A wind turbine blade damage detection method and system based on image data

Abstract

The application discloses a kind of wind power blade damage detection method and system based on image data, the method includes obtaining wind power blade surface original image data and standard image data are obtained by preprocessing;Extraction blade surface area mask, establish coordinate system and divide grid to form surface grid image;Damage detection is carried out to grid area using multi-type damage identification model, and output damage area and type;Edge extraction and geometric feature parameter calculation are carried out to damage area, and damage degree data are obtained;Damage score is calculated in combination with damage degree data, and grade label is labeled and damage density distribution is calculated accordingly;Historical detection data are acquired and historical damage evolution trend is analyzed, and then future damage evolution trend is integrated;Finally, complete detection result including damage area, type, degree data, grade label, damage density distribution and future evolution trend is output.The application can significantly improve the precision of wind power blade damage detection.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, and in particular to a method and system for detecting damage to wind turbine blades based on image data. Background Technology

[0002] As a core component of wind turbine generators, the integrity of wind turbine blades directly affects power generation efficiency and equipment safety. However, during long-term operation, wind turbine blades inevitably suffer various surface damages, such as cracks, corrosion, and peeling. If these damages are not detected and addressed in a timely manner, they can lead to serious safety accidents and economic losses.

[0003] In one existing technology, sensors are mainly used to detect damage to wind turbine blades. Sensors are installed at regular intervals in easily accessible or critical areas of the blades, such as near the main beam, to form a sparse detection network. The sensor data is then acquired and analyzed to obtain the detection results.

[0004] However, due to the fixed installation location of the sensors, the detection range is difficult to cover the entire blade, making it difficult to accurately locate and detect damage in uncovered areas or between two sensors. Therefore, existing technologies cannot provide sufficient detection accuracy for wind turbine blade damage detection. Summary of the Invention

[0005] This invention provides a method and system for wind turbine blade damage detection based on image data, in order to solve the problem that existing technologies cannot provide sufficient detection accuracy for wind turbine blade damage detection.

[0006] In a first aspect, to address the aforementioned technical problems, the present invention provides a wind turbine blade damage detection method based on image data, comprising:

[0007] The raw image data of the wind turbine blade surface is acquired and preprocessed to obtain standard image data;

[0008] Extract the mask of the blade surface region from the standard image data, establish a blade surface coordinate system and divide it into grids based on the boundary information of the mask to obtain a surface grid image;

[0009] A pre-trained multi-type damage recognition model is used to perform damage detection on each grid region in the surface grid image, and outputs multiple damage regions and corresponding damage types.

[0010] Edge extraction and geometric feature parameter calculation are performed on all damaged areas to obtain multiple damage level data;

[0011] Based on all the damage severity data, calculate the damage score for each damage location, label each damage location with a level label according to the damage score, and calculate the damage density distribution of each grid area of ​​the blade.

[0012] Historical leaf inspection data is obtained, historical damage evolution trends are analyzed, and future damage evolution trends are integrated based on the damage severity data and the historical damage evolution trends.

[0013] The damage area, damage type, damage severity data, grade label, damage density distribution, and future damage evolution trend are taken as the final detection results and output.

[0014] Preferably, the step of acquiring raw image data of the wind turbine blade surface and preprocessing it to obtain standard image data includes:

[0015] Complete raw image data of the blade surface was obtained by shooting from multiple angles;

[0016] The original image data is subjected to illumination compensation and contrast enhancement to obtain clear image data;

[0017] Feature point matching is performed on the clear image data to determine the overlapping areas of adjacent images;

[0018] Image fusion is performed on the overlapping regions to obtain fused image data;

[0019] The fused image data is subjected to geometric correction and color normalization to obtain standard image data.

[0020] Preferably, the step of extracting the mask of the blade surface region from the standard image data, establishing a blade surface coordinate system and dividing it into a grid based on the boundary information of the mask, to obtain a surface grid image includes:

[0021] The standard image data is divided into regions based on the grayscale thresholding method, and the mask of the blade surface region is extracted.

[0022] Edge enhancement processing is performed on the blurred boundary areas in the mask to determine the range of clear mask;

[0023] Based on the defined mask range, boundary information is obtained, and a blade surface coordinate system is constructed based on the boundary information.

[0024] The blade surface is divided into grids according to the blade surface coordinate system to obtain a surface grid image.

[0025] Preferably, the step of performing damage detection on each grid region in the surface grid image using a pre-trained multi-type damage recognition model, and outputting multiple damage regions and corresponding damage types, includes:

[0026] Each grid region in the surface grid image is divided into blocks to obtain multiple local image segments;

[0027] For the local image fragment, a pre-trained multi-type damage recognition model is used to analyze cracks, corrosion and spalling, extract the features of the local image fragment, determine whether a specific damage type exists, and record the corresponding confidence value.

[0028] The model detection results are comprehensively evaluated, and damage types and damage areas with confidence values ​​higher than the preset confidence threshold are labeled.

[0029] Preferably, the training process of the pre-trained multi-type damage recognition model includes:

[0030] A raw image dataset containing three damage types—cracks, corrosion, and spalling—is constructed. The raw image dataset is then subjected to random rotation, brightness perturbation, and noise injection to generate a training sample dataset.

[0031] Based on the YOLOv5 architecture, dedicated detectors for cracks, corrosion, and peeling were built, pre-trained weights were loaded, and the parameters of the preset number of layers in the backbone network were frozen.

[0032] Each dedicated detector is trained by transfer learning, and an adaptive learning rate strategy is adopted. An initial learning rate is set, and the learning rate is reduced when the validation set loss does not decrease after a preset number of consecutive training iterations. The Adam optimizer is used to optimize the loss function.

[0033] The confidence thresholds for each damage type are dynamically adjusted based on the validation set results. Once the detection recall rates for cracks, corrosion, and spalling reach the preset targets, training is stopped and the optimal model parameters are saved.

[0034] The three trained dedicated detectors are integrated, and the detection results are fused through a weighted voting mechanism to obtain a multi-type damage detection model.

[0035] Preferably, the step of performing edge extraction and geometric feature parameter calculation on all damaged areas to obtain multiple damage degree data includes:

[0036] The image of the damaged area is converted to grayscale, and Gaussian filtering is applied to suppress noise interference to obtain preliminary edge contour lines;

[0037] A morphological closing operation is performed on the initial edge contour lines to fill in the broken or blurred areas of the contour and generate a continuous closed damage boundary.

[0038] The area of ​​the damaged region is calculated and the depth of the damaged region is estimated through three-dimensional reconstruction. The length, width and orientation angle of the crack-like damage are measured.

[0039] Based on the mapping relationship of the blade surface coordinate system, pixel-level geometric parameters are converted into actual physical dimensions, generating multiple damage level data.

[0040] Preferably, the step of calculating a damage score for each damage location by combining all the damage severity data, labeling each damage location with a level label based on the damage score, and calculating the damage density distribution of each grid region of the blade includes:

[0041] Calculate the damage score for each location based on the damage severity data;

[0042] Different severity levels are assigned to different injury types, and each injury is labeled with a severity level based on the injury score.

[0043] The total score of all damages within each grid area is calculated, and the damage density distribution is then determined.

[0044] Preferably, the damage severity data includes damage area, crack length, crack orientation angle, and damage depth; the damage score is divided into a damage score for crack damage and damage scores for corrosion damage and spalling damage.

[0045] The formula for calculating the damage score of the crack damage is:

[0046]

[0047] The damage scoring formulas for corrosion damage and exfoliation damage are as follows:

[0048]

[0049] Wherein, Score1 is the damage score for crack damage; Score2 is the damage score for corrosion damage and spalling damage; w type A is the damage weighting coefficient; A is the damage area; A base L is the reference area; L is the crack length; L base θ is the reference length; θ is the crack orientation angle; D is the depth of the damaged area; D base The reference depth; w location α is the regional sensitivity factor corresponding to the damage location; β is the weighting coefficient of the damage area; γ is the weighting coefficient of the crack length; δ is the weighting coefficient of the crack orientation angle; and δ is the weighting coefficient of the damage area depth.

[0050] Preferably, the step of acquiring historical leaf detection data, analyzing historical damage evolution trends, integrating the damage severity data and historical damage evolution trends to obtain future damage evolution trends, and outputting the final detection results includes:

[0051] After acquiring historical detection data of the blades and removing noise data, the spatiotemporal distribution data of historical damage is obtained.

[0052] Based on the spatiotemporal distribution data, the expansion rate and direction of historical damage are analyzed to obtain the evolution trend of historical damage;

[0053] Based on the historical damage evolution trend and the damage severity data, the future expansion rate and direction of the current damage are analyzed, and the future damage evolution trend is obtained by integration.

[0054] Secondly, the present invention provides a wind turbine blade damage detection system based on image data, comprising:

[0055] The image data acquisition module acquires raw image data of the wind turbine blade surface and performs preprocessing to obtain standard image data;

[0056] The image segmentation and coordinate system establishment module extracts the mask of the blade surface region from the standard image data, establishes the blade surface coordinate system based on the boundary information of the mask, and divides it into a grid to obtain a surface grid image.

[0057] The damage partitioning detection module uses a pre-trained multi-type damage recognition model to perform damage detection on each grid region in the surface grid image, and outputs multiple damage regions and corresponding damage types.

[0058] The damage quantification module performs edge extraction and geometric feature parameter calculation on all damaged areas to obtain multiple damage quantification data.

[0059] The scoring density calculation module calculates the damage score for each damage location by combining all the damage degree data, labels each damage location with a level label based on the damage score, and calculates the damage density distribution of each grid area of ​​the blade.

[0060] The damage trend prediction module acquires historical detection data of the blade, analyzes the historical damage evolution trend, and integrates the damage degree data and the historical damage evolution trend to obtain the future damage evolution trend.

[0061] The result integration and output module outputs the damage area, damage type, damage severity data, grade label, damage density distribution, and future damage evolution trend as the final detection result.

[0062] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the wind turbine blade damage detection method based on image data described in any one of the above.

[0063] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the wind turbine blade damage detection method based on image data described in any one of the above.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. This invention acquires raw image data of the wind turbine blade surface from multiple angles and performs preprocessing, effectively reducing the impact of environmental noise and image distortion. Through regional illumination compensation and adaptive histogram equalization, it eliminates uneven illumination and enhances edge details, allowing microcracks and other minor damage to be clearly revealed. Furthermore, by extracting the blade surface mask, establishing a coordinate system, and performing meshing, it achieves full surface coverage detection, effectively avoiding blind spots present in traditional sensor detection. In the damage identification stage, it employs a multi-type damage identification model integrated with the YOLOv5 architecture, fusing dedicated detection results for cracks, corrosion, and spalling through transfer learning and weighted voting mechanisms, significantly improving recall. Finally, by combining 3D reconstruction and physical size mapping, it accurately quantifies damage geometric parameters, eliminating pixel-level errors and achieving high-precision detection of wind turbine blade surface damage, effectively solving the problem of insufficient detection accuracy in existing technologies.

[0066] 2. This invention reduces the need for manual parameter tuning and improves image preprocessing efficiency by using adaptive histogram equalization and automated image fusion technology. In the model detection stage, the use of a lightweight YOLOv5 model and pre-trained weight transfer reduces computational resource consumption and accelerates detection speed. Furthermore, this invention can automatically generate reports containing grade labels, density distribution, and trend predictions through structured result output, replacing the traditional manual data analysis process. For example, the system can directly output specific damage conclusions and predictions, such as "an L3-level crack exists in the A3 region of the mesh, and it is predicted to extend to the critical depth within 3 months," thereby significantly improving detection efficiency and decision response speed.

[0067] 3. This invention trains specialized detectors for cracks, corrosion, and spalling by type and integrates a weighted voting mechanism to ensure that multiple types of damage can be identified simultaneously. It employs morphological closing operations and 3D reconstruction technology to accurately extract damage boundaries and calculate parameters such as area, aspect ratio, and depth. Combined with standardized scoring formulas, it scientifically quantifies the impact of damage, achieving a comprehensive assessment of damage type, geometric features, and safety impact. It overcomes the limitations of traditional methods that can only detect a single type or provide a rough location, providing a more comprehensive and accurate basis for wind turbine blade maintenance decisions.

[0068] 4. Through image enhancement, edge extraction, and feature parameter calculation, this invention can effectively address the influence of different lighting conditions and blade surface conditions, enabling accurate detection and assessment of various types of damage. Furthermore, this invention can adjust detection parameters and models according to actual needs, exhibiting strong flexibility and adaptability. It can reliably and stably detect wind turbine blade damage under various conditions, and has broad application prospects. Attached Figure Description

[0069] Figure 1 This is a schematic flowchart of a wind turbine blade damage detection method based on image data provided in an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of a wind turbine blade damage detection system based on image data provided in an embodiment of the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] Reference Figure 1 The first embodiment of the present invention provides a schematic flowchart of a wind turbine blade damage detection method based on image data, including the following steps:

[0073] S11: Acquire raw image data of the wind turbine blade surface and preprocess it to obtain standard image data;

[0074] S12, extract the mask of the blade surface region from the standard image data, establish the blade surface coordinate system and divide it into grids according to the boundary information of the mask, and obtain the surface grid image;

[0075] S13, using a pre-trained multi-type damage recognition model to perform damage detection on each grid region in the surface grid image, and output multiple damage regions and corresponding damage types;

[0076] S14, perform edge extraction and geometric feature parameter calculation on all damaged areas to obtain multiple damage degree data;

[0077] S15, calculate the damage score for each damage by combining all the damage degree data, label the damage level of each damage according to the damage score and calculate the damage density distribution of each grid area of ​​the blade.

[0078] S16, acquire historical detection data of the blade, analyze the historical damage evolution trend, and integrate the damage degree data and the historical damage evolution trend to obtain the future damage evolution trend.

[0079] S17, the damaged area, the damaged type, the damaged degree data, the grade label, the damaged density distribution, and the future damage evolution trend are taken as the final detection results and output.

[0080] In step S11, acquiring raw image data of the wind turbine blade surface and preprocessing it to obtain standard image data includes:

[0081] Complete raw image data of the blade surface was obtained by shooting from multiple angles;

[0082] The original image data is subjected to illumination compensation and contrast enhancement to obtain clear image data;

[0083] Feature point matching is performed on the clear image data to determine the overlapping areas of adjacent images;

[0084] Image fusion is performed on the overlapping regions to obtain fused image data;

[0085] The fused image data is subjected to geometric correction and color normalization to obtain standard image data.

[0086] Specifically, the method of acquiring complete raw image data of the blade surface through multi-angle shooting typically employs a high-definition camera array arranged in a ring to simultaneously capture images of the rotating wind turbine blade.

[0087] For example, three 4K resolution industrial cameras can be arranged at 120° intervals around the wind turbine tower. The camera is triggered when the blade rotates to a preset angle (e.g., every 30° rotation), thus ensuring that image data of the entire blade surface area is acquired.

[0088] In one possible implementation, abnormal lighting areas are identified based on the brightness histogram of the original image data. If the brightness distribution uniformity is lower than a preset uniformity threshold, regional lighting compensation is performed on the original image data to obtain a uniformly illuminated leaf surface image.

[0089] For example, when sunlight shines from one side, the brightness value of the light-facing side of the leaf may be concentrated in the range of 200-255, while that of the shaded side may be concentrated in the range of 0-50. At this time, the standard deviation of the brightness histogram is calculated as a uniformity index. If it is lower than the preset uniformity threshold of 60, regional illumination compensation is required. Preferably, the image can be divided into grids, and adaptive gamma correction can be applied to each grid individually to make the brightness distribution more uniform.

[0090] It should be noted that the image after uniform illumination may have insufficient contrast. Therefore, adaptive histogram equalization is applied to the uniformly illuminated blade surface image to improve contrast. If the contrast is lower than the set threshold, the contrast enhancement parameter is increased.

[0091] For example, fine cracks on the blade surface appear as areas with similar gray values ​​in an image. In this case, applying Adaptive Histogram Equalization (CLAHE) can effectively improve contrast. If the calculated contrast value is lower than 0.4, the contrast enhancement parameter can be increased, such as increasing cliplimit from 2.0 to 3.5. For areas with blurred edges, sharpening filters, such as the Laplacian operator, can be applied to enhance edge details, making the tiny defects on the blade surface more clearly visible.

[0092] In one embodiment, stitching adjacent images is a crucial step in obtaining a complete image of the wind turbine blade surface. Overlapping regions can be identified using the SIFT feature point matching algorithm, and a progressive fusion method using multi-band fusion can be applied to these regions to eliminate stitching artifacts. The geometric correction can employ perspective transformation, while color normalization can be achieved using methods such as white balance correction, ultimately yielding standard image data of the wind turbine blade surface.

[0093] In step S12, the step of extracting the mask of the blade surface region from the standard image data, establishing a blade surface coordinate system based on the boundary information of the mask, and dividing it into a grid to obtain a surface grid image includes:

[0094] The standard image data is divided into regions based on the grayscale thresholding method, and the mask of the blade surface region is extracted.

[0095] Edge enhancement processing is performed on the blurred boundary areas in the mask to determine the range of clear mask;

[0096] Based on the defined mask range, boundary information is obtained, and a blade surface coordinate system is constructed based on the boundary information.

[0097] The blade surface is divided into grids according to the blade surface coordinate system to obtain a surface grid image.

[0098] In one possible implementation, the leaf region in an image is distinguished from the background region based on a grayscale thresholding method because the leaf region usually exhibits a relatively consistent grayscale range, while the background may include regions with different grayscale characteristics, such as the sky or towers. Specifically, a grayscale threshold of 100 can be set, and regions with grayscale values ​​higher than this threshold are initially marked as leaf regions. Using image processing software or algorithms, the standard image data is traversed pixel by pixel, and the image is divided into leaf regions and background regions according to the set grayscale threshold, extracting the initial mask of the leaf region. The position and outline of the leaf in the image can be roughly outlined based on the initial mask, but due to factors such as image quality and lighting conditions, the boundary may be blurred or unclear.

[0099] Preferably, a gradient-based edge detection method can be used to identify gray-scale abrupt change regions on the edge of the leaf. Then, a gradient-sharpening edge enhancement algorithm is used to process these boundary pixels, highlighting the edge details of the leaf and making the boundary of the leaf region clearer and more defined, thereby obtaining a clear mask range.

[0100] In one possible implementation, the boundary information processing for the clear mask range can be done using the findContours function in the OpenCV library, which can find the contour boundary of the mask region and output a series of boundary points in the form of pixel coordinates.

[0101] Preferably, the boundary points on the blade surface can be used as a reference to construct a two-dimensional coordinate system with the blade root as the origin. For example, the blade root can be set as the origin, the x-axis can be along the blade length direction, and the y-axis can be perpendicular to the length direction.

[0102] Specifically, the process of dividing the blade surface into grids can divide the blade surface into multiple uniform units; for example, the blade surface can be divided into 10x10 grid units, each unit representing a small area of ​​the blade surface, or the grid units can be divided according to the unit length of the coordinate system.

[0103] In one embodiment, assuming the blade is 50 meters long and 5 meters wide, the size of each grid cell is approximately 5 meters x 0.5 meters. A surface grid image is then obtained.

[0104] In step S13, the pre-trained multi-type damage recognition model is used to perform damage detection on each grid region in the surface grid image, outputting multiple damage regions and corresponding damage types, including:

[0105] Each grid region in the surface grid image is divided into blocks to obtain multiple local image segments;

[0106] For the local image fragment, a pre-trained multi-type damage recognition model is used to analyze cracks, corrosion and spalling, extract the features of the local image fragment, determine whether a specific damage type exists, and record the corresponding confidence value.

[0107] The model detection results are comprehensively evaluated, and damage types and damage areas with confidence values ​​higher than the preset confidence threshold are labeled.

[0108] In this process, each grid region in the surface grid image is divided into blocks to obtain multiple local image fragments. Each grid can be further divided into 10x10 or more refined local image fragments, and each fragment is stored and numbered independently. For example, each grid unit with a size of approximately 5 meters x 0.5 meters can be further divided into 10x10 image fragments. After block division, tens of thousands of local image fragments can be obtained, and each fragment can be analyzed for damage independently.

[0109] In one embodiment, assuming the model learns the linear features of cracks, the spot features of corrosion, and the block features of spalling through a large amount of labeled data, in actual detection, it will output a confidence value for the damage type for each segment; for example, a segment with a crack confidence value of 0.85, corrosion of 0.3, and spalling of 0.1 indicates that the crack is the most likely.

[0110] Preferably, the feature extraction of the local image segment can utilize an edge detection algorithm to extract the outline of the crack, or identify the color difference of the corroded area through color analysis; for example, in a local segment, edge detection found a linear anomaly about 0.5 meters long, which was determined to be a crack based on the confidence value.

[0111] It should be noted that the comprehensive evaluation of the model detection results can be carried out using a weighted average or a voting mechanism. The weighted average method is to weight the confidence values ​​of different damage types according to the model's performance indicators (such as accuracy and recall) and then calculate the weighted average confidence value. The voting mechanism allows multiple models or multiple versions of a model to detect the same image segment, with each model voting on the damage type, and the damage type is finally determined based on the voting results.

[0112] If the preset confidence threshold is 0.8, when the confidence value of a certain damage type exceeds this threshold, the system will determine that the segment has corresponding damage and mark the specific location. For example, if the confidence value of a crack in a segment is 0.85, exceeding the threshold, the system will mark the crack location on the image, making it easier for subsequent maintenance personnel to accurately locate it. Furthermore, it can combine the overall coordinate system of the blade to map the damage location of a local segment onto the overall blade structure. For example, if the crack is located 5 meters to the left of the center of the blade, this precise marking helps to develop targeted maintenance plans and improve maintenance efficiency.

[0113] In one implementation, the training process of the pre-trained multi-type damage recognition model includes:

[0114] A raw image dataset containing three damage types—cracks, corrosion, and spalling—is constructed. The raw image dataset is then subjected to random rotation, brightness perturbation, and noise injection to generate a training sample dataset.

[0115] Based on the YOLOv5 architecture, dedicated detectors for cracks, corrosion, and peeling were built, pre-trained weights were loaded, and the parameters of the preset number of layers in the backbone network were frozen.

[0116] Each dedicated detector is trained by transfer learning, and an adaptive learning rate strategy is adopted. An initial learning rate is set, and the learning rate is reduced when the validation set loss does not decrease for several consecutive rounds of training. The Adam optimizer is used to optimize the loss function.

[0117] The confidence thresholds for each damage type are dynamically adjusted based on the validation set results. Once the detection recall rates for cracks, corrosion, and spalling reach the preset targets, training is stopped and the optimal model parameters are saved.

[0118] The three trained dedicated detectors are integrated, and the detection results are fused through a weighted voting mechanism to obtain a multi-type damage detection model.

[0119] Preferably, the construction of the original image dataset containing three damage types—cracks, corrosion, and spalling—can be achieved by collecting images of wind turbine blades taken from different angles and under different lighting conditions at different wind farms, ensuring the diversity of the dataset. The collected images are labeled to clarify the type of damage. OpenCV library functions are used to randomly rotate, brightness-perturb, and inject noise into the dataset. The data-augmented images and the original images are combined to form a training sample dataset, ensuring a balanced number of images in the dataset, i.e., the number of images for the three damage types—cracks, corrosion, and spalling—is approximately equal. For example, after data augmentation, the dataset contains 2000 crack images, 2000 corrosion images, and 2000 spalling images.

[0120] It should be noted that the loading of pre-trained weights can be loading pre-trained weights from the COCO dataset. The COCO dataset is a large dataset used for tasks such as object detection, segmentation, and caption generation, containing hundreds of thousands of images and a large amount of annotation information. The YOLOv5 model is pre-trained on the COCO dataset, which can learn rich feature representations and general object detection capabilities. The pre-trained weights provide a good initialization for the model, enabling the model to converge faster and perform better in detection accuracy when training new detection tasks.

[0121] The parameter for the preset number of layers of the frozen backbone network can be set to the parameters of the first 20 layers to prevent the general features learned in the pre-training weights from being destroyed in the early stage of training due to unreasonable learning rate settings or large differences between the new dataset and the pre-training dataset.

[0122] In one embodiment, the initial learning rate ranges from 0.001 to 0.01. For example, the initial learning rate can be set to 0.001. A relatively small value ensures that the model will not update parameters too rapidly in the early stages of training due to an excessively large learning rate, thereby maintaining training stability. The preset number of rounds is usually set to 3 to 5 rounds. For example, it can be set to 3 rounds. This setting allows the model to fluctuate normally on the validation set for one or two rounds without adjusting the learning rate due to occasional increases in loss. The common practice for reducing the learning rate is to reduce it to 1 / 3 of the original value. For example, if the initial learning rate is 0.001, it can be reduced to 0.0003. This allows the model to learn patterns and features in the data more finely in the later stages of training, while avoiding oscillations near local optima.

[0123] Preferably, training stops once the detection recall reaches a preset threshold. Detection recall is the percentage of damage correctly detected by the model, reflecting the model's accuracy in identifying and detecting damage. Specifically, the crack detection recall can be preset to 90%. Cracks typically have a significant impact on the structural integrity of wind turbine blades, potentially leading to serious consequences such as blade breakage. A high recall ensures that most cracks are detected, allowing for timely maintenance and preventing potential safety accidents. The corrosion detection recall can be preset to 85%. Corrosion gradually weakens the strength and durability of blades, affecting their performance. An 85% recall aims to effectively identify corroded areas for timely treatment and to prevent further corrosion. The spalling detection recall can be preset to 80%. Spalling affects the aerodynamic performance and efficiency of blades. An 80% recall requires the model to identify most spalling areas, ensuring the normal operation and efficiency of the blades.

[0124] The weighted voting mechanism for fusing detection results involves assigning weights to each detector, calculating a weighted score for its detection results, fusing the results to determine the damage type and confidence level, and generating the final detection result.

[0125] In step S14, edge extraction and geometric feature parameter calculation are performed on all damaged areas to obtain multiple damage level data, including:

[0126] The image of the damaged area is converted to grayscale, and Gaussian filtering is applied to suppress noise interference to obtain preliminary edge contour lines;

[0127] A morphological closing operation is performed on the initial edge contour lines to fill in the broken or blurred areas of the contour and generate a continuous closed damage boundary.

[0128] The area of ​​the damaged region is calculated and the depth of the damaged region is estimated through three-dimensional reconstruction. The length, width and orientation angle of the crack-like damage are measured.

[0129] Based on the mapping relationship of the blade surface coordinate system, pixel-level geometric parameters are converted into actual physical dimensions, generating multiple damage level data.

[0130] In one embodiment, an RGB image containing the damaged area, such as a 200×200 pixel crack area, is cropped from the initial detection result. The RGB three channels are converted into a single-channel grayscale image using a weighted average method. A Gaussian blur is applied to the grayscale image using a Gaussian kernel size of 5×5 and a standard deviation of 1.5 to suppress random noise, such as camera noise or texture interference. For example, if a damaged area has salt noise (grayscale value 255), after Gaussian filtering, the noise is smoothed by the surrounding pixels (mean 120), and the grayscale value is reduced to below 140, resulting in edge contour lines.

[0131] In one possible implementation, for the processing of the initial edge contour lines, morphological operations repair broken or blurred boundaries through dilation and erosion; for example, when the boundary is discontinuous due to shadows in the crack image, the dilation operation can connect the broken parts, while the erosion operation removes redundant noise points; assuming that in a local image of a blade, a crack contour line with 3 small gaps is detected, these gaps are filled after morphological operations to form a continuous closed damage boundary.

[0132] It should be noted that the calculation of the damaged area area can be performed by reading the image data of the damaged boundary using OpenCV; calling the area calculation function to perform pixel-level statistics on the damaged area and calculate the area of ​​the damaged area in units of pixels; and converting the pixel area into the actual physical area in units of square meters according to the image resolution or scale factor.

[0133] Specifically, the method of estimating the depth of the damaged area through three-dimensional reconstruction can be achieved by acquiring three-dimensional point cloud data of the blade surface using structured light projection or binocular stereo vision. Based on images taken from multiple angles, a three-dimensional point cloud of the damaged area is generated through feature matching and triangulation. Using the intact surface as a reference plane, the maximum height difference between the point cloud of the damaged area and the reference plane is calculated to obtain the estimated depth of the damaged area.

[0134] It should be noted that, for the damaged area of ​​a crack type, the length, width, and orientation angle of the crack can be measured through linear fitting. For example, in the detection of a crack on the blade surface, a fitted straight line can determine its length as 15 cm, its width as 0.5 cm, and its orientation angle as 30 degrees, indicating that the crack extends along the direction of stress on the blade. This damage level data helps to determine the extent of the crack's impact on the blade structure.

[0135] It is worth noting that the geometric feature parameters in pixels should be converted into multiple damage level data in actual physical size for subsequent analysis.

[0136] In step S15, the damage score for each damage location is calculated by combining all the damage severity data, and a damage level label is assigned to each damage location based on the damage score, and the damage density distribution of each grid region of the blade is calculated, including:

[0137] Calculate the damage score for each location based on the damage severity data;

[0138] Different severity levels are assigned to different injury types, and each injury is labeled with a severity level based on the injury score.

[0139] The total score of all damages within each grid area is calculated, and the damage density distribution is then determined.

[0140] The damage severity data includes damage area, crack length, crack orientation angle, and damage depth. Based on all the damage severity data, combined with damage weighting coefficients set according to damage type and damage area sensitivity factors, a damage score is calculated for each location of damage. The damage weighting coefficients can be set according to the impact of damage type on the wind turbine blade; for example, 1.0 for cracks, 0.8 for corrosion, and 0.6 for spalling. The area sensitivity factor can be set according to the impact of damage location on the wind turbine blade; for example, 2.5 for the blade root, 1.0 for the blade middle, and 0.8 for the blade tip.

[0141] The severity level thresholds for different damage types can be divided according to damage scores. For example: for crack damage, 0-100 is low risk, 101-300 is medium risk, and ≥300 is high risk; for corrosion damage, 0-50 is low risk, 51-150 is medium risk, and ≥150 is high risk; for spalling damage, 0-80 is low risk, 81-200 is medium risk, and ≥200 is high risk.

[0142] For example, the wind turbine blade has corrosion at the blade root, with an area of ​​0.5 square meters and a depth of 15 millimeters, and an area sensitivity of 2.5, resulting in a damage score of 150; flaking at the blade tip, with an area of ​​0.08 square meters and a depth of 5 millimeters, and an area sensitivity of 0.8, resulting in a damage score of 10.75; and a crack in the middle of the blade, with a length of 3 meters and an angle of 60°, and an area sensitivity of 1.0, resulting in a damage score of 167.32. Based on the severity level threshold, the corrosion at the blade root is marked as high risk, the flaking at the blade tip is marked as low risk, and the crack in the middle of the blade is marked as medium risk.

[0143] Preferably, after calculating the damage score for all damages, the scores for all damages within each grid region are obtained and summed; the area of ​​each grid region is calculated based on the grid size; and the damage density is obtained by dividing the sum of the damage scores for each grid region by the area of ​​that grid.

[0144] For example, the damage density calculation results of all grid areas can be visualized to generate a damage density distribution map, which intuitively shows the damage density distribution on the blade surface. Through the damage density distribution, areas with more concentrated damage (high-density areas) on the blade can be quickly identified. Moreover, the damage density distribution not only reflects the number of damages but also the severity of the damages, which helps to more accurately assess the overall health of the blade. Maintenance personnel can develop more targeted maintenance plans based on the damage density distribution map.

[0145] It is worth noting that the damage score calculation formula for crack damage is as follows:

[0146]

[0147] The damage score calculation formulas for corrosion damage and exfoliation damage are as follows:

[0148]

[0149] Wherein, Score1 is the damage score for crack damage; Score2 is the damage score for corrosion damage and spalling damage; w type A represents the damage weighting coefficient; A represents the damage area in square meters; A base The base area is 1 square meter; L is the crack length in meters; L base The reference length is 1 meter; θ is the crack orientation angle; D is the depth of the damaged area in millimeters; D base The reference depth is 10 mm; w location α is the area sensitivity factor corresponding to the damage location; β is the damage area weighting coefficient of 30; γ is the crack length weighting coefficient of 50; γ is the weighting coefficient of the crack orientation angle sinusoidal output of 20; δ is the damage area depth weighting coefficient of 40.

[0150] The above embodiments calculate the corresponding damage scores according to the damage characteristics of different damages, eliminate the influence of different dimensions through normalization, and divide the corresponding severity level thresholds according to the damage scores of different damages, so that the damage scores of different types of damage can more accurately reflect the severity of the impact of damage on the structural integrity of the blade.

[0151] In step S16, the acquisition of historical leaf detection data, analysis of historical damage evolution trends, and integration of the damage severity data and historical damage evolution trends to obtain future damage evolution trends include:

[0152] After acquiring historical detection data of the blades and removing noise data, the spatiotemporal distribution data of historical damage is obtained.

[0153] Based on the spatiotemporal distribution data, the expansion rate and direction of historical damage are analyzed to obtain the evolution trend of historical damage;

[0154] Based on the historical damage evolution trend and the damage severity data, the future expansion rate and direction of the current damage are analyzed, and the future damage evolution trend is obtained by integration.

[0155] The process of acquiring historical blade detection data involves obtaining detection data of wind turbine blades at multiple different time points from a database or historical record file, including information such as damage location, type, and degree of damage. The collected historical data is then organized and cleaned to ensure consistent data format and remove obvious errors or abnormal data points, resulting in spatiotemporal distribution data of historical damage.

[0156] Preferably, time-series comparison tools such as Python's Pandas library and MATLAB's time-series analysis toolbox are used to compare damage data at different time points. For each damage point or region, the changes in damage severity data at different time points, such as area, depth, and length, are calculated. The damage propagation rate is obtained by calculating the parameter change rate between adjacent time points. The direction of damage propagation on the blade surface is analyzed, which can be determined by comparing the geometric changes of the damage boundary at different time points. The propagation rate and direction of each damage point or region are summarized to plot a damage evolution curve or graph over time. Statistical methods can be used to comprehensively analyze the data from multiple damage points to obtain the overall historical damage evolution trend.

[0157] It should be noted that, based on historical damage evolution trends and combined with current damage severity data, suitable time series forecasting tools or algorithms should be selected, such as linear regression, multinomial regression, exponential smoothing, ARIMA models, or LSTM neural networks, to predict the future expansion rate and direction of the current damage. After summarizing, damage evolution curves or charts over time should be plotted and comprehensive analysis should be performed to obtain the overall future damage evolution trend.

[0158] For example, suppose historical data on crack damage to a wind turbine blade shows that the damaged area at the past four time points was 0.001 square meters, 0.0015 square meters, 0.0022 square meters, and 0.003 square meters, respectively, with an accelerating rate of area expansion. Using a quadratic polynomial regression model to extrapolate the trend and assuming that the damage area changes quadratically with time, the extrapolation predicts that the damaged area will increase to 0.00362 square meters in the next 3 months and to 0.00545 square meters in the next 6 months. This indicates a rapid increase in the future damage evolution trend.

[0159] In step S17, the damaged area, the damaged type, the damaged degree data, the grade label, the damaged density distribution, and the future damage evolution trend are taken as the final detection results and output.

[0160] It should be noted that the final test results can be formatted into an easy-to-understand and display format, such as using text documents, tables, image annotations, etc., to ensure the clarity and readability of the data; or professional visualization tools, such as Python's Matplotlib, Seaborn, or MATLAB's plotting function, can be used to plot information such as damage area, type, degree data, grade labels, density distribution, and evolution trend on images or charts, making the results more intuitive and easy to understand.

[0161] In summary, this invention discloses a wind turbine blade damage detection method based on image data. It acquires raw image data of the wind turbine blade surface from multiple angles and preprocesses it to reduce the impact of environmental noise and image distortion. Illumination compensation and edge enhancement ensure the visibility of minor damage such as microcracks. The method extracts the blade surface mask, establishes a blade surface coordinate system, and divides it into a grid to achieve full-coverage detection of the blade surface, avoiding blind spots caused by the sparse layout of traditional sensors. It employs a multi-type damage recognition model integrated with the YOLOv5 architecture, fusing dedicated detection results for cracks, corrosion, and spalling through transfer learning and weighted voting mechanisms, significantly improving recall. Combining 3D reconstruction and physical size mapping, it accurately quantifies damage severity data, eliminating pixel-level errors. This improves the detection accuracy of wind turbine blade surface damage.

[0162] Reference Figure 2The second embodiment of the present invention provides a wind turbine blade damage detection system based on image data, comprising:

[0163] The image data acquisition module acquires raw image data of the wind turbine blade surface and performs preprocessing to obtain standard image data;

[0164] The image segmentation and coordinate system establishment module extracts the mask of the blade surface region from the standard image data, establishes the blade surface coordinate system based on the boundary information of the mask, and divides it into a grid to obtain a surface grid image.

[0165] The damage partitioning detection module uses a pre-trained multi-type damage recognition model to perform damage detection on each grid region in the surface grid image, and outputs multiple damage regions and corresponding damage types.

[0166] The damage quantification module performs edge extraction and geometric feature parameter calculation on all damaged areas to obtain multiple damage quantification data.

[0167] The scoring density calculation module calculates the damage score for each damage location by combining all the damage degree data, labels each damage location with a level label based on the damage score, and calculates the damage density distribution of each grid area of ​​the blade.

[0168] The damage trend prediction module acquires historical detection data of the blade, analyzes the historical damage evolution trend, and integrates the damage degree data and the historical damage evolution trend to obtain the future damage evolution trend.

[0169] The result integration and output module outputs the damage area, damage type, damage severity data, grade label, damage density distribution, and future damage evolution trend as the final detection result.

[0170] It should be noted that the wind turbine blade damage detection system based on image data provided in this embodiment of the invention is used to execute all the process steps of the wind turbine blade damage detection method based on image data in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0171] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an edge extraction program. When the processor executes the computer program, it implements the steps in the various embodiments of the image data-based wind turbine blade damage detection method described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the score density calculation module.

[0172] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0173] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0174] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0175] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0176] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0177] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0178] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for detecting wind turbine blade damage based on image data, characterized in that, include: The raw image data of the wind turbine blade surface is acquired and preprocessed to obtain standard image data; Extract the mask of the blade surface region from the standard image data, establish a blade surface coordinate system and divide it into grids based on the boundary information of the mask to obtain a surface grid image; A pre-trained multi-type damage recognition model is used to perform damage detection on each grid region in the surface grid image, outputting multiple damage regions and corresponding damage types. The training process of the pre-trained multi-type damage recognition model includes: constructing an original image dataset containing three damage types: cracks, corrosion, and spalling; randomly rotating, brightness perturbing, and injecting noise into the original image dataset to generate a training sample dataset; constructing dedicated detectors for cracks, corrosion, and spalling based on the YOLOv5 architecture, loading pre-trained weights, and freezing the parameters of a preset number of layers in the backbone network; performing transfer learning training on each dedicated detector using an adaptive learning rate strategy, setting an initial learning rate, and reducing the learning rate when the validation set loss does not decrease after a preset number of iterations; optimizing the loss function using the Adam optimizer; dynamically adjusting the confidence thresholds for each damage type based on the validation set results, stopping training when the detection recall rates for cracks, corrosion, and spalling reach preset indicators, and saving the optimal model parameters; integrating the three trained dedicated detectors and fusing the detection results through a weighted voting mechanism to obtain the multi-type damage detection model. Edge extraction and geometric feature parameter calculation are performed on all damaged areas to obtain multiple damage level data; Based on all the damage severity data, calculate the damage score for each damage location, label each damage location with a level label according to the damage score, and calculate the damage density distribution of each grid area of ​​the blade. Historical blade inspection data is acquired, and the historical damage evolution trend is analyzed. Based on the damage severity data and the historical damage evolution trend, the future damage evolution trend is obtained. Specifically, historical blade inspection data is acquired, and after removing noise data, the spatiotemporal distribution data of historical damage is obtained. Based on the spatiotemporal distribution data, the expansion rate and direction of historical damage are analyzed to obtain the historical damage evolution trend. Based on the historical damage evolution trend and the damage severity data, the future expansion rate and direction of current damage are analyzed, and the future damage evolution trend is obtained. The damage area, damage type, damage severity data, grade label, damage density distribution, and future damage evolution trend are taken as the final detection results and output.

2. The wind turbine blade damage detection method based on image data according to claim 1, characterized in that, The process of acquiring raw image data of the wind turbine blade surface and preprocessing it to obtain standard image data includes: Complete raw image data of the blade surface was obtained by shooting from multiple angles; The original image data is subjected to illumination compensation and contrast enhancement to obtain clear image data; Feature point matching is performed on the clear image data to determine the overlapping areas of adjacent images; Image fusion is performed on the overlapping regions to obtain fused image data; The fused image data is subjected to geometric correction and color normalization to obtain standard image data.

3. The wind turbine blade damage detection method based on image data according to claim 1, characterized in that, The process of extracting the mask of the blade surface region from the standard image data, establishing a blade surface coordinate system based on the boundary information of the mask, and dividing the surface into a grid to obtain a surface grid image includes: The standard image data is divided into regions based on the grayscale thresholding method, and the mask of the blade surface region is extracted. Edge enhancement processing is performed on the blurred boundary areas in the mask to determine the range of clear mask; Based on the defined mask range, boundary information is obtained, and a blade surface coordinate system is constructed based on the boundary information. The blade surface is divided into grids according to the blade surface coordinate system to obtain a surface grid image.

4. The wind turbine blade damage detection method based on image data according to claim 1, characterized in that, The method utilizes a pre-trained multi-type damage recognition model to perform damage detection on each grid region in the surface grid image, outputting multiple damage regions and corresponding damage types, including: Each grid region in the surface grid image is divided into blocks to obtain multiple local image segments; For the local image fragment, a pre-trained multi-type damage recognition model is used to analyze cracks, corrosion and spalling, extract the features of the local image fragment, determine whether a specific damage type exists, and record the corresponding confidence value. The model detection results are comprehensively evaluated, and damage types and damage areas with confidence values ​​higher than the preset confidence threshold are labeled.

5. The wind turbine blade damage detection method based on image data according to claim 1, characterized in that, The process of edge extraction and geometric feature parameter calculation for all damaged areas yields multiple damage level data, including: The image of the damaged area is converted to grayscale, and Gaussian filtering is applied to suppress noise interference to obtain preliminary edge contour lines; A morphological closing operation is performed on the initial edge contour lines to fill in the broken or blurred areas of the contour and generate a continuous closed damage boundary. The area of ​​the damaged region is calculated and the depth of the damaged region is estimated through three-dimensional reconstruction. The length, width and orientation angle of the crack-like damage are measured. Based on the mapping relationship of the blade surface coordinate system, pixel-level geometric parameters are converted into actual physical dimensions, generating multiple damage level data.

6. The wind turbine blade damage detection method based on image data according to claim 1, characterized in that, The process involves calculating a damage score for each location by combining all the damage severity data, assigning a level label to each damage location based on the damage score, and calculating the damage density distribution in each grid region of the blade, including: Calculate the damage score for each location based on the damage severity data; Different severity levels are assigned to different injury types, and each injury is labeled with a severity level based on the injury score. The total score of all damages within each grid area is calculated, and the damage density distribution is then determined.

7. The wind turbine blade damage detection method based on image data according to claim 6, characterized in that, The damage severity data includes damage area, crack length, crack orientation angle, and damage depth; the damage score is divided into damage score for crack damage and damage score for corrosion damage and spalling damage. The formula for calculating the damage score of the crack damage is: ； The damage scoring formulas for corrosion damage and exfoliation damage are as follows: ； in, Damage score for crack damage; Damage scores for corrosion damage and exfoliation damage; This refers to the damage weighting coefficient. The area of ​​damage; Used as the reference area; The length of the crack; As the reference length; The angle of the crack direction; The depth of the damaged area; Reference depth; This refers to the area sensitivity factor corresponding to the location of the damage. This is a weighting coefficient for the damaged area; This is the weighting coefficient for crack length; The weighting coefficient for the crack orientation angle; This is the weighting coefficient for the depth of the damaged area.

8. A wind turbine blade damage detection system based on image data, characterized in that, include: The image data acquisition module acquires raw image data of the wind turbine blade surface and performs preprocessing to obtain standard image data; The image segmentation and coordinate system establishment module extracts the mask of the blade surface region from the standard image data, establishes the blade surface coordinate system based on the boundary information of the mask, and divides it into a grid to obtain a surface grid image. The damage partitioning detection module uses a pre-trained multi-type damage recognition model to perform damage detection on each grid region in the surface grid image, outputting multiple damage regions and their corresponding damage types. The training process of the pre-trained multi-type damage recognition model includes: constructing an original image dataset containing three damage types: cracks, corrosion, and spalling; randomly rotating, brightness-perturbing, and injecting noise into the original image dataset to generate a training sample dataset; constructing dedicated detectors for cracks, corrosion, and spalling based on the YOLOv5 architecture, loading pre-trained weights, and freezing the parameters of a preset number of layers in the backbone network; performing transfer learning training on each dedicated detector using an adaptive learning rate strategy, setting an initial learning rate, and reducing the learning rate when the validation set loss does not decrease after a preset number of iterations; optimizing the loss function using the Adam optimizer; dynamically adjusting the confidence thresholds for each damage type based on the validation set results, stopping training when the detection recall rates for cracks, corrosion, and spalling reach preset indicators, and saving the optimal model parameters; integrating the three trained dedicated detectors and fusing the detection results through a weighted voting mechanism to obtain the multi-type damage detection model. The damage quantification module performs edge extraction and geometric feature parameter calculation on all damaged areas to obtain multiple damage quantification data. The scoring density calculation module calculates the damage score for each damage location by combining all the damage degree data, labels each damage location with a level label based on the damage score, and calculates the damage density distribution of each grid area of ​​the blade. The damage trend prediction module acquires historical blade inspection data, analyzes it to obtain historical damage evolution trends, and integrates the damage severity data and historical damage evolution trends to obtain future damage evolution trends. Specifically, it acquires historical blade inspection data, removes noise data to obtain spatiotemporal distribution data of historical damage, analyzes the expansion rate and direction of historical damage based on the spatiotemporal distribution data to obtain historical damage evolution trends, and combines the historical damage evolution trends with the damage severity data to analyze the future expansion rate and direction of current damage, integrating these to obtain future damage evolution trends. The result integration and output module outputs the damage area, damage type, damage severity data, grade label, damage density distribution, and future damage evolution trend as the final detection result.

Images