Fan blade surface defect detection method and system based on visual image of unmanned aerial vehicle

Abstract

The invention provides a fan blade surface defect detection method and system based on an unmanned aerial vehicle visual image. The invention relates to the technical field of defect detection, and the method comprises the following steps: acquiring a surface image sequence of a high-speed rotating blade, and synchronously acquiring blade movement track data; performing space-time correlation on the motion trail data and the image sequence, and compensating blade edge distortion frame by frame; dynamically distributing contrast enhancement weights for the windward side and the leeward side of the blade in the compensated image sequence according to the rotating speed, and generating a fused texture feature map; the knowledge distillation attention weight is adjusted according to the distribution proportion of high-frequency vibration and a static region in the feature map, and high-precision model features are migrated to a lightweight model; and embedding a blade motion state vector in the lightweight model, dynamically adjusting a detection sensitivity threshold, and outputting a defect result matched with the motion state. The problems of dynamic and static characteristic unbalance and high false detection rate of blade defect detection in a high-speed rotation scene can be solved.

Description

Technical Field

[0001] This application relates to the technical field of defect detection, and particularly to a method and system for detecting surface defects of wind turbine blades based on unmanned aerial vehicle (UAV) vision images. Background Art

[0002] Under high-speed flight conditions, dynamic blades generate complex surface deformations and instantaneous defects due to intense aerodynamic loads and ultra-high-speed rotation. It is necessary to achieve real-time online detection of sub-millimeter defects under strong vibration, high-speed motion blur, and non-steady illumination conditions, and ensure the robustness of the detection system against dynamic deformations and extreme working conditions.

[0003] Currently, a typical solution for this scenario is the dynamic surface reconstruction technology based on multi-spectral laser scanning. By scanning the blade surface with a high-frequency laser beam, combining reflection spectrum analysis to reconstruct the three-dimensional shape of the blade, using a deformation tracking algorithm to extract surface abnormal regions, and determining the defect type through 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 during ultra-high-speed rotation, resulting in missed detection of micro-cracks; Non-steady illumination and aerodynamic turbulence cause a sharp drop in the signal-to-noise ratio of the reflection spectrum, resulting in artifacts interfering in the reconstructed shape, and a significant decrease in the defect determination accuracy; The iterative calculations of three-dimensional reconstruction and deformation tracking are time-consuming, unable 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] This application provides a method and system for detecting surface defects of wind turbine blades based on UAV vision images to solve the problems of unbalanced detection of dynamic and static features and high false detection rate in the prior art.

[0006] In a first aspect, this application provides a method for detecting surface defects of wind turbine blades based on UAV vision images, including: Collecting a sequence of surface images of a high-speed rotating wind turbine blade by a UAV, and synchronously obtaining the blade motion trajectory data output by an inertial navigation module; Performing spatio-temporal association on the blade motion trajectory data and the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the pose of the imaging device extracted from the spatio-temporal association result, compensating frame by frame for the blade edge distortion caused by motion blur in the surface image sequence through the dynamic matching relationship; Enhancing the texture contrast difference between the windward side and the leeward side of the blade in the compensated surface image sequence at different rotational speeds through dynamic weight assignment, and generating a blade surface feature map in combination with the spatio-temporal continuity constraint of the blade motion trajectory data; Adjust the attention weights of the knowledge distillation framework based on the dynamic distribution ratio of the high-frequency vibration region and the static region in the blade surface feature map, and transfer the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weights; 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 matching the blade motion state vector by adjusting the sensitivity thresholds of the lightweight detection model for blade crack and coating peeling defects in different rotational speed intervals.

[0007] Optionally, the spatio-temporal correlation of the blade motion trajectory data and the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the imaging device pose extracted from the spatio-temporal correlation result, to perform frame-by-frame compensation on the blade edge distortion caused by motion blur in the surface image sequence, and obtain a compensated surface image sequence, including: Extract 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 generate a spatio-temporal index aligned with the time stamps of the surface image sequence; Obtain the difference in the rotation angle and the offset of the imaging device pose between adjacent frames according to the spatio-temporal index, and calculate the distortion direction and intensity ratio of the blade edge in the continuous motion process according to the spatial geometric superposition relationship between the difference and the offset; Generate a pixel displacement compensation vector for the blade edge region of the current frame based on the distortion direction and intensity ratio, and the pixel displacement compensation vector performs layer-by-layer offset correction on the pixel positions in the reverse direction of blade rotation; Perform edge contour sharpening on the blade edge distortion caused by motion blur in the surface image sequence according to the pixel displacement compensation vector, eliminate the residual blurred pixel points in the blade edge distortion, and generate a compensated surface image sequence.

[0008] Optionally, strengthening the texture contrast difference between the windward side and the leeward side of the blade in the compensated surface image sequence at different rotational speeds through dynamic weight allocation, and generating a blade surface feature map in combination with the spatio-temporal continuity constraint of the blade motion trajectory data, including: Dynamically allocate the contrast enhancement weight values of the windward side and the leeward side of the blade in the compensated surface image sequence according to the change amount of the blade rotation speed, where 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 rotational angle difference between consecutive frames in the blade motion trajectory data, accumulate the contrast enhancement weight values in adjacent frame images of the same blade region to generate a weight accumulation sequence consistent with the blade rotation direction; According to the numerical distribution of the weight accumulation sequence, perform positive gain adjustment and reverse suppression adjustment on the pixel intensities of the windward side and the leeward side respectively, so that the intensity difference of the pixel intensities increases with the rotational speed; Overlay the adjusted pixel intensities across frames in the time sequence of the blade motion trajectory data, and generate a blade surface feature map containing the dynamic texture features of the blade surface based on the contour of the intensity mutation region formed after the overlay.

[0009] Optionally, based on the dynamic distribution ratio of the high-frequency vibration region and the static region in the blade surface feature map, adjust the attention weights of the knowledge distillation framework, and transfer the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weights, including: Within adjacent time intervals, extract the fluctuation amplitude of the blade surface texture intensity in the blade surface feature map, and divide the boundary range between the high-frequency vibration region and the static region according to the fluctuation amplitude; Statistically calculate the area ratios of the high-frequency vibration region and the static region in the blade surface feature map to generate a numerical relationship representing the regional dynamic distribution ratio; According to the numerical relationship, proportionally amplify the attention weight coefficient corresponding to the high-frequency vibration region in the knowledge distillation framework, and simultaneously reduce the attention weight coefficient corresponding to the static region in the reverse direction; Pre-train a high-precision defect detection model according to the knowledge distillation framework, obtain the feature response pattern in the high-precision defect detection model that matches the high-frequency vibration region and the static region, and map the adjusted attention weight coefficient layer by layer to the lightweight detection model based on the feature response pattern.

[0010] Optionally, the step of performing positive gain adjustment and reverse suppression adjustment on the pixel intensities of the windward side and the leeward side respectively according to the numerical distribution of the weight accumulation sequence, so that the intensity difference of the pixel intensities increases with the rotational speed, includes: Associate the accumulation value of each frame in the weight accumulation sequence with the current rotational speed of the blade to generate dynamic gain coefficients and dynamic suppression coefficients for the windward side and the leeward side respectively; In the compensated surface image sequence, mark the dividing line between the windward side and the leeward side along the extension direction of the blade surface texture, and expand a set width on both sides of the dividing line to form a gradient adjustment area; Within the windward side range and the leeward side range of the gradient adjustment area, the intensity values of each pixel are respectively superimposed and enhanced and subtracted and weakened according to the dynamic gain coefficient and the dynamic suppression coefficient; Calculate the difference between the dynamic gain coefficient and the dynamic suppression coefficient based on the current rotational speed of the blade, and perform global intensity adjustment on the overall windward side and leeward side areas outside the gradient adjustment area according to the difference, so that the difference amplitude of the pixel intensities on the windward side and the leeward side increases with the increase in rotational speed.

[0011] Optionally, the method for obtaining the difference in the rotation angle and the offset of the imaging device pose in adjacent frames according to the spatio-temporal index, and calculating the distortion direction and intensity ratio of the blade edge during continuous movement according to the spatial geometric superposition relationship between the difference and the offset includes: Decompose the difference in the rotation angle in adjacent frames into a horizontal angle change component and a vertical angle change component along the blade rotation plane according to the spatio-temporal index; Decompose the offset of the imaging device pose into a horizontal displacement component parallel to the blade rotation direction and a vertical displacement component perpendicular to the rotation direction; Superimpose the horizontal angle change component and the horizontal displacement component, and the vertical angle change component and the vertical displacement component respectively according to a preset ratio to generate a total horizontal distortion offset and a total vertical distortion offset of the blade edge; Calculate the distortion direction and intensity ratio of the blade edge during continuous movement according to the vector synthesis direction and the sum of absolute values of the total horizontal distortion offset and the total vertical distortion offset respectively.

[0012] Optionally, the method for embedding the blade motion state vector generated from inertial navigation data into the lightweight detection model, and outputting a defect detection result matching the blade motion state vector by adjusting the sensitivity thresholds of the lightweight detection model for blade crack and coating peeling defects in different rotational speed intervals includes: Divide the rotational speed intervals according to the real-time rotational speed in the blade motion state vector generated from inertial navigation data, and the rotational speed intervals include a steady-state operation interval and a variable-speed operation interval; Set the sensitivity threshold for coating peeling defects in the steady-state operation interval to be higher than that for blade crack defects, and the sensitivity threshold for blade crack defects in the variable-speed operation interval to be higher than that for coating peeling defects; 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 defect features; According to the adjusted response weights, filter the defect probability values output by the lightweight detection model, 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.

[0013] In a second aspect, the present application provides a wind turbine blade surface defect detection system based on UAV vision images, including: An acquisition module, configured to acquire a surface image sequence of a high-speed rotating wind turbine blade through a UAV, and synchronously obtain blade motion trajectory data output by an inertial navigation module; A compensation module, configured to perform spatio-temporal association on the blade motion trajectory data and the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the pose of the imaging device extracted from the spatio-temporal association result, frame-by-frame compensate for the blade edge distortion caused by motion blur in the surface image sequence through the dynamic matching relationship; A strengthening module, configured to strengthen the texture contrast difference between the windward side and the leeward side of the blade in the compensated surface image sequence at different rotational speeds through dynamic weight allocation, and generate a blade surface feature map in combination with the spatio-temporal continuity constraint of the blade motion trajectory data; A migration module, configured to adjust the attention weights of the knowledge distillation framework based on the dynamic distribution ratio of the high-frequency vibration region and the static region 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 weights; An output module, configured 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 thresholds of the lightweight detection model for blade cracks and coating peeling defects in different rotational speed intervals.

[0014] In a third aspect, the present application provides a computing device, including 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 method for detecting wind turbine blade surface defects based on UAV vision images as described in the first aspect above.

[0015] In a fourth aspect, the present application provides a computer storage medium storing a computer program, and when the computer program is executed by a computer, it implements a method for detecting wind turbine blade surface defects based on UAV vision images as described in the first aspect.

[0016] In this application, a surface image sequence of a high-speed rotating fan blade is collected by a drone, and the blade motion trajectory data output by an inertial navigation module is synchronously obtained; the blade motion trajectory data is spatio-temporally correlated with the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the imaging device pose extracted from the spatio-temporal correlation result, frame-by-frame compensation is performed on the blade edge distortion caused by motion blur in the surface image sequence through the dynamic matching relationship; the texture contrast difference between the windward side and the leeward side of the blade in the surface image sequence after compensation is enhanced through dynamic weight allocation, and a blade surface feature map is generated in combination with the spatio-temporal continuity constraint of the blade motion trajectory data; based on the dynamic distribution ratio of the high-frequency vibration region and the static region 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 migrated to the lightweight detection model according to the adjusted attention weight; the inertial navigation module is embedded in the lightweight detection model, a blade motion state vector is generated in real time according to the inertial navigation data output by the inertial navigation module, and the sensitivity threshold of the lightweight detection model to blade crack and coating peeling defects in different rotational speed intervals is adjusted to output a defect detection result matching the blade motion state vector.

[0017] The technical solution of this application has the following beneficial effects: By synchronously collecting HDR images and inertial navigation data by a drone, spatio-temporal precise alignment of the blade surface image and the motion trajectory under high-speed rotation is achieved, and problems of overexposure / underexposure of light and positioning drift caused by high-speed motion are suppressed; motion blur distortion is compensated based on the dynamic matching relationship between the rotation angle and the imaging device pose, 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 side and the leeward side is enhanced through dynamic weight allocation, and a feature map integrating dynamic lighting characteristics is generated in combination with the spatio-temporal continuity of the motion trajectory, solving the problem of weakening of defect features under changing rotational speeds; the knowledge distillation attention weight is adjusted based on the distribution ratio of high-frequency vibration and static regions, realizing the adaptive migration of the features of the high-precision model to the lightweight model, and balancing the detection accuracy and the calculation efficiency; a real-time motion state vector is embedded and the sensitivity threshold is dynamically adjusted, enabling the detection model to adapt to the dynamic changes of defect features at different rotational speeds and reducing the false detection rate and the missed detection rate in high-speed scenarios.

[0018] Further, a spatio-temporal index is extracted from the blade motion trajectory data. Based on the spatial geometric superposition relationship between the rotation angle difference of adjacent frames and the pose offset of the imaging device, the distortion direction and intensity ratio are calculated to generate a pixel displacement compensation vector along the reverse rotation direction. Through layer-by-layer offset correction and edge sharpening processing, the residual pixels of motion blur are eliminated, and a compensated clear image sequence is output. Through the geometric correlation modeling of the motion trajectory and image data, the blade edge deformation and blur caused by high-speed rotation are accurately offset, the visibility of the surface defect contour and the quality of the input data of the detection algorithm are improved, and a high-fidelity image basis is provided for subsequent dynamic feature extraction and defect classification.

[0019] These aspects or other aspects of the present application will be more clearly understood in the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0021] Figure 1 FIG. shows a flowchart of a method for detecting surface defects of a wind turbine blade based on unmanned aerial vehicle vision images provided by the present application; Figure 2 FIG. shows a schematic structural diagram of a system for detecting surface defects of a wind turbine blade based on unmanned aerial vehicle vision images provided by the present application; Figure 3 FIG. shows a schematic structural diagram of a computing device provided by the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] In order to enable those skilled in the art to better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

[0023] In some of the processes described in the specification and claims of the present application and the above-mentioned drawings, there are multiple operations that appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. 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 such as "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0024] Researchers found that there are several difficult problems in the surface defect detection of high-speed rotating fan blades, such as severe motion blur, uneven dynamic illumination, weakening of defect features, and poor dynamic adaptability of the detection model. Traditional methods are difficult to balance detection accuracy and real-time performance under complex working conditions. Based on this, a method for detecting surface defects of fan blades based on UAV vision images is provided. Specifically, the method synchronously collects the surface image sequence of the blade and inertial navigation data through the UAV, constructs a spatio-temporal correlation model to compensate for the edge distortion caused by high-speed motion; adopts a dynamic weight allocation strategy to enhance the texture contrast difference between the windward side and the leeward side of the blade, and combines the motion trajectory constraint to generate an anti-interference feature map; based on the dynamic distribution ratio of the feature map, adjusts the attention weight of the knowledge distillation framework, and transfers the features of the high-precision model to the lightweight model; and embeds the real-time motion state vector to realize the dynamic adaptation of the defect sensitivity threshold. This method can solve the core problems of imbalance between dynamic and static features and high false detection rate in high-speed rotating scenarios through multi-source data collaboration and dynamic feature enhancement, and significantly improve the detection accuracy of defects such as cracks and coating peeling and the system robustness.

[0025] The technical solution of this application is applicable to the dynamic blade defect detection scenario in the high-speed flight state.

[0026] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present application.

[0027] Figure 1 The flowchart of a method for detecting surface defects of fan blades based on UAV vision images provided for the embodiments of the present application is as follows Figure 1 shown, and the method includes: 101. Collect the surface image sequence of the high-speed rotating fan blade through the UAV, and synchronously obtain the blade motion trajectory data output by the inertial navigation module; In this step, the HDR dynamic imaging technology is a technology that dynamically adjusts the exposure parameters of multiple frames according to the real-time rotation speed of the blade, which is used to suppress the overexposed and underexposed areas caused by high-speed motion and ensure the integrity of image details.

[0028] In the embodiment of the present application, the real-time rotation speed of the blades of a high-speed rotating fan is obtained through the HDR dynamic imaging technology, and a multi-frame exposure parameter combination required for HDR dynamic imaging is determined. The high exposure parameter is used to capture the low-light area on the leeward side of the blade, and the low exposure parameter is used to capture the high-light area on the windward side of the blade. The HDR dynamic imaging device continuously acquires the blade surface images according to the multi-frame exposure parameter combination, and at the same time triggers the inertial navigation module to generate corresponding synchronous timestamp marks during the acquisition of the blade surface images. The synchronous timestamp marks are bound to the three-dimensional space coordinates and rotation angle data of the blade output by the inertial navigation module to generate the blade motion trajectory data corresponding to each frame of the surface image. The multi-frame different exposure images within the same rotation period are subjected to light area fusion to eliminate the halo artifacts generated due to high-speed movement in the junction area between the windward side and the leeward side of the blade, and a surface image sequence with overexposure and underexposure suppressed is obtained.

[0029] Assume that the drone acquires blade images at a rate of 150 frames per second. The HDR dynamic imaging adjusts the exposure parameters to 1 / 200 s (windward side) and 1 / 800 s (leeward side) according to a rotational speed of 2500 rpm. The inertial navigation module synchronously generates the coordinates (X = 10.3 m, Y = 2.8 m, Z = 30°) and rotation angle (30°) of the timestamp mark. After fusing 5 frames of images, the halo intensity at the junction is reduced by 60%, and a clear image sequence is output.

[0030] 102. Perform spatio-temporal association on the blade motion trajectory data and the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the pose of the imaging device extracted from the spatio-temporal association result, frame-by-frame compensation is performed on the blade edge distortion caused by motion blur in the surface image sequence through the dynamic matching relationship. In this step, the dynamic matching relationship is the superposition association in spatial geometry between the blade rotation angle difference and the pose offset of the imaging device, and is used to describe the distortion direction and intensity of motion blur.

[0031] In the embodiment of the present application, first, the timestamp, three-dimensional coordinates, and rotation angle corresponding to each frame of image are extracted from the blade motion trajectory data to generate a spatio-temporal index table. Secondly, based on the blade rotation angle difference between adjacent frames (such as frame 1: 30°, frame 2: 35°, difference 5°) and the pose offset of the imaging device (such as horizontal displacement 2 mm) in the spatio-temporal index table, the spatio-temporal geometric superposition algorithm is used to calculate the blade edge distortion direction (such as horizontal left deviation 45°) and intensity ratio (horizontal displacement ratio 70%). Thirdly, a pixel displacement compensation vector (such as left shift 4 pixels, downward shift 1 pixel) is generated according to the distortion direction and intensity ratio, and the reverse motion compensation algorithm is used to perform layer-by-layer reverse offset correction on the edge pixels of the current frame. Finally, the residual blur is eliminated according to the correction result through the local gradient sharpening algorithm (such as Sobel operator to enhance the edge gradient), and a compensated image sequence is generated.

[0032] For example, continuing with the previous example, the difference in the rotation angle between adjacent frames in the spatio-temporal index is 8°, the horizontal displacement of the imaging device is 4 mm, the geometric superposition algorithm is used to calculate that the distortion direction is 25° to the lower right, and the intensity ratio is 1:0.6. A compensation vector (shift 6 pixels to the right and 3 pixels down) is generated to correct the edge pixels. After sharpening by the Sobel operator, the crack edge width is restored from a blurred 6 pixels to a clear 3 pixels.

[0033] 103. Strengthen the texture contrast difference between the windward side and the leeward side of the blade in the compensated surface image sequence at different rotational speeds through dynamic weight allocation, and generate a blade surface feature map in combination with the spatio-temporal continuity constraint of the blade motion trajectory data; In this step, the dynamic weight allocation dynamically adjusts the contrast enhancement weight values of the windward side and the leeward side regions according to the rotational speed change, and the weight values increase / decrease with the rotational speed. The spatio-temporal continuity constraint is a weight accumulation sequence generated by cross-frame accumulation of the weight values based on the temporal order and spatial position relationship of the motion trajectory data.

[0034] In the 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 rotational 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 increasing 2° per frame), the cross-frame accumulation algorithm is used to accumulate the weight values of the same blade region (the weight of the windward side accumulates to 1.3); thirdly, the pixel intensity of the windward side is adjusted by non-linear gain according to the accumulation result (such as intensity × 1.3), and the leeward side is suppressed (such as intensity × 0.6); finally, through the multi-frame spatio-temporal superposition algorithm, the intensity mutation contour between the windward side and the leeward side (such as crack width > 2 pixels) is extracted to generate a blade surface feature map.

[0035] For example, continuing with the previous example, for the compensated image sequence, when the rotational speed is 3000 rpm, the initial weight of the windward side is set to 0.85, and after cross-frame accumulation (the cumulative rotation difference of 5 frames is 15°), the weight is increased to 1.25; the pixel intensity of the windward side is increased by 25%, and the leeward side is decreased 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.

[0036] 104. Based on the dynamic distribution ratio of the high-frequency vibration region and the static region in the blade surface feature map, adjust the attention weight of the knowledge distillation framework, and transfer the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weight; In this step, the knowledge distillation framework migrates the feature response pattern of the high-precision defect detection model to the training framework of the lightweight model, and realizes feature selective migration through attention weight adjustment.

[0037] In the embodiment of the present application, first, the ratio of the area of the high-frequency vibration region (fluctuation amplitude > 50 gray levels) to the static region in the blade surface feature map is statistically calculated (such as 3:7) to generate a dynamic distribution ratio; secondly, according to the dynamic distribution ratio (such as 30% of the high-frequency region), the attention weight coefficient of the high-frequency region in the knowledge distillation framework is enlarged in equal proportion (such as ×1.5), and at the same time, the weight of the static region is reduced in reverse (such as ×0.5); thirdly, through the feature mapping technology, the feature response pattern of the high-frequency vibration region (such as crack edge gradient feature) in the pre-trained high-precision model is migrated to the lightweight model according to the adjusted weight; finally, the basic texture features of the static region (such as coating uniformity) are retained to complete the lightweight adaptation of the model.

[0038] For example, continuing the above example, the high-frequency vibration region (fluctuation amplitude > 60 gray levels, accounting for 28%) is divided in the blade surface feature map, and its attention weight is enlarged to 1.4 times at a ratio of 1:2.6; the crack edge gradient feature in the high-precision model is migrated to the lightweight model, and the crack response value after migration is increased from 0.7 to 0.92, and the weight of the static region remains 0.8 times.

[0039] 105. 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 matching the blade motion state vector by adjusting the sensitivity thresholds of the lightweight detection model for blade crack and coating peeling defects in different rotational speed intervals.

[0040] In this step, the blade motion state vector is a set of dynamic parameters generated from inertial navigation data, including key motion state information such as real-time rotational speed and three-dimensional space offset. The sensitivity threshold is the critical value of the feature response intensity for the lightweight detection model to determine defects, and it is dynamically adjusted with the rotational speed to adapt to different working conditions.

[0041] In the embodiment of the present application, first, the real-time rotational speed (such as 3000 rpm) and the three-dimensional offset (such as horizontal displacement 4 mm) are 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 reference sensitivity threshold is set (such as crack threshold 0.8); thirdly, the reference sensitivity threshold is dynamically increased based on the offset (threshold = 0.8 + 0.1 × 4 = 1.2); finally, the feature response value (such as crack response value 1.0) output by the lightweight model is compared with the increased threshold to output a defect result matching the motion state vector.

[0042] For example, continuing with the previous example, a motion state vector is generated based on a real-time rotational speed of 3200 rpm and a 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, triggers an alarm after exceeding the threshold, synchronously suppresses the false detection signal on the leeward side, and outputs the detection result with the defect position matching the rotational speed.

[0043] To solve the problems 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 deficiency of the accuracy of traditional compensation methods in scenarios of strong vibration and rapid light switching. In some embodiments, the spatio-temporal correlation of the blade motion trajectory data and the surface image sequence is performed, and based on the dynamic matching relationship between the blade rotation angle and the pose of the imaging device extracted from the spatio-temporal correlation result, 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: 201. Extract the three-dimensional spatial positioning data and rotation angle of the blade corresponding to the acquisition moment of each frame of image from the blade motion trajectory data, and generate a spatio-temporal index aligned with the time stamp of the surface image sequence; In step 201, the spatio-temporal index is a data structure composed of the three-dimensional spatial positioning data (X / Y / Z coordinates) and rotation angle of the blade aligned with the time stamp, and is used to associate the image frame with the motion trajectory.

[0044] In the embodiment of the present application, first, the three-dimensional spatial positioning data (such as X = 12.3 m, Y = 4.7 m, Z = 30°) and rotation angle (such as 45°) of the blade corresponding to each frame of image are extracted from the blade motion trajectory data, and the time stamps involved in the three-dimensional spatial positioning data and rotation angle of the blade are aligned with the image frame through a time stamp matching algorithm (such as the nearest neighbor interpolation method) to generate a spatio-temporal index table. Each row in the table records the frame number, time stamp, coordinates, and rotation angle for subsequent steps to call.

[0045] 202. Obtain the difference in the rotation angle and the offset of the pose of the imaging device between adjacent frames according to the spatio-temporal index, and calculate 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; In step 202, the spatial geometric superposition relationship is the vector synthesis rule of the rotation angle difference and the imaging device pose offset in three-dimensional space.

[0046] In the embodiments of the present application, first, the rotation angle differences of adjacent frames (e.g., frame 1: 30°, frame 2: 35°, difference 5°) and the pose offset of the imaging device (e.g., 2 mm displacement in the X direction) are obtained from the spatio-temporal index table; second, the difference is decomposed into a horizontal rotation component (e.g., 3°) and a vertical rotation component (e.g., 2°), and the pose offset is decomposed into a horizontal displacement (e.g., 1.5 mm) and a vertical displacement (e.g., 0.5 mm); third, the components are superimposed according to a preset weight (e.g., 0.7 for horizontal and 0.3 for vertical), and the distortion direction (e.g., 40° to the left) and the intensity ratio (e.g., horizontal intensity 1.05, vertical intensity 0.45) are calculated.

[0047] 203. Generate a pixel displacement compensation vector for the blade edge region of the current frame based on the distortion direction and the intensity ratio, and the pixel displacement compensation vector corrects the pixel positions layer by layer in the opposite direction of the blade rotation. In step 203, the layer-by-layer offset correction is to translate the pixel positions 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.

[0048] In the embodiments of the present application, first, according to the distortion direction (e.g., 40° to the left) and the intensity ratio (e.g., 1.05 for horizontal and 0.45 for vertical), a pixel displacement compensation vector (e.g., 7 pixels to the left and 3 pixels up) is generated; second, the pixel displacement compensation vector is split into 3 layers for step-by-step offset correction (e.g., 2 pixels to the left and 1 pixel up in the first layer) to avoid image tearing caused by a single large offset; finally, the vector after layer-by-layer correction is written into a configuration file for the image processing module to call.

[0049] 204. Perform edge contour sharpening on the blade edge distortion caused by motion blur in the surface image sequence according to the pixel displacement compensation vector, eliminate the remaining blurred pixel points in the blade edge distortion, and generate a compensated surface image sequence.

[0050] In step 204, the remaining blurred pixel points are the gray-scale gradient regions in the compensated image where the motion blur is not completely eliminated. Edge contour sharpening enhances the gray-scale mutation characteristics of the defect boundary through a gradient enhancement algorithm.

[0051] In the embodiments of the present application, first, an adaptive gradient threshold algorithm (e.g., Canny edge detection) is applied to the surface image sequence according to the pixel displacement compensation vector, and the regions with a gray-scale change rate lower than the threshold (e.g., <15%) are marked as remaining blurred pixels; second, a directional sharpening filter (e.g., anisotropic diffusion) is used to enhance the edge gradient of the surface image sequence along the blade texture direction to eliminate the remaining blurred pixels; finally, the sharpened surface image sequence is binarized (e.g., Otsu algorithm) to generate a compensated surface image sequence.

[0052] The following is a specific example: Suppose a drone detects cracks on the surface of the turbine blade of an aircraft in an environment with a flight speed of 200 km / h and a dive angle of 15°, and the turbine speed is 3200 rpm. In step 201, a spatio-temporal index table is generated, recording that the time stamp of the 100th frame is 163000 μs, the coordinates are X = 15.2 m, Y = 3.8 m, Z = 25°, and the rotation angle is 30°; in step 202, the rotation angle difference between adjacent frames (the 100th - 101st frames) is 5°, the horizontal displacement of the device is 2 mm, and the calculated distortion direction is 35° to the left, with the intensity ratio horizontal 0.8 and vertical 0.2; through step 203, a compensation vector is generated to shift 6 pixels to the left and 1 pixel upward, and is corrected in 3 layers (2 pixels are shifted to the left for each layer); through step 204, the width of the crack edge after sharpening is reduced from 4 blurred pixels to 1 clear pixel, and the contrast of the crack area in the binary surface image sequence reaches 8:1.

[0053] Steps 201 - 204 accurately associate images with motion data through spatio-temporal indexing, quantify the distortion direction and intensity based on the geometric superposition relationship, eliminate motion blur through layered compensation, and enhance the defect contour through directional sharpening. Finally, in a high-speed rotation scenario, while improving the crack detection accuracy and reducing the false detection rate, the compensation processing time is accurately controlled to meet the real-time detection requirements.

[0054] To solve the problem of the imbalance of texture contrast between the windward side and the leeward side caused by uneven dynamic illumination in the detection of surface defects of high-speed rotating fan blades, especially the technical bottleneck of the weakening of defect features and the increase of false detection rate at high rotational speeds. In some embodiments, the texture contrast difference between the windward side and the leeward side of the blade in the surface image sequence after enhanced compensation is strengthened through dynamic weight allocation, and a surface feature map of the blade is generated in combination with the spatio-temporal continuity constraint of the blade motion trajectory data, including: 301. According to the change amount of the blade rotation speed, dynamically allocate the contrast enhancement weight values of the windward side and the leeward side of the blade in the surface image sequence after compensation, where 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; In step 301, the contrast enhancement weight value is a numerical parameter dynamically allocated to the windward side and the leeward side according to the blade rotation speed, and is used to adjust the pixel intensity difference.

[0055] In the embodiments of the present application, first, the blade rotation speed (such as 2500 rpm) is obtained in real time, and according to the preset rotation speed and weight mapping table (for example, for every 100 rpm increase in rotation speed, the windward side weight +0.05, the leeward side weight -0.03), the contrast enhancement weight value of the current frame is dynamically allocated, and the allocated weight value is written into the image metadata for subsequent accumulation and adjustment calls.

[0056] 302. Based on the rotation angle difference between consecutive frames in the blade motion trajectory data, accumulate the contrast enhancement weight values in adjacent frame images of the same blade area to generate a weight accumulation sequence consistent with the blade rotation direction; In step 302, the weight accumulation sequence is the accumulation result of the contrast enhancement weight values in the same blade area in consecutive frames, reflecting the weight accumulation effect in the rotation direction.

[0057] In the embodiment of the present application, first, extract the rotation angle difference between adjacent frames (such as 5°) from the blade motion trajectory data, and calculate the displacement ratio in the rotation direction of the same blade area based on the rotation angle difference (such as 0.2 pixel displacement per degree); secondly, according to the displacement ratio, perform cross-frame accumulation of the contrast enhancement weight values for skipping adjacent frames (such as the upwind side weight 0.7 is accumulated to 0.7 + 0.05×5 = 0.95), to generate a weight accumulation sequence consistent with the rotation direction; finally, store the accumulation result as a time-series weight array.

[0058] 303. According to the numerical distribution of the weight accumulation sequence, perform positive gain adjustment and negative suppression adjustment on the pixel intensities of the upwind side and the leeward side respectively, so that the intensity difference of the pixel intensities increases with the rotational speed; In step 303, the positive gain adjustment is to non-linearly amplify the pixel intensity of the upwind side according to the weight accumulation value. The negative suppression adjustment is to proportionally reduce the pixel intensity of the leeward side according to the weight accumulation value.

[0059] In the embodiment of the present application, first, read the weight values of the upwind side and the leeward side of the current frame from the weight accumulation sequence (such as 1.1 and 0.6); secondly, set the gain coefficient and the suppression coefficient, and use the piecewise linear transformation algorithm to perform gain adjustment on the pixel intensity of the upwind side (such as intensity × 1.1), and perform suppression adjustment on the pixel intensity of the leeward side (such as intensity × 0.6); finally, generate the adjusted image data according to the adjustment result, where the contrast of the crack area on the upwind side is increased to 2.5 times that of the background.

[0060] 304. Stack the adjusted pixel intensities across frames according to the time sequence of the blade motion trajectory data, and generate a blade surface feature map containing the dynamic texture features of the blade surface based on the contour of the intensity mutation area formed after the stacking.

[0061] In step 304, the cross-frame stacking is to stack the pixels of multiple frames of adjusted images pixel by pixel in time sequence to enhance the dynamic texture features. The contour of the intensity mutation area is a continuous boundary area where the pixel intensity difference in the stacked image exceeds the set threshold.

[0062] In the embodiments of the present application, first, the adjusted multi-frame images are arranged in the chronological order of the blade motion trajectory data; secondly, the 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 (for example, 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 contour of the intensity mutation area in the accumulation result (for example, the crack width > 2 pixels), and a blade surface feature map containing dynamic texture features is generated.

[0063] The following is a specific example: Suppose a drone detects a micro-crack with a width of 1.2 mm on the windward side of the turbine blade of an aircraft in an environment with a flight speed of 220 km / h and a yaw angle of 10°. The turbine speed is 3200 rpm. In step 301, the windward side weight 0.85 and the leeward side weight 0.25 are assigned according to the speed of 3200 rpm; in step 302, the rotation angle difference between adjacent frames is 8°, and the windward side weight is accumulated to 0.85 + 0.05×8 = 1.25; in step 303, the pixel intensity on the windward side is increased by 25% (×1.25), and the leeward side is decreased by 30% (×0.7), and the contrast difference in the crack area reaches 3:1; in step 304, after stacking 10 frames of images, the crack contour is extracted, and the crack width in the generated blade surface feature map is clearly shown as 2 pixels, and the background noise intensity is reduced by 40%.

[0064] Steps 301-304 strengthen the contrast difference between the two sides of the blade through dynamic weight distribution, combine cross-frame accumulation and superposition to generate a high-discrimination feature map, significantly improve the visibility of micro-cracks and coating peeling defects in the high-speed rotation scenario, reduce the false detection rate caused by uneven illumination, and at the same time ensure the stability and adaptability of the detection system during drastic changes in speed through real-time weight adjustment and superposition processing.

[0065] To solve the problem of insufficient model generalization ability caused by the confusion between dynamic vibration and static area features in the detection of surface defects of high-speed rotating fan blades, especially the technical bottleneck that the lightweight detection model has low sensitivity to micro-defects under high-frequency vibration interference. In some embodiments, 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 mode in the pre-trained high-precision defect detection model is migrated to the lightweight detection model according to the adjusted attention weight, including: 401. Extract the fluctuation amplitude of the blade surface texture intensity in the blade surface feature map within adjacent time intervals, and divide the boundary range between the high-frequency vibration area and the static area according to the fluctuation amplitude; In step 401, the boundary range is the spatial demarcation line between the high-frequency vibration area and the static area in the feature map.

[0066] In the embodiments of the present application, first, five 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 the pixels with a fluctuation amplitude exceeding the threshold are marked as high-frequency vibration regions, and the rest are static regions; finally, adjacent high-frequency pixels in the high-frequency vibration regions are merged through a spatial clustering algorithm to generate a closed boundary range.

[0067] 402. Statistically calculate the area ratios of the high-frequency vibration regions and the static regions in the blade surface feature map to generate a numerical relationship representing the dynamic distribution ratio of the regions; In step 402, the numerical relationship is a ratio value with the area of the high-frequency region as the numerator and the area of the static region as the denominator.

[0068] In the embodiments of the present application, first, the connected components of the closed boundary range are marked to statistically calculate the total number of pixels in the high-frequency vibration regions (such as 1200 pixels) and the total number of pixels in the static regions (such as 3600 pixels); secondly, calculate the area ratios of the high-frequency and static regions. For example, the area ratio of the high-frequency region = the total number of pixels in the high-frequency vibration region / (the total number of pixels in the high-frequency vibration region + the total number of pixels in the static region); finally, quantify the ratio value of the area ratio into a numerical relationship (1:3) and write it into the configuration file of the knowledge distillation framework.

[0069] 403. According to the numerical relationship, proportionally amplify the attention weight coefficients corresponding to the high-frequency vibration regions in the knowledge distillation framework, and simultaneously inversely reduce the attention weight coefficients corresponding to the static regions; In step 403, extract the ratio of the area ratio of the high-frequency vibration region to the area ratio of the static region from the numerical relationship (such as 1:3), and convert the ratio into a weight adjustment proportionality factor; establish a mapping relationship between the spatial positions of the blade surface feature map and the attention weight coefficients in the knowledge distillation framework (such as coordinates X = 100 - 200, Y = 50 - 150 corresponding to the weight index group), and locate the distribution ranges of the weight coefficients corresponding to the high-frequency vibration regions and the static regions according to the mapping relationship; based on the weight adjustment proportionality factor, proportionally amplify all the attention weight coefficients within the distribution range of the high-frequency vibration regions layer by layer, where the amplification multiple is the square root of the weight adjustment proportionality factor (such as the original weight 0.6 → 0.6 × 1.732 ≈ 1.04); simultaneously inversely reduce all the attention weight coefficients within the distribution range of the static regions, where the reduction ratio is in an inverse relationship with the weight adjustment proportionality factor (such as the original weight 0.9 → 0.9 × 1 / 3 = 0.3).

[0070] 404. Pre-train a high-precision defect detection model according to the knowledge distillation framework, obtain the feature response pattern in the high-precision defect detection model that matches the high-frequency vibration region and the static region, and map the adjusted attention weight coefficients layer by layer to the lightweight detection model based on the feature response pattern.

[0071] In step 404, the feature response pattern is the feature activation pattern (such as the histogram of oriented gradients feature) for specific regions (such as cracks and coating peeling) in the high-precision defect detection model.

[0072] In the embodiment of the present application, first, extract the feature response pattern corresponding to the high-frequency vibration region (such as the gradient direction distribution at the crack edge) from the pre-trained high-precision defect detection model; secondly, based on the adjusted attention weight coefficients (such as 1.04 for high frequency and 0.3 for static), use the feature mapping technology to transfer the extracted feature response pattern to the corresponding layer (such as convolutional layer 3) of the lightweight model according to the weight ratio; finally, freeze the weight parameters of the static region in the corresponding layer of the lightweight detection model, and only update the parameters of the high-frequency vibration region according to the transferred feature response pattern to complete the model adaptation.

[0073] The following is a specific example: The drone detects a micro-crack with a width of 1.2 mm on the windward side of the turbine blade of the aircraft at an environment where the flight speed is 240 km / h and the roll angle is 5°. The turbine speed is 3500 rpm. In step 401, the standard deviation of the texture intensity in the crack region in 5 consecutive frames of images reaches 65 gray levels, which is marked as the high-frequency vibration region (the area accounts for 20%), and the static region accounts for 80%; in step 402, generate the numerical relationship 1:4 (20%:80%), and write it into the configuration file; in step 403, extract the scaling factor 4, calculate the square root magnification factor 2.0, and the reciprocal reduction ratio 0.25; the weight of the high-frequency region is amplified from 0.5 to 1.0 (0.5×2.0), and the weight of the static region is reduced from 0.8 to 0.2 (0.8×0.25); in step 404, transfer the gradient direction feature of the crack region in the migrated high-precision model to the third convolutional layer of the lightweight model. After migration, the crack response value is increased from 0.5 to 0.95, and the weight of the static region remains unchanged at 0.2.

[0074] Steps 401-404 generate a weight adjustment factor by dynamically dividing the high-frequency vibration region and quantifying the distribution ratio, combine the weight balance strategy of square root magnification and reciprocal reduction, directionally strengthen the defect feature migration of the high-frequency region, and at the same time suppress the interference signals of the static region, significantly improving the detection accuracy of the lightweight model in the high-speed vibration scenario. Through feature mapping and hierarchical adaptation, ensure the sensitivity and generalization ability of the model to micro-cracks in complex dynamic environments, reduce the false detection rate and meet the real-time requirements.

[0075] To solve the problem of weakened texture contrast between the windward side and the leeward side caused by uneven dynamic illumination and motion blur in the detection of surface defects of high-speed rotating fan blades, especially the technical defect that defect features are submerged by background noise at high rotational speeds. In some embodiments, according to the numerical distribution of the weight accumulation sequence, positive gain adjustment and negative suppression adjustment are respectively performed on the pixel intensities of the windward side and the leeward side, so that the intensity difference of the pixel intensities increases with the rotational speed, including: 501. Associate the accumulation value of each frame in the weight accumulation sequence with the current rotational speed of the blade, and generate a dynamic gain coefficient and a dynamic suppression coefficient for the windward side and the leeward side respectively; In step 501, the dynamic gain coefficient is an amplification parameter for the pixel intensity on the windward side generated based on the weight accumulation value and the blade rotational speed. The dynamic suppression coefficient is a suppression parameter for the pixel intensity on the leeward side generated based on the weight accumulation value and the blade rotational speed.

[0076] In the embodiments of the present application, first, the accumulation values of the windward side and the leeward side of the current frame are extracted from the weight accumulation sequence (such as the windward side accumulation value 1.5 and the leeward side accumulation value 0.8), and combined with the real-time rotational speed (such as 2800 rpm), and the dynamic gain coefficient on the windward side (1.5×0.001×2800≈1.32) and the dynamic suppression coefficient on the leeward side (0.8−0.0005×2800≈0.66) are generated through a non-linear interpolation algorithm; then the two coefficients are written into the image processing parameter table for subsequent adjustment and call.

[0077] 502. In the compensated surface image sequence, mark the dividing line between the windward side and the leeward side along the extension direction of the blade surface texture, and form a gradient adjustment area by expanding a set width to both sides based on the dividing line; In step 502, the gradient adjustment area is a pixel adjustment transition zone that expands to both sides based on the dividing line (such as a smooth transition area with a width of 20 pixels).

[0078] In the embodiments of the present application, first, the main texture direction of the blade surface (such as along the axis direction of the blade) is extracted from the compensated surface image sequence, and the morphological dilation algorithm is used to extend along the main texture direction to generate the dividing line; secondly, with the dividing line as the center, a set width (such as 10 pixels for each side) is expanded to the windward side and the leeward side respectively 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.

[0079] 503. In the windward side range and the leeward side range of the gradient adjustment area, respectively, according to the dynamic gain coefficient and the dynamic suppression coefficient, perform superposition enhancement and subtraction weakening on the intensity value of each pixel; In step 503, the superposition enhancement non-linearly boosts the pixel intensity on the windward side within the gradient adjustment area according to the dynamic gain coefficient. The subtraction attenuation proportionally reduces the pixel intensity on the leeward side within the gradient adjustment area according to the dynamic suppression coefficient.

[0080] In the embodiment of the present application, first, the coordinates of the gradient adjustment area are read from the mask file (e.g., X = 100 - 120, Y = 50 - 70); secondly, the pixel intensity is gradient-enhanced within the windward side range of the gradient adjustment area coordinates according to the dynamic gain coefficient (1.32) (from the demarcation line to the outer edge of the windward side, the gain amplitude linearly increases from 1.0 to 1.32), and the pixel intensity is gradient-suppressed within the leeward side range according to the dynamic suppression coefficient (0.66) (from the demarcation line to the outer edge of the leeward side, the suppression amplitude linearly decreases from 1.0 to 0.66).

[0081] 504. Calculate the difference between the dynamic gain coefficient and the dynamic suppression coefficient based on the current rotational speed of the blade, and perform global intensity adjustment on the overall areas of the windward side and the leeward side outside the gradient adjustment area according to the difference, so that the difference amplitude of the pixel intensities on the windward side and the leeward side increases as the rotational speed increases.

[0082] In step 504, the global intensity adjustment globally adjusts the intensity of the overall areas of the windward side and the leeward side outside the gradient adjustment area.

[0083] In the embodiment of the present application, first, 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 non-linearly corrected based on the real-time rotational speed (e.g., difference × rotational speed correction factor, rotational speed correction factor = 1 + 0.0001 × rotational speed) to obtain the final adjustment difference; finally, the intensity of the overall area of the windward side outside the gradient adjustment area is proportionally boosted according to the final adjustment difference, and at the same time, the intensity of the overall area of the leeward side is proportionally suppressed, so that the bilateral intensity difference dynamically expands as the rotational speed increases.

[0084] The following is a specific example: Suppose a drone detects surface cracks with a width of 2.0 mm on the windward side of the turbine blade of an aircraft in an environment with a flight speed of 250 km / h and a climb angle of 8°, and the turbine speed is 3500 rpm. In step 501, the weight accumulation value on the windward side is 1.8 and on the leeward side is 0.7. Combining with the rotational speed of 3500 rpm, a dynamic gain coefficient of 1.89 and an inhibition coefficient of 0.15 are generated; in step 502, a demarcation line is calibrated along the extension direction of the main crack to form a gradient adjustment area with a width of 30 pixels; in step 503, the intensity gradient on the windward side in the gradient area is increased to 1.89 times, and the leeward side is inhibited to 0.15 times; in step 504, a difference calculation is performed: gain coefficient 1.89 - inhibition coefficient 0.15 = 1.74, and then a difference correction is performed: 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 inhibited by 2.35×8% = 18.8%. Finally, the bilateral difference amplitude is expanded to 22:1.

[0085] Steps 501 - 504 generate dynamic gain and inhibition coefficients, accurately demarcate the gradient adjustment area, and coordinately adjust the local and global intensities, significantly enhancing the texture contrast difference between the windward side and the leeward side in the high-speed rotation scenario, effectively suppressing the weakening of defect features caused by uneven dynamic illumination, and improving the visibility of micro-cracks and coating peeling defects. Combining with the global intensity adaptation mechanism driven by differences, it ensures the stability of the detection system during drastic changes in rotational speed, reduces background noise interference, and realizes high-robustness dynamic defect detection.

[0086] To solve the problem of inaccurate edge distortion modeling caused by complex motion coupling in the detection of surface defects of high-speed rotating fan blades, especially the technical defect of cumulative compensation error under the combined action of dynamic rotation and equipment pose offset. In some embodiments, obtaining the difference in the rotation angle and the offset of the imaging device pose in adjacent frames according to the spatio-temporal 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: 601. Decompose the difference in the rotation angle in adjacent frames into a horizontal angle change component and a vertical angle change component along the blade rotation plane according to the spatio-temporal index; In step 601, the horizontal angle change component is the projection component of the difference in the blade rotation angle between adjacent frames in the blade rotation plane. The vertical angle change component is the component of the difference in the blade rotation angle between adjacent frames in the direction perpendicular to the rotation plane.

[0087] In the embodiments of the present application, first, obtain the rotation angle difference between adjacent frames from the spatio-temporal index (e.g., frame 1: 30°, frame 2: 35°, difference 5°), and establish a blade rotation plane coordinate system; secondly, project the angle difference vector into the blade rotation plane coordinate system, calculate the horizontal angle change component (e.g., 4.3°) and the vertical angle change component (e.g., 2.8°), and generate a component data table for subsequent calls.

[0088] 602. Decompose the offset of the imaging device pose into a horizontal displacement component parallel to the blade rotation direction and a vertical displacement component perpendicular to the rotation direction; In step 602, the horizontal displacement component is the projection of the imaging device pose offset in the blade rotation direction. The vertical displacement component is the projection of the imaging device pose offset in the direction perpendicular to the rotation direction.

[0089] In the embodiments of the present application, first, extract the imaging device pose offset from the spatio-temporal index (e.g., 4 mm displacement in the X direction, 2 mm displacement in the Y direction), and based on the blade rotation direction vector (e.g., unit vector [0.8, 0.6]), use the vector projection algorithm to calculate the dot product of the imaging device pose offset and the vector parallel to / perpendicular to the blade rotation direction, and obtain the horizontal displacement component (e.g., 4×0.8 + 2×0.6 = 4.4 mm) and the vertical displacement component (e.g., 4×0.6 - 2×0.8 = 0.8 mm), and generate a displacement component table.

[0090] 603. Superimpose the horizontal angle change component and the horizontal displacement component, and the vertical angle change component and the vertical displacement component respectively according to a preset ratio to generate the total horizontal distortion offset and the total vertical distortion offset of the blade edge; 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.

[0091] In the embodiments of the present application, first, read the horizontal angle change component (e.g., 4.3°) and the horizontal displacement component (e.g., 4.4 mm) in the component data table, and superimpose them according to a preset weight (e.g., angle weight 0.5, displacement weight 0.2) to generate the total horizontal distortion offset (e.g., 4.3×0.5 + 4.4×0.2 ≈ 3.15); secondly, perform the same weight superposition on the vertical components (2.8° and 0.8 mm) to generate the total vertical distortion offset (2.8×0.5 + 0.8×0.2 ≈ 1.56), and write it into the distortion parameter table.

[0092] 604. Calculate the distortion direction and intensity ratio of the blade edge during continuous movement according to the vector synthesis direction and the sum of the absolute values of the total horizontal distortion offset and the total vertical distortion offset respectively.

[0093] In step 604, the vector synthesis direction is the synthesis 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.

[0094] In the embodiments of the present application, first, the total horizontal distortion offset (3.15) and the total vertical distortion offset (1.56) are read from the distortion parameter table generated in step 603, and the distortion direction is calculated using the vector synthesis algorithm (through the inverse trigonometric function arctan(1.56 / 3.15)≈26.5°); secondly, the sum of the absolute values (3.15 + 1.56 = 4.71) is calculated as the intensity ratio; finally, the distortion direction and the intensity ratio are written into the compensation configuration file.

[0095] The following is a specific example: Suppose a drone detects cracks at the root of the turbine blade of an aircraft in an environment with a flight speed of 210 km / h and a sideslip angle of 12°, and the turbine speed is 3200 rpm. In step 601, the difference in the rotation angle 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° (pitch direction); in step 602, the pose offset of the imaging device is X = 5 mm and Y = 3 mm, which is decomposed into a horizontal displacement of 4.1 mm (along the rotation direction) and a vertical displacement of 2.2 mm (radial); the total horizontal offset obtained in step 603 = 5.2°×0.5 + 4.1 mm×0.2≈3.42, and the total vertical offset = 3.0°×0.5 + 2.2 mm×0.2≈1.94; in step 604, the vector synthesis direction angle arctan(1.94 / 3.42)≈29.6° is calculated as the distortion direction, and the intensity ratio = 3.42 + 1.94 = 5.36, which is used for subsequent motion blur compensation.

[0096] Steps 601 - 604 accurately quantify the dynamic coupling effect of blade edge distortion under high-speed rotation through multi-dimensional decomposition, weighted superposition, and vector synthesis of the rotation angle and pose offset, solve the problem of insufficient accuracy of traditional single-dimensional compensation models, and significantly improve the accuracy of motion blur compensation. Combining the collaborative calculation of horizontal and vertical components ensures the geometric deformation correction and intensity recovery of defect edges in complex motion scenarios, providing a reliable data basis for high-precision defect detection.

[0097] In order to solve the problem of misjudgment of defect types caused by sudden changes in rotational speed and changes in flight attitude during dynamic defect detection of turbine blades in a high-speed flight state, especially the technical defect of confusion between crack and coating peeling characteristics during variable-speed operation. In some embodiments, a blade motion state vector generated by embedding inertial navigation data in the lightweight detection model is used to output a defect detection result matching the blade motion state vector by adjusting the sensitivity thresholds of the lightweight detection model for blade crack and coating peeling defects in different rotational speed intervals, including: 701. Divide the rotational speed range according to the real-time rotational speed in the blade motion state vector generated from inertial navigation data. The rotational speed range includes a steady-state operation range and a variable-speed operation range. In step 701, the steady-state operation range is the stable operating state range where the turbine rotational speed fluctuation range is less than a preset threshold. The variable-speed operation range is the dynamic adjustment state range where the turbine rotational speed change rate exceeds the preset threshold.

[0098] In the embodiment of the present application, first, extract the real-time rotational speed (such as 3050 rpm) and its change rate (such as 80 rpm / s) from the blade motion state vector, and use the sliding window statistical method to calculate the rotational speed standard deviation within 10 consecutive seconds (such as the standard deviation < 30 rpm); secondly, if the standard deviation is lower than the threshold, it is determined as the steady-state range, otherwise it is classified into the variable-speed range; finally, write the mark of the range determination result into the detection model parameter table.

[0099] 702. Set 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 the sensitivity threshold of the blade crack defect in the variable-speed operation range to be higher than that of the coating peeling defect. In step 702, the sensitivity threshold is the lowest confidence score for the detection model to determine the existence of a defect.

[0100] In the embodiment of the present application, first, based on the characteristic that the coating peeling defect is more significant in the stable air flow, set the coating peeling threshold (0.85) to be higher than the crack threshold (0.75) in the steady-state operation range; secondly, because the sudden change in rotational speed is likely to cause crack propagation, set the crack threshold (0.8) to be higher than the coating peeling threshold (0.7) in the variable-speed operation range; finally, embed the set threshold setting strategy into the model decision layer.

[0101] 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. In step 703, the three-dimensional coordinate change rate is the change rate of the spatial position in the blade motion state vector. The feature fusion layer is the multi-modal data fusion network layer in the lightweight detection model that integrates motion parameters and image features.

[0102] In the embodiment of the present application, first, extract the three-dimensional coordinate change rate (such as X = 1.5 mm / s, Y = 0.8 mm / s) and rotational acceleration (50 rpm²) from the blade motion state vector; secondly, input the extracted parameters into the fully connected nodes of the feature fusion layer, and use the adaptive weighting algorithm to adjust the response weights of the fully connected nodes to the crack and coating features (such as the crack weight +0.2, the coating weight -0.1), and finally generate a dynamic weight configuration file for the lightweight detection model to call.

[0103] 704. According to the adjusted response weights, screen the defect probability values output by the lightweight detection model, retain the defect types and defect positions that exceed the sensitivity threshold, and generate a defect detection result that matches the blade motion state vector.

[0104] In step 704, the defect probability value is the confidence score of each defect type output by the lightweight detection model.

[0105] In the embodiment of the present application, first, obtain the probability values of cracks and coating peeling from the output of the lightweight detection model (such as 0.82 and 0.73); secondly, screen whether the crack probability value exceeds the threshold (0.8) according to the current rotational speed range (such as the variable-speed operation range); finally, write the qualified defect type (such as the crack type) and its position coordinates (such as X = 120, Y = 50) into the detection report, and generate a defect detection result that matches the blade motion state vector after excluding the unqualified results.

[0106] The following is a specific example: Suppose a drone detects the surface defects of the turbine blades of an aircraft in an environment with a flight speed of 250 km / h and a roll angle of 8°. The real-time rotational speed of the turbine accelerates from 2800 rpm to 3200 rpm (variable-speed range), the three-dimensional coordinate change rate X = 2.1 mm / s, and the rotational acceleration is 60 rpm². Step 701: Calculate the rotational speed standard deviation of 45 rpm (> threshold 30 rpm), and determine it as the variable-speed operation range; step 702 set the crack threshold to 0.8 and the coating peeling threshold to 0.7; step 703 input the coordinate change rate and rotational acceleration into the feature fusion layer, and after adjustment, the crack weight is +0.25 and the coating weight is -0.15; in step 704, the lightweight detection model outputs a crack probability of 0.83 and a coating peeling probability of 0.68, retain the crack defect position (X = 115, Y = 48), and generate a defect detection result.

[0107] Steps 701 - 704 significantly improve the discrimination ability of cracks and coating peeling defects on turbine blades in high-speed flight scenarios through dynamic division of rotational speed ranges, reverse sensitivity thresholds, fusion of motion parameters to adjust model weights, and threshold matching screening, solve the problem of false detection caused by feature confusion in variable-speed working conditions, ensure the strict matching of the detection result with the dynamic state of the aircraft, and enhance the environmental adaptability of the system.

[0108] Figure 2 The following is a schematic structural diagram of a wind turbine blade surface defect detection system based on UAV vision images provided by the embodiment of the present application, as Figure 2 shown, the system includes: The acquisition module 21 is configured to continuously acquire a surface image sequence of the blades of a high-speed rotating fan through HDR dynamic imaging technology, and synchronously obtain the blade motion trajectory data output by the inertial navigation module. The HDR dynamic imaging technology dynamically adjusts the multi-frame exposure parameters according to the blade rotation speed to suppress the overexposed and underexposed areas caused by high-speed motion. The compensation module 22 is configured to perform spatio-temporal correlation on the blade motion trajectory data and the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the pose of the imaging device extracted from the spatio-temporal correlation result, perform frame-by-frame compensation on the blade edge distortion caused by motion blur in the surface image sequence. The enhancement module 23 is configured to enhance the texture contrast difference between the windward side and the leeward side of the blade in the compensated surface image sequence at different rotation speeds through dynamic weight allocation, and generate a blade surface feature map in combination with the spatio-temporal continuity constraint of the blade motion trajectory data. The migration module 24 is configured to adjust the attention weights 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 weights. The output module 25 is configured to embed the blade motion state vector generated from the inertial navigation data into the lightweight detection model, and output the defect detection result matching the blade motion state vector by adjusting the sensitivity thresholds of the lightweight detection model for blade cracks and coating peeling defects in different rotation speed intervals.

[0109] Figure 2 The described fan blade surface defect detection system based on UAV vision images can execute Figure 1 The described fan blade surface defect detection method based on UAV vision images in the illustrated embodiment, and its implementation principle and technical effects will not be elaborated. For the fan blade surface defect detection system based on UAV vision images in the above embodiment, the specific ways for each module and unit to perform operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0110] In a possible design, Figure 2 The fan blade surface defect detection system in the illustrated embodiment can be implemented as a computing device, such as Figure 3 shown, and this computing device can include a storage component 31 and a processing component 32; The storage component 31 stores one or more computer instructions, and among them, the one or more computer instructions are called and executed by the processing component 32.

[0111] The processing component 32 is used for the aboveFigure 1 A method for detecting surface defects of a wind turbine blade based on drone vision images in the described embodiment.

[0112] Among them, 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 by 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 for executing the above method.

[0113] The storage component 31 is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage 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.

[0114] Of course, the computing device may also necessarily include other components, such as input / output interfaces, display components, communication components, etc.

[0115] The input / output interface provides an interface between the processing component and the peripheral interface module, and the above peripheral interface module may be an output device, an input device, etc.

[0116] The communication component is configured to facilitate communication between the computing device and other devices in a wired or wireless manner, etc.

[0117] Among them, the computing device may be a physical device or an elastic computing host provided by a cloud computing platform, etc. At this time, the computing device may refer to a cloud server, and the above processing component, storage component, etc. may be basic server resources leased or purchased from a cloud computing platform.

[0118] The embodiment of the present application also provides a computer storage medium storing a computer program, and when the computer program is executed by a computer, it can implement the above Figure 1 A method for detecting surface defects of a wind turbine blade based on drone vision images in the shown embodiment.

[0119] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0120] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

[0121] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate 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 a wind turbine blade based on drone vision images, characterized in that, Including: Collecting a surface image sequence of a high-speed rotating fan blade by a drone, and synchronously obtaining the blade motion trajectory data output by an inertial navigation module; Performing spatio-temporal correlation on the blade motion trajectory data and the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the imaging device pose extracted from the spatio-temporal correlation result, compensating the blade edge distortion caused by motion blur in the surface image sequence frame by frame through the dynamic matching relationship; Strengthening the texture contrast difference between the windward side and the leeward side of the blade in the compensated surface image sequence at different rotational speeds through dynamic weight allocation, and generating a blade surface feature map in combination with the spatio-temporal continuity constraint of the blade motion trajectory data; Based on the dynamic distribution ratio of the high-frequency vibration region and the static region in the blade surface feature map, adjusting the attention weight of the knowledge distillation framework, 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; Embedding the inertial navigation module in the lightweight detection model, generating a blade motion state vector in real time according to the inertial navigation data output by the inertial navigation module, and outputting a defect detection result matching the blade motion state vector by adjusting the sensitivity threshold of the lightweight detection model to blade crack and coating peeling defects in different rotational speed intervals.

2. The method according to claim 1, wherein The performing spatio-temporal correlation on the blade motion trajectory data and the surface image sequence, and based on the dynamic matching relationship between the blade rotation angle and the imaging device pose extracted from the spatio-temporal correlation result, compensating the blade edge distortion caused by motion blur in the surface image sequence frame by frame through the dynamic matching relationship, includes: Extracting the blade three-dimensional space positioning data and rotation angle corresponding to each frame image acquisition moment from the blade motion trajectory data, and generating a spatio-temporal index aligned with the time stamp of the surface image sequence; Obtaining the difference in the rotation angle and the offset of the imaging device pose in adjacent frames according to the spatio-temporal index, and calculating the distortion direction and intensity ratio of the blade edge in the continuous motion process according to the spatial geometric superposition relationship between the difference and the offset; Generating a pixel displacement compensation vector for the blade edge region of the current frame based on the distortion direction and intensity ratio, and the pixel displacement compensation vector performs layer-by-layer offset correction on the pixel position in the reverse direction of blade rotation; Performing edge contour sharpening processing on the blade edge distortion caused by motion blur in the surface image sequence according to the pixel displacement compensation vector, eliminating the residual blurred pixel points in the blade edge distortion, and generating a compensated surface image sequence.

3. The method according to claim 1, wherein The strengthening the texture contrast difference between the windward side and the leeward side of the blade in the compensated surface image sequence at different rotational speeds through dynamic weight allocation, and generating a blade surface feature map in combination with the spatio-temporal continuity constraint of the blade motion trajectory data, includes: Dynamically allocate the contrast enhancement weight values for the windward side and leeward side of the blade in the compensated surface image sequence according to the change amount of the blade rotation speed, where 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, accumulate the contrast enhancement weight values 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 weight accumulation sequence, perform positive gain adjustment and negative suppression adjustment on the pixel intensities of the windward side and leeward side respectively, so that the intensity difference of the pixel intensities increases with the rotation speed; Stack the adjusted pixel intensities across frames in the time sequence of the blade motion trajectory data, and generate a blade surface feature map containing the dynamic texture features of the blade surface based on the contour of the intensity mutation region formed after stacking; 4. The method according to claim 1, wherein Adjust the attention weights of the knowledge distillation framework based on the dynamic distribution ratio of the high-frequency vibration region and the static region in the blade surface feature map, and transfer the feature response pattern in the pre-trained high-precision defect detection model to the lightweight detection model according to the adjusted attention weights, including: Extract the fluctuation amplitude of the blade surface texture intensity in the blade surface feature map within adjacent time intervals, and divide the boundary range between the high-frequency vibration region and the static region according to the fluctuation amplitude; Statistically calculate the area ratios of the high-frequency vibration region and the static region in the blade surface feature map to generate a numerical relationship representing the dynamic distribution ratio of the regions; According to the numerical relationship, proportionally amplify the attention weight coefficient corresponding to the high-frequency vibration region in the knowledge distillation framework, and simultaneously reduce the attention weight coefficient corresponding to the static region in the opposite direction; Pre-train a high-precision defect detection model according to the knowledge distillation framework, obtain the feature response pattern in the high-precision defect detection model that matches the high-frequency vibration region and the static region, and map the adjusted attention weight coefficients layer by layer to the lightweight detection model based on the feature response pattern; 5. The method according to claim 3, characterized in that The step of performing positive gain adjustment and negative suppression adjustment on the pixel intensities of the windward side and leeward side respectively according to the numerical distribution of the weight accumulation sequence, so that the intensity difference of the pixel intensities increases with the rotation speed, includes: Associate the accumulation value of each frame in the weight accumulation sequence with the current rotation speed of the blade, and generate dynamic gain coefficients and dynamic suppression coefficients for the windward side and leeward side respectively; In the compensated surface image sequence, mark the dividing line between the windward side and the leeward side along the extension direction of the blade surface texture, and expand a set width on both sides of the dividing line to form a gradient adjustment area; In the windward side range and leeward side range of the gradient adjustment area, superimpose and enhance and subtract and weaken the intensity value of each pixel according to the dynamic gain coefficient and the dynamic suppression coefficient respectively; Calculate the difference between the dynamic gain coefficient and the dynamic suppression coefficient based on the current rotational speed of the blade, and perform global intensity adjustment on the overall areas of the windward side and the leeward side outside the gradient adjustment area according to the difference, so that the difference amplitude of the pixel intensities of the windward side and the leeward side expands as the rotational speed increases.

6. The method according to claim 2, characterized in that, Obtain the difference in the rotation angle and the offset of the imaging device pose in adjacent frames according to the spatio-temporal index, and calculate the distortion direction and intensity ratio of the blade edge during continuous movement according to the spatial geometric superposition relationship between the difference and the offset, including: Decompose the difference in the rotation angle in adjacent frames into a horizontal angle change component and a vertical angle change component along the blade rotation plane according to the spatio-temporal index; Decompose the offset of the imaging device pose into a horizontal displacement component parallel to the blade rotation direction and a vertical displacement component perpendicular to the rotation direction; Superimpose the horizontal angle change component and the horizontal displacement component, and the vertical angle change component and the vertical displacement component respectively according to a preset ratio to generate a total horizontal distortion offset and a total vertical distortion offset of the blade edge; Calculate the distortion direction and intensity ratio of the blade edge during continuous movement according to the vector synthesis direction and the sum of the absolute values of the total horizontal distortion offset and the total vertical distortion offset respectively.

7. The method according to claim 1, characterized in that Embed the blade motion state vector generated from inertial navigation data in the lightweight detection model, and output a defect detection result matching the blade motion state vector by adjusting the sensitivity thresholds of the lightweight detection model for blade crack and coating peeling defects in different rotational speed intervals, including: Divide the rotational speed intervals according to the real-time rotational speed in the blade motion state vector generated from inertial navigation data, and the rotational speed intervals include a steady-state operation interval and a variable-speed operation interval; Set the sensitivity threshold for coating peeling defects in the steady-state operation interval to be higher than that for blade crack defects, and the sensitivity threshold for blade crack defects in the variable-speed operation interval to be higher than that for coating peeling defects; 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 to dynamically adjust the response weights of the nodes of the lightweight detection model to defect features; According to the adjusted response weights, screen the defect probability values output by the lightweight detection model, and retain the defect types and defect positions exceeding the sensitivity threshold to generate a defect detection result matching the blade motion state vector.

8. A fan blade surface defect detection system based on UAV vision images, characterized in that, Including: An acquisition module, configured to continuously acquire a surface image sequence of a high-speed rotating fan blade through HDR dynamic imaging technology, and synchronously obtain blade motion trajectory data output by an inertial navigation module, and the HDR dynamic imaging technology dynamically adjusts multi-frame exposure parameters according to the blade rotational speed to suppress overexposed and underexposed areas caused by high-speed movement; A compensation module, configured to perform spatio-temporal association on the blade motion trajectory data and the surface image sequence, and perform frame-by-frame compensation on the blade edge distortion caused by motion blur in the surface image sequence based on the dynamic matching relationship between the blade rotation angle and the imaging device pose extracted from the spatio-temporal association result; Reinforcement module, which is used to enhance the texture contrast difference between the windward side and the leeward side of the blade in the compensated surface image sequence at different rotational speeds through dynamic weight allocation, and generate a blade surface feature map by combining the spatio-temporal continuity constraint of the blade motion trajectory data; Migration module, which is used to adjust the attention weights 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 weights; Output module, which 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 result matching the blade motion state vector by adjusting the sensitivity thresholds of the lightweight detection model to blade crack and coating peeling defects in different rotational speed intervals.

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 method for detecting surface defects of a wind turbine blade based on UAV vision 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, it implements a method for detecting surface defects of a wind turbine blade based on UAV vision images as described in any one of claims 1 to 7.

Images