CN-121981977-A - BD-Net-based two-stage detection and positioning method for blade defects of wind driven generator

CN121981977ACN 121981977 ACN121981977 ACN 121981977ACN-121981977-A

Abstract

The invention discloses a BD-Net based two-stage detection and positioning method for blade defects of a wind driven generator, which comprises the steps of firstly, utilizing GSS-Det to perform preliminary detection on a high-resolution image of the blade obtained by an unmanned aerial vehicle, and outputting a defect candidate region and confidence level thereof; the method comprises the steps of carrying out a dynamic decision mechanism according to confidence coefficient, directly entering a fine screening stage for a high confidence coefficient region, carrying out high resolution re-cutting for a medium confidence coefficient region, carrying out secondary detection, filtering for a low confidence coefficient region, carrying out fine screening by adopting a dual-branch attention mechanism network D2A-Net, classifying and accurately dividing the fine granularity of the blade defect, outputting the defect type, the severity level and the corresponding division mask, establishing a camera internal parameter calibration model, calculating the mapping relation between pixel coordinates and blade physical coordinates, mapping the defect pixel coordinates to the blade physical coordinates, and outputting the accurate physical position of the defect. The invention can obviously improve the accuracy, efficiency and positioning precision of blade defect detection.

Inventors

LIU ZHOU
ZHENG SHIJIN
ZHOU FEI
DING HONG
Ye Xingpei
CAO WENKAI

Assignees

江苏省国信研究院有限公司

Dates

Publication Date: 20260505
Application Date: 20260108

Claims (5)

1. The BD-Net-based two-stage detection and positioning method for the blade defects of the wind driven generator is characterized by comprising the following steps of: Constructing a dual-stage wind driven generator blade defect detection integrated network model BD-Net, wherein the model BD-Net comprises a GSS-Det target detection network and a dual-branch attention mechanism network D2A-Net; coarse detection is carried out on the blade image acquired by the unmanned aerial vehicle based on the GSS-Det, and a defect boundary box and corresponding confidence coefficient are output; performing dynamic decision making mechanism according to the confidence coefficient, and directly entering a fine screening stage for the high confidence coefficient region; fine screening is carried out by adopting a double-branch attention mechanism network D2A-Net, fine granularity classification and accurate segmentation of blade defects are carried out, and defect types, severity levels and corresponding segmentation masks are output; And establishing a camera internal parameter calibration model, calculating a mapping relation between pixel coordinates and physical coordinates of the blade, mapping defective pixel coordinates to the physical coordinates of the blade, and outputting the accurate physical position of the defect.
2. The BD-Net based wind turbine blade defect two-stage detection and localization method of claim 1, wherein the GSS-Det target detection network is a lightweight detection network formed by structural improvement on YOLOv s frame, and the lightweight design comprises: S21, replacing the traditional standard convolution by using a Ghost convolution module, and reducing the redundant feature calculation amount by using a feature redundancy utilization mechanism to realize the light weight of the network structure; And S22, introducing dynamic sparse training in a training stage, automatically removing network connection with smaller influence on a detection result through iterative pruning, and further reducing the model parameter quantity and the calculation complexity.
3. The BD-Net based wind turbine blade defect two-stage detection and localization method of claim 1, wherein the dynamic decision mechanism according to the confidence level comprises: Triggering high-resolution re-clipping and performing secondary detection on the re-clipped region when the confidence coefficient is in the (0.3,0.8) interval; 1) When the confidence coefficient is 0.8< confidence coefficient, marking as detecting the defect, and sending the defect map into a subsequent classification and segmentation dual-branch network for defect fine classification; 2) When the confidence coefficient is less than or equal to 0.3 and less than or equal to 0.8, triggering a high-resolution re-cutting strategy, and performing 4 times high-resolution re-cutting on the ROI area to perform secondary detection; 3) When the confidence coefficient is less than or equal to 0.3, marking the image as a normal image, and not processing the normal image.
4. The BD-Net based wind turbine blade defect two-stage detection and localization method of claim 1, wherein the dual branch attention mechanism network D2A-Net is an optimized dual branch attention mechanism network D2A-Net, and the defect classification and segmentation is synchronously completed through the optimized dual branch attention mechanism network D2A-Net, and the defect type, severity level and segmentation mask are output, comprising: The D2A-Net dual-branch attention mechanism network adopts a parallel structure, and classification branches focus on defect surface characteristics to classify defect types such as cracks, pores, surface corrosion and the like; The classification branches take VGG16 as characteristics to extract a main network, and a CBAM channel-space attention mechanism is introduced to enhance the expression capability of the surface textures and the intensity characteristics of the defects; the split branches adopt U-Net as a main network and combine an SE channel attention mechanism with CARAFE dynamic up-sampling modules for identifying complex crack structures and narrow defects.
5. The two-stage detection and positioning method for blade defects of a BD-Net-based wind turbine according to claim 1, wherein the unmanned aerial vehicle is provided with an RTK-GPS+IMU, the shooting position is recorded in real time, the shooting position comprises longitude, latitude, altitude and attitude angle, a camera internal parameter calibration model is established, and the mapping relation between pixel coordinates and physical coordinates of the blade is calculated by the following steps: Performing de-distortion correction on image pixel coordinates based on camera internal parameters; Homography transformation matrix for calculating camera to blade plane by using pose data of unmanned aerial vehicle ; By inverse homography transformation Mapping the corrected pixel coordinates to a blade physical coordinate system; The physical distance from the defect point to the reference point of the blade root along the length direction of the blade is calculated by combining with the actual length of the blade, and the calculation formula is as follows: Wherein, the As the physical coordinates of the defect point, Is the coordinates of a reference point of the blade root, For the actual length of the blade, Is the length of the blade in the projection plane.

Description

BD-Net-based two-stage detection and positioning method for blade defects of wind driven generator Technical Field The invention belongs to the technical field of computer vision, and particularly relates to a BD-Net-based two-stage detection and positioning method for blade defects of a wind driven generator. Background With the rapid development of new energy industry, wind power generation has been widely used in the global field as an important component of green clean energy. The fan blade is used as one of the most critical bearing components in the wind generating set, and the running state of the fan blade directly influences the generating efficiency and the equipment safety. However, the blade is exposed to a complex natural environment for a long period of time, and various structural defects such as cracks, delamination, air holes, corrosion, lightning damage and the like are easily generated. Traditional blade inspection mode, like artifical climbing inspection, ground telescope observe and take a photograph artifical check, although can discover showing the damage to a certain extent, there is obvious not enough in the aspect of security, efficiency and accuracy, especially when facing large-scale wind-powered electricity generation field, large-scale blade inspection demand, more be difficult to satisfy real-time, high-efficient, automated detection's requirement. Therefore, the development of the fan blade defect detection technology which is efficient, accurate and capable of being applied in engineering has important practical significance. In the prior literature, li Zhe provides an intelligent fan blade detection method based on Unmanned Aerial Vehicle (UAV) inspection in a text of intelligent inspection and defect detection of fan blades based on unmanned aerial vehicles. According to the system, the unmanned aerial vehicle is used for shooting the blade image, and the identification and the positioning of defects such as cracks and corrosion are realized by utilizing the image enhancement, threshold segmentation and a characteristic point matching method based on acceleration robust features (SURF). The method breaks through the limit of manual detection to a certain extent, and can realize automatic extraction of the defects on the surface of the blade. However, as the SURF and other traditional feature descriptors are sensitive to illumination change, blade scale difference and blurred images, the method is easy to produce false detection and omission under a complex outdoor environment, and has insufficient performance in the aspect of small target defect identification, so that the detection precision and stability are difficult to meet engineering requirements. In another related study, wang Jun in a method for detecting surface defects of fan blades by improving YOLOv s algorithm, a progressive feature pyramid network (AFPN) and a channel space attention mechanism (CBAM) are introduced based on YOLOv s, and MPDIoU is adopted as a positioning loss function to improve the detection capability of defects such as cracks, air holes and the like. The method is superior to the traditional algorithm in terms of detection precision and speed, and the effectiveness of the deep learning model in the blade defect detection task is proved. However, this method relies on a single-stage detection structure of uniform scale, and erroneous judgment may still occur for small targets with unstable confidence. Meanwhile, the method only outputs the image coordinates of the defects, and the pose and the geographic information of the unmanned aerial vehicle are not combined, so that the real space position of the defects cannot be directly provided, and the subsequent operation and maintenance positioning and maintenance efficiency are affected. In summary, the existing fan blade defect detection technology still has obvious defects in the aspects of complex scene adaptability, small target recognition capability, detection stability, space positioning capability and the like. Therefore, the two-stage fan blade defect detection and positioning method capable of realizing defect detection, defect classification, severity assessment and defect accurate positioning is particularly necessary, and has a remarkable engineering application prospect. Disclosure of Invention In view of the above, the invention aims to provide a BD-Net based two-stage detection and positioning method for blade defects of a wind driven generator, which solves the problems of low detection efficiency, inaccurate defect type identification, difficult physical positioning and the like in the prior art. The technical scheme provided by the invention is that the two-stage detection and positioning method for the blade defect of the wind driven generator based on BD-Net comprises the following steps: Constructing a dual-stage wind driven generator blade defect detection integrated network model BD-Net, wherein the model BD-Net compris