Intelligent detection method for surface defects of wind turbine blade based on image processing

Abstract

The application relates to the field of image processing, and discloses an intelligent detection method for surface defects of a wind turbine blade based on image processing, which is used for improving the accuracy of surface defect detection of the wind turbine blade. The method comprises the following steps: fusing pulse laser thermal excitation infrared thermal imaging and a multi-angle polarized light source array polarization imaging technology, constructing a dual-mode data set, extracting thermal anomaly features from a dynamic thermal image sequence, and obtaining photoelastic stress features through phase analysis and local energy filtering. The features are fused by using a thermal conduction blocking and stress concentration space coincidence reinforcement mechanism, a fusion defect indication map is generated, a defect probability distribution map is obtained, a risk value is calculated and predicted, a repair strategy library is matched according to the risk value, a repair scheme code is output, and a three-dimensional repair guide map is generated. The application realizes intelligent and accurate detection and repair guidance of the surface defects of the blade, and improves the comprehensiveness and accuracy of defect detection.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to an intelligent detection method for surface defects on wind turbine blades based on image processing. Background Technology

[0002] With the transformation of the global energy structure, wind energy, as a clean and renewable energy source, is playing an increasingly important role in energy supply. As the core equipment for wind energy development and utilization, the safe and stable operation of wind turbines directly affects the efficiency and reliability of wind power generation. Wind turbine blades, as key components for capturing wind energy, are exposed to a complex natural environment for extended periods, enduring aerodynamic loads, gravity, temperature variations, and other factors, making them highly susceptible to defects such as surface cracks, wear, and corrosion. These defects not only reduce the aerodynamic performance of the blades and affect power generation efficiency but may also lead to serious accidents such as blade breakage, threatening the safe operation of the entire wind turbine. Therefore, timely and accurate detection and assessment of surface defects on wind turbine blades are of great significance for ensuring the reliable operation of wind turbines, extending their service life, and reducing operation and maintenance costs.

[0003] Traditional methods for detecting surface defects on wind turbine blades mainly rely on manual visual inspection and periodic shutdown inspections. Manual visual inspection is limited by factors such as the experience and eyesight of the inspectors, as well as the inspection environment, making it prone to missed or false detections. Furthermore, its low efficiency makes it difficult to meet the rapid inspection needs of large-scale wind farms. While periodic shutdown inspections can provide a more comprehensive check of the blade condition, they increase the downtime of the wind turbine, resulting in power generation losses. Frequent start-up and shutdown operations can also cause damage to the turbine equipment.

[0004] Therefore, we propose an intelligent detection method for surface defects of wind turbine blades based on image processing to solve the above problems. Summary of the Invention

[0005] This invention provides an intelligent detection method for surface defects on wind turbine blades based on image processing, which improves the accuracy of surface defect detection on wind turbine blades.

[0006] The first aspect of this invention provides an intelligent detection method for surface defects on wind turbine blades based on image processing. The method includes: generating instantaneous thermal excitation on the blade surface, synchronously triggering the acquisition of a dynamic thermal image sequence, simultaneously activating a polarized light source array to illuminate the blade surface, acquiring a set of polarized images, and establishing a dual-modal dataset; performing anisotropic diffusion filtering on the dynamic thermal image sequence to generate a thermal conduction gradient tensor matrix, extracting principal curvature extrema to form a thermal anomaly feature map, performing phase analysis on the polarized image set to obtain the photoelastic phase distribution, and combining this with local energy filtering. A photoelastic stress feature map is generated; based on the spatial overlap enhancement mechanism between the thermal conduction obstruction region and the stress concentration region, a fusion defect indication map is generated according to the pixel intensity of the thermal anomaly feature map and the gradient magnitude of the photoelastic stress feature map; in a logarithmic continuous scale space, phase consistency fluctuation analysis is performed on the fusion defect indication map to generate a defect probability distribution map; based on the phase distribution data resolved from the polarization image group and the spatial coordinates of the defect probability distribution map, a three-dimensional depth map of the defect is reconstructed; based on the area of ​​the connected region of the defect probability distribution map and the maximum depth change rate of the three-dimensional depth map, a predicted risk value is obtained.

[0007] Optionally, in a first implementation of the first aspect of the present invention, the method includes: emitting a laser pulse to act on a target area on the blade surface, and simultaneously sending a hardware-level trigger signal to an infrared thermal imager; starting acquisition based on the trigger signal after the laser pulse acts, and generating a time-coded dynamic thermal image sequence; simultaneously with the laser pulse emission, activating 0°, 45°, 90°, and 135° polarized LED light sources in a multi-angle polarization light source array, and capturing four sets of polarization images through a single exposure; performing time axis registration between the dynamic thermal image sequence and the polarization image groups based on the timestamp of the trigger signal, and completing pixel-level spatial alignment using a pre-calibrated spatial transformation matrix, and outputting a dual-modal dataset.

[0008] Optionally, in a second implementation of the first aspect of the present invention, the method includes: performing anisotropic diffusion filtering on the dynamic thermal image sequence to generate a denoised thermal conduction image sequence; calculating the second-order partial derivatives of the denoised thermal conduction image sequence in the spatiotemporal domain to construct a thermal conduction gradient tensor matrix; solving for the eigenvalues ​​of the thermal conduction gradient tensor matrix, extracting the set of maximum principal curvature extrema points, and generating a thermal anomaly feature map; performing photoelastic phase analysis on the polarization image group, calculating the arctangent function value through the four-directional polarization intensity, and generating an original phase distribution map; performing local energy filtering on the original phase distribution map, extracting stress concentration regions, and outputting a photoelastic stress feature map.

[0009] Optionally, in a third implementation of the first aspect of the present invention, the phase angle of the original phase distribution map is: ;in, Intensity in the 0° polarization direction. Intensity at a polarization direction of 45° Intensity in the 90° polarization direction. Intensity at 135° polarization direction, coefficient This is because the phase angle and stress difference exhibit a birefringence effect.

[0010] Optionally, in a fourth implementation of the first aspect of the present invention, the method includes: identifying the heat conduction obstruction region in the thermal anomaly feature map and extracting its pixel intensity distribution as a first physical field input; calculating the gradient magnitude of the photoelastic stress feature map and marking the stress concentration region as a second physical field input; performing a Hadamard product operation on the pixel intensity distribution and the gradient magnitude to generate an initial fusion response map; performing a physical field co-enhancement operation on the initial fusion response map to enhance the response intensity of the spatially overlapping region of heat conduction obstruction and stress concentration, and outputting a fusion defect indication map.

[0011] Optionally, in a fifth implementation of the first aspect of the present invention, the method includes: constructing a logarithmically distributed continuous scale space, with a scale range covering 0.5 pixels to 8.0 pixels, and generating a multi-scale filter bank; performing phase consistency fluctuation analysis on the fused defect indication map in the continuous scale space to obtain the local phase response at each scale; performing scale normalization processing on the local phase response to generate a normalized phase response map set; and integrating the normalized phase response map set along the scale dimension to generate a defect probability distribution map.

[0012] Optionally, in a sixth implementation of the first aspect of the present invention, the method includes: extracting the phase gradient magnitude of the defect region based on the phase distribution data parsed from the polarization image group and the spatial coordinates of the defect probability distribution map; converting the phase gradient magnitude into a depth value using a phase-depth mapping model to generate a three-dimensional depth map of the defect; marking connected regions in the defect probability distribution map to obtain the geometric area; extracting the maximum depth change rate from the three-dimensional depth map of the defect; and obtaining a predicted risk value based on the product of the geometric area and the maximum depth change rate.

[0013] Optionally, in the seventh implementation of the first aspect of the present invention, the method further includes generating maintenance decision parameters: matching the predicted risk value with a pre-stored maintenance strategy library to output maintenance plan code; generating a material removal thickness distribution map based on the spatial distribution characteristics of the three-dimensional depth map of the defect; and combining the maintenance plan code with the material removal thickness distribution map to output a three-dimensional maintenance guidance map.

[0014] The mechanism of this invention is as follows: through the spatiotemporal synchronous coupling of thermal-optical dual physical fields and the image processing chain driven by physical mechanisms, machine learning-free intelligent detection and quantitative decision-making of blade defects are realized.

[0015] Beneficial effects: By fusing the two modes, pulsed laser thermally excited infrared thermography can capture abnormal heat conduction on the blade surface, while multi-angle polarized light source array polarization imaging technology can obtain stress distribution information. The combination of the two provides binocular vision for detection, enabling more comprehensive and accurate discovery of different types of defects. Whether it is a defect caused by obstructed heat conduction or stress concentration, it will be impossible to hide, greatly improving the coverage and accuracy of defect detection.

[0016] Based on the spatial overlap enhancement mechanism between the thermal conduction obstruction region and the stress concentration region, the pixel intensity of the thermal anomaly feature map and the gradient magnitude of the photoelastic stress feature map are multiplied by the Hadamard product to generate a fused defect indication map. This fully considers the coupling relationship between different physical fields, enhances the feature response of the defect region, and makes the fused feature map more clearly indicate the defect location. Compared with the traditional simple fusion method, this greatly improves the sensitivity and accuracy of defect detection.

[0017] In the log-continuous scale space, phase consistency fluctuation analysis is performed on the fused defect indicator map. The defect probability distribution map is generated by the scale-normalized response integral. The defect characteristics are analyzed from multiple scales, which can more comprehensively consider the performance of defects at different scales. The generated defect probability distribution map can more accurately reflect the possibility of defect existence and provide more detailed data support for risk assessment.

[0018] Based on the predicted risk value, the pre-stored maintenance strategy library is matched, and maintenance plan code is output. Targeted maintenance suggestions are provided for each defect. Different maintenance strategies are corresponding to defects with different risk levels, realizing personalized customization of maintenance plans and improving the pertinence and effectiveness of maintenance.

[0019] Based on the spatial distribution characteristics of the 3D depth map of the defect, a material removal thickness distribution map is generated. By combining the repair plan code with the material removal thickness distribution map, a 3D repair guidance map is output. Repair personnel can intuitively see the location and depth of the defect and the specific requirements of the repair operation, thus obtaining an accurate navigation map. This greatly reduces the difficulty and uncertainty of repair, improves repair efficiency and quality, and reduces secondary damage caused by improper repair. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an embodiment of the intelligent detection method for surface defects of wind turbine blades based on image processing in this invention.

[0021] Figure 2 This is a schematic diagram of another embodiment of the intelligent detection method for surface defects of wind turbine blades based on image processing in this invention.

[0022] Figure 3This is a schematic diagram of an embodiment of the intelligent detection device for surface defects of wind turbine blades based on image processing in this invention. Detailed Implementation

[0023] This invention provides an intelligent detection method for surface defects on wind turbine blades based on image processing, which improves the accuracy of surface defect detection on wind turbine blades. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent detection method for surface defects of wind turbine blades based on image processing in this invention includes:

[0025] 101. Dual-physics field image synchronous acquisition: A pulsed laser thermal excitation device generates instantaneous thermal excitation on the blade surface, synchronously triggering an infrared thermal imager to acquire a dynamic thermal image sequence with a time resolution of 500Hz or higher; at the same time, a multi-angle polarization light source array is activated to illuminate the blade surface, and a four-way polarization camera synchronously acquires a group of polarization images containing polarization states of 0°, 45°, 90°, and 135° to establish a dual-modal dataset;

[0026] It is understood that the executing entity of this invention can be an intelligent detection device for surface defects on wind turbine blades based on image processing, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0027] It should be noted that the equipment selection and configuration, specifically the pulsed laser thermal excitation device, uses a pulsed laser with a wavelength of 1064nm (peak power 5kW) to generate instantaneous thermal excitation on the blade surface within 0.5ms, with a spot diameter of 10mm and an energy density of 5J / cm². 2 This ensures that the heat penetration depth reaches 2mm on the surface of the blade composite material; the laser trigger signal outputs a synchronous pulse (rising edge <10ns) through the BNC interface to trigger the infrared thermal imager and polarization camera to start.

[0028] Infrared thermal imager acquisition system: Employs a high-speed mid-wave infrared camera (FLIRX8500sc), with a resolution of 640×512 pixels, a sampling rate set to 500Hz or higher (520Hz in this example), and a noise equivalent temperature difference (NETD) of <20mK. The dynamic thermal image sequence acquisition time is 1.5 seconds, generating 780 thermal images with a temporal resolution of 1.92ms / frame, recording the temperature decay process on the blade surface.

[0029] Polarization optical system: Light source: Annular LED polarization light source array (wavelength 850nm), illuminating the leaf with four-way polarization angles of 0°, 45°, 90°, and 135°, providing an illuminance of 100klux and covering an area of ​​1m². 2 Camera: Four-directional polarization CMOS camera (LUCIDPHX050S), resolution 2448×2048, pixel size 3.45μm, which captures four sets of polarization state images simultaneously through time-division exposure (exposure time 200μs / state).

[0030] Spatial registration and synchronization control: Spatial registration: Four high-reflectivity targets (5mm in diameter) are preset on the blade surface, with a laser thermal excitation point 20cm away from the target center. The infrared and polarization cameras achieve spatial alignment through target coordinate mapping, with a registration error of <0.1 pixels.

[0031] Timing synchronization: The laser trigger signal synchronously starts the infrared thermal imager (internal clock) and the polarization camera (external trigger mode). The time deviation between the first frame of the infrared sequence and the first frame of the polarization image group is <50μs, verified by the timestamp of the acquisition card (ADLINK PCIe-CPL64).

[0032] The data acquisition process involves the laser emitting a pulse at position S1 on the blade, simultaneously triggering the infrared thermal imager to record a thermal sequence (780 frames, 520Hz) and the polarization camera to capture a four-state image group (4 images per group, 1ms interval). The translation device moves to the adjacent position S2 (15cm apart), and the above steps are repeated to cover the entire 60m length of a single blade (400 detection points in total).

[0033] Data volume per collection: Hot sequence: MB; Polarization group: MB; Total size of the bimodal dataset: GB.

[0034] Output bimodal dataset structure:

[0035]

[0036] 102. Collaborative extraction of physical field features: After anisotropic diffusion filtering of the dynamic thermal image sequence, the second-order partial derivatives in the spatiotemporal domain are calculated to generate the thermal conduction gradient tensor matrix, the principal curvature extrema are extracted to form a thermal anomaly feature map, the polarization image group is phase-analyzed, the photoelastic phase distribution is calculated by the arctangent function, and the photoelastic stress feature map is generated by combining local energy filtering.

[0037] It should be noted that the thermal physical field feature extraction (dynamic thermal image sequence processing) and anisotropic diffusion filtering are performed on the dynamic thermal image sequence (780 frames, 640×512 pixels) acquired at 520Hz. The thermal conductivity is set. Smoothing coefficient The iteration count is 20. Noise in uniform regions is suppressed through heat conduction while preserving defect edges (regions with gradient abrupt changes > 30℃ / px). Heat conduction gradient tensor generation: The second-order partial derivatives in the spatiotemporal domain of the filtered sequence are calculated to generate the heat conduction gradient tensor matrix at time points. At time s, extract the spatial gradient components. , and time gradient This forms a 3×3 tensor matrix. Thermal anomaly feature map generation: The principal curvature extrema of the tensor matrix (curvature radius < 0.5 mm⁻¹) are calculated and marked as thermal anomaly regions. Actual measurements show that the curvature extrema of the crack region reaches 12.8, significantly higher than the background value (0.3~1.2).

[0038] Photophysical field feature extraction (polarization image group processing), phase resolution: Photoelastic phase calculation is performed on a four-axis polarization image group (2448×2048 pixels, 0° / 45° / 90° / 135°). The arctangent function is used. ;

[0039] Generate a phase distribution map (range 0~π), where the phase jump in the crack region is >1.2 rad.

[0040] Photoelastic stress characteristic map generation: Local energy filtering (window size 5×5) is applied to the phase map to enhance stress concentration areas. The measured gradient modulus of the bubble defect reaches 85 MPa / px, while the normal region is <5 MPa / px. Output example:

[0041]

[0042] 103. Fusion of thermal-optical physical features: Based on the spatial overlap enhancement mechanism of the thermal conduction obstruction region and the stress concentration region, the pixel intensity of the thermal anomaly feature map and the gradient magnitude of the photoelastic stress feature map are subjected to Hadamard product operation to generate a fused defect indication map.

[0043] It should be noted that the input feature map alignment is as follows: Thermal anomaly feature map (640×512 pixels): generated in step 102, with the extreme value of the principal curvature of the crack region ≥12.8mm. -1 (Normalized intensity 0.85), background area ≤ 1.2mm -1 (Intensity 0.05-0.15). Photoelastic stress feature map (2448×2048 pixels): Bubble defect gradient modulus ≥85MPa / px (normalized value 0.92), normal area <5MPa / px (normalized value 0.08). By mapping the coordinates of four preset high reflectivity targets, the thermal anomaly feature map is downsampled to 2448×2048 resolution through bicubic interpolation, spatially aligned with the photoelastic stress feature map, with a registration error <0.3 pixels.

[0044] The Hadamard product fusion operation employs a spatial overlap enhancement mechanism: In the crack region, internal structural fracture leads to impaired heat conduction (high thermal anomaly intensity), while stress concentration occurs (high gradient modulus), resulting in a spatial overlap >90%. Surface contaminants only cause optical scattering (medium to high gradient modulus) but no thermal conduction anomaly (low thermal anomaly intensity), with an overlap <15%.

[0045] Pixel-level fusion calculation: Perform Hadamard product (i.e., multiply the corresponding pixel intensities) on each pixel of the two registered feature maps: Crack region example: Thermal anomaly intensity 0.85 × stress gradient magnitude 0.92 → fusion value 0.782 Background region example: Thermal anomaly intensity 0.12 × stress gradient magnitude 0.10 → fusion value 0.012;

[0046] Output a fused defect indication map (2448×2048, float32 format), with the intensity of the actual defect area increased to above 0.75 and background noise suppressed to below 0.05. Output:

[0047]

[0048] 104. Multi-scale defect boundary segmentation: In the logarithmic continuous scale space of 0.5-8.0 pixels, phase consistency fluctuation analysis is performed on the fused defect indication map, and a defect probability distribution map is generated by the scale normalized response integral.

[0049] It should be noted that the logarithmic continuous scale space is constructed, with a scale range of 0.5 to 8.0 pixels on the fusion defect indicator map (2448×2048 pixels), and a total of 12 scale levels (scale parameters). Spatial sampling: Each scale layer is convolved using a Gaussian kernel function to generate a multi-scale image pyramid. At this scale, the gradient response intensity at the crack edge reaches its peak (approximately 0.78), while the background noise response is <0.05.

[0050] Phase consistency fluctuation analysis, directional filtering: An 8-directional Gabor filter bank (angle interval 22.5°) is used to extract phase consistency features at each scale. Taking the crack region as an example, the phase consistency value reaches 0.92 in the 45° direction (main crack direction), while it is only 0.15 in the vertical direction (135°). Multi-scale response integration: The phase consistency fluctuation energy of each pixel is calculated at all scales and directions: Crack region: Energy peaks are concentrated in... Scale, average fluctuation energy 0.85; background region: energy is dispersed and the mean is <0.10.

[0051] Scale-normalized response integral, weight allocation: weights are allocated based on scale sensitivity, small scale ( ) Weighting accounts for 60% (capturing minute cracks), large scale ( The defect probability distribution is 20% (to suppress structural texture interference). Probability map generation: Weighted integral of the phase consistency response at 12 scales outputs a defect probability distribution map (2448×2048, float32 format). Measured data: Crack region probability value ≥ 0.90 (confidence interval); Stain / texture region probability value ≤ 0.15; Defect-background signal-to-noise ratio improved to 28:1 (15:1 when inputting the fused map). Output verification:

[0052]

[0053] 105. Three-dimensional morphology quantification and risk assessment: Based on the phase distribution data resolved from the polarization image group and the spatial coordinates of the defect probability distribution map, a three-dimensional depth map of the defect is reconstructed through a phase-depth mapping model. Based on the area of ​​the connected region of the defect probability distribution map and the maximum depth change rate of the three-dimensional depth map, the predicted risk value is calculated.

[0054] It should be noted that the 3D depth reconstruction of the defect involved input data alignment: Polarization phase data: the photoelastic phase distribution was analyzed from a four-directional polarization image group (0° / 45° / 90° / 135°), with a phase jump value in the crack region ≥1.2 rad (spatial resolution 2448×2048). Defect probability distribution map: generated in step 104, with a crack region probability value ≥0.90 and coordinate range (x:1200-1250, y:800-850).

[0055] Phase-depth mapping: Establishing a linear relationship between phase difference and depth: a 1 rad increase in phase transition corresponds to a 0.8 mm increase in depth. At the crack center (x:1225, y:825), a 1.5 rad phase transition corresponds to a 1.2 mm depth; at the edge (x:1200, y:800), a 0.3 rad phase transition corresponds to a 0.24 mm depth. Output a 3D depth map (2448×2048, float32), with a maximum crack depth of 1.2 mm and a minimum depth of 0.2 mm (background noise <0.05 mm).

[0056] Risk quantification calculation, parameter extraction: Connected region area: The area of ​​connected regions in the defect probability map above the threshold of 0.85 is 800 mm. 2 (Equivalent rectangle 40mm × 20mm). Maximum depth change rate: The peak value of the depth change rate per millimeter along the principal crack direction (45°) is 0.25 (calculated as follows: ).

[0057] Risk Value Calculation: Risk Model: ( , ).

[0058] This example:

[0059] (Risk threshold: >500 is considered high risk). Output report:

[0060]

[0061] In this embodiment of the invention, defect detection is performed by combining thermophysical and optical physical fields. The thermophysical field reflects changes in heat conduction caused by defects in the internal structure of the blade, while the optical physical field reflects the surface stress distribution. The fusion of these two fields overcomes the limitations of single-field detection; surface stains may appear in the optical physical field but not in the thermophysical field. This fusion allows for accurate differentiation between defects and interference, significantly improving the accuracy and reliability of defect detection. A unique method is employed in the feature extraction stage. In thermophysical field feature extraction, anisotropic diffusion filtering can preserve defect edges while suppressing noise in uniform regions. Optical physical field feature extraction calculates the photoelastic phase distribution using the arctangent function, accurately capturing phase jumps in crack regions, enabling extracted features to more precisely reflect defect characteristics and providing a solid foundation for subsequent accurate analysis. Feature fusion is performed based on the spatial overlap enhancement mechanism between heat conduction-impeded regions and stress concentration regions. A fused defect indicator map is generated through Hadamard product operations, fully utilizing the correlation between the two physical field features in the defect region. This significantly enhances the intensity of the real defect region, suppresses background noise, and greatly improves the contrast between the defect and the background, facilitating accurate defect boundary segmentation. Phase consistency fluctuation analysis in a logarithmic continuous scale space enables multi-scale defect boundary segmentation. By constructing a multi-scale image pyramid, extracting phase consistency features using a multi-directional Gabor filter bank, and generating a defect probability distribution map through scale-normalized response integration, defect features at different scales can be captured. At small scales, fine cracks are captured; at large scales, structural texture interference is suppressed, effectively improving defect boundary positioning accuracy and reducing edge positioning errors. This not only achieves defect detection but also establishes a three-dimensional morphology quantification and risk assessment system. Based on the phase distribution data of polarization image groups and the defect probability distribution map, a three-dimensional depth map of the defect is reconstructed, and the predicted risk value is calculated by combining the area of ​​the connected region and the maximum depth change rate. This system can comprehensively assess the risk from multiple aspects such as the geometric morphology and physical characteristics of the defect, providing a scientific basis for blade maintenance decisions, helping to promptly identify high-risk defects, and ensuring the safe operation of wind turbine units.

[0062] Please see Figure 2 Another embodiment of the intelligent detection method for surface defects of wind turbine blades based on image processing in this invention includes:

[0063] 201. Instantaneous thermal excitation is generated on the blade surface to synchronously trigger the acquisition of dynamic thermal image sequences; at the same time, a polarization light source array is activated to illuminate the blade surface, and a polarization image group is acquired to establish a dual-modal dataset.

[0064] It should be noted that the pulsed laser thermal excitation device emits a laser pulse with a duration of 10μs to act on the target area on the blade surface, and simultaneously sends a hardware-level trigger signal to the infrared thermal imager. Based on the trigger signal, the infrared thermal imager starts acquisition within 1ms after the laser pulse is applied, and generates a time-coded dynamic thermal image sequence at a frame rate of not less than 500Hz. At the same time as the laser pulse is emitted, the 0°, 45°, 90°, and 135° polarization LED light sources of the multi-angle polarization light source array are activated, and four sets of polarization images are captured by a single exposure through the beam splitter prism of the four-directional polarization camera. Based on the timestamp of the trigger signal, the dynamic thermal image sequence and the polarization image group are time-axis registered, and pixel-level spatial alignment is completed using a pre-calibrated spatial transformation matrix, outputting a spatiotemporally unified dual-modal dataset.

[0065] It should be noted that the hardware configuration is as follows: Pulsed laser excitation: A pulsed laser with a wavelength of 1064nm and a peak power of 2kW is used to apply a laser pulse with a duration of 10μs to the target area (50mm×50mm) on the blade surface, with an energy density of 5J / cm². 2 .

[0066] Infrared thermal imager: The Gaode Zhigan PT series flagship thermal imager (resolution 1280×1024 pixels, temperature measurement accuracy ±2%) is selected. The acquisition is started within 1ms after the laser pulse ends by a hardware trigger signal (TTL level), and the image is continuously recorded for 2 seconds at a frame rate of 500Hz to generate a dynamic thermal image sequence of 1000 frames (time encoding accuracy 0.1ms).

[0067] Polarization Imaging System: Light Source: Circular polarized LED array (0°, 45°, 90°, 135° polarization states), wavelength 520nm, illuminance 5000lux, illuminated synchronously with laser pulses (delay <1μs). Camera: Four-way beam splitter polarization camera (IMX250MYR sensor), capturing four sets of polarization images simultaneously in a single exposure (100μs) (resolution 2048×1536).

[0068] Synchronization and Alignment Process: Timing Control: The laser controller sends a trigger signal to the thermal imager and polarization source controller, with a timestamp synchronization error ≤50ns. The start frame of the thermal imaging sequence is marked as... ms, polarization image group labeled as μs (laser application time). Spatial calibration: Pre-calibration stage: A checkerboard calibration plate is attached to the blade surface. Images are acquired by the thermal imager and polarization camera respectively, and the spatial transformation matrix is ​​calculated (affine transformation, accuracy ±0.5 pixels). Alignment operation: The polarization image group is registered to the thermal image sequence coordinate system through bilinear interpolation, and a 512×512 pixel dual-modal dataset is output (including thermal image AD values, polarization phase angle, and intensity).

[0069] Output dataset example: Thermal image sequence: Each frame contains a temperature field matrix (unit: AD value), time axis interval 2ms, spatial resolution 0.1mm / pixel. Polarization image group: Four-channel polarization intensity matrix (I0°, I... 45 °、I 90 °、I 135 °), after registration, it shares the coordinate system with the thermal image sequence. Metadata: timestamp sequence, ambient temperature (25℃), emissivity (0.95), laser energy parameters.

[0070] 202. After performing anisotropic diffusion filtering on the dynamic thermal image sequence, a thermal conduction gradient tensor matrix is ​​generated, the principal curvature extrema are extracted to form a thermal anomaly feature map, phase analysis is performed on the polarization image group to obtain the photoelastic phase distribution, and a photoelastic stress feature map is generated by combining local energy filtering.

[0071] Specifically, anisotropic diffusion filtering is performed on the dynamic thermal image sequence to generate a denoised heat conduction image sequence; the second-order partial derivatives in the spatiotemporal domain of the denoised heat conduction image sequence are calculated to construct a heat conduction gradient tensor matrix; the eigenvalues ​​of the heat conduction gradient tensor matrix are solved to extract the set of maximum principal curvature extrema points and generate a thermal anomaly feature map; photoelastic phase analysis is performed on the polarized image group, and the arctangent function value is calculated through the four-directional polarization intensity to generate an original phase distribution map; local energy filtering is performed on the original phase distribution map to extract stress concentration regions and output a photoelastic stress feature map;

[0072] It should be noted that for the 50mm×50mm inspection area on the surface of the wind turbine blade (including artificially pre-made defects: 0.2mm deep microcracks and 3mm diameter debonding area), dynamic thermal imaging sequences (500Hz frame rate, 2 seconds duration) and polarization image groups (0°, 45°, 90°, 135° four-way polarization states) were simultaneously acquired.

[0073] Thermal anomaly feature map generation process, dynamic noise reduction by thermal conduction: anisotropic diffusion filtering (conduction coefficient 0.15, iteration count 10) is performed on 1000 frames of thermal image sequence to suppress noise while retaining defect edges, and output the noise-reduced thermal image sequence (signal-to-noise ratio improved to 35dB).

[0074] Construction of thermal conduction gradient tensor: Calculate the second-order partial derivatives of the denoised sequence in the spatiotemporal domain (time step 2ms, spatial step 0.1mm / pixel) to generate a 5×5×5 thermal conduction gradient tensor matrix (dimension: spatial x×y×time t).

[0075] Principal curvature extremum extraction: Solve the tensor matrix eigenvalues, extract the maximum principal curvature extremum point (threshold > 0.8), mark the heat conduction obstruction area (debonding area), and output the thermal anomaly feature map (resolution 512×512 pixels, defect area intensity value 120-255).

[0076] Photoelastic stress feature map generation process, photoelastic phase analysis: based on four-directional polarization intensity (I0°, I... 45 °、I 90 °、I 135 °), calculate the original phase distribution map according to the formula:

[0077] ;coefficient This is because the phase angle and stress difference exhibit a birefringence effect;

[0078] The phase angle jump in the microcrack region is ±60° (within ±10° of the background).

[0079] Local energy filtering enhancement: A 5×5 local energy filter (weight kernel: Gaussian-Laplace hybrid) is used to enhance the stress concentration area and output a photoelastic stress characteristic map (the area with gradient modulus > 0.6 is marked as stress anomaly, corresponding to the crack location).

[0080] Thermal anomaly characteristic map: The debonding zone exhibits high-temperature retention (average pixel intensity 180), while microcracks show low-temperature dark bands (intensity value 50) due to heat flow obstruction. Photoelastic stress characteristic map: Significant stress concentration at the microcrack edges (peak gradient modulus value 0.92), while the stress distribution in the debonding zone is uniform (gradient modulus value <0.3).

[0081] 203. Based on the spatial overlap enhancement mechanism of the heat conduction obstruction region and the stress concentration region, a fusion defect indication map is generated according to the pixel intensity of the thermal anomaly feature map and the gradient magnitude of the photoelastic stress feature map.

[0082] Specifically, the heat conduction obstruction region in the thermal anomaly feature map is identified, and its pixel intensity distribution is extracted as the first physical field input; the gradient magnitude of the photoelastic stress feature map is calculated, and the stress concentration region is marked as the second physical field input; based on the thermal-stress spatial coupling mechanism, the pixel intensity distribution and the gradient magnitude are subjected to Hadamard product operation to generate an initial fusion response map; physical field cooperative enhancement operation is performed on the initial fusion response map to enhance the response intensity of the spatially overlapping region of heat conduction obstruction and stress concentration, and a fusion defect indication map is output;

[0083] It should be noted that for the 50mm×50mm inspection area on the surface of the wind turbine blade (including artificially pre-made defects: 0.2mm deep microcracks and 3mm diameter debonding area), a thermal anomaly feature map (resolution 512×512 pixels) and a photoelastic stress feature map (same resolution) have been generated through step 202.

[0084] The process for generating the defect indication map involves the following physical field inputs: Thermal anomaly feature map: Extracting regions with impeded heat conduction (pixel intensity of 180±10 in the debonding area, and low intensity of 50±5 in the microcrack area due to thermal blockage) as the first physical field input. Photoelastic stress feature map: Calculating the gradient magnitude (Sobel operator), and marking stress concentration regions (gradient magnitude of 0.92±0.05 at the microcrack edge, and background gradient magnitude of <0.3 in the debonding area) as the second physical field input.

[0085] Hadamard product operation: Perform pixel-by-pixel multiplication on the two feature maps: Debonding zone: thermal intensity (180) × stress gradient modulus (0.25) → initial response value 45; Microcrack: thermal intensity (50) × stress gradient modulus (0.92) → initial response value 46;

[0086] An initial fusion response map (pixel value range 0~255) is generated. At this time, the response intensity of the microcrack and the debonding zone is similar and difficult to distinguish.

[0087] Physical field synergistic enhancement: Spatial overlapping region identification: Locating overlapping regions of heat conduction obstruction and stress concentration (overlap of microcrack regions >85%, debonding regions <30%). Enhancement operation: Applying a 3x weight enhancement to overlapping regions: the microcrack response value increases from 46 to 138; the response value of the debonding region remains at 45 (no enhancement due to low overlap); the response value of non-overlapping regions decreases by 50%, suppressing background noise.

[0088] Output results, fusion defect indication map: microcrack areas are significantly highlighted after reinforcement (average pixel intensity 138±12), debonding areas have weak response (intensity 45±8), and background noise intensity <10.

[0089] 204. In a log-continuous scale space, perform phase consistency fluctuation analysis on the fused defect indication map to generate a defect probability distribution map;

[0090] Specifically, a continuous scale space with a logarithmic distribution is constructed, covering a scale range from 0.5 pixels to 8.0 pixels, to generate a multi-scale filter bank; in the continuous scale space, phase consistency fluctuation analysis is performed on the fused defect indication map to calculate the local phase response at each scale; the local phase response is scale-normalized to generate a normalized phase response map set; the normalized phase response map set is integrated along the scale dimension to generate a defect probability distribution map.

[0091] It should be noted that the input data is: a fused defect indication map (resolution 512×512 pixels), containing a 0.2mm deep microcrack (response intensity 138±12) and a 3mm diameter debonding zone (response intensity 45±8), with a background noise intensity <10.

[0092] Implementation details, multi-scale filter bank construction: Construct a logarithmic scale space with a scale range of 0.5~8.0 pixels, generating 6 scale layers (0.5, 1.0, 2.0, 4.0, 6.4, 8.0 pixels) at exponential intervals. The corresponding filter sizes for each scale are: minimum scale 0.5 pixels (3×3 Gaussian kernel), maximum scale 8.0 pixels (41×41 Gaussian kernel), used to capture microscopic cracks to macroscopic debonding defects.

[0093] Phase consistency fluctuation analysis: At each scale, the local phase response of the fused image was calculated: 0.5 pixel scale: sensitive to microcrack edges, local phase response peak value 0.85 (only 0.15 in the debonding zone); 8.0 pixel scale: captures the overall shape of the debonding zone, response peak value 0.70 (0.25 for microcracks). The phase response calculation used a Log-Gabor odd-symmetric filter bank with a directional resolution of 22.5° (8 directions in total).

[0094] Scale normalization: Energy normalization is performed on the local phase response at each scale: microcrack region: 0.5 pixel scale weight 0.6, 8.0 pixel scale weight 0.15; debonding region: 0.5 pixel scale weight 0.1, 8.0 pixel scale weight 0.5. Output normalized response atlas to eliminate scale-sensitive differences.

[0095] Defect probability distribution generation: Integral normalized response atlas along the scale dimension: Microcrack region: integral value 0.92 (close to 1.0 indicates high defect probability); Debonding region: integral value 0.35 (medium to low probability); Background region: integral value <0.1. The defect probability distribution map is output in heatmap form, with a probability threshold set to 0.6 (>0.6 indicates high-risk defects). Scale parameter table:

[0096]

[0097] 205. Based on the phase distribution data resolved from the polarization image group and the spatial coordinates of the defect probability distribution map, reconstruct a three-dimensional depth map of the defect. Based on the area of ​​the connected region of the defect probability distribution map and the maximum depth change rate of the three-dimensional depth map, obtain the predicted risk value.

[0098] Specifically, based on the phase distribution data parsed from the polarization image group and combined with the spatial coordinates of the defect probability distribution map, the phase gradient magnitude of the defect region is extracted; the phase gradient magnitude is converted into a depth value through a phase-depth mapping model to generate a three-dimensional depth map of the defect; connected regions are marked in the defect probability distribution map and their geometric areas are calculated; the maximum depth change rate is extracted from the three-dimensional depth map of the defect; and the predicted risk value is calculated based on the product of the geometric area and the maximum depth change rate.

[0099] It should be noted that the input data includes: polarization image group: intensity matrix of four-directional polarization states (0°, 45°, 90°, 135°) (resolution 2048×1536), phase gradient magnitude at the microcrack edge 0.95±0.03, and phase gradient magnitude at the center of the debonding zone 0.15±0.05. The defect probability distribution map is 512×512 pixels, with a probability value of 0.92 (high risk) for the microcrack area, 0.35 (medium to low risk) for the debonding zone, and <0.1 for the background area.

[0100] The implementation process involves phase-depth mapping: extracting the phase gradient magnitude of regions with a defect probability > 0.6 (microcracks: 0.95, debonding zone: 0.15). This is achieved by calibrating the phase-depth model (empirical formula: ...). , The maximum depth of the microcrack is 0.20 mm (error ±0.02 mm), and the depth of the debonding zone is 0.03 mm (error ±0.01 mm).

[0101] Defect region quantification: Connected regions in the marked probability map: Microcracks: Linear regions, area 2.5mm. 2 (Length × Width: 5mm × 0.5mm); Debonding area: Circular area, 7.1mm² 2 (3mm in diameter).

[0102] Maximum depth change rate (unit: mm / pixel) extracted from 3D depth map: microcrack edge: 0.08 (sharp change area); debonding zone center: 0.005 (gentle change area).

[0103] Predicted risk value calculation: Microcracks Debonding zone: Normalization (range 0-1): Microcrack risk level 0.82 (high risk), debonding zone 0.15 (low risk). Parameter comparison table:

[0104]

[0105] 206. Maintenance Decision Parameter Generation: Based on the predicted risk value, match the pre-stored maintenance strategy library and output maintenance plan code; based on the spatial distribution characteristics of the defect 3D depth map, generate a material removal thickness distribution map; combine the maintenance plan code and the material removal thickness distribution map to output a 3D maintenance guidance map.

[0106] It should be noted that the input data includes: predicted risk values: 0.82 for the microcrack area (high risk) and 0.15 for the debonding area (low risk). The 3D depth map of the defects shows: maximum microcrack depth of 0.20 mm and debonding area depth of 0.03 mm (spatial resolution 0.1 mm / pixel).

[0107] The process involves steps and matching maintenance solutions: a pre-stored maintenance strategy library defines three types of solutions: High-risk solution (H): Immediate shutdown for maintenance, code "H001" (including carbon fiber reinforcement + epoxy resin filling process). Medium-risk solution (M): Planned maintenance, code "M002" (resin filling only). Low-risk solution (L): Handled during annual maintenance, code "L003" (surface grinding). A microcrack risk value of 0.82 matches solution code H001; a debonding zone risk value of 0.15 matches solution code L003.

[0108] Material removal thickness distribution map generation: Spatial distribution characteristics based on 3D depth map: Microcrack region: Depth 0.20mm → mm (leaf tip area). De-adhesion zone: depth 0.03mm → mm (leaf root region), output thickness distribution heat map.

[0109] 3D Repair Guidance Map Synthesis: Overlay repair scheme codes and thickness distribution maps: micro-crack areas are marked H001, and a 0.24mm layered grinding path is displayed simultaneously; debonding areas are marked L003. Output 3D guidance map, supporting AR device visualization operation (positioning accuracy ±2mm).

[0110] Repair plan and parameter correspondence table:

[0111]

[0112] In this embodiment of the invention, transient thermal excitation and multi-angle polarization imaging are triggered synchronously, and spatiotemporal alignment is achieved through hardware-level timestamps to construct a spatiotemporally unified dual-modal dataset. A Hadamard product operation and physical field co-enhancement algorithm are proposed, and the response weights are dynamically adjusted by the spatial overlap between the heat conduction obstruction region and the stress concentration region. A 0.5~8.0 pixel logarithmic scale space is constructed, and multi-scale features from micro-cracks to macro-debonding are captured through a Log-Gabor filter bank. Scale normalization processing is combined to eliminate scale-sensitive differences. Based on the phase-depth mapping model, the phase gradient magnitude is converted into a depth value, and the predicted risk value is calculated by combining the area of ​​the connected region and the maximum depth change rate. A maintenance strategy library is pre-stored, and the solution code is automatically output according to the risk value. The maintenance solution code and the material removal thickness distribution map are superimposed to generate a three-dimensional guidance map supported by AR devices.

[0113] Figure 3This is a schematic diagram of the structure of an intelligent detection device for surface defects of wind turbine blades based on image processing, provided in an embodiment of the present invention. The intelligent detection device 300 for surface defects of wind turbine blades based on image processing can vary considerably due to differences in configuration or performance. The device 300 includes a transmitter 301, a receiver 302, and a processor 303. The processor 303 can also be a controller. Figure 3 The device is designated as "controller / processor 303". Optionally, the device 300 may also include a modem processor 305, which may include an encoder 306, a modulator 307, a decoder 308, and a demodulator 309.

[0114] In one example, transmitter 301 modulates (e.g., analog-to-analog conversion, filtering, amplification, and up-conversion, etc.) the output sample and generates an uplink signal, which is transmitted via an antenna to an access network device. On the downlink, the antenna receives the downlink signal transmitted by the access network device. Receiver 302 modulates (e.g., filtering, amplification, down-conversion, and digitization, etc.) the signal received from the antenna and provides an input sample. In modem processor 305, encoder 306 receives service data and signaling messages to be transmitted on the uplink and processes (e.g., formatting, encoding, and interleaving) the service data and signaling messages. Modulator 307 further processes (e.g., symbol mapping and modulation) the encoded service data and signaling messages and provides an output sample. Demodulator 309 processes (e.g., demodulates) the input sample and provides a symbol estimate. Decoder 308 processes (e.g., deinterleaving and decoding) the symbol estimate and provides decoded data and signaling messages to device 300. Encoder 306, modulator 307, demodulator 309, and decoder 308 can be implemented by a combined modem processor 305. These units perform processing according to the radio access technology adopted by the radio access network (e.g., LTE and other evolved systems access technologies). It should be noted that when device 300 does not include modem processor 305, the above-mentioned functions of modem processor 305 can also be performed by processor 303.

[0115] The processor 303 controls and manages the operation of the device 300, and is used to execute the processing procedures performed by the device 300 in the above embodiments of this disclosure. For example, the processor 303 is also used to execute various steps of the transmitting or receiving device in the above method embodiments, and / or other steps of the technical solutions described in the embodiments of this disclosure.

[0116] Furthermore, the device 300 may also include a memory 304 for storing program code and data for the device 300.

[0117] Understandable Figure 3Only a simplified design of device 300 is shown. In practical applications, device 300 can include any number of transmitters, receivers, processors, modem processors, memory, etc., and all devices that can implement the embodiments of this disclosure are within the protection scope of the embodiments of this disclosure.

[0118] The present invention also provides an intelligent detection device for surface defects of wind turbine blades based on image processing. The intelligent detection device for surface defects of wind turbine blades based on image processing includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the intelligent detection method for surface defects of wind turbine blades based on image processing in the above embodiments.

[0119] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the image processing-based intelligent detection method for surface defects of wind turbine blades.

[0120] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0121] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions 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 invention.

Claims

1. A method for intelligent detection of surface defects in wind turbine blades based on image processing, characterized in that, The intelligent detection method for surface defects of wind turbine blades based on image processing includes: Instantaneous thermal excitation is generated on the blade surface, which synchronously triggers the acquisition of dynamic thermal image sequences. At the same time, a polarization light source array is activated to illuminate the blade surface, acquire a group of polarization images, and establish a dual-modal dataset. After anisotropic diffusion filtering is applied to the dynamic thermal image sequence, a thermal conduction gradient tensor matrix is ​​generated. The principal curvature extrema are extracted to form a thermal anomaly feature map. Phase analysis is performed on the polarization image group to obtain the photoelastic phase distribution. Combined with local energy filtering, a photoelastic stress feature map is generated. Based on the spatial overlap enhancement mechanism between the thermal conduction obstruction region and the stress concentration region, a fusion defect indication map is generated according to the pixel intensity of the thermal anomaly feature map and the gradient modulus of the photoelastic stress feature map. In a log-continuous scale space, phase consistency fluctuation analysis is performed on the fused defect indication map to generate a defect probability distribution map; Based on the phase distribution data parsed from the polarization image group and the spatial coordinates of the defect probability distribution map, a three-dimensional depth map of the defect is reconstructed. Based on the area of ​​the connected region in the defect probability distribution map and the maximum depth change rate of the three-dimensional depth map, a predicted risk value is obtained.

2. The intelligent detection method for surface defects of wind turbine blades based on image processing according to claim 1, characterized in that, include: A laser pulse is emitted and applied to the target area on the blade surface, while a hardware-level trigger signal is simultaneously sent to the infrared thermal imager. The acquisition is initiated based on a trigger signal after the laser pulse is applied, generating a time-coded dynamic thermal image sequence; Simultaneously with the laser pulse emission, the 0°, 45°, 90°, and 135° polarized LED light sources of the multi-angle polarization light source array are activated, and four sets of polarization images are captured through a single exposure. Based on the timestamp of the trigger signal, the dynamic thermal image sequence and the polarization image group are time-axis registered, and pixel-level spatial alignment is completed using a pre-calibrated spatial transformation matrix to output a dual-modal dataset.

3. The intelligent detection method for surface defects of wind turbine blades based on image processing according to claim 2, characterized in that, include: Anisotropic diffusion filtering is performed on the dynamic thermal image sequence to generate a denoised heat conduction image sequence; Calculate the second-order partial derivatives in the spatiotemporal domain of the denoised heat conduction image sequence, and construct the heat conduction gradient tensor matrix; Solve for the eigenvalues ​​of the heat conduction gradient tensor matrix, extract the set of extreme points of maximum principal curvature, and generate a thermal anomaly feature map; Photoelastic phase analysis is performed on the polarization image group, and the arctangent function value is calculated by the four-way polarization intensity to generate the original phase distribution map; The original phase distribution map is subjected to local energy filtering to extract the stress concentration region and output a photoelastic stress feature map.

4. The intelligent detection method for surface defects of wind turbine blades based on image processing according to claim 3, characterized in that, The phase angle of the original phase distribution map is: ； in, Intensity in the 0° polarization direction. Intensity at a polarization direction of 45° Intensity in the 90° polarization direction. Intensity at 135° polarization direction, coefficient This is because the phase angle and stress difference exhibit a birefringence effect.

5. The intelligent detection method for surface defects of wind turbine blades based on image processing according to claim 3, characterized in that, include: Identify the heat conduction obstruction region in the thermal anomaly feature map and extract its pixel intensity distribution as the first physical field input; Calculate the gradient magnitude of the photoelastic stress feature map and mark the stress concentration region as the second physical field input; The pixel intensity distribution and the gradient magnitude are subjected to a Hadamard product operation to generate an initial fused response map. A physical field co-enhancement operation is performed on the initial fusion response map to enhance the response intensity of the spatially overlapping region of heat conduction obstruction and stress concentration, and a fusion defect indication map is output.

6. The intelligent detection method for surface defects of wind turbine blades based on image processing according to claim 5, characterized in that, include: Construct a continuous scale space with a logarithmic distribution, covering a scale range from 0.5 pixels to 8.0 pixels, and generate a multi-scale filter bank; In the continuous scale space, phase consistency fluctuation analysis is performed on the fusion defect indicator map to obtain the local phase response at each scale; The local phase response is scaled and normalized to generate a normalized phase response atlas. Integrate the normalized phase response atlas along the scale dimension to generate a defect probability distribution map.

7. The intelligent detection method for surface defects of wind turbine blades based on image processing according to claim 6, characterized in that, include: Based on the phase distribution data analyzed from the polarization image group, and combined with the spatial coordinates of the defect probability distribution map, the phase gradient magnitude of the defect region is extracted. The phase gradient magnitude is converted into a depth value using a phase-depth mapping model to generate a three-dimensional depth map of the defect. By marking connected regions in the defect probability distribution map, the geometric area is obtained; Extract the maximum depth change rate from the three-dimensional depth map of the defect; The predicted risk value is obtained by multiplying the geometric area by the maximum rate of change of depth.

8. The intelligent detection method for surface defects of wind turbine blades based on image processing according to claim 1, characterized in that, It also includes the generation of maintenance decision parameters: Based on the predicted risk value, a pre-stored maintenance strategy library is matched, and maintenance plan code is output; Based on the spatial distribution characteristics of the defect 3D depth map, a material removal thickness distribution map is generated. By combining the repair scheme code with the material removal thickness distribution map, a three-dimensional repair guidance map is output.

Images