CN-121544619-B - Wind power generation blade surface defect detection method and system based on TFPN-YOLO model

CN121544619BCN 121544619 BCN121544619 BCN 121544619BCN-121544619-B

Abstract

The invention relates to a wind power generation blade surface defect detection method and system based on TFPN-YOLO model, which are used for constructing an improved defect detection network, wherein the improved defect detection network takes a YOLOv network as a basic network, a wind power generation blade surface image dataset is obtained, the dataset is constructed after preprocessing, the improved defect detection network is trained by the dataset, a wind power generation blade surface diagram is collected, the trained improved defect detection network is input, defect type, position and confidence information are obtained, and the system is set by the method. The invention effectively solves the key problems in the field of wind power generation blade surface defect detection, obviously improves the detection precision and efficiency, reduces the consumption of calculation resources, improves the applicability and stability of a model, provides a new solution for the blade surface defect detection in the wind power generation industry, and promotes the technical progress and development of the field.

Inventors

ZHANG FANGTAO
XU YABIN
WANG FAN
LI QINCHUAN

Assignees

浙江理工大学

Dates

Publication Date: 20260512
Application Date: 20260119

Claims (7)

1. A wind power generation blade surface defect detection method based on TFPN-YOLO model is characterized by constructing an improved defect detection network, wherein the improved defect detection network takes YOLOv network as a base network; The improved defect detection network comprises a backup module, a Neck module and a Head module which are sequentially matched; The method comprises the steps of replacing all feature extraction blocks of a basic network with feature extraction blocks modified by Fourier gating bottleneck convolution, wherein the feature extraction blocks modified by Fourier gating bottleneck convolution comprise CBS blocks and splitting blocks which are sequentially connected, parallel identical branches and refining branches are arranged behind the splitting blocks and are fused through a full connection layer and then output through a spatial attention network and the CBS blocks, the refining branches comprise 3 Fourier gating bottleneck convolution modules which are sequentially connected, each Fourier gating bottleneck convolution module comprises Fourier gating convolution blocks which are cooperatively arranged, each Fourier gating convolution block comprises a bottleneck convolution layer, an activation function layer and a Fourier enhancement block, in the Fourier enhancement block, the output of a Fourier transform layer is output to a channel splicing layer after passing through an average pooling layer and a channel dimension average layer in amplitude spectrum branches and phase spectrum branches, a convolution block is arranged behind the channel splicing layer, and the output of the convolution block is multiplied with the output element of the Fourier transform layer; setting a self-supervision visual Backbone network after a feature extraction block of a second Fourier gating bottleneck convolution modification of the backhaul module, wherein the output of the self-supervision visual Backbone network is associated with the Neck module; Setting a dynamic multi-scale feature fusion module after A2C2f of the back bone module and outputting the dynamic multi-scale feature fusion module to a Neck module; replacing all the full connection blocks of the basic network with a bidirectional feature pyramid network, and arranging a plurality of convolution layers in cooperation with the bidirectional feature pyramid network for cross-channel information integration; setting a detection Head for multi-scale feature fusion by matching with the Head module; acquiring a wind power generation blade surface image dataset, constructing a dataset after preprocessing, and training the improved defect detection network by using the dataset; And acquiring a wind power generation blade surface diagram, inputting the trained improved defect detection network, and acquiring defect type, position and confidence information.
2. The method for detecting the surface defects of the wind power generation blade based on TFPN-YOLO model as claimed in claim 1, wherein the self-supervision visual backbone network comprises a recombination layer, an up-sampling layer and a convolution layer which are sequentially connected.
3. The method for detecting the surface defects of the wind power generation blade based on TFPN-YOLO model according to claim 1, wherein the dynamic multiscale feature fusion module comprises multiscale receptive field branches, cascaded large nuclear attention branches and longitudinal and transverse-global context branches which are arranged in parallel; the multiscale receptive field branches comprise CBS blocks and 3 largest pooling layers which are connected in sequence; The cascade large-core attention branch comprises 4 large-core attention blocks which are connected in sequence, wherein the output of a CBS block is the input of a first large-core attention block, and the input of each large-core attention block is the sum of the output of a corresponding maximum pooling layer and the output of the last large-core attention block; The vertical and horizontal-global context branches comprise an average pooling layer, a dynamic convolution layer, a grouping vertical and horizontal convolution group and a bilinear upsampling layer which are connected in sequence; and the outputs of the multiscale receptive field branches, the cascade large nuclear attention branches and the longitudinal and transverse global context branches are spliced and then used as the outputs of the dynamic multiscale feature fusion module.
4. The method for detecting the surface defects of the wind power generation blade based on TFPN-YOLO model according to claim 1, wherein a TFPN feature fusion channel is established based on the bidirectional feature pyramid network; The TFPN feature fusion channel fuses the output of the self-supervision visual Backbone network and the subsequent A2C2f in the back bone module with the first bidirectional feature pyramid network in the Neck module, fuses the output of the feature extraction block modified by the self-supervision visual Backbone network and the first Fourier gate bottleneck convolution in the back bone module with the second bidirectional feature pyramid network in the Neck module, fuses the output of the 2A 2C2f after the first bidirectional feature pyramid network and the second bidirectional feature pyramid network with the third bidirectional feature pyramid network in the Neck module, and fuses the output of the A2C2f after the dynamic multiscale feature fusion module and the third bidirectional feature pyramid network with the fourth bidirectional feature pyramid network in the Neck module.
5. The method for detecting the surface defects of the wind turbine blade based on TFPN-YOLO model according to claim 1, wherein the detection head for multi-scale feature fusion comprises a boundary box regression branch, a classification branch and a boundary box prediction branch which are arranged in parallel; The boundary box regression branch, the classification branch and the boundary box prediction branch all comprise a convolution layer and a plurality of parallel CBS blocks which are sequentially arranged, and the outputs of the plurality of parallel CBS blocks are added and then output through the CBS blocks and the convolution layer.
6. The method for detecting surface defects of wind turbine blade based on TFPN-YOLO model as set forth in claim 1, wherein a loss function is constructed The improved defect detection network is trained to, , Wherein K is the total number of scales used in the multi-scale analysis; Degradation to single scale SIoU if k=1; Representing the fusion weights for the i-th scale, , In order for the parameters to be able to be learned, Expressed in scale The cross-over ratio of the two phases is lower, Expressed in scale The distance below the cost is matched to the cost, Expressed in scale The following shape matches the cost.
7. A wind power generation blade surface defect detection system based on TFPN-YOLO model is characterized by comprising: at least one processor, and A memory communicatively coupled to at least one of the processors; Wherein the memory stores instructions executable by the processor for execution by the processor to implement the TFPN-YOLO model-based wind turbine blade surface defect detection method of one of claims 1-6.

Description

Wind power generation blade surface defect detection method and system based on TFPN-YOLO model Technical Field The invention relates to the technical field of general image data processing or generation, in particular to a wind power generation blade surface defect detection method and system based on TFPN-YOLO model. Background Wind power generation plays a key role in global energy structure transformation as an important component of clean energy. The fan blade is a core component of the wind turbine generator, and the structural integrity of the fan blade directly influences the power generation efficiency, the service life of equipment and the operation safety. However, due to long-term exposure of the blade to complex natural environments such as strong wind, rain erosion, ultraviolet radiation, salt spray corrosion and the like, and the limitation of the manufacturing process, the surface of the blade is easy to generate defects such as cracks, flaking, bulges, coating shedding, lightning damage and the like. If not found and handled in time, these defects may spread rapidly, resulting in blade breakage, unit downtime, and even safety accidents, resulting in significant economic losses. Currently, the defect detection of fan blades mainly depends on manual visual inspection, detection of an unmanned aerial vehicle-mounted visible light/thermal imager, detection based on acoustic or vibration signals, ultrasonic detection and the like. The manual visual inspection is that a technician closely observes the surface of the blade through a telescope, an aerial working platform or a hanging basket, and has low accuracy and high cost. The unmanned aerial vehicle carries visible light/thermal imaging instrument to detect is to obtain the blade image through taking photo by plane, combine the image processing technology to discern the defect, this method has the problem that is easy to be interfered by environmental factor, thermal imaging is effective only to specific defect such as inside debonding, image resolution and flight stability direct constraint detection precision. The detection based on acoustic or vibration signals captures the excited vibration of the blade or acoustic reflection signals through the sensor to diagnose internal damage, is insensitive to small defects, has high omission rate, needs to arrange a complex sensor network, and has high implementation difficulty and high cost. Although the ultrasonic detection can identify the tiny defects of the blade, the coverage is wide, the ultrasonic detection can be performed only by applying external force to the equipment, and the accurate condition of the defect detection target cannot be reflected. With the rapid development of artificial intelligence technology, the wind power generation blade surface defect detection method based on deep learning gradually becomes a research hotspot, and particularly the deep learning technology is widely applied to the field of target detection. According to the method, a deep learning model is built, a large number of wind driven generator blade surface defect images are trained and learned, and accurate identification and positioning of defects are achieved. However, existing deep learning-based wind turbine blade surface defect detection algorithms still have some problems. For example, part of algorithms have large calculation amount and low detection speed due to complex model structure, and are difficult to meet the requirement of real-time detection, and meanwhile, some algorithms still need to improve the detection precision, and especially have unsatisfactory defect detection effects under small targets and complex backgrounds. Along with the development of deep learning technology, YOLO series algorithms are widely used with the characteristics of high efficiency and accuracy. YOLOv12 as one of the latest versions of the YOLO series, performs well in the object detection task. However, when the method is applied to the detection of the surface defects of the fan blade, the problems of insufficient detection precision and large parameter amount still exist. Specifically, when the existing YOLOv network processes the fan blade surface defect data set, the small defect characteristics of the blade surface are difficult to fully extract due to the limitation of the model grid structure, so that the detection precision is not high. Meanwhile, the larger parameter increases the complexity of the model and the demand for computing resources, which is unfavorable for deployment and popularization in practical application. The traditional detection method has the problems of low efficiency and insufficient precision, and the existing detection method based on deep learning has the limitations of complex model, low detection speed and the like. Aiming at the problems and the defects in the prior art, a lightweight and high-precision wind power generation blade surface defect detection algorithm is expect