CN-121982385-A - Wind power blade production defect analysis method based on image recognition

CN121982385ACN 121982385 ACN121982385 ACN 121982385ACN-121982385-A

Abstract

The invention relates to the technical field of wind power blade manufacturing, and particularly discloses a wind power blade production defect analysis method based on image recognition. The analysis method is based on blade production process images which are arranged in high altitude and captured in real time by an industrial camera and blade production process images which are captured in real time by a handheld photographic device, and independent analysis processes are carried out on the images of two scenes, wherein the analysis processes of the images comprise S1, segmentation and normalization pretreatment of the images, S2, multi-scale feature extraction and key region attention enhancement treatment of the images, S3, multi-scale feature complementation and screening treatment of the images, S4, defect accurate positioning and classification treatment of the images, and S5, judging the existence state of the defects, the type and position coordinates of the current defects. The invention improves the quality detection efficiency of the wind power blade in the production process, ensures the quality detection precision and reduces the degree of dependence on manpower.

Inventors

BIE CHUNHUA
LI HAOLIANG
LING LE
YU LI
ZHAO LISHAN
PU XIAOMIN
CHEN YANG
SONG WEI
Sun Lingze
SHI WENBIN
WU BO

Assignees

东方电气集团数字科技有限公司
东方电气股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260116

Claims (10)

1. A wind power blade production defect analysis method based on image recognition is characterized by comprising the following steps of: the analysis method is based on the blade production process images which are captured in real time by the high-altitude arrangement industrial camera and the blade production process images which are captured in real time by the handheld photographic equipment, and an independent analysis process is carried out on the images of the two scenes; the image analysis method comprises the following steps: s1, carrying out segmentation and normalization pretreatment on an image; S2, taking YOLOv S as a baseline model, constructing a backstone network structure by adopting a CSPDARKNET53 network structure and combining an SE Block attention mechanism, and carrying out multi-scale feature extraction and key region attention enhancement treatment on the image; S3, constructing a Neck module by adopting a PAN module, an FPN network structure and a CBAM module, and performing multi-scale feature complementation and screening treatment on the image; s4, constructing a multi-scale detection Head module, optimizing a loss function, and carrying out defect accurate positioning and classification treatment on the image; S5, judging the existence state of the defect and the type and position coordinates of the current defect by combining the boundary frame coordinates, the target confidence level and the type confidence level output by the Head module and a threshold value; Outputting a classification result of 'large wrinkles' or 'production sequelae' through the high-altitude scene image; Multiple classification results including "dry yarn", "wrinkles", "hypoperfusion", "blushing", "air bubbles" are output through the hand-held scene image.
2. The method for analyzing the production defects of the wind power blade based on the image recognition according to claim 1, wherein in the step S1, the preprocessing process of the image is as follows: Firstly, dividing an original image into a plurality of subareas by using a multi-size window; then unifying the resolution of the segmented sub-region images; then converting the BGR format into an RGB format; then, carrying out normalization processing on the image; Finally, carrying out standardization treatment according to the following relation: ; Wherein I represents pre-processing image data; Mu represents the mean; Sigma represents standard deviation.
3. The method for analyzing the production defects of the wind power blades based on image recognition according to claim 1 or 2, wherein before preprocessing the images of the handheld scene, interference contents including unit name identification watermarks and frames included in the images are removed, and only a core area containing the defects is reserved.
4. The method for analyzing the production defects of the wind power blade based on the image identification according to claim 1, wherein in the step S2, the image is subjected to multi-scale feature extraction and key region attention enhancement, and the method specifically comprises the following process steps: Step 1, dividing an input image characteristic diagram into two parts through CSPNet network structures, wherein one part is processed by a convolution layer and a residual block, and the other part is directly transmitted; finally splicing and fusing the two parts of features, and sequentially extracting deep semantic features of three scales P3, P4 and P5; Step 2, after each layer of convolution of CSPDARKNET network structure, adding a Squeeze-and-specification Block module; firstly, global average pooling processing is carried out on the convolution output feature map according to the following relation to obtain channel-level feature vectors: ; In the formula, Representing a global average pooling value; representing the height of the feature map; representing the width of the feature map; representing the number of channels; Representing the characteristic diagram in the height direction A position index on the table; representing the characteristic diagram in the width direction A position index on the table; generating channel weights through two layers of fully connected networks of a first layer of ReLU activation function and a second layer of Sigmoid activation function, and finally acting the weights on the original feature map through element-by-element multiplication to enhance the attention of the model to the defect area, wherein the following relational expression is satisfied: ; In the formula, The feature map weighted by the attention of the channel is represented, so that the attention degree of a defect area is enhanced; a feature map representing an input; Representing a Sigmoid activation function; representing a ReLU activation function; a weight matrix representing a first fully connected layer; a weight matrix representing a second fully connected layer; And 3, training the first 5 layers of parameters of the initial frozen backhaul network structure, and training only the subsequent network layers, so as to reduce the consumption of computing resources and accelerate the convergence of the model.
5. The method for analyzing defects in wind power blade production based on image recognition according to claim 4, wherein in step 1, parameters of the CSPNet network structure are set as follows: An initial convolution layer, namely an input channel 3, an output channel 64, convolution kernels 6 multiplied by 6, a step distance 2 and a padding2; The subsequent convolution layer is combined with the C3 modules, namely the convolution layer of the output channel 128 is followed by 3C 3 modules, the convolution layer of the output channel 256 is followed by 6C 3 modules, the convolution layer of the output channel 512 is followed by 9C 3 modules, the convolution layer of the output channel 1024 is followed by 3C 3 modules, and finally the feature extraction effect is enhanced by the SPPF module.
6. The method for analyzing the production defects of the wind power blade based on the image identification according to claim 1, wherein in the step S3, the image is subjected to multi-scale feature complementation and screening, and the method specifically comprises the following process steps of; step A, performing 1X 1 convolution dimension reduction on a P5 feature layer output by a backhaul network structure, and performing multi-channel data up-sampling; step B, after the up-sampling in the step A is completed, splicing and fusing the up-sampling with the P4 characteristic layer subjected to the same 1X 1 convolution dimension reduction; After the fused feature layer is processed, performing 1×1 convolution dimension reduction and up-sampling again, and performing splicing fusion with the P3 feature layer subjected to the 1×1 convolution dimension reduction to enhance the small defect detection capability; c, performing 3X 3 convolution downsampling on the P3 feature layer after the fusion of the FPN network structure, and performing splicing fusion processing on the P4 feature layer after the processing in the FPN network structure process; Then 3X 3 convolution downsampling is carried out on the feature layer, and splicing and fusion processing is carried out on the feature layer and the P5 feature layer which is output by a Backbone network structure and subjected to 1X 1 convolution dimension reduction, so that feature flow and information transmission are enhanced; And D, after the P3, P4 and P5 feature layers are fused, sequentially introducing a convolution block attention module to carry out attention screening.
7. The method for analyzing the production defects of the wind power blade based on the image recognition according to claim 6, wherein in the step D, the attention screening comprises the following specific steps: Step a, generating channel weights through a channel attention module, respectively executing global average pooling and global maximum pooling on an input feature map, splicing the obtained results, sequentially processing through a 1X 1 convolution, a ReLU activation function, a 1X 1 convolution and a Sigmoid activation function, and finally outputting weight coefficients of all channels; The following relation is satisfied: ; In the formula, Representing a channel weight matrix; representing a feature map; step b, respectively carrying out average pooling and maximum pooling on the feature images in the channel dimension, carrying out splicing treatment on the two obtained two-dimensional feature images, carrying out 7×7 convolution treatment on the splicing result, and finally outputting a space weight matrix after activating by a Sigmoid activation function; the following relation is satisfied: ; In the formula, Representing a spatial weight matrix; multiplying the channel weight and the space weight with the feature map in sequence by using element-by-element multiplication, and integrating the attention screening effect; the following relation is satisfied: ; In the formula, The attention screening result of the feature map is shown.
8. The method for analyzing the production defects of the wind power blade based on the image recognition according to claim 1, wherein in the step S4, the image is subjected to the accurate positioning and classification of the defects, and the method specifically comprises the following process steps of; Taking the P3, P4 and P5 feature layers processed by the Neck module as detection head input, and outputting 6-dimensional vectors including the coordinates of the central point of the boundary frame, the width and height of the boundary frame, the target confidence coefficient and the category confidence coefficient by each scale; Wherein the P3 scale focuses on detecting small defects including bubbles; the P4 scale focuses on detecting mid-defects including wrinkles; the P5 scale focuses on detecting large defects including production drifts; Step II, calculating the regression loss of the boundary frame by adopting a Complete-IoU loss function, and comprehensively considering the overlapping area, the center distance and the length-width ratio; the following relation is satisfied: ; In the formula, Representing Complete-IoU loss function; Representing the intersection ratio of the prediction frame and the real frame; a is the area of a prediction frame; b is the real frame area; representing the Euclidean distance between the predicted frame and the center point of the real frame; Representing the minimum bounding box diagonal length; Representing weight coefficients for balancing the effects of aspect ratio loss terms; a consistency metric representing aspect ratio of the predicted frame to the real frame; the real frame width; is a real frame height; the predicted frame width; Is the prediction frame height; step III, the total loss function is formed by weighted summation of classification loss and target confidence loss; the following relation is satisfied: ; In the formula, Representing a total loss function; Weight coefficients representing the regression loss of the bounding box; a weight coefficient representing a target confidence loss; representing a target confidence loss function; a weight coefficient representing a classification loss; representing a classification loss function.
9. The method for analyzing the production defects of the wind power blades based on image recognition according to claim 1, wherein the industrial cameras are a plurality of industrial-grade CMOS cameras which are distributed above a blade production station and are provided with vision overlapping areas, each industrial camera is respectively connected with a high-altitude edge computer in a signal manner, and the high-altitude edge computer completes the recognition process of the input corresponding images; The handheld photographing equipment is a mobile phone which is held by a constructor, the mobile phone is connected with a mobile phone server in a signal manner, and the mobile phone server completes the identification process of the input corresponding image.
10. The method for analyzing the production defects of the wind power blades based on image recognition according to claim 1, wherein the blade production process comprises a process node before blade glue filling, a process node after blade glue filling and a process node after blade die closing.

Description

Wind power blade production defect analysis method based on image recognition Technical Field The invention relates to the technical field of wind power blade manufacturing, in particular to a wind power blade production defect analysis method based on image recognition. Background In the structure of the wind driven generator, the wind power blade is one of core components, is a source for capturing wind energy by the unit, and plays a vital role in the running process of the unit. The production process route of the wind power blade is longer, and the production process flow of the wind power blade can be summarized into seven links of mould preparation, core material placement, pressure maintaining completion, pouring solidification, mould closing first adhesion, mould closing second adhesion and heating solidification. Each link has various defects caused by improper manual operation, such as wrinkling and layering foreign matters caused by improper layering operation in the links of core material placement, die assembly and the like, and whitening defects caused by improper resin proportion or temperature control in the links of pouring, curing, heating and the like. Various defects will directly affect the quality and production efficiency of the final blade product to different extents. Therefore, irreversible heat curing reaction can occur in the production process of the wind power blade, and quality control in the production process of the wind power blade directly determines the standard rate of the wind power blade and the operation achievements of upgrading, reducing cost and enhancing efficiency of enterprises. At present, quality control in the wind power blade production process is mainly performed manually, and because manual detection is limited by factors such as subjective experience difference and physiological fatigue of operators, the technical problems of low identification consistency, high omission rate and the like exist in identification of defects such as wrinkles, tiny foreign matters and hidden blushing, and the omission rate is usually up to more than 15% according to incomplete statistics. Therefore, the current quality control mainly based on manual work has the technical problems of high quality control difficulty, low production efficiency, low precision and the like. Disclosure of Invention Aiming at the special property of the wind power blade production process, the technical requirement on quality control and the defects of the prior art, the invention provides the wind power blade production defect analysis method for analyzing and identifying the image of the wind power blade production process based on YOLOv algorithm so as to improve the quality detection efficiency of the wind power blade in the production process, ensure the quality detection precision and reduce the manual dependency. The technical logic of the invention is that Yolo v s algorithm is used as a baseline model, and single-Stage target detection architecture (One-Stage Detector) is adopted to realize end-to-end defect positioning and classification. The backbone Network in the original Yolo v algorithm is a Cross STAGE PARTIAL Network (CSPNet) part Network, which is used for extracting deep semantic features of an image, and the method adds a Squeeze-and-Excitation (SE) Block after each layer of convolution so as to enhance the attention of a model to a key region. SE Block improves the expressive power and generalization performance of the model by automatically learning which channels are more important. Because factors such as shooting angles and illumination conditions can influence the quality of shot pictures, a neck network adds Convolutional Block Attention Module (CBAM) behind each feature pyramid layer on the basis of an original FPN+PAN structure so as to strengthen the sensitivity of a model to defect features and defect positions and prevent feature loss. The technical aim of the invention is achieved by the following technical scheme, namely the wind power blade production defect analysis method based on image recognition, wherein the analysis method is based on blade production process images captured in real time by an aerial arrangement industrial camera and blade production process images captured in real time by handheld photographic equipment, and an independent analysis process is carried out on the images of two scenes; the image analysis method comprises the following steps: s1, carrying out segmentation and normalization pretreatment on an image; S2, taking YOLOv S as a baseline model, constructing a backstone network structure by adopting a CSPDARKNET53 network structure and combining an SE Block attention mechanism, and carrying out multi-scale feature extraction and key region attention enhancement treatment on the image; S3, constructing a Neck module by adopting a PAN module, an FPN network structure and a CBAM module, and performing multi-scale feature complementation