module. The data is then fused using a multi-modal fusion detection module with spatiotemporal alignment and a cross-modal attention mechanism, resulting in confidence-weighted fusion. Compared to existing single-modal detection methods, the three modalities complement each other: visible light images provide surface texture details, infrared thermal imaging captures temperature anomalies, and lidar point clouds reflect geometric deformations. This comprehensive approach significantly improves the overall comprehensiveness and accuracy of defect detection, effectively solving the problem that single-modal detection struggles to address various types of defects.

[0053] This application utilizes the data feedback submodule of the data interaction and visualization module to feed back the defect distribution data, detection accuracy evaluation data, and route execution data from this inspection to the route optimization submodule of the route planning and autonomous navigation module. Based on this, the route optimization submodule performs waypoint density reallocation, gimbal parameter adjustment, and route smoothing optimization, achieving a complete data closed loop of "collection-detection-decision-optimization." With the increase in the number of inspections, the rationality of route planning and the effectiveness of image acquisition gradually improve, realizing continuous self-improvement of the data-driven intelligent operation and maintenance strategy.

[0054] This application utilizes a built-in blade structure importance zoning model within its defect comprehensive analysis and decision-making module. Based on the importance of the structural zoning where the defect is located and the defect size parameters, it comprehensively classifies the severity level and, combined with tower sweep risk assessment results, generates priority-ranked maintenance decision recommendations based on a maintenance urgency index. Compared to existing systems that only output detection results, this application provides maintenance personnel with a complete decision-making basis from defect discovery to maintenance decisions, significantly improving the scientific nature and efficiency of maintenance management. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the module connection of an intelligent identification system for defects in wind turbine blades provided in one embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0057] This application provides an intelligent defect identification system for unmanned aerial vehicle (UAV) inspection of wind turbine blades, such as... Figure 1 As shown, it includes: route planning and autonomous navigation module, multi-source data acquisition module, data preprocessing and feature enhancement module, multimodal fusion detection module, defect comprehensive analysis and decision-making module, and data interaction and visualization module;

[0058] The route planning and autonomous navigation module includes: a wind turbine parameter input submodule, a hemispherical inspection route generation submodule, a dynamic obstacle avoidance submodule, and a route optimization submodule.

[0059] The wind turbine parameter input submodule receives and stores structural and geographic information parameters of the wind turbine, including the latitude and longitude coordinates of the turbine center point, the altitude of the hub center, the total blade length, the number of blades, the rated speed range of the blades, and the yaw angle of the turbine. This submodule provides a human-machine interface, allowing maintenance personnel to input or import wind farm survey data through ground station software. It also supports the automatic acquisition of real-time wind turbine operating parameters from the wind farm's SCADA system, providing fundamental data support for flight path planning.

[0060] The hemispherical inspection route generation submodule generates a hemispherical inspection route with the center of the wind turbine hub as the center and a preset safety distance as the radius, based on the wind turbine structural parameters provided by the wind turbine parameter input submodule.

[0061] The specific method for generating the flight path includes: firstly, establishing a local coordinate system for the wind turbine with the spatial coordinates of the hub center as the origin;

[0062] Then, based on the blade length and the preset safe observation distance (preferably 15-25 meters), several track points are evenly distributed on the hemispherical surface. The spatial interval between adjacent track points is constrained to ensure that adjacent acquired images have an overlap rate of not less than 30%. The track points include five-dimensional attitude information such as longitude, latitude, altitude, and gimbal camera pitch and yaw angles.

[0063] Finally, the above waypoint sequence is converted into a KML or MissionPlanner compatible flight path file and output to the UAV flight control system;

[0064] This hemispherical flight path enables the UAV to maintain a standard observation distance and angle with various parts of the blade surface during the inspection process, without the need to obtain the real-time attitude information of the wind turbine blades in advance, or to perform shutdown or yaw operations on the wind turbine, thus having good versatility and inspection efficiency.

[0065] In addition, when the wind turbine is running and the blades are continuously rotating, this submodule sets the exposure timing parameters for image acquisition according to the rated speed range of the blades, so that the camera shutter trigger time matches the time window when the blades pass through the preset acquisition area, ensuring the clarity of the acquired images of the blades in motion.

[0066] The dynamic obstacle avoidance submodule receives multi-source perception data from the UAV’s onboard obstacle avoidance sensors in real time. The onboard obstacle avoidance sensors include, but are not limited to, a front-facing binocular vision sensor, an infrared perception sensor, and a lidar ranging sensor.

[0067] This submodule has a built-in local route replanning unit. When it is determined from sensor data that there are obstacles (including static obstacles such as nearby wind turbine towers, overhead lines, and substation buildings, as well as dynamic obstacles such as birds that intrude into the inspection airspace during flight) within the preset safe distance of the current planned route, the local route correction process is immediately triggered.

[0068] The local flight path correction process specifically includes: starting from the UAV's current position and ending at the next unobstructed track point in the original flight path, a new local avoidance path is planned around the obstacle using an improved fast random search tree algorithm; after the UAV bypasses the obstacle and returns to the original flight path, it resumes the inspection task along the originally planned route. This submodule also monitors the UAV's battery status and communication link quality. When the battery level falls below a preset return-to-home threshold or the communication link is interrupted for more than a set time, it autonomously generates an emergency return-to-home command and outputs it to the flight control system.

[0069] The route optimization submodule receives historical inspection data and defect distribution information from the data interaction and visualization module, and iteratively optimizes the route planning strategy for subsequent inspections accordingly.

[0070] Specifically, this submodule performs statistical analysis on the types, spatial locations, and severity levels of defects detected in historical inspections, and generates a heat map of the probability distribution of defects on the blade surface.

[0071] Based on the heat map of defect probability distribution, the spatial distribution density of track points is dynamically adjusted in the route planning of subsequent inspections. The density of track points is increased and the hovering collection time is extended in areas with high defect incidence (such as the leading edge of the blade, the tip of the blade, and the windward side of the middle section of the blade). The density of track points is appropriately reduced in areas with no defect records for a long time to reduce redundant flight mileage.

[0072] Simultaneously, this submodule also fine-tunes and optimizes the shooting angle parameters of the gimbal camera based on the image quality evaluation results of historical inspection data, in order to improve the imaging quality of defective areas in subsequent inspections. After the optimized flight path parameters are confirmed by the ground station operators, they are updated to the flight path template library of the hemispherical inspection flight path generation submodule for use in subsequent inspection tasks.

[0073] The data flow of the route planning and autonomous navigation module is as follows: the wind turbine parameter input submodule transmits the wind turbine structural parameters to the hemispherical inspection route generation submodule, which generates the initial planned route and outputs it to the UAV flight control system for execution.

[0074] The multi-source data acquisition module is deployed on an unmanned aerial vehicle (UAV) platform and includes: a visible light image acquisition submodule, an infrared thermal imaging acquisition submodule, a lidar point cloud acquisition submodule, an acquisition control submodule, and a data caching submodule.

[0075] The visible light image acquisition submodule is equipped with a high-resolution three-axis stabilized gimbal camera, used to acquire visible light image data of the wind turbine blade surface during the UAV's flight along the planned route. This submodule integrates a dual-optical-path system of a wide-angle camera and a zoom camera. The wide-angle camera uses a large-size CMOS image sensor with an effective pixel count of no less than 40 million, used for wide-area shooting of the entire blade; the zoom camera has an optical zoom capability of no less than 20x, used for long-distance close-up acquisition of details in local areas of the blade.

[0076] The gimbal system features three-axis mechanical stabilization, with angular vibration suppression accuracy of no less than ±0.01° across the pitch, roll, and yaw axes, ensuring clear and stable images even in complex airflow environments. This submodule incorporates an autofocus and exposure control unit, capable of adjusting focus and exposure parameters in real time based on ambient lighting conditions and shooting distance. It also supports multiple exposure modes, including shutter priority, aperture priority, and full manual, to adapt to different weather conditions and shooting needs at different times of day.

[0077] The infrared thermal imaging acquisition submodule is equipped with an uncooled infrared thermal imager for simultaneously acquiring thermal imaging data of the blade surface and capturing temperature anomalies caused by defects. This submodule's infrared detector has a resolution of no less than 640×512 pixels, a thermal sensitivity better than 50mK (i.e., 0.05℃), and a temperature measurement range covering -20℃ to +150℃. The thermal imager is equipped with a fixed-focus or continuously zoom infrared lens, and its field of view is matched with that of a visible light zoom camera to ensure that the infrared and visible light images have a matched field of view.

[0078] This submodule supports switching between multiple pseudo-color modes (including iron red, white-hot, black-hot, rainbow, etc.), making it easier for maintenance personnel to identify temperature differences from different visual angles. It also supports isotherm analysis and point, line, and surface area temperature measurement functions, enabling quantitative measurement of the highest, lowest, and average temperatures of any area on the blade surface. For localized temperature rise anomalies caused by lightning strikes, frictional overheating, or internal structural damage, this submodule can capture and locate them with high temperature and spatial resolution.

[0079] The lidar point cloud acquisition submodule is equipped with a lidar scanner to collect three-dimensional point cloud data of the wind turbine blades and their surrounding environment, obtaining the geometric contour information and spatial attitude information of the blades. The lidar in this submodule has a scanning frequency of no less than 300,000 points per second, a ranging accuracy better than centimeter level, and an effective ranging range of no less than 200 meters.

[0080] By emitting laser pulses and receiving echo signals, this submodule can acquire the position coordinates (X, Y, Z) of each point on the blade surface relative to the UAV, forming dense three-dimensional point cloud data. This point cloud data is used not only for the subsequent construction of the blade's three-dimensional model and the precise measurement of defect geometric parameters (such as crack length, depth, and spalling area), but also for real-time perception of the relative distance and attitude relationship between the UAV and the blade surface, providing spatial reference information for flight control. Furthermore, when the wind turbine is in operation, this submodule can capture the dynamic point cloud sequence of rotating blades, providing basic data for monitoring blade flapping, oscillation, and other motion states.

[0081] The acquisition control submodule receives the planned route data output by the route planning and autonomous navigation module, and synchronously controls the visible light image acquisition submodule, infrared thermal imaging acquisition submodule and lidar point cloud acquisition submodule to perform data acquisition according to the preset acquisition points, acquisition time and acquisition parameters in the route.

[0082] This submodule incorporates a high-precision clock synchronization unit, which uses the pulse-per-second (PPS) signal from the Global Navigation Satellite System (GNSS) as a unified clock reference to synchronize the data acquisition times of each sensor at the nanosecond level. The clock synchronization unit generates a duration synchronization trigger signal, which simultaneously sends trigger pulses to the visible light camera, infrared thermal imager, and lidar when the UAV arrives at the preset acquisition point, ensuring that the three sensors acquire data from the same spatial area of ​​the blade at the same time.

[0083] This submodule simultaneously receives attitude and position data from the UAV's onboard inertial navigation unit (IMU) and real-time dynamic differential positioning (RTK) module. It encapsulates the UAV's three-dimensional spatial coordinates, three-axis attitude angles (pitch, roll, and yaw) and the timestamp of the acquisition time into the metadata of each frame of acquired data, providing a precise pose reference for the spatial registration and temporal alignment of subsequent multimodal data.

[0084] The acquisition and control submodule also calculates the dynamic triggering delay of the camera shutter based on the blade rotation speed parameters provided by the flight path planning module, so that the image acquisition time is precisely matched with the time window when the blade rotates to the best observation angle, thereby obtaining high-quality blade images under the condition that the wind turbine does not stop running.

[0085] The data caching submodule is used to timestamp-align and locally cache the collected multi-source data, forming a multimodal data packet containing visible light images, infrared thermal imaging, and lidar point clouds. This submodule is equipped with a large-capacity solid-state storage medium with a storage capacity of no less than 256GB, which can meet the local storage requirements of all multi-source data generated in a single complete inspection task (approximately 20 minutes).

[0086] The data caching submodule aligns and indexes visible light image frames, infrared thermal imaging frames, and lidar point cloud frames according to a unified timeline, based on the timestamps and spatial pose information provided by the acquisition control submodule. This forms a multimodal data record with the acquisition time as the primary key. Each record contains visible light image data, infrared thermal image data, 3D point cloud data, and corresponding UAV pose metadata (latitude and longitude, altitude, three-axis attitude angles, acquisition timestamp, etc.) for the same spatial region of the blade at the same time.

[0087] During the inspection task, the data caching submodule monitors the remaining capacity of the storage medium in real time. When the remaining capacity falls below a preset threshold, it sends a storage warning signal to the acquisition control submodule, which then dynamically adjusts the acquisition frequency or resolution based on task priority. After the inspection task is completed, the data caching submodule outputs the multimodal data packets completely to the data preprocessing and feature enhancement module via a high-speed data link between the UAV and the ground station (including but not limited to 4G / 5G cellular networks, microwave point-to-point communication, or physical transfer via large-capacity memory cards).

[0088] The data preprocessing and feature enhancement module includes: an image denoising submodule, a contrast enhancement submodule, a multi-scale detail enhancement submodule, and a data normalization submodule;

[0089] The image denoising submodule receives visible light image data output by the multi-source data acquisition module and uses the Non-Local Means (NLM) algorithm to remove background noise from the image while smoothing defect edges. The NLM algorithm utilizes the non-local self-similarity redundancy information present in the image. For the current pixel to be processed, a search window (preferably 21×21 pixels) is set in the image centered on it. All pixels are traversed within this search window, and neighboring windows (preferably 7×7 pixels) are taken centered on the current pixel and the traversed candidate pixels, respectively. The Gaussian weighted Euclidean distance between the two neighboring windows is calculated to measure the structural similarity between the two pixels.

[0090] The weight value of each candidate pixel is calculated based on similarity; the higher the similarity, the greater the weight. Finally, the gray values ​​of all pixels within the search window are weighted and averaged to obtain the denoised estimate of the current pixel. The smoothing coefficient h in the weight calculation is determined using an adaptive strategy: first, the noise standard deviation σ is estimated through the spatial gradient distribution of the image; then, h = k × σ is set, where k is an empirical scaling factor with a value ranging from 0.8 to 1.2, which can be dynamically adjusted according to the actual noise level of the image.

[0091] The search window size and the neighborhood window size are also adaptively adjusted according to the image resolution and noise level. For high-resolution images (above 40 million pixels), a larger search window (31×31 pixels) is used to make full use of the image's redundant information, while for low-resolution images, the window size is reduced accordingly to ensure computational efficiency.

[0092] Compared to traditional local denoising methods such as Gaussian filtering and median filtering, the NLM algorithm can effectively suppress background noise while preserving the edge structure and texture details of defects such as microcracks and sand holes on the blade surface to the greatest extent, avoiding the loss of defect features due to excessive smoothing.

[0093] The contrast enhancement submodule performs adaptive contrast enhancement processing on the denoised image, employing the Contrast-Limited Adaptive Histogram Equalization (CLAHE) algorithm to improve the overall image quality. Unlike traditional histogram equalization methods that uniformly adjust the grayscale distribution across the entire image, the CLAHE algorithm first divides the image into several equally sized rectangular regions (called "patches," preferably 8×8 pixels), calculates the grayscale histogram for each patch, and then performs histogram equalization transformation.

[0094] Before histogram equalization, a contrast limit threshold (preferably 2.0–4.0) is set, and the pixel count at each gray level is cropped below this threshold. The cropped pixels are then evenly redistributed to all gray levels, effectively suppressing the noise over-amplification problem common in traditional histogram equalization methods. After each patch completes its independent equalization transformation, bilinear interpolation is used to smoothly fuse the transformation results between adjacent patches, eliminating possible artificial boundary effects between patches.

[0095] For the problem of insufficient local contrast in visible light images of wind turbine blades caused by uneven lighting, shadow occlusion, or shooting in foggy weather, the CLAHE algorithm can adaptively enhance different regions of the image, making the grayscale difference between defective and normal regions on the blade surface more significant, thus creating good input conditions for subsequent defect feature extraction and recognition.

[0096] The multi-scale detail enhancement submodule performs multi-scale detail enhancement processing on the contrast-enhanced image to further highlight the visual salience of minute defects on the blade surface. This submodule first performs multi-level average pooling on the input image, constructing a feature pyramid representation with four scale levels (original scale, 1 / 2 scale, 1 / 4 scale, and 1 / 8 scale) from fine to coarse. As the pooling levels deepen, the feature responses within the local receptive field gradually smooth out, and high-frequency detail information in the image is gradually filtered out.

[0097] For each pooling level, the feature map of the previous level is upsampled to restore the original scale, and then element-wise differencing is performed with the feature map at the original scale to obtain the detail feature map at that scale. In the difference result, the positive value region represents the region in the original image with a gray value higher than the smooth background (such as raised defects or stains), and the negative value region represents the region with a gray value lower than the smooth background (such as cracks, depressions or corrosion pits). The absolute value reflects the contrast intensity between the defect and the background.

[0098] The aforementioned difference operations can explicitly extract local detail changes at different scales. Small-scale difference is sensitive to fine structures such as microcracks, while large-scale difference has a better response to larger defects such as detached areas. After completing the difference calculations at each scale, the detail feature maps at all scales are concatenated and fused with the original feature map through the channel dimension to generate an enhanced feature map containing rich multi-scale detail information. This enhanced feature map not only preserves the overall structural information of the original image but also fully integrates the detail difference responses at various scales. This results in a significant enhancement in the response of minute defect features that are easily overlooked in conventional image processing, greatly improving the ability of subsequent detection models to perceive minute defects.

[0099] The data normalization submodule performs unified normalization processing on the enhanced visible light image, infrared thermal imaging data, and lidar point cloud data to meet the input format requirements of the multimodal fusion detection module. For visible light image data, the RGB three-channel values ​​of each pixel are linearly mapped from the original integer range of 0 to 255 to the floating-point range of 0.0 to 1.0, and further standardized according to the input statistical characteristics of the pre-trained detection model (subtracting the training set mean and dividing by the standard deviation).

[0100] For infrared thermal imaging data, the original temperature values ​​(unit: degrees Celsius) of each pixel are min-max normalized to linearly map the temperature values ​​of the entire image to the range of 0.0 to 1.0. At the same time, the original temperature value corresponding to each pixel is retained as metadata for subsequent anomaly detection based on the actual temperature threshold.

[0101] For LiDAR point cloud data, the three-dimensional spatial coordinates (X, Y, Z) of each point are translated and scaled for normalization. A local coordinate system is established with the center of the wind turbine hub as the origin, and all point cloud coordinates are transformed to this local coordinate system. Then, the values ​​on each coordinate axis are standardized to eliminate the coordinate offset caused by different flight positions and altitudes of the UAV. After normalization, the data of the three modalities are converted into a unified tensor format (H×W×3 tensor for visible light images, H×W×1 tensor for infrared images, and N×3 tensor for point clouds), and each is attached with its own spatial location label and timestamp label. These are then output to the multimodal fusion detection module through the data packaging interface.

[0102] The multimodal fusion detection module includes: a visible light target detection submodule, an infrared anomaly detection submodule, a point cloud geometric analysis submodule, a multimodal feature fusion submodule, and a defect comprehensive identification submodule;

[0103] The visible light target detection submodule employs an improved YOLO detection network to detect defect targets in the preprocessed visible light image. This submodule uses YOLOv8n or YOLOv11n as the baseline model and introduces receptive-field attention convolution (RFAConv) into the backbone network to replace some of the traditional 3×3 convolution modules;

[0104] The calculation process of RFAConv consists of three steps: First, average pooling is performed on the input feature map to obtain global feature information. After 1×1 convolution to promote feature interaction, an attention feature map is generated through the Softmax activation function. At the same time, 3×3 convolution is performed on the input feature map to obtain spatial features with a larger receptive field. After batch normalization and ReLU activation, the receptive field spatial feature map is obtained. Finally, the attention feature map and the receptive field spatial feature map are multiplied element-wise. After shape adjustment, the weighted final feature is output through a 3×3 convolution with a stride of 3.

[0105] Through the above mechanism, RFAConv can dynamically adjust the weights of different receptive fields, enabling the network to enhance its perception of global context information while maintaining its ability to capture local details. In the detection head part, this submodule uses an improved loss function to replace the original CIoU loss;

[0106] Specifically, the concept of FocalLoss is introduced into the bounding box regression loss to construct Focal-CIoU or Focal-EIoU loss functions. This loss function dynamically adjusts the loss contribution of different samples by introducing a modulation factor (1−pt)γ. For easily classifiable samples (pt close to 1), the modulation factor approaches 0, and the loss contribution is significantly suppressed; for difficult-to-classify samples (pt far from 1), the modulation factor approaches 1, and the loss contribution is preserved and strengthened. This mechanism effectively alleviates the problem of scarce samples of minute defects and severe imbalance between positive and negative samples in the wind turbine blade defect dataset, allowing the model to focus more on minute defects that are difficult to detect, such as cracks and pinholes.

[0107] The output of this submodule is a set of visible light defect detection results. Each detection result includes a defect category label (such as crack, corrosion, peeling, sand hole, dirt, etc.), the coordinates of the upper left and lower right corners of the horizontal rectangular bounding box (x1, y1, x2, y2), and the corresponding confidence score (a floating-point number between 0 and 1).

[0108] The infrared anomaly detection submodule employs a method combining temperature threshold analysis and deep learning to detect abnormal regions in the preprocessed infrared thermal imaging data. This submodule first analyzes the temperature values ​​of the input infrared thermal image, converting the grayscale values ​​of each pixel into corresponding temperature values ​​(unit: degrees Celsius).

[0109] Then, a two-level anomaly detection process is executed: the first level is global adaptive threshold detection, which statistically analyzes the temperature distribution of the entire infrared image, calculates the temperature mean μ and standard deviation σ, and uses μ+k×σ (k is a preset coefficient with a value range of 2.5 to 3.5) as the global anomaly temperature threshold. Pixel areas with temperatures exceeding this threshold are initially marked as thermal anomaly candidate areas.

[0110] The second level is local contrast verification. For each candidate thermal anomaly area, the difference between the average temperature of the area and the average temperature of the surrounding normal area is calculated. When the temperature difference exceeds the preset local temperature difference threshold (preferably 2 to 5°C, which can be dynamically adjusted according to the ambient temperature and wind speed), the area is confirmed as an effective thermal anomaly area.

[0111] For localized temperature rise anomalies caused by lightning strikes, frictional overheating, or internal structural damage, this method can capture and locate them with high temperature resolution. Furthermore, this submodule incorporates a lightweight convolutional neural network-based thermal anomaly classifier to further classify confirmed thermal anomaly regions, distinguishing between genuine thermal anomalies caused by surface defects and false thermal anomalies caused by direct sunlight, environmental reflection, etc., effectively reducing the false alarm rate. The output of this submodule is a thermal anomaly detection result set, including the center coordinates of each thermal anomaly region, the region's extent (represented by a rectangle), the highest temperature, the average temperature, and the anomaly type label.

[0112] The point cloud geometry analysis submodule performs 3D reconstruction and geometric analysis on the preprocessed lidar point cloud data to extract the geometric deformation features of the blade surface. This submodule first performs outlier removal and downsampling on the input point cloud data, reducing the point cloud density to a level suitable for subsequent processing while maintaining the overall geometric contour of the blade.

[0113] Subsequently, the submodule performs the following three geometric analyses: First, the blade's three-dimensional contour is reconstructed by using an implicit surface reconstruction method based on radial basis functions or a Poisson surface reconstruction algorithm to reconstruct the discrete point cloud into a continuous three-dimensional surface mesh model of the blade, thereby obtaining the complete geometric shape of the blade.

[0114] Second, geometric deformation detection involves registering and comparing the blade point cloud data acquired during the current inspection with historical point cloud data or blade design CAD models of the same wind turbine. Registration employs the Iterative Closest Point (ICP) algorithm to align the two point clouds to the same coordinate system. After alignment, the spatial distance deviation between corresponding points is calculated to generate a deformation deviation distribution map for the entire blade. When the deformation deviation in a certain area exceeds a preset threshold (preferably 5–10 mm, set according to the allowable tolerances of different blade parts), that area is marked as a geometric anomaly.

[0115] Third, blade attitude analysis involves fitting the blade's spatial central axis and torsion angle based on point cloud data to determine if the blade exhibits flapping, oscillation, or torsional deformation exceeding permissible limits. This submodule outputs a geometric anomaly detection result set, including the spatial location (3D coordinate range), deformation type (protrusion, depression, twist, etc.), and deformation value (unit: millimeters) for each geometric anomaly region.

[0116] The multimodal feature fusion submodule receives the output results from the visible light target detection submodule, the infrared anomaly detection submodule, and the point cloud geometric analysis submodule. It uses an attention-based multimodal feature fusion method to perform spatiotemporal alignment and confidence-weighted fusion of the detection results of different modalities, generating a unified fusion feature tensor.

[0117] This submodule first performs a spatiotemporal alignment operation: based on the timestamp and spatial pose information of the data collected by the UAV, it transforms the two-dimensional defect locations in the visible light image, the two-dimensional thermal anomaly locations in the infrared thermal image, and the three-dimensional geometric anomaly locations in the point cloud data to a unified local coordinate system of the wind turbine, establishing a spatial correspondence between the three modal detection results. After the spatiotemporal alignment is completed, this submodule constructs a three-input feature fusion network. This network first extracts feature vectors from the visible light detection results, infrared detection results, and point cloud geometric analysis results through three parallel feature encoding branches, respectively;

[0118] Then, a cross-modal attention mechanism is introduced, using the features of the visible light modality as the query and the features of the infrared and point cloud modalities as the key and value. The correlation weights between the features of each modality are calculated by scaling dot product attention. Finally, the features of each modality are adaptively weighted and summed according to the attention weights to obtain the fused unified feature tensor.

[0119] During the fusion process, if the confidence level of a certain modality's detection result is low while that of another modality is high, the attention mechanism will automatically increase the fusion weight of the high-confidence modality, and vice versa, thereby achieving complementary advantages of different modal information. The output of this submodule is the fused feature tensor and the corresponding spatial location label;

[0120] The defect comprehensive identification submodule comprehensively determines the existence, type, severity, and spatial location of blade defects based on the fused feature tensor output by the multimodal feature fusion submodule, and outputs a comprehensive defect detection result.

[0121] This submodule has a built-in multi-task classification and regression network that takes the fused feature tensor as input and performs three tasks simultaneously: binary classification of defect existence (whether a defect exists), multi-classification of defect types (cracks, corrosion, flaking, sand holes, dirt, etc.), and regression of defect severity level (light, medium, and severe, based on a comprehensive assessment of the defect's size, depth, temperature anomaly, and geometric variables).

[0122] When only one or two of the three modes of visible light, infrared and point cloud are detected as abnormal while the other modes are not detected, this submodule makes a comprehensive judgment based on the confidence level of each mode and the preset mode confidence weight. When the detection confidence level of a single mode is higher than the preset high confidence threshold (preferably 0.85), the detection conclusion of that mode is directly adopted.

[0123] When the detection results of each modality are inconsistent and the confidence levels are all at a moderate level, a comprehensive judgment is made using the majority voting principle or Bayesian decision rule. For abnormal regions that are detected simultaneously by all three modalities and have the same spatial location, they are judged as high-confidence defects, and the defect conclusion is directly output.

[0124] The output of this submodule is a comprehensive defect detection result set. Each result includes the defect's unique number, spatial location (three-dimensional coordinates and range in the local coordinate system of the wind turbine), defect type, severity level, comprehensive confidence score, and references to the original detection results for each modality.

[0125] The defect comprehensive analysis and decision-making module includes: a defect classification and grading submodule, a tower sweep risk assessment submodule, and an operation and maintenance decision generation submodule.

[0126] The defect classification and grading submodule receives the comprehensive defect detection results output by the multimodal fusion detection module. Based on preset defect classification standards, it categorizes defects into types such as cracks, corrosion, flaking, pinholes, and dirt, and classifies the severity level of defects according to parameters such as size, depth, and location. This submodule incorporates a blade structure importance partitioning model. Based on T / CIA0011-2026 "Technical Specification for Damage Grading and Repair of Wind Turbine Blades in Wind Farms," ​​the blade is divided along its span into several structural partitions: the blade root region, the main load-bearing region in the middle of the blade (near the maximum chord length), the blade leading edge, the blade trailing edge, and the blade tip region. Defects of the same size within different partitions have different severity level weights.

[0127] Specifically, for crack-type defects in the main load-bearing area, the severity level weight is set to the highest because this area bears the main bending load of the blade; for detachment-type defects in the blade tip area, although this area is not a main load-bearing part, detachment will lead to a decrease in aerodynamic performance and local stress concentration, so the severity level weight is set to medium to high; for minor sand holes or surface dirt in non-critical areas, the severity level weight is set to a lower level.

[0128] Regarding defect size parameters, this submodule classifies defects into three severity levels based on the two-dimensional projected area of ​​the defect, the maximum linear size (crack length or pinhole diameter) provided in the multimodal fusion detection results, and the defect depth or protrusion height provided by point cloud geometric analysis: minor level (defect length less than 10mm or depth less than 0.5mm, located in a non-critical area, not affecting structural load-bearing capacity), medium level (defect length 10-50mm or depth 0.5-2mm, or located in a moderately important area of ​​the structure, possibly affecting local load-bearing capacity), and severe level (defect length greater than 50mm or depth greater than 2mm, or located in the main load-bearing area, already affecting or about to affect the structural integrity and operational safety of the blade).

[0129] This submodule also dynamically adjusts the defect severity level based on the blade's service life and cumulative operating hours. For older blades that have been in operation for more than 10 years, the defect severity level for blades of the same size is increased by one level to reflect the impact of material aging and fatigue accumulation. After classification and grading, this submodule outputs a structured defect list. Each record includes a unique defect number, defect type label, defect spatial location (including the structural partition to which it belongs), severity level (levels 1 to 3), and detailed parameters for grading (size, depth, location weight, etc.).

[0130] The tower sweep risk assessment submodule calculates the minimum angle between the blade tip and the tower based on the blade spatial position data output by the multimodal fusion detection module and the blade geometric information provided by the point cloud geometric analysis submodule. This submodule first extracts the spatial axis fitting parameters for each of the three blades from the point cloud geometric analysis results to determine the azimuth and elevation angles of each blade in space. Simultaneously, it extracts the axial direction and cross-sectional profile of the tower from the point cloud data, establishing a local three-dimensional coordinate system with the center of the tower bottom as the origin.

[0131] Based on this, the submodule calculates the minimum angle between the blade and the tower through the following steps: First, based on the spatial axis of each blade and the blade tip position coordinates, calculate the three-dimensional coordinates of the blade tip point in the tower coordinate system.

[0132] The second step is to establish the spatial distance equation from the tower surface to the blade tip—the tower is approximately a cylinder, and its surface equation is determined by the tower radius and the direction of the tower axis.

[0133] The third step is to calculate the spatial distance from the blade tip to the nearest point on the tower surface. When the blade rotates to the azimuth angle range that overlaps with the tower (usually 140° to 210°), this distance is the clearance distance of the blade at the current moment.

[0134] The fourth step is to convert the clearance distance into the angle between the blade and the tower. Given the blade length L and the clearance distance d, the angle between the blade tip and the tower is θ = arcsin(d / L). When the blade length is much larger than the tower radius, this angle is approximately equal to the minimum angle between the blade and the tower axis.

[0135] This submodule uses a preset safety angle threshold (preferably 5°, corresponding to a clearance distance of approximately 10.5 meters for a blade length of approximately 120 meters) as the criterion. When the calculated minimum blade-to-tower angle is greater than the safety threshold, it is determined that there is no risk of tower sweeping, and a "safe" status is output.

[0136] When the minimum included angle is between the safety threshold and the warning threshold (preferably 3° to 5°, corresponding to a clearance distance of about 6.3 to 10.5 meters), a first-level tower sweep risk warning signal (yellow warning) is generated to remind operation and maintenance personnel to pay attention and strengthen monitoring;

[0137] When the minimum included angle is less than the warning threshold (preferably 3°, corresponding to a clearance distance of approximately 6.3 meters), a secondary tower sweep risk warning signal (red warning) is generated, indicating the presence of a real-time tower sweep hazard. This submodule also receives wind speed sensor and blade pitch angle feedback data as auxiliary judgment parameters. When the wind speed exceeds 120% of the rated wind speed or the blade pitch angle actuator malfunctions, the values ​​of the safety threshold and warning threshold are adjusted accordingly to reflect the negative impact of increased blade flapping amplitude on the clearance distance under extreme operating conditions.

[0138] The risk assessment results output by this submodule include the minimum included angle value of each blade at the current moment, the corresponding clearance distance, the risk level (safe / yellow warning / red warning) and the suggested countermeasures (continuous monitoring / arrange special inspection / immediate shutdown).

[0139] The operation and maintenance decision generation submodule generates operation and maintenance decision suggestions based on the defect classification results of the defect classification and grading submodule and the risk assessment results of the tower sweeping risk assessment submodule. These suggestions include defect repair priority ranking, suggested maintenance time windows, and recommended maintenance strategies.

[0140] This submodule has a built-in maintenance decision model based on risk priority. The model comprehensively considers four dimensions: the severity level of the defect, the importance of the structural partition where the defect is located, the development trend prediction of the defect, and the current tower sweep risk level, and calculates the comprehensive maintenance urgency index (RUI) for each defect.

[0141] The formula for calculating RUI is: Where S is the defect severity level score (1-3 points), Z is the structural zoning weight coefficient (critical area 1.5, general area 1.0, non-critical area 0.6), T is the defect development trend factor (calculated based on the size change rate of the same defect in historical inspection data, with a change rate greater than 20% / year taking 1.2, a change rate of 10%-20% / year taking 1.0, and a change rate less than 10% / year taking 0.8), and R is the tower sweep risk level score (red warning takes 2.0, yellow warning takes 1.2, no risk takes 0).

[0142] Based on the RUI value, this submodule sorts all detected defects from highest to lowest priority. Defects with an RUI greater than 3.5 are listed as the highest priority (it is recommended to complete the repair within 7 days), defects with an RUI between 2.5 and 3.5 are listed as medium priority (it is recommended to complete the repair within 30 days), and defects with an RUI less than 2.5 are listed as regular priority (to be included in the next planned maintenance).

[0143] Regarding the recommended maintenance time window, this submodule recommends specific maintenance time windows for each high-priority and medium-priority defect based on the wind farm's meteorological forecast data (wind speed forecast, precipitation probability, and temperature range for the next 7 days) and the wind turbine's power generation plan (whether there is a planned shutdown window). Priority is given to selecting periods with wind speeds below 8 m / s, no precipitation, and temperatures between 5℃ and 30℃, and these periods should coincide with the wind turbine's planned maintenance or low-wind-speed shutdown windows as much as possible to reduce power generation loss due to maintenance.

[0144] Regarding recommended maintenance strategies, this submodule outputs standardized repair plan templates based on defect type and severity level. For minor surface cracks or pinholes, surface grinding and gel coat repair processes are recommended.

[0145] For moderate corrosion or flaking defects, it is recommended to use laminated resin injection and fiberglass cloth layup repair process; for severe structural cracks or large-area flaking, it is recommended to use structural reinforcement or segmented replacement scheme, and indicate that the blade manufacturer or professional maintenance team needs to intervene.

[0146] This submodule also generates a comprehensive operation and maintenance decision report, which lists detailed information on all defects, corresponding repair plans, suggested repair time windows, and estimated repair man-hours and material requirements in order of priority, providing actionable decision-making basis for wind farm operation and maintenance management personnel.

[0147] The data interaction and visualization module includes: a data storage submodule, a 3D visualization submodule, an inspection report generation submodule, and a data feedback submodule;

[0148] The data storage submodule is used to store the comprehensive diagnostic report, defect detection results, and operation and maintenance decision data output by the defect comprehensive analysis and decision module into the database, establishing a defect archive for the entire life cycle of the wind turbine blades. This submodule adopts a distributed database architecture, including structured and unstructured data storage units. The structured data storage unit stores structured fields such as the defect's unique ID, defect type label, spatial location coordinates (three-dimensional coordinates in the wind turbine's local coordinate system), severity level, detection timestamp, corresponding wind turbine ID, and blade ID. It is managed using a relational database (such as PostgreSQL) and supports rapid retrieval and statistical analysis based on multiple dimensions such as wind turbine ID, detection time, defect type, and severity level.

[0149] The unstructured data storage unit is used to store the original visible light image slices, infrared thermal imaging slices, local point cloud data fragments, and intermediate result files in the detection process corresponding to the defects. It is managed using object storage (such as MinIO or cloud storage service). Each defect record is indexed and associated with the corresponding unstructured data through a unique identifier.

[0150] This submodule establishes an independent defect file database table for each blade, using the blade manufacturing number and commissioning date as the primary key of the file. It records all defect information detected in each inspection and subsequent maintenance records (including maintenance date, maintenance method, and post-maintenance re-inspection results) in chronological order, forming a complete traceability chain from defect discovery, maintenance to re-inspection closure.

[0151] The data storage submodule also includes a built-in data backup and recovery mechanism, which synchronizes incremental data to an off-site disaster recovery storage node daily to ensure the security and durability of defect archive data. This submodule also provides a standardized data access interface (RESTful API), allowing the 3D visualization submodule and inspection report generation submodule to access defect data from the database as needed.

[0152] The 3D visualization submodule, based on the wind turbine's 3D model, spatially locates and annotates the defect detection results on the 3D model, presenting the type, location, and severity of the defects in a visual manner. This submodule first loads the corresponding parametric wind turbine 3D model from the 3D model library based on the wind turbine's model parameters (blade length, hub height, tower diameter, blade airfoil data, etc.). The model includes the geometry of the tower, hub, three blades, and nacelle, with each component possessing an independent local coordinate system and spatial transformation matrix.

[0153] The defect spatial location mapping unit converts the three-dimensional spatial coordinates (located in the local coordinate system of the wind turbine) output by the defect comprehensive identification submodule into corresponding points in the three-dimensional model. For each defect, a three-dimensional annotation anchor point is generated at its precise location in the model space. The annotation style is rendered differently according to the defect type and severity level: crack-type defects are annotated with a three-dimensional curve along the crack direction using red lines, while the crack extension range is indicated by a semi-transparent red curved surface.

[0154] Corrosion and flaking defects are marked with orange or yellow translucent spheres or ellipsoids, with the radius of the sphere proportional to the actual size of the defect; pinhole defects are marked with purple dots on the model surface for precise location.

[0155] Dirt-related defects are indicated by a gray semi-transparent area to show the extent of the contamination. Each marker is accompanied by an information label; when the operator hovers the mouse over or clicks the marker, an information card pops up displaying the defect's detailed attributes (defect type, severity level, detection date, dimensional parameters, repair recommendations, etc.).

[0156] This submodule supports multi-view browsing operations, including overall model rotation, scaling, translation, and section view. The section view function allows operators to cut sections at any position along the blade span to observe the distribution depth of defects inside the blade (based on depth information provided by point cloud geometric analysis).

[0157] This submodule also supports a time-series comparison view, which overlays defect data from multiple inspections onto the same 3D model. Different colors are used to distinguish defects detected in different periods (newly detected defects are highlighted and flashed, while historically repaired defects are displayed in gray), allowing maintenance personnel to intuitively assess defect trends and repair effectiveness. The 3D visualization rendering engine is based on WebGL or Unity3D technology and supports access via a web browser, requiring no dedicated client software installation.

[0158] The inspection report generation submodule automatically generates a visually appealing inspection report based on defect detection results and maintenance decision data. The report includes basic wind turbine information, inspection time, flight path, defect type and location, risk assessment conclusions, and handling recommendations. This submodule has multiple built-in report templates suitable for three scenarios: routine daily inspection reports, detailed defect investigation reports, and emergency fault notification reports.

[0159] Routine inspection reports use a concise template, mainly consisting of a defect statistics summary table and a defect distribution diagram, highlighting the comparison of the number of defects between the current and previous inspections. Detailed defect investigation reports use a comprehensive template, providing a side-by-side comparison of the original visible light image, infrared thermal image, and local point cloud reconstruction image for each detected defect, along with the defect's size measurement data and the criteria for severity level determination. Emergency fault notification reports use a briefing template, highlighting severe defects and tower sweep risk information with a red warning icon; the report's first page provides clear emergency handling suggestions and contact channels.

[0160] The report generation engine first retrieves all defect records and corresponding unstructured data (image slices, point cloud fragments, etc.) from the data storage submodule. Then, it automatically arranges the content according to the selected template's layout, filling in the wind turbine's basic information and inspection parameters at the report header. Defect statistical charts (pie chart of defect type distribution, bar chart of severity level distribution, and line graph of defect location along the blade spanwise) are inserted in the middle of the report. Detailed information and image data for each defect are listed item by item. The maintenance urgency ranking and handling suggestions output by the operation and maintenance decision generation submodule are summarized in the conclusions and recommendations section at the end of the report. All images in the report are automatically cropped and labeled, with colored rectangles marking defect locations on the original images, along with defect numbers and type labels, ensuring that report readers can quickly locate and understand defect information. After the report is generated, it is output in both PDF and Word formats and automatically sent to the data storage submodule for archiving. Designated wind farm operation and maintenance personnel are also notified via email or instant messaging.

[0161] The data feedback submodule feeds back the defect distribution data, detection accuracy data, and route execution data of this inspection to the route optimization submodule of the route planning and autonomous navigation module for iterative optimization of subsequent inspection routes, realizing a complete data closed loop of "collection-detection-decision-optimization".

[0162] This submodule first extracts three types of feedback data from the complete data stream of this inspection: The first type is defect distribution statistics, including the spatial distribution frequency of each defect in three dimensions: blade span (expressed as a percentage of blade length, such as the distribution in the range of 0% to 100%), tangential (leading edge, middle, trailing edge) and blade surface (windward side, leeward side), as well as the number of defects detected and the distribution of severity levels for each defect type;

[0163] The second category is the detection accuracy evaluation data, including the image sharpness score of each image acquisition point (calculated based on image gradient energy and Laplacian variance), the deviation between the gimbal camera attitude angle and the ideal observation angle at each track point, and the number of invalid detection samples and the corresponding track point positions caused by image blurring or improper exposure.

[0164] The third category is route execution data, including the three-dimensional spatial deviation (root mean square error) between the actual flight trajectory and the planned route, the deviation between the actual arrival time and the planned time of each waypoint, and the position and length of the route deviation segment caused by obstacle avoidance.

[0165] After receiving the above feedback data, the route optimization submodule performs the following three optimization calculations: (1) Based on the defect distribution statistics, the track point density is redistributed, and the track point density in the high defect incidence area (such as the leading edge middle section and the blade tip area) is increased by 20% to 50%, while the track point density in the long-term defect-free detection area is reduced accordingly.

[0166] (2) Based on the gimbal parameter adjustment of the detection accuracy evaluation data, analyze the reasons for the deviation of the track point whose image clarity score is lower than the preset threshold (poor lighting angle, shooting distance too far, gimbal pitch angle deviation, etc.) and adjust the shooting time, observation distance or gimbal attitude angle setting value of the track point accordingly in subsequent inspections.

[0167] (3) Based on the flight path execution data, the flight path is smoothed. For flight path segments with large deviations from the actual flight path, the planned flight path is made smoother by adding intermediate waypoints or adjusting the spacing between waypoints. This reduces the large attitude adjustments of the UAV between waypoints and improves flight stability and image acquisition quality. The optimized flight path parameters are converted and written back to the flight path template library of the hemispherical inspection flight path generation submodule. The templates are stored with the wind turbine number and inspection date as version identifiers for automatic recall by subsequent inspection tasks of the same wind turbine.

[0168] This feedback mechanism enables the system's inspection strategy to continuously improve itself based on historical inspection results and actual execution. As the number of inspections increases, the rationality of route planning and the effectiveness of image acquisition gradually improve, forming a data-driven intelligent operation and maintenance closed loop.

[0169] The four sub-modules are connected by the data storage sub-module as the core hub. The 3D visualization sub-module and the inspection report generation sub-module read data from it, and the data feedback sub-module extracts statistical information from it. At the same time, the output of the data feedback sub-module is directly transmitted to the route planning module without passing through the data storage sub-module, so as to ensure the real-time nature of the feedback data and the priority response to the route optimization task.

[0170] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A smart defect identification system for unmanned aerial vehicle (UAV) inspection of wind turbine blades, characterized in that, include: The system includes a route planning and autonomous navigation module, a multi-source data acquisition module, a data preprocessing and feature enhancement module, a multimodal fusion detection module, a defect comprehensive analysis and decision-making module, and a data interaction and visualization module. The route planning and autonomous navigation module generates a hemispherical inspection route with the wind turbine hub center as the center and a preset safety distance as the radius based on the input wind turbine structural parameters, and outputs the planned route data to the multi-source data acquisition module. The multi-source data acquisition module is mounted on the UAV platform. Based on the received planned flight path data, it synchronously controls the execution of visible light image acquisition, infrared thermal imaging acquisition, and lidar point cloud acquisition, and outputs the acquired multi-source data to the data preprocessing and feature enhancement module. The data preprocessing and feature enhancement module performs noise reduction, contrast enhancement, multi-scale detail enhancement and normalization on the received multi-source data, and outputs the processed visible light image data, infrared thermal imaging data and lidar point cloud data to the multimodal fusion detection module. The multimodal fusion detection module performs defect target detection on visible light image data, thermal anomaly detection on infrared thermal imaging data, and geometric deformation detection on lidar point cloud data. Then, it performs spatiotemporal alignment and confidence weighted fusion on the detection results of the three modes through multimodal feature fusion based on attention mechanism. Based on the fused features, it comprehensively determines the existence, type, severity, and spatial location of the defect, and outputs the comprehensive defect detection result to the defect comprehensive analysis and decision module. The defect comprehensive analysis and decision-making module classifies and grades defects based on the received comprehensive defect detection results, calculates the minimum angle between the blade and the tower based on the blade spatial position data to assess the tower sweep risk, and then generates operation and maintenance decision suggestions by combining the defect classification results and the tower sweep risk assessment results, and outputs the operation and maintenance decision data to the data interaction and visualization module. The data interaction and visualization module stores, visualizes in three dimensions, and generates reports on the received operation and maintenance decision data. It also feeds back the defect distribution data, detection accuracy data, and route execution data of this inspection to the route planning and autonomous navigation module for iterative optimization of subsequent inspection routes.

2. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 1, characterized in that, The route planning and autonomous navigation module includes: a wind turbine parameter input submodule, a hemispherical inspection route generation submodule, a dynamic obstacle avoidance submodule, and a route optimization submodule; The wind turbine parameter input submodule receives and stores the wind turbine's structural parameters and geographic information parameters, and transmits them to the hemispherical inspection route generation submodule. The hemispherical inspection route generation submodule generates wind turbine structural parameters based on the wind turbine parameter input submodule, establishes a local coordinate system of the wind turbine with the spatial coordinates of the hub center as the origin, evenly distributes several track points on the hemispherical surface, and converts the track point sequence into a route file for output to the UAV flight control system; the submodule also sets the exposure timing parameters for image acquisition based on the rated speed range of the blades, so that the camera shutter trigger time matches the time window when the blades pass through the preset acquisition area; The dynamic obstacle avoidance submodule receives multi-source perception data from the UAV's onboard obstacle avoidance sensor in real time. It has a built-in local path replanning unit. When it is determined that there is an obstacle within the preset safe distance of the current planned route, it triggers a local route correction process. After the UAV bypasses the obstacle and returns to the original route, it resumes the inspection task along the original planned route. The route optimization submodule receives historical inspection data and defect distribution information from the data interaction and visualization module. It performs statistical analysis on the defect types, spatial locations, and severity levels detected in the historical inspections to generate a heat map of the defect probability distribution on the blade surface. Based on this heat map, it dynamically adjusts the spatial distribution density of the track points in the route planning of subsequent inspections. The optimized route parameters are then updated in the route template library of the hemispherical inspection route generation submodule.

3. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 2, characterized in that, The multi-source data acquisition module is deployed on an unmanned aerial vehicle (UAV) platform and includes: a visible light image acquisition submodule, an infrared thermal imaging acquisition submodule, a lidar point cloud acquisition submodule, an acquisition control submodule, and a data cache submodule; The data acquisition and control submodule receives the planned route data output by the route planning and autonomous navigation module. Based on the preset acquisition points, acquisition times, and acquisition parameters in the route, it synchronously controls the visible light image acquisition submodule, infrared thermal imaging acquisition submodule, and lidar point cloud acquisition submodule to perform data acquisition. This submodule has a built-in high-precision clock synchronization unit, which uses the second pulse signal of the Global Navigation Satellite System as a unified clock reference to synchronize the data acquisition time of each sensor. When the UAV arrives at the preset acquisition point, it simultaneously sends trigger pulses to the three sensors. The data caching submodule aligns and indexes visible light image frames, infrared thermal imaging frames, and lidar point cloud frames according to a unified time axis, based on the timestamp and spatial pose information provided by the acquisition control submodule. This forms a multimodal data record with the acquisition time as the primary key. Each record contains visible light image data, infrared thermal image data, 3D point cloud data, and corresponding UAV pose metadata for the same spatial region of the blade at the same time. After the inspection task is completed, this submodule outputs the complete multimodal data package to the data preprocessing and feature enhancement module.

4. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 3, characterized in that, The data preprocessing and feature enhancement module includes: an image denoising submodule, a contrast enhancement submodule, a multi-scale detail enhancement submodule, and a data normalization submodule; The image denoising submodule receives visible light image data output by the multi-source data acquisition module, uses a non-local mean algorithm to remove background noise in the image, and uses the non-local self-similarity redundancy information in the image to perform a weighted average estimation of each pixel. The contrast enhancement submodule uses a limited contrast adaptive histogram equalization algorithm on the denoised image to divide the image into several rectangular regions, perform histogram equalization on each region, and then perform smooth fusion. The multi-scale detail enhancement submodule constructs a multi-scale feature pyramid for the contrast-enhanced image, and splices and fuses the detail feature maps at each scale with the original feature map to generate an enhanced feature map containing multi-scale detail information. The data normalization submodule performs unified normalization processing on the enhanced visible light image, infrared thermal imaging data, and lidar point cloud data, converting the data of the three modalities into a unified tensor format before outputting them to the multimodal fusion detection module.

5. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 4, characterized in that, The multimodal fusion detection module includes: a visible light target detection submodule, an infrared anomaly detection submodule, a point cloud geometric analysis submodule, a multimodal feature fusion submodule, and a defect comprehensive identification submodule; The visible light target detection submodule uses an improved YOLO detection network to detect defect targets in the preprocessed visible light image. It introduces receptive field attention convolution to replace some traditional convolution modules in the backbone network, dynamically adjusts the weights of different receptive fields, and uses an improved loss function with a modulation factor in the detection head to alleviate the problem of imbalance between positive and negative samples. The infrared anomaly detection submodule performs a two-level anomaly determination process on the preprocessed infrared thermal imaging data: the first level is global adaptive threshold detection, which statistically analyzes the temperature distribution of the entire infrared image and initially marks the candidate thermal anomaly region with a global anomaly temperature threshold; the second level is local contrast verification, which calculates the difference between the average temperature of the candidate region and the average temperature of the surrounding normal region, and confirms the region as a valid thermal anomaly when the temperature difference exceeds a preset threshold. The point cloud geometric analysis submodule performs three-dimensional reconstruction and geometric analysis on the pre-processed lidar point cloud data, registers and compares the blade point cloud data obtained from the current inspection with historical point cloud data or blade design models, and calculates the spatial distance deviation between corresponding points to generate a deformation deviation distribution map. The multimodal feature fusion submodule receives the output results of the visible light target detection submodule, the infrared anomaly detection submodule, and the point cloud geometric analysis submodule. Based on the timestamp and spatial pose information when the UAV collects data, it converts the detection results of the three modes into a unified local coordinate system of the wind turbine to establish a spatial correspondence. Then, it uses a cross-modal attention mechanism to adaptively weight and sum the features of each mode to obtain the unified feature tensor after fusion. The defect comprehensive identification submodule, based on the fused feature tensor output by the multimodal feature fusion submodule, simultaneously performs binary classification of defect existence, multi-class classification of defect type, and regression of defect severity level through the built-in multi-task classification and regression network, and outputs a comprehensive defect detection result.

6. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 5, characterized in that, The defect comprehensive analysis and decision-making module includes: a defect classification and grading submodule, a tower sweeping risk assessment submodule, and an operation and maintenance decision generation submodule; The defect classification and grading submodule receives the comprehensive defect detection results output by the multimodal fusion detection module, classifies the defects according to the preset defect classification criteria, and classifies the severity level of the defects according to the size, depth, and location parameters of the defects combined with the blade structure importance partitioning model. The tower sweep risk assessment submodule calculates the minimum angle between the blade and the tower based on the blade spatial position data and the blade geometric information provided by the point cloud geometric analysis submodule. The preset safe angle threshold is used as the judgment criterion. When the minimum angle is greater than the safe threshold, it is determined that there is no tower sweep risk. When the minimum angle is between the safe threshold and the warning threshold, a first-level tower sweep risk warning signal is generated. When the minimum angle is less than the warning threshold, a second-level tower sweep risk warning signal is generated. The operation and maintenance decision generation submodule calculates the maintenance urgency index of each defect based on the defect classification and grading submodule and the risk assessment results of the tower sweeping risk assessment submodule. Based on the risk priority maintenance decision model, the submodule sorts all detected defects by priority and generates corresponding operation and maintenance decision suggestions.

7. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 6, characterized in that, The data interaction and visualization module includes: a data storage submodule, a 3D visualization submodule, an inspection report generation submodule, and a data feedback submodule; The data storage submodule stores the comprehensive diagnostic report, defect detection results, and operation and maintenance decision data output by the defect comprehensive analysis and decision module into the database, establishes an independent defect file database table for each blade, and records all defect information detected in each inspection and subsequent maintenance records in chronological order to form a complete traceability chain. The three-dimensional visualization submodule is based on the three-dimensional model of the wind turbine. It spatially positions and marks the defect detection results on the three-dimensional model, presents the type, location and severity of the defects in a visual way, and supports time-series comparison view to overlay and display defect data from multiple inspections on the same three-dimensional model. The inspection report generation submodule automatically generates a graphic inspection report based on defect detection results and operation and maintenance decision data; The data feedback submodule feeds back the defect distribution statistics, detection accuracy evaluation data, and route execution data of this inspection to the route optimization submodule of the route planning and autonomous navigation module. The route optimization submodule then performs trackpoint density redistribution, gimbal parameter adjustment, and route smoothing optimization based on the feedback data, thus realizing a data closed loop of collection-detection-decision-optimization.

8. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 3, characterized in that, The visible light image acquisition submodule is equipped with a high-resolution three-axis stabilized gimbal camera, integrating a dual-optical-path system of a wide-angle camera and a zoom camera. The wide-angle camera is used for wide-area shooting of the entire blade, while the zoom camera has an optical zoom capability of no less than 20x for capturing long-distance close-up details of local areas of the blade. The gimbal system has a three-axis mechanical stabilization function. The infrared thermal imaging acquisition submodule is equipped with an uncooled infrared thermal imager, which is used to simultaneously acquire thermal imaging data of the blade surface and capture temperature anomaly areas caused by defects. The lidar point cloud acquisition submodule is equipped with a lidar scanner, which is used to acquire three-dimensional point cloud data of the wind turbine blades and the surrounding environment, and obtain the geometric contour information and spatial attitude information of the blades.

9. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 4, characterized in that, The multi-scale detail enhancement submodule performs multi-level average pooling operations on the input image to construct a feature pyramid representation with four scale levels from fine to coarse. For each pooling level, the feature map of the previous level is upsampled to restore it to the original scale and then subjected to element-wise difference operation with the feature map of the original scale to obtain the detail feature map at that scale. After completing the difference calculation at each scale, the detail feature maps at all scales are concatenated and fused with the original feature map through the channel dimension to generate an enhanced feature map containing multi-scale detail information.

10. The intelligent identification system for defects in wind turbine blades by unmanned aerial vehicle (UAV) inspection according to claim 5, characterized in that, The multimodal feature fusion submodule constructs a three-input feature fusion network, extracting feature vectors from visible light detection results, infrared detection results, and point cloud geometric analysis results through three parallel feature encoding branches. A cross-modal attention mechanism is introduced, using features of the visible light modality as queries and features of the infrared and point cloud modalities as keys and values. The correlation weights between the features of each modality are calculated by scaling dot product attention, and the features of each modality are adaptively weighted and summed according to the attention weights to obtain the unified feature tensor after fusion.

Images