A wind turbine blade surface defect detection method and system based on UAV visual images

Abstract

The present application provides a method and system for detecting surface defects of wind turbine blades based on UAV visual images. Among them, the present application relates to the field of defect detection technology. The method acquires surface image sequences of high-speed rotating blades and synchronously obtains blade motion trajectory data; spatially and temporally associates the motion trajectory data with the image sequence and compensates for blade edge distortion frame by frame; dynamically allocates contrast enhancement weights to the windward and leeward sides of the blades in the compensated image sequence according to the rotation speed to generate a fused texture feature map; adjusts the knowledge distillation attention weights according to the distribution ratio of high-frequency vibration and static areas in the feature map, and transfers high-precision model features to a lightweight model; embeds the blade motion state vector in the lightweight model, dynamically adjusts the detection sensitivity threshold, and outputs defect results that match the motion state. The present application can solve the problem of imbalance of dynamic and static features and high false detection rate in blade defect detection under high-speed rotating scenarios.

Description

Technical Field

[0001] The present application relates to the field of defect detection technology, and in particular to a method and system for detecting surface defects of wind turbine blades based on drone visual images. Background Art

[0002] Under high-speed flight conditions, dynamic blades produce complex surface deformations and transient defects due to severe aerodynamic loads and ultra-high-speed rotation. It is necessary to achieve real-time online detection of submillimeter defects under strong vibration, high-speed motion blur and non-steady-state lighting conditions, and ensure the robustness of the detection system to dynamic deformation and extreme working conditions.

[0003] The typical solution for this scenario is dynamic surface reconstruction technology based on multispectral laser scanning. It scans the blade surface with a high-frequency laser beam, reconstructs the three-dimensional morphology of the blade in combination with reflection spectrum analysis, extracts abnormal surface areas using a deformation tracking algorithm, and determines the defect type using a preset deformation threshold.

[0004] The existing solutions have the following defects: the sampling frequency of laser scanning is limited by hardware, making it difficult to capture the micron-level deformation characteristics of transient defects in ultra-high-speed rotation, resulting in missed detection of tiny cracks; non-steady-state illumination and aerodynamic turbulence cause the signal-to-noise ratio of the reflection spectrum to drop sharply, artifact interference appears in the reconstructed morphology, and the accuracy of defect judgment is significantly reduced; the iterative calculation of three-dimensional reconstruction and deformation tracking takes too long to meet the millisecond-level online detection requirements in high-speed flight scenarios, and the system power consumption is too high, making it difficult to integrate into a lightweight detection platform. Summary of the Invention

[0005] The present application provides a method and system for detecting surface defects of wind turbine blades based on UAV visual images, which is used to solve the problems of imbalance in dynamic and static feature detection and high false detection rate in the prior art.

[0006] In a first aspect, the present application provides a method for detecting surface defects of wind turbine blades based on drone visual images, comprising:

[0007] The surface image sequence of high-speed rotating wind turbine blades is collected by drone, and the blade motion trajectory data output by the inertial navigation module is simultaneously obtained;

[0008] Performing spatiotemporal correlation between the blade motion trajectory data and the surface image sequence, and based on a dynamic matching relationship between the blade rotation angle and the imaging device posture extracted from the spatiotemporal correlation result, compensating for blade edge distortion caused by motion blur in the surface image sequence frame by frame through the dynamic matching relationship;

[0009] The blade surface feature map is generated by combining the spatiotemporal continuity constraint of the blade motion trajectory data with the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the surface image sequence after strengthening the compensation by dynamic weight allocation;

[0010] Based on the dynamic distribution ratio of high-frequency vibration areas and static areas in the blade surface feature map, the attention weight of the knowledge distillation framework is adjusted, and the feature response pattern of the pre-trained high-precision defect detection model is migrated to the lightweight detection model according to the adjusted attention weight;

[0011] The inertial navigation module is embedded in the lightweight detection model, and a blade motion state vector is generated in real time according to the inertial navigation data output by the inertial navigation module. By adjusting the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges, a defect detection result matching the blade motion state vector is output.

[0012] Optionally, the spatiotemporally correlating the blade motion trajectory data with the surface image sequence, and based on a dynamic matching relationship between the blade rotation angle and the imaging device posture extracted from the spatiotemporal correlation result, compensating for blade edge distortion caused by motion blur in the surface image sequence frame by frame through the dynamic matching relationship to obtain a compensated surface image sequence, comprising:

[0013] Extracting the blade's three-dimensional spatial positioning data and rotation angle corresponding to each frame image acquisition moment from the blade motion trajectory data, and generating a spatiotemporal index aligned with the surface image sequence timestamp;

[0014] Obtaining the difference in rotation angles in adjacent frames and the offset of the imaging device posture according to the spatiotemporal index, and calculating the distortion direction and intensity ratio of the blade edge during continuous motion according to the spatial geometric superposition relationship between the difference and the offset;

[0015] Generate a pixel displacement compensation vector for the edge area of the blade in the current frame based on the distortion direction and intensity ratio, and perform layer-by-layer offset correction on the pixel position along the opposite direction of blade rotation using the pixel displacement compensation vector;

[0016] The edge contour sharpening process is performed on the leaf edge distortion caused by motion blur in the surface image sequence according to the pixel displacement compensation vector, residual blurred pixel points in the leaf edge distortion are eliminated, and a compensated surface image sequence is generated.

[0017] Optionally, the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the surface image sequence after the dynamic weight distribution enhancement compensation is combined with the spatiotemporal continuity constraint of the blade motion trajectory data to generate a blade surface feature map, including:

[0018] According to the change of the blade rotation speed, the contrast enhancement weight values of the windward side and the leeward side of the blade in the compensated surface image sequence are dynamically allocated, wherein the contrast enhancement weight value of the windward side increases with the increase of the blade rotation speed, and the contrast enhancement weight value of the leeward side decreases with the increase of the blade rotation speed;

[0019] Based on the rotation angle difference between consecutive frames in the blade motion trajectory data, the contrast enhancement weight values are accumulated in adjacent frame images of the same blade area to generate a weight accumulation sequence consistent with the blade rotation direction;

[0020] According to the numerical distribution of the weighted accumulation sequence, positive gain adjustment and negative suppression adjustment are respectively performed on the pixel intensities on the windward side and the leeward side, so that the intensity difference of the pixel intensities increases with the rotation speed;

[0021] The adjusted pixel intensities are superimposed across frames in the time sequence of the leaf motion trajectory data, and a leaf surface feature map containing dynamic texture features of the leaf surface is generated based on the intensity mutation area contour formed after the superposition.

[0022] Optionally, adjusting the attention weight of the knowledge distillation framework based on the dynamic distribution ratio of the high-frequency vibration area and the static area in the blade surface feature map, and migrating the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weight, includes:

[0023] Extracting the fluctuation amplitude of the blade surface texture intensity in the blade surface feature map in adjacent time intervals, and dividing the boundary range between the high-frequency vibration area and the static area according to the fluctuation amplitude;

[0024] Counting the area proportions of the high-frequency vibration region and the static region in the blade surface characteristic map, and generating a numerical relationship representing the dynamic distribution ratio of the region;

[0025] According to the numerical relationship, the attention weight coefficient corresponding to the high-frequency vibration area in the knowledge distillation framework is geometrically amplified, and the attention weight coefficient corresponding to the static area is reversely reduced;

[0026] A high-precision defect detection model is pre-trained according to the knowledge distillation framework, and a characteristic response pattern matching the high-frequency vibration area and the static area in the high-precision defect detection model is obtained. Based on the characteristic response pattern, the adjusted attention weight coefficient is mapped layer by layer to the lightweight detection model.

[0027] Optionally, performing positive gain adjustment and negative suppression adjustment on the pixel intensities on the windward side and the leeward side respectively according to the numerical distribution of the weighted accumulation sequence so that the intensity difference of the pixel intensities increases with the rotation speed includes:

[0028] Associating the accumulated value of each frame in the weighted accumulation sequence with the current rotation speed of the blade to generate the dynamic gain coefficient and dynamic suppression coefficient of the windward side and the leeward side respectively;

[0029] In the compensated surface image sequence, the boundary line between the windward side and the leeward side is marked along the extension direction of the blade surface texture, and a set width is extended to both sides based on the boundary line to form a gradual adjustment area;

[0030] In the windward side range and the leeward side range of the gradual adjustment zone, the intensity value of each pixel is subjected to superposition enhancement and subtraction weakening according to the dynamic gain coefficient and the dynamic suppression coefficient respectively;

[0031] The difference between the dynamic gain coefficient and the dynamic suppression coefficient is calculated based on the current rotation speed of the blade, and the entire area of the windward side and the leeward side outside the gradient adjustment area is subjected to global intensity adjustment according to the difference, so that the difference in pixel intensity between the windward side and the leeward side increases with the increase in rotation speed.

[0032] Optionally, obtaining the difference in rotation angles in adjacent frames and the offset of the imaging device posture according to the spatiotemporal index, and calculating the distortion direction and intensity ratio of the blade edge during continuous motion according to the spatial geometric superposition relationship between the difference and the offset, includes:

[0033] Decomposing the difference in rotation angles in adjacent frames into a horizontal angle change component and a vertical angle change component along the blade rotation plane according to the spatiotemporal index;

[0034] Decompose the offset of the imaging device posture into a horizontal displacement component parallel to the blade rotation direction and a vertical displacement component perpendicular to the rotation direction;

[0035] The horizontal angle change component and the horizontal displacement component, and the vertical angle change component and the vertical displacement component are respectively superimposed in a preset ratio to generate a total horizontal distortion offset and a total vertical distortion offset of the blade edge;

[0036] The distortion direction and intensity ratio of the blade edge during the continuous movement are calculated according to the sum of the vector synthesis direction and absolute value of the total horizontal distortion offset and the total vertical distortion offset.

[0037] Optionally, embedding a blade motion state vector generated by inertial navigation data in the lightweight detection model, adjusting the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges, and outputting a defect detection result matching the blade motion state vector, includes:

[0038] Dividing the rotation speed intervals according to the real-time rotation speed in the blade motion state vector generated by the inertial navigation data, wherein the rotation speed intervals include a steady-state operation interval and a variable speed operation interval;

[0039] The sensitivity threshold of the coating shedding defect in the steady-state operation range is set to be higher than that of the blade crack defect, and the sensitivity threshold of the blade crack defect in the variable speed operation range is set to be higher than that of the coating shedding defect;

[0040] Inputting the three-dimensional coordinate change rate and rotational acceleration parameters in the blade motion state vector into the feature fusion layer of the lightweight detection model, and dynamically adjusting the response weights of the nodes of the lightweight detection model to the defect features;

[0041] According to the adjusted response weights, the defect probability values output by the lightweight detection model are screened, the defect types and defect locations exceeding the sensitivity threshold are retained, and a defect detection result matching the blade motion state vector is generated.

[0042] In a second aspect, the present application provides a wind turbine blade surface defect detection system based on drone visual images, comprising:

[0043] An acquisition module is used to collect surface image sequences of high-speed rotating wind turbine blades through a UAV and simultaneously obtain blade motion trajectory data output by an inertial navigation module;

[0044] a compensation module for performing spatiotemporal correlation between the blade motion trajectory data and the surface image sequence, and performing frame-by-frame compensation for blade edge distortion caused by motion blur in the surface image sequence based on a dynamic matching relationship between the blade rotation angle and the imaging device posture extracted from the spatiotemporal correlation result;

[0045] An enhancement module is configured to enhance the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the compensated surface image sequence by dynamic weight allocation, and generate a blade surface feature map in combination with the spatiotemporal continuity constraint of the blade motion trajectory data;

[0046] A migration module, configured to adjust the attention weight of the knowledge distillation framework based on the dynamic distribution ratio of the high-frequency vibration area and the static area in the blade surface feature map, and migrate the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weight;

[0047] An output module is used to embed the inertial navigation module in the lightweight detection model, generate a blade motion state vector in real time according to the inertial navigation data output by the inertial navigation module, and output a defect detection result that matches the blade motion state vector by adjusting the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges.

[0048] In a third aspect, the present application provides a computing device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a wind turbine blade surface defect detection method based on drone visual images as described in the first aspect above.

[0049] In a fourth aspect, the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements a method for detecting surface defects of wind turbine blades based on drone visual images as described in the first aspect.

[0050] In this application, a surface image sequence of a high-speed rotating wind turbine blade is collected by a drone, and blade motion trajectory data output by an inertial navigation module is synchronously obtained; the blade motion trajectory data is temporally and spatially correlated with the surface image sequence, and a dynamic matching relationship between the blade rotation angle and the imaging device posture extracted from the temporal and spatial correlation result is used to compensate for the blade edge distortion caused by motion blur in the surface image sequence frame by frame through the dynamic matching relationship; the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the surface image sequence after compensation is enhanced by dynamic weight allocation, and the temporal and spatial continuity of the blade motion trajectory data is combined with the temporal and spatial continuity of the blade motion trajectory data. A blade surface feature map is generated by a beam; based on the dynamic distribution ratio of the high-frequency vibration area and the static area in the blade surface feature map, the attention weight of the knowledge distillation framework is adjusted, and the feature response pattern in the pre-trained high-precision defect detection model is transferred to the lightweight detection model according to the adjusted attention weight; the inertial navigation module is embedded in the lightweight detection model, and the blade motion state vector is generated in real time according to the inertial navigation data output by the inertial navigation module, and the defect detection result matching the blade motion state vector is output by adjusting the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges.

[0051] The technical solution of this application has the following beneficial effects:

[0052] By synchronously collecting HDR images and inertial navigation data from drones, precise spatiotemporal alignment of the blade surface image and motion trajectory under high-speed rotation is achieved, suppressing the problems of light overexposure / underexposure and positioning drift caused by high-speed motion; based on the dynamic matching relationship between the rotation angle and the imaging device posture, motion blur distortion is compensated, the geometric deformation of the blade edge caused by high-speed rotation is eliminated, and the contour clarity of the defect area is improved; the texture contrast difference between the windward and leeward sides is enhanced through dynamic weight distribution, and the spatiotemporal continuity of the motion trajectory is combined to generate a feature map that integrates dynamic lighting characteristics, solving the problem of weakening defect features under speed changes; based on the ratio of high-frequency vibration and static area distribution, the knowledge distillation attention weight is adjusted to achieve adaptive migration of high-precision model features to lightweight models, balancing detection accuracy and computational efficiency; embedding real-time motion state vectors and dynamically adjusting the sensitivity threshold enables the detection model to adapt to the dynamic changes of defect features under different speeds, reducing the false detection rate and missed detection rate in high-speed scenarios.

[0053] Furthermore, a spatiotemporal index is extracted from the blade motion trajectory data. The distortion direction and intensity ratio are calculated based on the spatial geometric superposition of the rotation angle differences between adjacent frames and the imaging device's position offset. A pixel displacement compensation vector is generated in the opposite direction of rotation. Layer-by-layer offset correction and edge sharpening are used to eliminate residual motion blur pixels, resulting in a compensated, clear image sequence. By modeling the geometric association between motion trajectory and image data, the blade edge deformation and blur caused by high-speed rotation are precisely offset, improving the visibility of surface defect outlines and the quality of input data for the detection algorithm, providing a high-fidelity image foundation for subsequent dynamic feature extraction and defect classification.

[0054] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0056] Figure 1 A flow chart of a method for detecting surface defects of wind turbine blades based on UAV visual images provided by the present application is shown;

[0057] Figure 2 A schematic diagram of the structure of a wind turbine blade surface defect detection system based on drone visual images provided by the present application is shown;

[0058] Figure 3 A schematic structural diagram of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0059] In order to enable people skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0060] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0061] Researchers have found that surface defect detection in high-speed rotating wind turbine blades faces challenges such as severe motion blur, uneven dynamic illumination, weakened defect features, and poor dynamic adaptability of detection models. Traditional methods struggle to balance detection accuracy and real-time performance under complex operating conditions. Based on this, a method for detecting surface defects in wind turbine blades based on drone visual images is proposed. Specifically, a drone synchronously collects blade surface image sequences and inertial navigation data, constructs a spatiotemporal correlation model to compensate for edge distortion caused by high-speed motion. A dynamic weight allocation strategy is used to enhance the texture contrast difference between the windward and leeward sides of the blade, and motion trajectory constraints are combined to generate an anti-interference feature map. The attention weights of the knowledge distillation framework are regulated based on the dynamic distribution ratio of the feature map, migrating high-precision model features to a lightweight model. A real-time motion state vector is embedded to achieve dynamic adaptation of the defect sensitivity threshold. This method can address the core issues of imbalance between dynamic and static features and high false detection rates in high-speed rotating scenarios through multi-source data collaboration and dynamic feature enhancement, significantly improving the detection accuracy and system robustness of defects such as cracks and coating peeling.

[0062] The technical solution of the present application can be applied to dynamic blade defect detection scenarios under high-speed flight conditions.

[0063] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0064] Figure 1A flow chart of a method for detecting surface defects of wind turbine blades based on drone visual images is provided for the embodiment of the present application. Figure 1 As shown, the method includes:

[0065] 101. Use a drone to collect surface image sequences of high-speed rotating wind turbine blades and simultaneously obtain blade motion trajectory data output by an inertial navigation module;

[0066] In this step, HDR dynamic imaging technology is a technology that dynamically adjusts multi-frame exposure parameters according to the real-time rotation speed of the blades. It is used to suppress over-exposure and under-exposure areas caused by high-speed movement to ensure the integrity of image details.

[0067] In an embodiment of the present application, the real-time rotation speed of a high-speed rotating wind turbine blade is obtained by HDR dynamic imaging technology, and a multi-frame exposure parameter combination required for HDR dynamic imaging is determined, wherein a high exposure parameter is used to capture a low-light area on the leeward side of the blade, and a low exposure parameter is used to capture a high-light area on the windward side of the blade; the blade surface image is continuously collected by an HDR dynamic imaging device according to the multi-frame exposure parameter combination, and the inertial navigation module is triggered to generate a corresponding synchronization timestamp mark when the blade surface image is collected; the synchronization timestamp mark is bound to the three-dimensional spatial coordinates and rotation angle data of the blade output by the inertial navigation module to generate blade motion trajectory data corresponding to each frame of the surface image; the illumination area of multiple frames of different exposure images within the same rotation cycle is fused to eliminate the halo artifacts caused by high-speed motion in the boundary area between the windward side and the leeward side of the blade, and obtain a surface image sequence with suppressed overexposure and underexposure.

[0068] Assuming the drone captures leaf images at a rate of 150 frames per second, HDR dynamic imaging adjusts exposure parameters to 1 / 200s (windward) and 1 / 800s (leeward) based on a rotational speed of 2500 rpm. The inertial navigation module synchronously generates timestamp coordinates (X=10.3m, Y=2.8m, Z=30°) and a rotation angle (30°). After fusing five frames, the halo intensity at the junction is reduced by 60%, resulting in a clear image sequence.

[0069] 102. Performing spatiotemporal correlation on the blade motion trajectory data and the surface image sequence, and based on a dynamic matching relationship between the blade rotation angle and the imaging device posture extracted from the spatiotemporal correlation result, compensating for blade edge distortion caused by motion blur in the surface image sequence frame by frame through the dynamic matching relationship;

[0070] In this step, the dynamic matching relationship is the superposition association between the blade rotation angle difference and the imaging device posture offset in spatial geometry, which is used to describe the distortion direction and intensity of motion blur.

[0071] In an embodiment of the present application, first, the timestamp, three-dimensional coordinates and rotation angle corresponding to each frame image are extracted from the blade motion trajectory data to generate a spatiotemporal index table; secondly, based on the rotation angle difference of adjacent frames in the spatiotemporal index table (such as frame 1: 30°, frame 2: 35°, difference 5°) and the imaging device posture offset (such as horizontal displacement of 2mm), the spatiotemporal geometric superposition algorithm is used to calculate the blade edge distortion direction (such as 45° horizontal left deviation) and intensity ratio (horizontal displacement accounts for 70%); again, according to the distortion direction and intensity ratio, a pixel displacement compensation vector (such as 4 pixels left, 1 pixel down) is generated, and the inverse motion compensation algorithm is used to perform reverse offset correction on the edge pixels of the current frame layer by layer; finally, a local gradient sharpening algorithm (such as Sobel operator to enhance edge gradient) is used to eliminate residual blur according to the correction result to generate a compensated image sequence.

[0072] For example, continuing with the above example, the rotation angle difference between adjacent frames in the spatiotemporal index is 8°, the horizontal displacement of the imaging device is 4mm, and the geometric superposition algorithm is used to calculate the distortion direction as 25° to the lower right, with an intensity ratio of 1:0.6. A compensation vector is generated (6 pixels to the right, 3 pixels to the bottom) to correct the edge pixels. After sharpening with the Sobel operator, the crack edge width is restored from a blurred 6 pixels to a clear 3 pixels.

[0073] 103. Generating a blade surface feature map by combining the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the surface image sequence after strengthening compensation by dynamic weight allocation and the spatiotemporal continuity constraint of the blade motion trajectory data;

[0074] In this step, dynamic weight allocation dynamically adjusts the contrast enhancement weights for the windward and leeward regions based on speed changes, with the weights increasing / decreasing with speed. The spatiotemporal continuity constraint generates a weighted accumulation sequence by accumulating weights across frames based on the temporal sequence and spatial position of the motion trajectory data.

[0075] In an embodiment of the present application, first, an increasing weight value (0.7→0.9) is assigned to the windward side in the compensated surface image sequence according to the real-time rotation speed (such as 2800 rpm), and a decreasing weight value (0.4→0.2) is assigned to the leeward side; secondly, based on the continuous rotation angle difference in the motion trajectory data (such as an increase of 2° per frame), a cross-frame accumulation algorithm is used to accumulate the weight values of the same blade area (the windward side weight is accumulated to 1.3); again, according to the accumulation result, a nonlinear gain adjustment is performed on the pixel intensity on the windward side (such as intensity × 1.3), and an inhibition adjustment is performed on the leeward side (such as intensity × 0.6); finally, a multi-frame spatiotemporal superposition algorithm is used to extract the intensity mutation contour between the windward side and the leeward side (such as crack width > 2 pixels) to generate a blade surface feature map.

[0076] For example, continuing with the above example, for the compensated image sequence, the initial weight of the windward side is set to 0.85 at a rotation speed of 3000 rpm. After cross-frame accumulation (the cumulative rotation difference of 5 frames is 15°), the weight is increased to 1.25; the pixel intensity on the windward side is increased by 25%, and the pixel intensity on the leeward side is reduced by 45%. After superimposing 12 frames of images, the crack contour line is extracted to generate a blade surface feature map with a contrast difference of 3.2 times.

[0077] 104. Based on the dynamic distribution ratio of the high-frequency vibration area and the static area in the blade surface feature image, adjust the attention weight of the knowledge distillation framework, and migrate the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weight;

[0078] In this step, the knowledge distillation framework is a training framework that transfers the feature response pattern of the high-precision defect detection model to the lightweight model, and realizes feature selective transfer through attention weight adjustment.

[0079] In an embodiment of the present application, first, the area ratio of high-frequency vibration areas (fluctuation amplitude > 50 gray levels) and static areas is counted in the blade surface feature map (e.g., 3:7) to generate a dynamic distribution ratio; secondly, according to the dynamic distribution ratio (e.g., the high-frequency area accounts for 30%), the attention weight coefficient of the high-frequency area in the knowledge distillation framework is proportionally amplified (e.g., ×1.5), and the weight of the static area is reversely reduced (e.g., ×0.5); again, through feature mapping technology, the characteristic response pattern of the high-frequency vibration area in the pre-trained high-precision model (e.g., crack edge gradient characteristics) is migrated to the lightweight model according to the adjusted weights; finally, the basic texture features of the static area (e.g., coating uniformity) are retained to complete the lightweight adaptation of the model.

[0080] For example, continuing with the above example, the high-frequency vibration area (fluctuation amplitude > 60 gray levels, accounting for 28%) is divided in the blade surface feature map, and its attention weight is amplified to 1.4 times at a ratio of 1:2.6; the crack edge gradient features in the high-precision model are migrated to the lightweight model. After migration, the crack response value increases from 0.7 to 0.92, and the static area weight remains at 0.8 times.

[0081] 105. The inertial navigation module is embedded in the lightweight detection model, and a blade motion state vector is generated in real time according to the inertial navigation data output by the inertial navigation module. By adjusting the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges, a defect detection result matching the blade motion state vector is output.

[0082] In this step, the blade motion state vector is a set of dynamic parameters generated from inertial navigation data. It contains key motion state information such as real-time rotational speed and three-dimensional spatial offset. The sensitivity threshold is the critical value of the characteristic response intensity used by the lightweight detection model to determine defects. It is dynamically adjusted with rotational speed to adapt to different operating conditions.

[0083] In an embodiment of the present application, the real-time rotational speed (e.g., 3000 rpm) and the three-dimensional offset (e.g., a horizontal displacement of 4 mm) are first extracted from the inertial navigation data to generate a motion state vector; secondly, the detection interval is divided according to the rotational speed (high speed > 2500 rpm), and a baseline sensitivity threshold is set (e.g., a crack threshold of 0.8); thirdly, the baseline sensitivity threshold is dynamically increased based on the offset (threshold = 0.8 + 0.1 × 4 = 1.2); finally, the characteristic response value output by the lightweight model (e.g., a crack response value of 1.0) is compared with the increased threshold, and a defect result matching the motion state vector is output.

[0084] For example, continuing with the above example, a motion state vector is generated based on the real-time rotational speed of 3200 rpm and the horizontal displacement of 5 mm, and the high-speed interval threshold is dynamically adjusted to 1.3 (benchmark 0.8+0.1×5). The lightweight model detects a crack response value of 1.28, which triggers an alarm after exceeding the threshold. The false detection signal on the leeward side is simultaneously suppressed, and the detection result of the defect position matching the rotational speed is output.

[0085] In order to solve the problem of edge geometric deformation and texture distortion caused by dynamic motion blur in the detection of surface defects of high-speed rotating fan blades, especially the lack of accuracy of traditional compensation methods in strong vibration and rapid lighting switching scenarios. In some embodiments, the blade motion trajectory data is temporally and spatially correlated with the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the imaging device posture extracted from the temporal and spatial correlation results, the blade edge distortion caused by motion blur in the surface image sequence is compensated frame by frame through the dynamic matching relationship to obtain a compensated surface image sequence, including:

[0086] 201. Extracting the three-dimensional spatial positioning data and rotation angle of the blade corresponding to each frame image acquisition moment from the blade motion trajectory data, and generating a spatiotemporal index aligned with the surface image sequence timestamp;

[0087] In step 201 , the spatiotemporal index is a data structure consisting of the three-dimensional spatial positioning data (X / Y / Z coordinates) of the blade aligned with timestamps and the rotation angle, and is used to associate image frames with motion trajectories.

[0088] In an embodiment of the present application, the three-dimensional spatial positioning data of the blade (such as X=12.3m, Y=4.7m, Z=30°) and the rotation angle (such as 45°) corresponding to each frame image are first extracted from the blade motion trajectory data, and the timestamps involved in the three-dimensional spatial positioning data and the rotation angle of the blade are aligned with the image frame through a timestamp matching algorithm (such as the nearest neighbor interpolation method) to generate a spatiotemporal index table. Each row in the table records the frame number, timestamp, coordinates and rotation angle for subsequent calls.

[0089] 202. Obtaining the difference in rotation angles in adjacent frames and the offset of the imaging device posture according to the spatiotemporal index, and calculating the distortion direction and intensity ratio of the blade edge during continuous motion according to the spatial geometric superposition relationship between the difference and the offset;

[0090] In step 202, the spatial geometric superposition relationship is a vector synthesis rule of the rotation angle difference and the imaging device posture offset in three-dimensional space.

[0091] In an embodiment of the present application, the rotation angle difference between adjacent frames (such as frame 1: 30°, frame 2: 35°, difference 5°) and the imaging device posture offset (such as 2mm displacement in the X direction) are first obtained from the spatiotemporal index table; secondly, the difference is decomposed into a horizontal rotation component (such as 3°) and a vertical rotation component (such as 2°), and the posture offset is decomposed into a horizontal displacement (such as 1.5mm) and a vertical displacement (such as 0.5mm); again, the components are superimposed according to preset weights (such as 0.7 for horizontal and 0.3 for vertical), and the distortion direction (such as 40° left deviation) and intensity ratio (such as horizontal intensity 1.05 and vertical intensity 0.45) are calculated.

[0092] 203. Generate a pixel displacement compensation vector for the edge region of the blade in the current frame based on the distortion direction and intensity ratio, wherein the pixel displacement compensation vector performs layer-by-layer offset correction on the pixel position in the opposite direction of blade rotation;

[0093] In step 203, the layer-by-layer offset correction is to shift the pixel position step by step in the opposite direction of the blade rotation, and the offset amount of each layer is 1 / N of the total compensation amount.

[0094] In the embodiment of the present application, first, a pixel displacement compensation vector is generated (such as 7 pixels left and 3 pixels up) based on the distortion direction (such as 40° to the left) and the intensity ratio (such as 1.05 horizontally and 0.45 vertically); secondly, the pixel displacement compensation vector is split into 3 layers for gradual offset correction (such as the first layer shifts 2 pixels left and 1 pixel up) to avoid image tearing caused by a single large offset; finally, the corrected vectors layer by layer are written into the configuration file for the image processing module to call.

[0095] 204. Perform edge contour sharpening processing on the leaf edge distortion caused by motion blur in the surface image sequence according to the pixel displacement compensation vector, eliminate residual blurred pixels in the leaf edge distortion, and generate a compensated surface image sequence.

[0096] In step 204, the residual blurred pixels are grayscale gradient areas in the compensated image that are not completely eliminated due to motion blur. Edge sharpening is to enhance the grayscale mutation characteristics of the defect boundary through a gradient enhancement algorithm.

[0097] In an embodiment of the present application, an adaptive gradient threshold algorithm (such as Canny edge detection) is first applied to the surface image sequence according to the pixel displacement compensation vector, and areas where the grayscale change rate is lower than the threshold (such as <15%) are marked as residual blurred pixels; secondly, a directional sharpening filter (such as anisotropic diffusion) is used to enhance the edge gradient of the surface image sequence along the leaf texture direction to eliminate residual blurred pixels; finally, the sharpened surface image sequence is binarized (such as the Otsu algorithm) to generate a compensated surface image sequence.

[0098] Here's a specific example:

[0099] Assume that a drone is inspecting surface cracks on an aircraft turbine blade at a speed of 200 km / h and a dive angle of 15°, with the turbine speed at 3200 rpm. Step 201 generates a spatiotemporal index table, recording the timestamp of frame 100 at 163,000 μs, coordinates X = 15.2 m, Y = 3.8 m, Z = 25°, and a rotation angle of 30°. In step 202, the rotation angle difference between adjacent frames (frames 100 and 101) is 5°, and the device is horizontally displaced 2 mm. The calculated distortion direction is 35° to the left, with an intensity ratio of 0.8 horizontally and 0.2 vertically. Step 203 generates a compensation vector that shifts the image left by 6 pixels and up by 1 pixel, and performs correction in three layers (each layer shifting 2 pixels to the left). Step 204 reduces the crack edge width from a blurry 4 pixels to a clear 1 pixel after sharpening, and the contrast ratio of the crack region in the binary surface image sequence reaches 8:1.

[0100] Steps 201-204 accurately associate images and motion data through spatiotemporal indexing, quantify the distortion direction and intensity based on geometric superposition relationships, eliminate motion blur through layered compensation, and enhance defect contours through directional sharpening. Ultimately, in high-speed rotation scenarios, the crack detection accuracy is improved and the false detection rate is reduced, while the compensation processing time is controlled with high precision to meet real-time detection requirements.

[0101] To address the problem of texture contrast imbalance between the windward and leeward sides of high-speed rotating fan blades caused by uneven dynamic illumination during surface defect detection, particularly the technical bottleneck of weakened defect features and increased false detection rates at high rotational speeds, in some embodiments, the texture contrast differences between the windward and leeward sides of the blade at different rotational speeds in the surface image sequence after dynamic weight allocation and compensation are combined with the spatiotemporal continuity constraints of the blade motion trajectory data to generate a blade surface feature map, including:

[0102] 301. Dynamically assign contrast enhancement weights to the windward and leeward sides of the blade in the compensated surface image sequence according to the change in blade rotation speed, wherein the contrast enhancement weight on the windward side increases as the blade rotation speed increases, and the contrast enhancement weight on the leeward side decreases as the blade rotation speed increases;

[0103] In step 301 , the contrast enhancement weight value is a numerical parameter dynamically assigned to the windward side and the leeward side according to the blade rotation speed, and is used to adjust the pixel intensity difference.

[0104] In an embodiment of the present application, the blade rotation speed (such as 2500 rpm) is first obtained in real time, and the contrast enhancement weight value of the current frame is dynamically allocated according to the preset speed and weight mapping table (for example, for every 100 rpm increase in speed, the windward side weight is +0.05, and the leeward side weight is -0.03), and the allocated weight value is written into the image metadata for subsequent accumulation and adjustment calls.

[0105] 302. Based on the rotation angle difference between consecutive frames in the blade motion trajectory data, the contrast enhancement weight values are accumulated in adjacent frame images of the same blade region to generate a weight accumulation sequence consistent with the blade rotation direction;

[0106] In step 302 , the weight accumulation sequence is the accumulation result of contrast enhancement weight values of the same leaf region in consecutive frames, reflecting the weight accumulation effect in the rotation direction.

[0107] In an embodiment of the present application, the rotation angle difference between adjacent frames (such as 5°) is first extracted from the blade motion trajectory data, and the displacement ratio of the same blade area in the rotation direction is calculated based on the rotation angle difference (such as 0.2 pixel displacement per degree); secondly, according to the displacement ratio, the contrast enhancement weight value is skipped across adjacent frames for accumulation (such as the windward side weight of 0.7 is accumulated to 0.7+0.05×5=0.95) to generate a weight accumulation sequence consistent with the rotation direction; finally, the accumulation result is stored as a time series weight array.

[0108] 303. Performing positive gain adjustment and negative suppression adjustment on the pixel intensities on the windward side and the leeward side respectively according to the numerical distribution of the weighted cumulative sequence, so that the intensity difference of the pixel intensities increases with the rotation speed;

[0109] In step 303, the forward gain adjustment is to nonlinearly amplify the windward side pixel intensity according to the weighted accumulated value, and the reverse suppression adjustment is to proportionally reduce the leeward side pixel intensity according to the weighted accumulated value.

[0110] In an embodiment of the present application, first, the weight values of the windward side and the leeward side of the current frame (such as 1.1 and 0.6) are read from the weight accumulation sequence; secondly, the gain coefficient and the suppression coefficient are set, and the piecewise linear transformation algorithm is used to perform gain adjustment on the windward side pixel intensity (such as intensity × 1.1), and to perform suppression adjustment on the leeward side pixel intensity (such as intensity × 0.6); finally, the adjusted image data is generated according to the adjustment result, in which the contrast of the crack area on the windward side is increased to 2.5 times that of the background.

[0111] 304. Superimpose the adjusted pixel intensities across frames according to the time sequence of the leaf motion trajectory data, and generate a leaf surface feature map containing dynamic texture features of the leaf surface based on the intensity mutation area contour formed after superposition.

[0112] In step 304, cross-frame overlay is to overlay multiple frames of adjusted images pixel by pixel in chronological order to enhance dynamic texture features. The intensity mutation area contour is a continuous boundary area in the overlayed image where the difference in pixel intensity exceeds a set threshold.

[0113] In an embodiment of the present application, the adjusted multi-frame images are first arranged in chronological order according to the blade motion trajectory data; secondly, a sliding window superposition algorithm (window size 5 frames) is used to accumulate the pixel intensities of the same blade area in the arranged chronological order (such as the cumulative intensity of the crack area on the windward side is increased by 120%); finally, the edge gradient detection algorithm is used to extract the contours of the intensity mutation area in the accumulation result (such as the crack width > 2 pixels) to generate a blade surface feature map containing dynamic texture features.

[0114] Here's a specific example:

[0115] Assume that a drone is inspecting a 1.2mm-wide microcrack on the windward side of an aircraft turbine blade at a speed of 220 km / h and a yaw angle of 10°. The turbine speed is 3200 rpm. Step 301 assigns a weight of 0.85 to the windward side and 0.25 to the leeward side based on the 3200 rpm speed. In step 302, the rotation angle difference between adjacent frames is 8°, and the windward weight is accumulated to 0.85 + 0.05 × 8 = 1.25. Step 303 increases the pixel intensity on the windward side by 25% (×1.25) and decreases it on the leeward side by 30% (×0.7), resulting in a 3:1 contrast difference in the crack area. Step 304 extracts the crack contour by superimposing 10 frames. The resulting blade surface feature map clearly shows the crack width as 2 pixels, and the background noise intensity is reduced by 40%.

[0116] Steps 301-304 enhance the contrast difference between the two sides of the blade through dynamic weight distribution, and combine cross-frame accumulation and overlay to generate a high-resolution feature map, significantly improving the visibility of tiny cracks and coating peeling defects in high-speed rotation scenarios, reducing the false detection rate caused by uneven lighting, and at the same time, through real-time weight adjustment and overlay processing, ensuring the stability and adaptability of the detection system when the rotation speed changes drastically.

[0117] To address the problem of insufficient model generalization capability in high-speed rotating fan blade surface defect detection due to the confusion between dynamic vibration and static area features, especially the technical bottleneck of low sensitivity of lightweight detection models to tiny defects under high-frequency vibration interference. In some embodiments, the attention weight of the knowledge distillation framework is adjusted based on the dynamic distribution ratio of high-frequency vibration areas and static areas in the blade surface feature map, and the feature response pattern of the pre-trained high-precision defect detection model is transferred to the lightweight detection model according to the adjusted attention weight, including:

[0118] 401. Extracting the fluctuation amplitude of the blade surface texture intensity in the blade surface feature map within adjacent time intervals, and dividing the boundary range between the high-frequency vibration region and the static region according to the fluctuation amplitude;

[0119] In step 401 , the boundary range is a spatial dividing line between a high-frequency vibration region and a static region in the feature map.

[0120] In an embodiment of the present application, first, 5 consecutive frames of images are selected from the blade surface feature map, and the sliding window standard deviation algorithm is used to calculate the fluctuation amplitude of the texture intensity of each pixel point; secondly, a dynamic threshold is set (such as the standard deviation threshold = 45), and pixels whose fluctuation amplitude exceeds the threshold are marked as high-frequency vibration areas, and the rest are static areas; finally, the adjacent high-frequency pixel points in the high-frequency vibration area are merged through the spatial clustering algorithm to generate a closed boundary range.

[0121] 402. Counting the area proportions of the high-frequency vibration region and the static region in the blade surface characteristic map to generate a numerical relationship representing the dynamic distribution ratio of the region;

[0122] In step 402 , the numerical relationship is a ratio value with the high-frequency region area as the numerator and the static region area as the denominator.

[0123] In an embodiment of the present application, the connected domains of the closed boundary range are first marked to count the total number of pixels in the high-frequency vibration area (e.g., 1200 pixels) and the total number of pixels in the static area (e.g., 3600 pixels); secondly, the area ratios of the high-frequency and static areas are calculated, for example, the area ratio of the high-frequency area = the total number of pixels in the high-frequency vibration area / (the total number of pixels in the high-frequency vibration area + the total number of pixels in the static area); finally, the ratio of the area ratio is quantified into a numerical relationship (1:3) and written into the configuration file of the knowledge distillation framework.

[0124] 403. Based on the numerical relationship, the attention weight coefficient corresponding to the high-frequency vibration area in the knowledge distillation framework is geometrically amplified, and the attention weight coefficient corresponding to the static area is reversely reduced;

[0125] In step 403, the ratio of the area proportion of the high-frequency vibration region to the area proportion of the static region is extracted from the numerical relationship (e.g., 1:3), and the ratio is converted into a weight adjustment scaling factor. A mapping relationship between the spatial position of the blade surface feature map and the attention weight coefficient is established in the knowledge distillation framework (e.g., coordinates X=100-200, Y=50-150 correspond to the weight index group), and the weight coefficient distribution range corresponding to the high-frequency vibration region and the static region is located according to the mapping relationship. Based on the weight adjustment scaling factor, all attention weight coefficients within the high-frequency vibration region distribution range are geometrically amplified layer by layer, where the amplification factor is the square root of the weight adjustment scaling factor (e.g., original weight 0.6 → 0.6×1.732≈1.04). Simultaneously, all attention weight coefficients within the static region distribution range are reversely reduced, where the reduction ratio is inversely related to the weight adjustment scaling factor (e.g., original weight 0.9 → 0.9×1 / 3=0.3).

[0126] 404. Pre-train a high-precision defect detection model according to the knowledge distillation framework, obtain a characteristic response pattern matching the high-frequency vibration area and the static area in the high-precision defect detection model, and map the adjusted attention weight coefficient layer by layer to the lightweight detection model based on the characteristic response pattern.

[0127] In step 404 , the characteristic response pattern is a characteristic activation pattern (eg, gradient direction histogram feature) for a specific region (eg, crack, coating peeling) in a high-precision defect detection model.

[0128] In an embodiment of the present application, first, the characteristic response pattern corresponding to the high-frequency vibration area (such as the gradient direction distribution of the crack edge) is extracted from the pre-trained high-precision defect detection model; secondly, based on the adjusted attention weight coefficient (such as 1.04 for high frequency and 0.3 for static), the feature mapping technology is used to migrate the extracted characteristic response pattern to the corresponding level of the lightweight model (such as convolutional layer 3) according to the weight ratio; finally, the weight parameters of the static area in the corresponding level of the lightweight detection model are frozen, and only the parameters of the high-frequency vibration area are updated according to the migrated characteristic response pattern to complete the model adaptation.

[0129] Here's a specific example:

[0130] A drone was flying at 240 km / h and a roll angle of 5° to detect microcracks 1.2 mm wide on the windward side of an aircraft turbine blade. The turbine speed was 3500 rpm. In step 401, the standard deviation of the texture intensity in the crack region across five consecutive image frames reached 65 grayscale levels, marking it as a high-frequency vibration region (20% area), while the static region accounted for 80%. In step 402, a numerical relationship of 1:4 (20%:80%) was generated and written to a configuration file. In step 403, a scaling factor of 4 was extracted, calculating the square root magnification factor of 2.0 and the reciprocal reduction factor of 0.25. The weight of the high-frequency region was amplified from 0.5 to 1.0 (0.5 × 2.0), and the weight of the static region was reduced from 0.8 to 0.2 (0.8 × 0.25). In step 404, the gradient direction features of the crack region from the high-precision model were transferred to the third convolutional layer of the lightweight model. After the transfer, the crack response value increased from 0.5 to 0.95, while the weight of the static region remained unchanged at 0.2.

[0131] Steps 401-404 dynamically divide high-frequency vibration regions and quantify the distribution ratio to generate weight adjustment factors. Combined with a weight balancing strategy of square root amplification and reciprocal reduction, this strategy specifically strengthens defect feature migration in high-frequency regions while suppressing interference signals in static regions. This significantly improves the detection accuracy of the lightweight model in high-speed vibration scenarios. Through feature mapping and hierarchical adaptation, the model's sensitivity and generalization capabilities for small cracks in complex dynamic environments are ensured, reducing false detection rates and meeting real-time requirements.

[0132] In order to solve the problem of weakened texture contrast between the windward and leeward sides caused by dynamic uneven illumination and motion blur in the detection of surface defects on high-speed rotating fan blades, especially the technical defect that the defect features are drowned out by background noise at high rotation speeds, in some embodiments, the pixel intensities on the windward and leeward sides are respectively adjusted in a positive gain and a negative suppression manner based on the numerical distribution of the weighted accumulation sequence, so that the intensity difference of the pixel intensities increases with the rotation speed, including:

[0133] 501. Associating the accumulated value of each frame in the weight accumulation sequence with the current rotation speed of the blade to generate a dynamic gain coefficient and a dynamic suppression coefficient for the windward side and the leeward side, respectively;

[0134] In step 501, the dynamic gain coefficient is a windward pixel intensity amplification parameter generated based on the weighted cumulative value and the blade rotation speed. The dynamic suppression coefficient is a leeward pixel intensity suppression parameter generated based on the weighted cumulative value and the blade rotation speed.

[0135] In the embodiment of the present application, the accumulated values of the windward side and the leeward side of the current frame are first extracted from the weight accumulation sequence (e.g., the accumulated value of the windward side is 1.5, and the accumulated value of the leeward side is 0.8), and combined with the real-time rotation speed (e.g., 2800 rpm), a nonlinear interpolation algorithm is used to generate the windward side dynamic gain coefficient (1.5×0.001×2800≈1.32) and the leeward side dynamic suppression coefficient (0.8−0.0005×2800≈0.66); then the two coefficients are written into the image processing parameter table for subsequent adjustment and call.

[0136] 502. In the compensated surface image sequence, mark the boundary line between the windward side and the leeward side along the extension direction of the blade surface texture, and expand a set width to both sides based on the boundary line to form a gradual adjustment area;

[0137] In step 502 , the gradual adjustment area is a pixel adjustment transition zone (eg, a smooth transition area with a width of 20 pixels) extending toward both sides based on the dividing line.

[0138] In an embodiment of the present application, first, the main texture direction of the blade surface (such as along the blade axis) is extracted from the compensated surface image sequence, and a morphological expansion algorithm is used to extend along the main texture direction to generate a dividing line; secondly, with the dividing line as the center, the set width is expanded to the windward side and the leeward side (such as 10 pixels each) to form a gradient adjustment area with a total width of 20 pixels; finally, the coordinates of the gradient adjustment area are written into the mask file.

[0139] 503. In the windward side range and the leeward side range of the gradual adjustment zone, perform superposition enhancement and subtraction reduction on the intensity value of each pixel according to the dynamic gain coefficient and the dynamic suppression coefficient respectively;

[0140] In step 503, superposition enhancement is to nonlinearly enhance the pixel intensity on the windward side of the gradient adjustment region according to the dynamic gain coefficient. Subtraction attenuation is to proportionally reduce the pixel intensity on the leeward side of the gradient adjustment region according to the dynamic suppression coefficient.

[0141] In the embodiment of the present application, the coordinates of the gradient adjustment area are first read from the mask file (such as X=100-120, Y=50-70); secondly, the pixel intensity is gradient enhanced according to the dynamic gain coefficient (1.32) within the windward side range of the gradient adjustment area coordinates (from the dividing line to the outer edge of the windward side, the gain amplitude increases linearly from 1.0 to 1.32), and the pixel intensity is gradient suppressed according to the dynamic suppression coefficient (0.66) within the leeward side range (from the dividing line to the outer edge of the leeward side, the suppression amplitude decreases linearly from 1.0 to 0.66).

[0142] 504. Calculate the difference between the dynamic gain coefficient and the dynamic suppression coefficient based on the current rotation speed of the blade, and perform global intensity adjustment on the entire area of the windward side and the leeward side outside the gradient adjustment area according to the difference, so that the difference in pixel intensity between the windward side and the leeward side increases with the increase in rotation speed.

[0143] In step 504, the global intensity adjustment is to perform a global intensity adjustment on the entire area of the windward side and the leeward side outside the gradual adjustment area.

[0144] In the embodiment of the present application, first, numerical values are extracted from the dynamic gain coefficient and the dynamic suppression coefficient generated in step 501, the difference between the two is directly calculated, and the difference is nonlinearly corrected based on the real-time rotational speed (such as difference × rotational speed correction factor, rotational speed correction factor = 1 + 0.0001 × rotational speed) to obtain the final adjustment difference; finally, according to the final adjustment difference, the intensity of the entire area on the windward side outside the gradual adjustment zone is proportionally increased, and at the same time, the intensity of the entire area on the leeward side is proportionally suppressed, so that the intensity difference between the two sides dynamically expands with the rotational speed.

[0145] Here's a specific example:

[0146] Assume that a UAV is testing a 2.0mm wide surface crack on the windward side of an aircraft turbine blade at a flight speed of 250km / h and a climb angle of 8°, with the turbine rotating at 3500rpm. In step 501, the weighted cumulative values are 1.8 on the windward side and 0.7 on the leeward side. Combined with a rotational speed of 3500 rpm, a dynamic gain coefficient of 1.89 and a suppression coefficient of 0.15 are generated. In step 502, a demarcation line is demarcated along the main crack extension direction to form a 30-pixel wide gradient adjustment zone. In step 503, the intensity gradient on the windward side within the gradient zone is increased to 1.89 times, and the intensity gradient on the leeward side is suppressed to 0.15 times. In step 504, the difference is calculated: gain coefficient 1.89 − suppression coefficient 0.15 = 1.74. This difference is then corrected: 1.74 × (1 + 0.0001 × 3500) = 1.74 × 1.35 ≈ 2.35. Finally, global adjustment is completed: the overall intensity on the windward side is increased by 2.35 × 5% = 11.75%, and the leeward side is suppressed by 2.35 × 8% = 18.8%. The final difference between the two sides is expanded to 22:1.

[0147] Steps 501-504 significantly enhance the texture contrast difference between the windward and leeward sides of the high-speed rotation scene by generating dynamic gain and suppression coefficients, precisely dividing the gradient adjustment zone, and coordinating local and global intensity adjustments. This effectively suppresses the weakening of defect features caused by dynamic uneven illumination, improving the visibility of small cracks and coating peeling defects. Combined with a difference-driven global intensity adaptation mechanism, this ensures the stability of the inspection system during drastic speed changes, reduces background noise interference, and achieves highly robust dynamic defect detection.

[0148] In order to solve the problem of inaccurate edge distortion modeling caused by complex motion coupling in the detection of surface defects in high-speed rotating fan blades, especially the technical defect of compensating for error accumulation under the combined effects of dynamic rotation and device posture offset, in some embodiments, the method of obtaining the difference in rotation angles in adjacent frames and the offset of the imaging device posture according to the spatiotemporal index, and calculating the distortion direction and intensity ratio of the blade edge during continuous motion according to the spatial geometric superposition relationship between the difference and the offset, includes:

[0149] 601. Decompose the difference in rotation angles in adjacent frames into a horizontal angle change component and a vertical angle change component along the blade rotation plane according to the spatiotemporal index;

[0150] In step 601, the horizontal angle variation component is the projection component of the blade rotation angle difference between adjacent frames in the blade rotation plane, and the vertical angle variation component is the projection component of the blade rotation angle difference between adjacent frames in the direction perpendicular to the rotation plane.

[0151] In an embodiment of the present application, the rotation angle difference between adjacent frames is first obtained from the spatiotemporal index (such as frame 1: 30°, frame 2: 35°, difference 5°), and the blade rotation plane coordinate system is established; secondly, the angle difference vector is projected into the blade rotation plane coordinate system, and the horizontal angle change component (such as 4.3°) and the vertical angle change component (such as 2.8°) are calculated to generate a component data table for subsequent call.

[0152] 602. Decompose the offset of the imaging device posture into a horizontal displacement component parallel to the blade rotation direction and a vertical displacement component perpendicular to the rotation direction;

[0153] In step 602, the horizontal displacement component is the projection of the imaging device posture offset in the blade rotation direction, and the vertical displacement component is the projection of the imaging device posture offset in the direction perpendicular to the rotation direction.

[0154] In an embodiment of the present application, the imaging device posture offset (such as 4mm displacement in the X direction and 2mm displacement in the Y direction) is first extracted from the spatiotemporal index. Based on the blade rotation direction vector (such as the unit vector [0.8, 0.6]), the vector projection algorithm is used to calculate the dot product of the imaging device posture offset and the vector parallel to / perpendicular to the blade rotation direction to obtain the horizontal displacement component (such as 4×0.8+2×0.6=4.4mm) and the vertical displacement component (such as 4×0.6−2×0.8=0.8mm), and generate a displacement component table.

[0155] 603. Superimpose the horizontal angle change component and the horizontal displacement component, and the vertical angle change component and the vertical displacement component according to a preset ratio to generate a total horizontal distortion offset and a total vertical distortion offset of the blade edge.

[0156] In step 603, the total horizontal distortion offset is the weighted superposition result of the horizontal angle change component and the horizontal displacement component. The total vertical distortion offset is the weighted superposition result of the vertical angle change component and the vertical displacement component.

[0157] In the embodiment of the present application, the horizontal angle change component (such as 4.3°) and the horizontal displacement component (such as 4.4mm) in the component data table are first read, and superimposed according to the preset weights (such as angle weight 0.5 and displacement weight 0.2) to generate a total horizontal distortion offset (such as 4.3×0.5+4.4×0.2≈3.15); secondly, the vertical components (2.8° and 0.8mm) are superimposed with the same weight to generate a total vertical distortion offset (2.8×0.5+0.8×0.2≈1.56), which is written into the distortion parameter table.

[0158] 604. Calculate the distortion direction and intensity ratio of the blade edge during the continuous motion process according to the sum of the vector synthesis direction and absolute value of the total horizontal distortion offset and the total vertical distortion offset.

[0159] In step 604, the vector composite direction is the composite direction angle of the total horizontal and vertical distortion offsets. The intensity ratio is the sum of the absolute values of the total horizontal and vertical distortion offsets.

[0160] In the embodiment of the present application, the total horizontal distortion offset (3.15) and the total vertical distortion offset (1.56) are first read from the distortion parameter table generated in step 603, and the distortion direction is calculated using a vector synthesis algorithm (using the inverse trigonometric function arctan(1.56 / 3.15)≈26.5°). Next, the sum of the absolute values (3.15+1.56=4.71) is calculated as the intensity ratio. Finally, the distortion direction and intensity ratio are written to the compensation configuration file.

[0161] Here's a specific example:

[0162] Assume that a drone is inspecting cracks in the root of an aircraft turbine blade at a speed of 210 km / h and a sideslip angle of 12°, with the turbine speed at 3200 rpm. In step 601, the rotation angle difference between adjacent frames is 6°, which is decomposed into a horizontal component of 5.2° (along the rotation plane) and a vertical component of 3.0° (in the pitch direction). In step 602, the imaging device pose offsets (X = 5 mm, Y = 3 mm) are decomposed into a horizontal displacement of 4.1 mm (along the rotation direction) and a vertical displacement of 2.2 mm (in the radial direction). In step 603, the total horizontal offset is calculated as 5.2° × 0.5 + 4.1 mm × 0.2 ≈ 3.42, and the total vertical offset is calculated as 3.0° × 0.5 + 2.2 mm × 0.2 ≈ 1.94. In step 604, the resulting vector direction angle, arctan(1.94 / 3.42) ≈ 29.6°, is calculated as the distortion direction, with an intensity ratio of 3.42 + 1.94 = 5.36, which is used for subsequent motion blur compensation.

[0163] Steps 601-604 precisely quantify the dynamic coupling effect of blade edge distortion under high-speed rotation through multidimensional decomposition, weighted superposition, and vector synthesis of rotation angle and pose offsets. This addresses the inaccuracy of traditional single-dimensional compensation models and significantly improves the accuracy of motion blur compensation. Combined with the coordinated calculation of horizontal and vertical components, this ensures geometric deformation correction and strength restoration of defect edges in complex motion scenarios, providing a reliable data foundation for high-precision defect detection.

[0164] To address the problem of misjudgment of defect types caused by sudden changes in speed and flight attitude during dynamic defect detection of turbine blades in high-speed flight, particularly the technical defect of confusion between crack and coating shedding characteristics during variable speed operation, in some embodiments, the lightweight detection model embeds a blade motion state vector generated by inertial navigation data, and adjusts the sensitivity threshold of the lightweight detection model to blade cracks and coating shedding defects in different speed ranges to output defect detection results that match the blade motion state vector, including:

[0165] 701. Divide the rotation speed intervals according to the real-time rotation speed in the blade motion state vector generated by inertial navigation data, wherein the rotation speed intervals include a steady-state operation interval and a variable speed operation interval;

[0166] In step 701, the steady-state operation range is a stable working state range in which the turbine speed fluctuation range is less than a preset threshold. The variable-speed operation range is a dynamic adjustment state range in which the turbine speed change rate exceeds a preset threshold.

[0167] In an embodiment of the present application, the real-time speed (such as 3050 rpm) and its rate of change (such as 80 rpm / s) are first extracted from the blade motion state vector, and the standard deviation of the speed within 10 consecutive seconds is calculated using the sliding window statistical method (such as standard deviation <30 rpm); secondly, if the standard deviation is lower than the threshold, it is determined to be a steady-state interval, otherwise it is classified as a variable speed interval; finally, the mark of the interval determination result is written into the detection model parameter table.

[0168] 702. Setting the sensitivity threshold of the coating peeling defect in the steady-state operation range to be higher than that of the blade crack defect, and setting the sensitivity threshold of the blade crack defect in the variable-speed operation range to be higher than that of the coating peeling defect;

[0169] In step 702 , the sensitivity threshold is the minimum confidence score for the detection model to determine that a defect exists.

[0170] In the embodiment of the present application, first, based on the characteristic that coating shedding defects are more prominent in stable airflow, the coating shedding threshold (0.85) is set higher than the crack threshold (0.75) in the steady-state operation range; secondly, because sudden changes in speed can easily cause crack expansion, the crack threshold (0.8) is set higher than the coating shedding threshold (0.7) in the variable speed operation range; finally, the set threshold setting strategy is embedded in the model decision layer.

[0171] 703. Input the three-dimensional coordinate change rate and rotational acceleration parameters in the blade motion state vector into the feature fusion layer of the lightweight detection model, and dynamically adjust the response weights of the nodes of the lightweight detection model to the defect features;

[0172] In step 703, the three-dimensional coordinate change rate is the rate of change of the spatial position in the blade motion state vector. The feature fusion layer is a multimodal data fusion network layer that integrates motion parameters and image features in the lightweight detection model.

[0173] In the embodiment of the present application, the three-dimensional coordinate change rate (such as X=1.5mm / s, Y=0.8mm / s) and the rotational acceleration (50rpm²) are first extracted from the blade motion state vector; secondly, the extracted parameters are input into the fully connected nodes of the feature fusion layer, and an adaptive weighting algorithm is used to adjust the response weights of the fully connected nodes to the crack and coating features (such as crack weight +0.2, coating weight -0.1). Finally, a dynamic weight configuration file is generated for the detection lightweight model to call.

[0174] 704. Filter the defect probability values output by the lightweight detection model according to the adjusted response weights, retain the defect types and defect locations that exceed the sensitivity threshold, and generate a defect detection result that matches the blade motion state vector.

[0175] In step 704 , the defect probability value is the confidence score of each defect type output by the lightweight detection model.

[0176] In an embodiment of the present application, first, the probability values of cracks and coating peeling (such as 0.82 and 0.73) are obtained from the output of the lightweight detection model; secondly, the crack probability value is screened based on the current speed range (such as the variable speed operation range) to see whether it exceeds the threshold value (0.8); finally, the qualified defect type (such as the crack type) and its position coordinates (such as X=120, Y=50) are written into the inspection report, and after eliminating the unqualified results, a defect detection result matching the blade motion state vector is generated.

[0177] Here's a specific example:

[0178] Assume that a drone is inspecting surface defects on an aircraft turbine blade at a speed of 250 km / h and a roll angle of 8°. The turbine's real-time speed accelerates from 2800 rpm to 3200 rpm (a variable speed range), with a three-dimensional coordinate change rate of X = 2.1 mm / s and a rotational acceleration of 60 rpm². Step 701: Calculate the speed standard deviation of 45 rpm (> the threshold of 30 rpm), identifying the variable speed range. Step 702: Set the crack threshold to 0.8 and the coating loss threshold to 0.7. Step 703: Input the coordinate change rate and rotational acceleration to the feature fusion layer, adjusting the crack weight to +0.25 and the coating weight to -0.15. In step 704, the lightweight detection model outputs a crack probability of 0.83 and a coating loss probability of 0.68, retains the crack defect location (X = 115, Y = 48), and generates the defect detection result.

[0179] Steps 701-704 significantly improve the ability to distinguish between turbine blade cracks and coating peeling defects in high-speed flight scenarios by dynamically dividing the speed range, reversing the sensitivity threshold, fusing motion parameters to adjust the model weight, and threshold matching screening. This solves the problem of false detection caused by feature confusion in variable speed conditions, ensures that the detection results strictly match the dynamic state of the aircraft, and enhances the environmental adaptability of the system.

[0180] Figure 2 The present invention provides a schematic diagram of a wind turbine blade surface defect detection system based on drone visual images, as shown in FIG. Figure 2 As shown, the system includes:

[0181] An acquisition module 21 is configured to continuously acquire a surface image sequence of a high-speed rotating wind turbine blade using HDR dynamic imaging technology, and simultaneously acquire blade motion trajectory data output by an inertial navigation module. The HDR dynamic imaging technology dynamically adjusts multi-frame exposure parameters based on the blade rotation speed to suppress over- and under-exposure caused by high-speed motion.

[0182] a compensation module 22 for performing spatiotemporal correlation between the blade motion trajectory data and the surface image sequence, and performing frame-by-frame compensation for blade edge distortion caused by motion blur in the surface image sequence based on a dynamic matching relationship between the blade rotation angle extracted from the spatiotemporal correlation result and the imaging device posture;

[0183] An enhancement module 23 is configured to enhance the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the compensated surface image sequence by dynamic weight allocation, and generate a blade surface feature map in combination with the spatiotemporal continuity constraint of the blade motion trajectory data;

[0184] A migration module 24 is configured to adjust the attention weight of the knowledge distillation framework based on the dynamic distribution ratio of the high-frequency vibration area and the static area in the blade surface feature map, and migrate the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weight;

[0185] The output module 25 is used to embed the blade motion state vector generated by the inertial navigation data in the lightweight detection model, and output the defect detection results matching the blade motion state vector by adjusting the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges.

[0186] Figure 2 The wind turbine blade surface defect detection system based on UAV visual images can be performed Figure 1 The implementation principles and technical effects of the wind turbine blade surface defect detection method based on drone visual imagery described in the illustrated embodiment are not further elaborated. The specific manner in which each module and unit performs operations in the wind turbine blade surface defect detection system based on drone visual imagery in the aforementioned embodiment has been described in detail in the relevant embodiments of the method and will not be further elaborated here.

[0187] In one possible design, Figure 2 The wind turbine blade surface defect detection system based on drone visual images of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0188] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0189] The processing component 32 is used for the above Figure 1 The embodiment provides a method for detecting surface defects of wind turbine blades based on drone visual images.

[0190] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0191] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0192] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0193] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0194] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0195] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0196] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment provides a method for detecting surface defects of wind turbine blades based on drone visual images.

[0197] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0198] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0199] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0200] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting surface defects of wind turbine blades based on UAV visual images, characterized in that: include: The surface image sequence of high-speed rotating wind turbine blades is collected by drone, and the blade motion trajectory data output by the inertial navigation module is simultaneously obtained; Performing spatiotemporal correlation between the blade motion trajectory data and the surface image sequence, and based on a dynamic matching relationship between the blade rotation angle and the imaging device posture extracted from the spatiotemporal correlation result, compensating for blade edge distortion caused by motion blur in the surface image sequence frame by frame through the dynamic matching relationship; The blade surface feature map is generated by combining the spatiotemporal continuity constraint of the blade motion trajectory data with the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the surface image sequence after strengthening the compensation by dynamic weight allocation; Based on the dynamic distribution ratio of high-frequency vibration areas and static areas in the blade surface feature map, the attention weight of the knowledge distillation framework is adjusted, and the feature response pattern of the pre-trained high-precision defect detection model is migrated to the lightweight detection model according to the adjusted attention weight; The inertial navigation module is embedded in the lightweight detection model, and a blade motion state vector is generated in real time according to the inertial navigation data output by the inertial navigation module. By adjusting the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges, a defect detection result matching the blade motion state vector is output.

2. The method according to claim 1, characterized in that The step of performing spatiotemporal correlation between the blade motion trajectory data and the surface image sequence, and performing frame-by-frame compensation for blade edge distortion caused by motion blur in the surface image sequence based on a dynamic matching relationship between the blade rotation angle and the imaging device posture extracted from the spatiotemporal correlation result, includes: Extracting the blade's three-dimensional spatial positioning data and rotation angle corresponding to each frame image acquisition moment from the blade motion trajectory data, and generating a spatiotemporal index aligned with the surface image sequence timestamp; Obtaining the difference in rotation angles in adjacent frames and the offset of the imaging device posture according to the spatiotemporal index, and calculating the distortion direction and intensity ratio of the blade edge during continuous motion according to the spatial geometric superposition relationship between the difference and the offset; Generate a pixel displacement compensation vector for the edge area of the blade in the current frame based on the distortion direction and intensity ratio, and perform layer-by-layer offset correction on the pixel position along the opposite direction of blade rotation using the pixel displacement compensation vector; The edge contour sharpening process is performed on the leaf edge distortion caused by motion blur in the surface image sequence according to the pixel displacement compensation vector, residual blurred pixel points in the leaf edge distortion are eliminated, and a compensated surface image sequence is generated.

3. The method according to claim 1, characterized in that The blade surface feature map is generated by combining the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the surface image sequence after dynamic weight distribution enhancement compensation with the spatiotemporal continuity constraint of the blade motion trajectory data, including: According to the change of the blade rotation speed, the contrast enhancement weight values of the windward side and the leeward side of the blade in the compensated surface image sequence are dynamically allocated, wherein the contrast enhancement weight value of the windward side increases with the increase of the blade rotation speed, and the contrast enhancement weight value of the leeward side decreases with the increase of the blade rotation speed; Based on the rotation angle difference between consecutive frames in the blade motion trajectory data, the contrast enhancement weight values are accumulated in adjacent frame images of the same blade area to generate a weight accumulation sequence consistent with the blade rotation direction; According to the numerical distribution of the weighted accumulation sequence, positive gain adjustment and negative suppression adjustment are respectively performed on the pixel intensities on the windward side and the leeward side, so that the intensity difference of the pixel intensities increases with the rotation speed; The adjusted pixel intensities are superimposed across frames in the time sequence of the leaf motion trajectory data, and a leaf surface feature map containing dynamic texture features of the leaf surface is generated based on the intensity mutation area contour formed after the superposition.

4. The method according to claim 1, wherein The method adjusts the attention weight of the knowledge distillation framework based on the dynamic distribution ratio of the high-frequency vibration area and the static area in the blade surface feature map, and migrates the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weight, including: Extracting the fluctuation amplitude of the blade surface texture intensity in the blade surface feature map in adjacent time intervals, and dividing the boundary range between the high-frequency vibration area and the static area according to the fluctuation amplitude; Counting the area proportions of the high-frequency vibration region and the static region in the blade surface characteristic map, and generating a numerical relationship representing the dynamic distribution ratio of the region; According to the numerical relationship, the attention weight coefficient corresponding to the high-frequency vibration area in the knowledge distillation framework is geometrically amplified, and the attention weight coefficient corresponding to the static area is reversely reduced; A high-precision defect detection model is pre-trained according to the knowledge distillation framework, and a characteristic response pattern matching the high-frequency vibration area and the static area in the high-precision defect detection model is obtained. Based on the characteristic response pattern, the adjusted attention weight coefficient is mapped layer by layer to the lightweight detection model.

5. The method according to claim 3, characterized in that The method further comprises: performing positive gain adjustment and negative suppression adjustment on the pixel intensities on the windward side and the leeward side respectively according to the numerical distribution of the weighted accumulation sequence, so that the intensity difference of the pixel intensities increases with the rotation speed, including: Associating the accumulated value of each frame in the weighted accumulation sequence with the current rotation speed of the blade to generate the dynamic gain coefficient and dynamic suppression coefficient of the windward side and the leeward side respectively; In the compensated surface image sequence, the boundary line between the windward side and the leeward side is marked along the extension direction of the blade surface texture, and a set width is extended to both sides based on the boundary line to form a gradual adjustment area; In the windward side range and the leeward side range of the gradual adjustment zone, the intensity value of each pixel is subjected to superposition enhancement and subtraction weakening according to the dynamic gain coefficient and the dynamic suppression coefficient respectively; The difference between the dynamic gain coefficient and the dynamic suppression coefficient is calculated based on the current rotation speed of the blade, and the entire area of the windward side and the leeward side outside the gradient adjustment area is subjected to global intensity adjustment according to the difference, so that the difference in pixel intensity between the windward side and the leeward side increases with the increase in rotation speed.

6. The method according to claim 2, characterized in that The step of obtaining the difference in rotation angles in adjacent frames and the offset of the imaging device posture according to the spatiotemporal index, and calculating the distortion direction and intensity ratio of the blade edge during the continuous motion according to the spatial geometric superposition relationship between the difference and the offset, includes: Decomposing the difference in rotation angles in adjacent frames into a horizontal angle change component and a vertical angle change component along the blade rotation plane according to the spatiotemporal index; Decompose the offset of the imaging device posture into a horizontal displacement component parallel to the blade rotation direction and a vertical displacement component perpendicular to the rotation direction; The horizontal angle change component and the horizontal displacement component, and the vertical angle change component and the vertical displacement component are respectively superimposed in a preset ratio to generate a total horizontal distortion offset and a total vertical distortion offset of the blade edge; The distortion direction and intensity ratio of the blade edge during the continuous movement are calculated according to the sum of the vector synthesis direction and absolute value of the total horizontal distortion offset and the total vertical distortion offset.

7. The method according to claim 1, characterized in that The blade motion state vector generated by the inertial navigation data is embedded in the lightweight detection model, and the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges is adjusted to output a defect detection result matching the blade motion state vector, including: Dividing the rotation speed intervals according to the real-time rotation speed in the blade motion state vector generated by the inertial navigation data, wherein the rotation speed intervals include a steady-state operation interval and a variable speed operation interval; The sensitivity threshold of the coating shedding defect in the steady-state operation range is set to be higher than that of the blade crack defect, and the sensitivity threshold of the blade crack defect in the variable speed operation range is set to be higher than that of the coating shedding defect; Inputting the three-dimensional coordinate change rate and rotational acceleration parameters in the blade motion state vector into the feature fusion layer of the lightweight detection model, and dynamically adjusting the response weights of the nodes of the lightweight detection model to the defect features; According to the adjusted response weights, the defect probability values output by the lightweight detection model are screened, the defect types and defect locations exceeding the sensitivity threshold are retained, and a defect detection result matching the blade motion state vector is generated.

8. A wind turbine blade surface defect detection system based on drone visual images, characterized in that: include: An acquisition module, configured to continuously capture a surface image sequence of high-speed rotating wind turbine blades using HDR dynamic imaging technology, and simultaneously acquire blade motion trajectory data output by an inertial navigation module. The HDR dynamic imaging technology dynamically adjusts multi-frame exposure parameters based on the blade rotation speed to suppress over- and under-exposure caused by high-speed motion; a compensation module for performing spatiotemporal correlation between the blade motion trajectory data and the surface image sequence, and performing frame-by-frame compensation for blade edge distortion caused by motion blur in the surface image sequence based on a dynamic matching relationship between the blade rotation angle extracted from the spatiotemporal correlation result and the imaging device posture; An enhancement module is configured to enhance the texture contrast difference between the windward side and the leeward side of the blade at different rotation speeds in the compensated surface image sequence by dynamic weight allocation, and generate a blade surface feature map in combination with the spatiotemporal continuity constraint of the blade motion trajectory data; A migration module, configured to adjust the attention weight of the knowledge distillation framework based on the dynamic distribution ratio of the high-frequency vibration area and the static area in the blade surface feature map, and migrate the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weight; An output module is used to embed the blade motion state vector generated by inertial navigation data in the lightweight detection model, and output defect detection results that match the blade motion state vector by adjusting the sensitivity threshold of the lightweight detection model to blade cracks and coating peeling defects in different speed ranges.

9. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a wind turbine blade surface defect detection method based on drone visual images as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, a method for detecting surface defects of wind turbine blades based on drone visual images as described in any one of claims 1 to 7 is implemented.

Images