A wind turbine blade surface defect detection method suitable for edge computing devices

Abstract

A wind turbine blade surface defect detection method suitable for edge computing devices, comprising the following steps: step 1: constructing a surface defect image dataset facing the physical scene of the wind turbine blade and performing data processing to form an original dataset; step 2: performing image enhancement and spatial mapping to simulate a real outdoor complex industrial environment on the original dataset, expanding the sample size to an effective training scale, and dividing it into a training set, a validation set and a test set according to a preset ratio; step 3: constructing a light-weight edge-guided multi-scale target detection network DEML-DEIM based on the DEIM paradigm; step 4: inputting the training set obtained in step 2 into the DEML-DEIM network constructed in step 3 for iterative training; step 5: deploying the trained model to the edge computing device for detecting the surface defects of the wind turbine blade.

Description

Technical Field

[0001] This invention relates to the fields of new energy and wind power generation technology, and specifically to a method for detecting surface defects on wind turbine blades suitable for edge computing devices. Background Technology

[0002] Wind energy is a core component of the global sustainable energy system, and wind turbine blades, as a critical wind-catching structure, are susceptible to defects such as cracks caused by dust accumulation, cyclic load damage, wind erosion, and environmental aging. Failure to detect these defects in a timely manner can lead to structural fractures, resulting in serious safety accidents and economic losses. Therefore, achieving high-precision, high-efficiency automated detection of blade surface defects is crucial for ensuring operational safety and reducing total life-cycle costs.

[0003] Current detection methods rely on manual inspections or drone / telescope-assisted imaging, which suffers from high labor intensity, low efficiency, and poor consistency in harsh high-altitude environments. While deep learning-driven automated detection offers some improvement, existing solutions still have shortcomings: 1) One-stage CNN detectors (such as the YOLO series, for example: real-time object detection based on hypergraph-enhanced adaptive visual perception YOLOv13) are prone to missed detections / false detections of irregular defects due to anchor boxes and NMS mechanism; 2) Two-stage CNNs (such as unbiased Faster R-CNN for single-source domain generalized object detection) are limited by their local receptive fields and have difficulty capturing long-distance dependencies; 3) Transformer-type models (such as DEIM: based on the DETR algorithm to achieve fast convergence through improved matching mechanism) eliminate anchor boxes and NMS and converge quickly, but they are not adaptable to complex backgrounds, dense scenes with small targets and resource constraints of edge devices. Industrial defect detection research has made many improvements, such as DETR-based improvements: RT-DETR for forging surface defect detection, YOLOv8-based improvements: its cross-scale feature extraction of forging surface cracks, DETR-PANet: a DETR-based method for accurate visual inspection in 3C assembly scenarios, etc. However, these existing technologies still have difficulties in balancing lightweight design and inspection accuracy, making them unsuitable for the surface defect detection of wind turbine blades based on edge computing devices.

[0004] To address the aforementioned issues, this invention proposes DEML-DEIM—a lightweight, high-precision defect detection method based on the DEIM-n baseline and specifically designed for wind turbine blade inspection. Summary of the Invention

[0005] The purpose of this invention is to address the technical shortcomings of existing target detection technologies, such as conventional CNN detectors and Transformer models. These shortcomings include feature extraction not being optimized for irregular defects, susceptibility to background noise interference, and difficulty in balancing high detection accuracy with low computational cost. Consequently, wind turbine blade surface defect detection suitable for edge computing devices suffers from low recognition accuracy, frequent false alarms and missed detections when facing defects with extreme scale variations, and overly bloated models that are difficult to deploy in real time and efficiently on edge devices with extremely limited computing resources. Therefore, this invention is proposed.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for detecting surface defects on wind turbine blades suitable for edge computing devices includes the following steps: Step 1: Construct a surface defect image dataset for the physical scenario of wind turbine blades and perform data processing to form the original dataset; Step 2: Perform image enhancement and spatial mapping on the original dataset to simulate a real complex outdoor industrial environment, expand the sample size to an effective training scale, and divide it into training set, validation set and test set according to a preset ratio; Step 3: Construct a lightweight edge-guided multi-scale target detection network DEML-DEIM based on the DEIM paradigm; Step 4: Input the training set obtained in Step 2 into the DEML-DEIM network constructed in Step 3 for iterative training; Step 5: Deploy the trained model to an edge computing device to detect surface defects on wind turbine blades.

[0007] In step 1, the specific steps are as follows: Acquire a dataset of wind turbine blade surface images taken by drones from multiple angles and distances, containing real outdoor physical scenes; clean the image data to meet the actual application requirements of defect detection in complex industrial environments, remove neglected areas that are prone to background physical noise interference, such as severely non-uniform lighting or non-defect texture areas, and retain only defect labels with clear physical geometric contours, including three types of defects: stains, damage, and cracks, to adapt to high-precision wind turbine blade defect detection tasks.

[0008] In step 2, the specific steps are as follows: The cleaned wind turbine blade image is converted into a digital matrix format, and image enhancement and preprocessing operations are performed to simulate the interference of a real outdoor complex industrial environment. The image enhancement and preprocessing operations include: random cropping, segmentation, and size normalization operations to simulate the drastic scale changes caused by different physical distances and defect morphologies during UAV inspection; color correction and dynamic brightness adjustment to simulate the interference of natural lighting changes; and expanding the spatial distribution characteristics of defects through geometric transformations such as random flipping and rotation. After completing the above spatial and color feature transformations, the model is divided into training set, validation set, and test set according to a preset ratio to improve the robustness of the model in open physical scenarios.

[0009] In step 4, the specific steps are as follows: The wind turbine blade image matrix in the training set is input into the lightweight edge-guided multi-scale object detection network DEML-DEIM constructed in step 3 for iterative training; during the training process, the optimal matching relationship between the bounding box set predicted by the model and the real defect annotation box set is calculated based on the bipartite graph matching strategy; based on the optimal matching result, the classification loss and the bounding box regression loss reflecting the degree of overlap in physical space are jointly calculated; the AdamW optimizer is used, the initial learning rate and weight decay are configured, and the network weights are continuously updated through the backpropagation algorithm until the loss function converges, thus obtaining the trained lightweight real-time defect detection model.

[0010] In step 5, the specific steps are as follows: Deploy the trained lightweight defect detection model to an edge computing device; acquire real-time images of the surface of the wind turbine blade to be detected and input them into the model; the model directly outputs the category confidence of each defect target in the image and the bounding box coordinates that accurately represent the actual physical spatial location of the defect through a single forward inference; finally, filter the prediction results according to the preset confidence threshold and output the final defect location and classification information to support subsequent downstream actions such as automated alarms and maintenance decisions.

[0011] In step 3, the constructed lightweight edge-guided multi-scale target detection network DEML-DEIM based on the DEIM paradigm includes a backbone network, an encoder network connected to the output of the backbone network, and a decoder network connected to the output of the encoder network; the backbone network is specifically as follows: The input image is fed into the Defect Adaptive Multi-Scale Backbone Module (DAMS). DAMS preserves fine-grained features during early downsampling using parallel receptive fields, addressing the technical issues of low recognition accuracy and susceptibility to false positives and missed detections when facing defects with extreme scale variations. The output of DAMS is connected to the input of the first-stage module Ino_Stage1. The output of Ino_Stage1 is connected to the input of the second-stage module Ino_Stage2. The output of Ino_Stage2 is connected to the input of the third-stage module Ino_Stage3 and the first convolutional module in the encoder. The output of Ino_Stage3 is connected to the input of the fourth-stage module Ino_Stage4 and the second convolutional module in the encoder. The output of Ino_Stage4 is connected to the input of the third convolutional module in the encoder.

[0012] The internal connection structure of the first stage module Ino_Stage1, the second stage module Ino_Stage2, the third stage module Ino_Stage3, and the fourth stage module Ino_Stage4 is as follows: Each stage module includes a pre-convolutional module and multiple cascaded internal blocks (Ino_Block). Within each Ino_Block, input features are passed to one input of the multi-branch reparameterized lightweight module (MBRL) and the addition module (Add). Multiple parallel branches of the MBRL are fused via the concatenation module (Concat) and then connected to the input of the edge-guided enhanced attention module (EGEA). The EGEA uses edge prior knowledge to guide the network to focus on high-frequency defect regions, and its output is connected to the other input of the addition module (Add). The output of the Add module (Add) serves as the output of this Ino_Block, thereby enhancing multi-scale and directional feature extraction capabilities while maintaining training stability and addressing the technical problem of overly bloated models that are difficult to deploy efficiently in real-time on edge devices with extremely limited computing resources.

[0013] In the encoder, the output of the third convolutional module is connected to the input of the attention interaction feature module AIFI. The attention interaction feature module AIFI enhances the high-level semantic feature representation within a single scale through a self-attention mechanism, thereby improving the network's ability to capture features of the irregular surface defect edges of wind turbine blades. The output of the attention interaction feature module AIFI is connected to the input of the fourth convolutional module in the encoder. The output of the fourth convolutional module in the encoder is connected to the input of the first upsampling module. The outputs of the first upsampling module and the second convolutional module in the encoder are connected to the input of the first lightweight spatial channel pixel collaborative fusion module LSCP. Meanwhile, the output of the fourth convolution module in the encoder is connected to the input of the second lightweight spatial channel pixel collaborative fusion module LSCP, and the output of the second lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the first high-efficiency layer aggregation network module ELAN. The output of the first upsampling module is connected to the input of the second lightweight spatial channel pixel co-fusion module LSCP. The output of the second lightweight spatial channel pixel co-fusion module LSCP is connected to the input of the second high-efficiency layer aggregation network module ELAN. The output of the second high-efficiency layer aggregation network module ELAN is connected to the input of the fifth convolution module in the encoder. The output of the fifth convolution module is connected to the input of the second upsampling module. The output of the second upsampling module and the output of the first convolution module in the encoder are connected to the input of the third lightweight spatial channel pixel co-fusion module LSCP. The output of the third lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the third high-efficiency layer aggregation network module ELAN. The output of the third high-efficiency layer aggregation network module ELAN is connected to the input of the first downsampling module. The outputs of the first downsampling module and the fourth convolution module in the encoder are connected to the input of the fourth lightweight spatial channel pixel collaborative fusion module LSCP. The output of the fourth lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the fourth high-efficiency layer aggregation network module ELAN. The output of the fourth high-efficiency layer aggregation network module ELAN is connected to the input of the second downsampling module. The output of the second downsampling module is connected to the input of the second lightweight spatial channel pixel collaborative fusion module LSCP. The encoder is equipped with multiple lightweight spatial channel pixel collaborative fusion modules (LSCP) to promote efficient multi-scale semantic alignment and fusion of shallow and deep features. This effectively solves the technical problems of extreme scale changes in wind turbine blade defects and the fact that tiny cracks are easily submerged by complex backgrounds, leading to missed detections. This ensures the detection accuracy of edge devices with extremely low parameter levels. The outputs of the first high-efficiency layer aggregation network module ELAN, the second high-efficiency layer aggregation network module ELAN, and the third high-efficiency layer aggregation network module ELAN are connected to the input of the flatten module in the decoder.

[0014] In the decoder, the output of the Flatten module is connected to the input of the initial feature sequence layer; the output of the initial feature sequence layer is divided into two flow branches: The first main branch is connected to the main input of the first decoder layer, Decoder_Layer1; The second branch serves as the object query branch, and after being led out, it is connected to the query input of the first decoder layer Decoder_Layer1, the query input of the second decoder layer Decoder_Layer2, the query input of the third decoder layer Decoder_Layer3, and so on, until it is finally connected to the query input of the Nth decoder layer Decoder_LayerN. The output of the first decoder layer Decoder_Layer1 is connected to the main input of the second decoder layer Decoder_Layer2, the output of the second decoder layer Decoder_Layer2 is connected to the main input of the third decoder layer Decoder_Layer3, and so on, until the output of the (N-1)th decoder layer Decoder_LayerN-1 is connected to the main input of the Nth decoder layer Decoder_LayerN at the end. The decoder utilizes the aforementioned hierarchical structure and object query mechanism to achieve end-to-end defect target prediction, completely eliminating cumbersome post-processing steps such as traditional nonmaximum suppression, and solving the problem of overly bloated models. This significantly improves the real-time inference efficiency of blade inspection on edge computing devices with extremely limited computing resources. The output of the Nth decoder layer, Decoder_LayerN, is connected to the prediction result output layer and finally outputs the surface defect detection results of wind turbine blades suitable for edge computing devices.

[0015] The Defect Adaptive Multi-Scale Backbone Module (DAMS) is specifically as follows: It includes an initial downsampling layer, a multi-scale feature extraction layer, a defect-aware attention fusion layer, and a final downsampling layer.

[0016] The input image is fed into the input of the initial downsampling layer; the initial downsampling layer uses a 3×3 convolution with a stride of 2, batch normalization, and ReLU activation function for spatial dimensionality reduction, and its output is connected to the input of the four parallel branches of the multi-scale feature extraction layer. The four parallel branches are: the first branch (3×3 convolution, batch normalization, ReLU), the second branch (5×5 convolution, batch normalization, ReLU), the third branch (depth-separable convolution consisting of 3×3 channel-wise convolution and 1×1 pointwise convolution, followed by batch normalization and ReLU), and the fourth branch (adaptive average pooling, batch normalization, ReLU, bilinear interpolation upsampling, and 1×1 convolution); the number of output channels of each branch is configured to be divisible by the number of subsequent convolutional groups. The outputs of the four branches are connected to the input of the defect-aware attention fusion layer. This fusion layer concatenates the input features along the channel dimension, compresses the spatial dimension through global average pooling, and then generates channel attention weights by sequentially passing two 1×1 convolutions, ReLU, and Sigmoid activation functions. These weights are then multiplied with the concatenated features channel by channel to enhance the defect region. The output of the defect-aware attention fusion layer is connected to the input of the final downsampling layer; the final downsampling layer uses a 3×3 convolution with a stride of 2 for secondary spatial downsampling, and its output is the final output of the DAMS module.

[0017] The Edge-Guided Enhanced Attention Module (EGEA) is as follows: It includes a basic feature preprocessing layer, a dual-branch complementary feature extraction layer, a guided attention enhancement layer, and a stage fusion output layer.

[0018] The input features of EGEA are fed into the input of the basic feature preprocessing layer; this preprocessing layer uses convolution operations to perform nonlinear transformations on the input features, and its output is connected to the input of the dual-branch complementary feature extraction layer. The dual-branch complementary feature extraction layer includes: a Scharr edge branch: extracting gradients in the x and y directions using the Scharr operator, calculating the edge magnitude, enhancing it through a bottleneck structure, and outputting edge-guided features; a Gaussian smoothing branch: dynamically generating Gaussian kernels and combining them with grouped convolutions for spatial smoothing and noise suppression, enhancing it through a bottleneck structure, and outputting smoothing-guided features; and a staged feature switching mechanism: selecting edge-guided features in the shallow Ino_Block and smoothing-guided features in the deep Ino_Block. The selected guiding features are input together with the original input features of EGEA into the guiding attention enhancement layer; the guiding attention enhancement layer is executed as follows: First, calculate X ⊙ X_guide + X, then refine it locally using a 3×3 convolution to obtain X_refine. Global average pooling and one-dimensional convolution are then applied to X_refine to generate channel attention weights. After Sigmoid normalization, these weights are multiplied channel-by-channel with X_refine. Finally, they are added to the original input feature X and batch normalized to obtain the attention-enhanced feature. The attention-enhanced feature is input to the stage fusion output layer. This layer uses a multilayer perceptron composed of two 1×1 convolutions for feature transformation, and introduces DropPath regularization and residual connections to output the final EGEA feature.

[0019] The multi-branch, heavily parameterized, lightweight module MBRL is specifically as follows: It includes hierarchical multi-scale convolution branches, anisotropic cross-convolution branches, and conditional residual connections.

[0020] The input features of MBRL are simultaneously passed to the following four parallel branches: a 5×5 large kernel convolution branch, a 3×3 standard convolution branch, a 1×1 pointwise convolution branch, and an anisotropic cross branch composed of 3×1 and 1×3 convolutions. The outputs of the above four branches are concatenated along the channel dimension and then input to a 1×1 convolutional fusion layer for dynamic reweighting fusion. The fusion result is then processed by batch normalization and SiLU activation function. When the input and output spatial dimensions and number of channels match, the original input of MBRL is connected to the addition module through an identity mapping; otherwise, it is connected to the addition module after being adjusted by a 1×1 convolution; the output of the 1×1 convolution fusion layer is added to the other input of the addition module to obtain the final output of the MBRL module.

[0021] The lightweight spatial channel pixel co-fusion module LSCP is as follows: It includes spatial attention components, channel attention components, pixel-level attention components, and a progressive fusion paradigm. Given two features X and Y, X and Y are first added element-wise to obtain the initial fused feature I. Based on the initial fusion feature I, spatial attention, channel attention, and pixel attention are constructed simultaneously: Spatial attention: I is subjected to min pooling, max pooling, and global average pooling respectively. The results are concatenated along the channel dimension and then subjected to 3×3 reflection-filled convolution to generate a spatial weight distribution map; Channel attention: I is subjected to global average pooling, and then passed through 1×1 convolution, ReLU, 1×1 convolution, and Sigmoid activation function in sequence to generate channel attention weights; The spatial attention output and channel attention output are used together as priors, and after tensor reshaping and group convolution, fine-grained pixel-level weights A_p are generated; The pixel-level weights are used for weighted fusion according to the following formula: I_att = I + A_p ⊙ X + (1-A_p) ⊙ Y, and the attention-guided fusion feature is output; Finally, I_att is channel-calibrated using a 1×1 convolution to output the final fused features.

[0022] Compared with the prior art, the present invention has the following technical effects: 1) This invention uses drones to capture high-resolution images of blades from multiple angles and distances, and constructs a dedicated dataset containing three typical defects: stains, damage, and cracks. Through enhancement operations such as random cropping, color correction, brightness adjustment, and geometric flipping, it simulates changes in outdoor lighting and differences in shooting perspective, expands the diversity of samples and spatial distribution coverage, and provides a more robust feature learning foundation for the model. Unlike the unchanging factory production line dataset, it significantly improves the model's generalization ability in complex natural scenes.

[0023] 2) This invention constructs an end-to-end detection network DEML-DEIM based on the DEIM-n paradigm, completely eliminating manual anchor frame design and NMS post-processing, simplifying the detection process and reducing the complexity of engineering deployment; the total number of model parameters is only 3.99M, the computation amount is 9.08GFLOPs, and the inference speed reaches 47.96FPS. While comprehensively surpassing mainstream models such as YOLOv5s / v8s / v11s / v13s, RT-DETR-r18 / r34 / r50 in terms of accuracy (P@90.7%, mAP@0.586.7%), it still maintains extremely low resource consumption and can be directly deployed on edge computing devices to achieve real-time inspection.

[0024] 3) This invention proposes a Defect Adaptive Multi-Scale Backbone Module (DAMS) to replace the original Stem layer. It covers multiple receptive fields from local micro-textures to global structures through four parallel branches: 3×3 convolution, 5×5 convolution, depthwise separable convolution, and global average pooling. Channel attention is introduced to recalibrate the spliced ​​features, forcing the network to retain the spatial details of small defects such as microcracks in the early downsampling stage. This effectively solves the problems of fine-grained feature submersion and incomplete multi-scale defect extraction caused by the traditional fixed receptive field.

[0025] 4) This invention proposes an Edge-Guided Enhanced Attention Module (EGEA) to replace the native SE attention. It innovatively constructs a dual-branch structure of Scharr edge extraction and Gaussian smoothing: the Scharr branch captures high-frequency details of crack and damage edges, while the smoothing branch filters out surface texture and lighting noise. The dual-branch features are used as priors to guide the lightweight one-dimensional convolutional channel attention for adaptive enhancement, and a staged switching strategy of shallow edge bias and deep smoothing bias is adopted to enable the network to focus on real defect areas, significantly enhancing the ability to identify low-contrast stains and thin cracks, and solving the deficiency of conventional attention lacking morphological guidance.

[0026] 5) This invention improves multi-scale directional feature extraction with zero inference cost through the MBRL module. The invention designs a multi-branch reparameterized lightweight convolution module (MBRL). During the training phase, it adopts a multi-branch structure with 5×5 large kernel convolution, 3×3 standard convolution, 1×1 pointwise convolution, and directional decoupling cross convolution to comprehensively capture features across scales and anisotropic directional features of crack-like defects. During inference, the multi-branch is folded into a single convolutional flow through structural reparameterization without increasing any computational overhead. At the same time, conditional residual connections are used to ensure stable gradient propagation. While maintaining the extremely lightweight model, the feature representation integrity of linear cracks and macroscopic damage is greatly enhanced.

[0027] 6) This invention achieves multi-dimensional attention collaborative fusion through the LSCP module. This invention proposes a lightweight spatial channel pixel collaborative fusion module (LSCP), which locates defect regions through spatial attention, recalibrates semantic responses through channel attention, and performs fine-grained weight mapping using grouped convolutions for pixel-level attention. It dynamically balances shallow details and deep semantics according to the progressive paradigm of initial fusion → attention guidance → dimensional calibration, which effectively solves the problem of missed detection of small defects (such as stains and fine lines) in feature pyramid networks due to the semantic information being submerged by deep features. At the same time, it significantly improves the ability to preserve the edges of long strip defects, with extremely low computational increment, ensuring the high efficiency of feature fusion. 7) This invention boasts superior overall performance and outstanding cross-scenario generalization ability. On a self-built dataset of 5,495 wind turbine blade images, compared to the baseline DEIM, the method of this invention achieves an 8.2 percentage point improvement in accuracy and a 2.2 percentage point improvement in mAP@0.5 with only 0.28 M parameters and 2.18 GFLOPs, outperforming all compared YOLO and RT-DETR series models. In zero-shot generalization experiments without any fine-tuning, the accuracy reaches 85.9% and the mAP reaches 81.1%, surpassing the existing best model for the same task. This demonstrates its excellent adaptability to different data distributions, shooting equipment, and environments, providing an efficient, reliable, and lightweight solution for the industrialization of wind turbine blade defect detection. Attached Figure Description

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a framework diagram of the lightweight edge-guided multi-scale target detection network based on the DEIM paradigm proposed in step 3 of this invention. Figure 3 yes Figure 2 Framework structure diagram of the DAMS module in China; Figure 4 yes Figure 2Framework structure diagram of the EGEA module in China; Figure 5 yes Figure 2 Framework structure diagram of the MBRL module in China; Figure 6 yes Figure 2 Framework structure diagram of the LSCP module; Figure 7 This is a qualitative visualization diagram of the detection results and feature responses in an embodiment of the present invention. Detailed Implementation

[0029] like Figure 1 As shown, a method for detecting surface defects on wind turbine blades suitable for edge computing devices includes the following steps: Step 1: Construct a surface defect image dataset for the physical scenario of wind turbine blades and perform data processing to form the original dataset; Step 2: Perform image enhancement and spatial mapping on the original dataset to simulate a real complex outdoor industrial environment, expand the sample size to an effective training scale, and divide it into training set, validation set and test set according to a preset ratio; Step 3: Construct a lightweight edge-guided multi-scale target detection network DEML-DEIM based on the DEIM paradigm; Step 4: Input the training set obtained in Step 2 into the DEML-DEIM network constructed in Step 3 for iterative training; Step 5: Deploy the trained model to an edge computing device to detect surface defects on wind turbine blades.

[0030] In step 1, the specific steps are as follows: Acquire a dataset of wind turbine blade surface images taken by drones from multiple angles and distances, containing real outdoor physical scenes; clean the image data to meet the actual application requirements of defect detection in complex industrial environments, remove neglected areas that are prone to background physical noise interference, such as severely non-uniform lighting or non-defect texture areas, and retain only defect labels with clear physical geometric contours, including three types of defects: stains, damage, and cracks, to adapt to high-precision wind turbine blade defect detection tasks.

[0031] In step 2, the specific steps are as follows: The cleaned wind turbine blade image is converted into a digital matrix format, and image enhancement and preprocessing operations are performed to simulate the interference of a real outdoor complex industrial environment. The image enhancement and preprocessing operations include: random cropping, segmentation, and size normalization operations to simulate the drastic scale changes caused by different physical distances and defect morphologies during UAV inspection; color correction and dynamic brightness adjustment to simulate the interference of natural lighting changes; and expanding the spatial distribution characteristics of defects through geometric transformations such as random flipping and rotation. After completing the above spatial and color feature transformations, the model is divided into training set, validation set, and test set according to a preset ratio to improve the robustness of the model in open physical scenarios.

[0032] In step 4, the specific steps are as follows: The wind turbine blade image matrix in the training set is input into the lightweight edge-guided multi-scale object detection network DEML-DEIM constructed in step 3 for iterative training; during the training process, the optimal matching relationship between the bounding box set predicted by the model and the real defect annotation box set is calculated based on the bipartite graph matching strategy; based on the optimal matching result, the classification loss and the bounding box regression loss reflecting the degree of overlap in physical space are jointly calculated; the AdamW optimizer is used, the initial learning rate and weight decay are configured, and the network weights are continuously updated through the backpropagation algorithm until the loss function converges, thus obtaining the trained lightweight real-time defect detection model.

[0033] In step 5, the specific steps are as follows: Deploy the trained lightweight defect detection model to an edge computing device; acquire real-time images of the surface of the wind turbine blade to be detected and input them into the model; the model directly outputs the category confidence of each defect target in the image and the bounding box coordinates that accurately represent the actual physical spatial location of the defect through a single forward inference; finally, filter the prediction results according to the preset confidence threshold and output the final defect location and classification information to support subsequent downstream actions such as automated alarms and maintenance decisions.

[0034] In step 3, the constructed lightweight edge-guided multi-scale target detection network DEML-DEIM based on the DEIM paradigm includes a backbone network, an encoder network connected to the output of the backbone network, and a decoder network connected to the output of the encoder network; the backbone network is specifically as follows: The input image is fed into the Defect Adaptive Multi-Scale Backbone Module (DAMS). DAMS preserves fine-grained features during early downsampling using parallel receptive fields, addressing the technical issues of low recognition accuracy and susceptibility to false positives and missed detections when facing defects with extreme scale variations. The output of DAMS is connected to the input of the first-stage module Ino_Stage1. The output of Ino_Stage1 is connected to the input of the second-stage module Ino_Stage2. The output of Ino_Stage2 is connected to the input of the third-stage module Ino_Stage3 and the first convolutional module in the encoder. The output of Ino_Stage3 is connected to the input of the fourth-stage module Ino_Stage4 and the second convolutional module in the encoder. The output of Ino_Stage4 is connected to the input of the third convolutional module in the encoder.

[0035] The internal connection structure of the first stage module Ino_Stage1, the second stage module Ino_Stage2, the third stage module Ino_Stage3, and the fourth stage module Ino_Stage4 is as follows: Each stage module includes a pre-convolutional module and multiple cascaded internal blocks (Ino_Block). Within each Ino_Block, input features are passed to one input of the multi-branch reparameterized lightweight module (MBRL) and the addition module (Add). Multiple parallel branches of the MBRL are fused via the concatenation module (Concat) and then connected to the input of the edge-guided enhanced attention module (EGEA). The EGEA uses edge prior knowledge to guide the network to focus on high-frequency defect regions, and its output is connected to the other input of the addition module (Add). The output of the Add module (Add) serves as the output of this Ino_Block, thereby enhancing multi-scale and directional feature extraction capabilities while maintaining training stability and addressing the technical problem of overly bloated models that are difficult to deploy efficiently in real-time on edge devices with extremely limited computing resources.

[0036] In the encoder, the output of the third convolutional module is connected to the input of the attention interaction feature module AIFI. The attention interaction feature module AIFI enhances the high-level semantic feature representation within a single scale through a self-attention mechanism, thereby improving the network's ability to capture features of the irregular surface defect edges of wind turbine blades. The output of the attention interaction feature module AIFI is connected to the input of the fourth convolutional module in the encoder. The output of the fourth convolutional module in the encoder is connected to the input of the first upsampling module. The outputs of the first upsampling module and the second convolutional module in the encoder are connected to the input of the first lightweight spatial channel pixel collaborative fusion module LSCP. Meanwhile, the output of the fourth convolution module in the encoder is connected to the input of the second lightweight spatial channel pixel collaborative fusion module LSCP, and the output of the second lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the first high-efficiency layer aggregation network module ELAN. The output of the first upsampling module is connected to the input of the second lightweight spatial channel pixel co-fusion module LSCP. The output of the second lightweight spatial channel pixel co-fusion module LSCP is connected to the input of the second high-efficiency layer aggregation network module ELAN. The output of the second high-efficiency layer aggregation network module ELAN is connected to the input of the fifth convolution module in the encoder. The output of the fifth convolution module is connected to the input of the second upsampling module. The output of the second upsampling module and the output of the first convolution module in the encoder are connected to the input of the third lightweight spatial channel pixel co-fusion module LSCP. The output of the third lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the third high-efficiency layer aggregation network module ELAN. The output of the third high-efficiency layer aggregation network module ELAN is connected to the input of the first downsampling module. The outputs of the first downsampling module and the fourth convolution module in the encoder are connected to the input of the fourth lightweight spatial channel pixel collaborative fusion module LSCP. The output of the fourth lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the fourth high-efficiency layer aggregation network module ELAN. The output of the fourth high-efficiency layer aggregation network module ELAN is connected to the input of the second downsampling module. The output of the second downsampling module is connected to the input of the second lightweight spatial channel pixel collaborative fusion module LSCP. The encoder is equipped with multiple lightweight spatial channel pixel collaborative fusion modules (LSCP) to promote efficient multi-scale semantic alignment and fusion of shallow and deep features. This effectively solves the technical problems of extreme scale changes in wind turbine blade defects and the fact that tiny cracks are easily submerged by complex backgrounds, leading to missed detections. This ensures the detection accuracy of edge devices with extremely low parameter levels. The outputs of the first high-efficiency layer aggregation network module ELAN, the second high-efficiency layer aggregation network module ELAN, and the third high-efficiency layer aggregation network module ELAN are connected to the input of the flatten module in the decoder.

[0037] In the decoder, the output of the Flatten module is connected to the input of the initial feature sequence layer; the output of the initial feature sequence layer is divided into two flow branches: The first main branch is connected to the main input of the first decoder layer, Decoder_Layer1; The second branch serves as the object query branch, and after being led out, it is connected to the query input of the first decoder layer Decoder_Layer1, the query input of the second decoder layer Decoder_Layer2, the query input of the third decoder layer Decoder_Layer3, and so on, until it is finally connected to the query input of the Nth decoder layer Decoder_LayerN. The output of the first decoder layer Decoder_Layer1 is connected to the main input of the second decoder layer Decoder_Layer2, the output of the second decoder layer Decoder_Layer2 is connected to the main input of the third decoder layer Decoder_Layer3, and so on, until the output of the (N-1)th decoder layer Decoder_LayerN-1 is connected to the main input of the Nth decoder layer Decoder_LayerN at the end. The decoder utilizes the aforementioned hierarchical structure and object query mechanism to achieve end-to-end defect target prediction, completely eliminating cumbersome post-processing steps such as traditional nonmaximum suppression, and solving the problem of overly bloated models. This significantly improves the real-time inference efficiency of blade inspection on edge computing devices with extremely limited computing resources. The output of the Nth decoder layer, Decoder_LayerN, is connected to the prediction result output layer and finally outputs the surface defect detection results of wind turbine blades suitable for edge computing devices.

[0038] like Figure 3 As shown, the Defect Adaptive Multi-Scale Backbone Module (DAMS) is specifically as follows: It includes an initial downsampling layer, a multi-scale feature extraction layer, a defect-aware attention fusion layer, and a final downsampling layer.

[0039] The input image is fed into the input of the initial downsampling layer; the initial downsampling layer uses a 3×3 convolution with a stride of 2, batch normalization, and ReLU activation function for spatial dimensionality reduction, and its output is connected to the input of the four parallel branches of the multi-scale feature extraction layer. The four parallel branches are: the first branch (3×3 convolution, batch normalization, ReLU), the second branch (5×5 convolution, batch normalization, ReLU), the third branch (depth-separable convolution consisting of 3×3 channel-wise convolution and 1×1 pointwise convolution, followed by batch normalization and ReLU), and the fourth branch (adaptive average pooling, batch normalization, ReLU, bilinear interpolation upsampling, and 1×1 convolution); the number of output channels of each branch is configured to be divisible by the number of subsequent convolutional groups. The outputs of the four branches are connected to the input of the defect-aware attention fusion layer. This fusion layer concatenates the input features along the channel dimension, compresses the spatial dimension through global average pooling, and then generates channel attention weights by sequentially passing two 1×1 convolutions, ReLU, and Sigmoid activation functions. These weights are then multiplied with the concatenated features channel by channel to enhance the defect region. The output of the defect-aware attention fusion layer is connected to the input of the final downsampling layer; the final downsampling layer uses a 3×3 convolution with a stride of 2 for secondary spatial downsampling, and its output is the final output of the DAMS module.

[0040] like Figure 4 As shown, the Edge-Guided Enhanced Attention Module (EGEA) is specifically as follows: It includes a basic feature preprocessing layer, a dual-branch complementary feature extraction layer, a guided attention enhancement layer, and a stage fusion output layer.

[0041] The input features of EGEA are fed into the input of the basic feature preprocessing layer; this preprocessing layer uses convolution operations to perform nonlinear transformations on the input features, and its output is connected to the input of the dual-branch complementary feature extraction layer. The dual-branch complementary feature extraction layer includes: a Scharr edge branch: extracting gradients in the x and y directions using the Scharr operator, calculating the edge magnitude, enhancing it through a bottleneck structure, and outputting edge-guided features; a Gaussian smoothing branch: dynamically generating Gaussian kernels and combining them with grouped convolutions for spatial smoothing and noise suppression, enhancing it through a bottleneck structure, and outputting smoothing-guided features; and a staged feature switching mechanism: selecting edge-guided features in the shallow Ino_Block and smoothing-guided features in the deep Ino_Block. The selected guiding features are input together with the original input features of EGEA into the guiding attention enhancement layer; the guiding attention enhancement layer is executed as follows: First, calculate X ⊙ X_guide + X, then refine it locally using a 3×3 convolution to obtain X_refine. Global average pooling and one-dimensional convolution are then applied to X_refine to generate channel attention weights. After Sigmoid normalization, these weights are multiplied channel-by-channel with X_refine. Finally, they are added to the original input feature X and batch normalized to obtain the attention-enhanced feature. The attention-enhanced feature is input to the stage fusion output layer. This layer uses a multilayer perceptron composed of two 1×1 convolutions for feature transformation, and introduces DropPath regularization and residual connections to output the final EGEA feature.

[0042] like Figure 5 As shown, the multi-branch, heavily parameterized lightweight module MBRL is specifically as follows: It includes hierarchical multi-scale convolution branches, anisotropic cross-convolution branches, and conditional residual connections.

[0043] The input features of MBRL are simultaneously passed to the following four parallel branches: a 5×5 large kernel convolution branch, a 3×3 standard convolution branch, a 1×1 pointwise convolution branch, and an anisotropic cross branch composed of 3×1 and 1×3 convolutions. The outputs of the above four branches are concatenated along the channel dimension and then input to a 1×1 convolutional fusion layer for dynamic reweighting fusion. The fusion result is then processed by batch normalization and SiLU activation function. When the input and output spatial dimensions and number of channels match, the original input of MBRL is connected to the addition module through an identity mapping; otherwise, it is connected to the addition module after being adjusted by a 1×1 convolution; the output of the 1×1 convolution fusion layer is added to the other input of the addition module to obtain the final output of the MBRL module.

[0044] like Figure 6 As shown, the Lightweight Spatial Channel Pixel Collaborative Fusion Module (LSCP) is specifically as follows: It includes spatial attention components, channel attention components, pixel-level attention components, and a progressive fusion paradigm. Given two features X and Y, X and Y are first added element-wise to obtain the initial fused feature I. Based on the initial fusion feature I, spatial attention, channel attention, and pixel attention are constructed simultaneously: Spatial attention: I is subjected to min pooling, max pooling, and global average pooling respectively. The results are concatenated along the channel dimension and then subjected to 3×3 reflection-filled convolution to generate a spatial weight distribution map; Channel attention: I is subjected to global average pooling, and then passed through 1×1 convolution, ReLU, 1×1 convolution, and Sigmoid activation function in sequence to generate channel attention weights; The spatial attention output and channel attention output are used together as priors, and after tensor reshaping and group convolution, fine-grained pixel-level weights A_p are generated; The pixel-level weights are used for weighted fusion according to the following formula: I_att = I + A_p ⊙ X + (1-A_p) ⊙ Y, and the attention-guided fusion feature is output; Finally, I_att is channel-calibrated using a 1×1 convolution to output the final fused features.

[0045] Example: like Figure 2 As shown, in order to address the challenges of wind turbine blade defect detection, this invention proposes a defect-oriented multi-branch lightweight DEIM model (DEML-DEIM). The proposed model retains the end-to-end detection framework of DEIM, while redesigning the task-related components in the backbone and neck networks to improve the feature representation capabilities for weak defects, slender structures, and cross-scale details in complex contexts.

[0046] Specifically, the original StemBlock is replaced with the DAMS module to enhance early multi-scale feature extraction before drastic downsampling. Within the backbone network, the EGEA module is introduced to improve morphology-aware feature representation. Simultaneously, the MBRL module strengthens orientation and multi-scale modeling capabilities through a multi-branch design with inference-time reparameterization. The backbone network outputs three levels of features {S3, S4, S5}, maintaining the same dimensionality as the original DEIM architecture, thus ensuring seamless integration with the hybrid encoder and decoder. In the neck network, an LSCP module is designed to optimize cross-scale feature fusion by jointly evaluating spatial, channel, and pixel-level feature importance. The fused features are then processed by the encoder and decoder to complete end-to-end object classification and bounding box prediction.

[0047] Compared to the baseline DEIM model, DEML-DEIM primarily optimizes its three core limitations: insufficient early feature extraction, weak feature representation learning within the backbone network, and poor multi-scale feature fusion. These improvements aim to enhance the model's ability to represent small, weak, and anisotropic leaf defects while maintaining efficient computational performance.

[0048] like Figure 3 As shown, in order to alleviate the limitations of single-path early feature extraction in HGNetv2, this invention introduces a dual-branch adaptive multi-scale backbone (DAMS) module to enhance the early feature representation capability of wind turbine blade defect detection. Given input features DAMS first performs spatial dimensionality reduction using stacked convolutions with a stride of 2:

[0049] Then, multiple parallel transformations are applied to capture complementary feature patterns:

[0050] in, This represents convolutional operations with different receptive fields, and a global context branch implemented by global average pooling followed by upsampling. These branches can extract local details, broader contextual information, and global statistical features in parallel.

[0051] The resulting features are then spliced ​​and merged:

[0052] Subsequently, a channel attention mechanism is used to adaptively recalibrate the feature responses:

[0053]

[0054] Finally, a second downsampling operation is applied to generate the output features:

[0055] Overall, DAMS enhances the diversity of early features while maintaining computational efficiency, providing richer feature representations for subsequent defect detection.

[0056] like Figure 4As shown, the original squeeze and excitation (SE) mechanism in HGNetv2 only performs channel-level recalibration without explicitly modeling spatial structural cues, which limits its ability to capture subtle defect patterns in complex contexts. To address this issue, this invention proposes an Edge-Guided Enhanced Attention (EGEA) module, which integrates structural priors into the feature refinement process through bi-branch guidance and stage-aware attention enhancement.

[0057] Given an input feature map First, a lightweight preprocessing block is applied: (7); Then, two complementary guiding branches are constructed. The first branch uses Scharr gradients to extract edge-aware features: (8); (9); Where Xx and Xy are the horizontal and vertical gradient responses, respectively. The second branch generates a smooth structural prior through Gaussian-guided convolution: (10); To match feature representation requirements across different network depths, guide features are selected in a stage-aware manner: (11); The selected guiding features are fused with the original features to obtain a refined feature representation: (12); Subsequently, the channel attention weights are generated using the following formula: (13); And the enhanced features are calculated: (14); Finally, the output of EGEA is obtained using the following formula: (15); By jointly utilizing edge-sensitive guided features, noise-suppressing structural priors, and adaptive channel recalibration, the EGEA module enhances defect-related responses and significantly improves feature representation capabilities for fine cracks and weak defect regions.

[0058] like Figure 5As shown, to enhance the HGNetv2 backbone network's ability to model blade defects with large scale variations and arbitrary orientations, this invention introduces a Multi-Branch Reparameterized Lightweight Convolution (MBRL) module. MBRL employs parallel 1×1, 3×3, and 5×5 convolutions to capture complementary receptive fields, while an additional cross branch consisting of 3×1 and 1×3 convolutions enhances sensitivity to elongated crack-like structures. The outputs of each branch are concatenated and fused using 1×1 convolutions. (16); To improve the stability of the optimization and preserve the original feature information, a residual shortcut was further introduced: (17); When the channel dimensions do not match, a 1×1 projection is used on the residual path.

[0059] A key advantage of MBRL is that its multi-branch training structure can be transformed into an equivalent single convolution during the inference phase through structural re-parameterization. Therefore, this module improves multi-scale and orientation-sensitive feature representations during training while maintaining inference efficiency comparable to standard convolutional layers after deployment. Thus, for wind turbine blade defect detection, MBRL offers a good balance between feature representation power and computational efficiency.

[0060] like Figure 6 As shown, to improve multi-scale feature fusion in hybrid encoders, this invention introduces a lightweight spatial-channel-pixel collaborative fusion (LSCP) module. This module jointly models the importance of spatial, channel, and pixel levels to achieve adaptive feature aggregation. Given two input features X and Y, LSCP first performs an initial fusion: (18); Based on the initial fusion feature I, spatial attention and channel attention are used to capture complementary, position-sensitive, and semantically important information, while an additional pixel attention branch further refines local structural details. These three attention cues are synergistically combined to generate a fine-grained fusion map, and the refined feature calculation is as follows: (19); Where Ap represents the pixel-level attention map. It is element-wise multiplication.

[0061] Finally, a 1×1 convolution is applied for channel calibration: (20); By progressively integrating spatial saliency, channel dependence, and pixel-level refinement, LSCP achieves more discriminative and robust feature fusion for small-scale, low-contrast, and elongated defects, while introducing only minimal computational overhead.

[0062] Experimental results and analysis: 1) Experimental setup: To address the lack of publicly available datasets for wind turbine blade defect detection, this invention constructs a proprietary dataset comprising 559 high-resolution images captured by unmanned aerial vehicles (UAVs) at various angles and distances. This dataset covers three defect categories: dirt, damage, and cracks. The original images were initially divided into a training set (70%), a validation set (15%), and a test set (15%) to prevent data leakage. Subsequently, data augmentation techniques (including random cropping, flipping, and rotation) were applied only to the training set, expanding the dataset to 5,495 samples.

[0063] All experiments were conducted on a high-performance workstation equipped with an Intel Core i7-8700 CPU, an NVIDIA RTX 4070 Super GPU (12GB VRAM), and a Windows 10 operating system. The deep learning framework used was PyTorch 2.5.1 and CUDA 11.8. The input image size was resized to 640×640 pixels, and the model was trained for 300 epochs using the AdamW optimizer (learning rate 0.0008, weight decay 0.0001). Due to GPU memory limitations, the batch size was set to 4.

[0064] This invention uses precision (P), recall (R), and mean precision (mAP@0.5) at an IoU threshold of 0.5 to evaluate detection performance. Efficiency is measured by GFLOPs and the number of parameters. To improve the reliability of the reported results, each experiment was repeated three times under different random seeds, and the average performance is reported in Tables 1-3. The observed differences were small and did not affect the overall ranking of the comparative methods.

[0065] 2) Ablation experiment: The effectiveness of key modules was evaluated using a wind turbine blade defect dataset, focusing on the DAMS backbone module, EGEA attention module, MBRL multi-scale convolution module, and LSCP encoder feature fusion module. Each module was tested individually and in combination with the baseline DEIM model. The ablation experiment results are summarized in Table 1:

[0066] Individual modules all showed significant improvements in detection performance: DAMS improved precision by 7.6%, EGEA by 6.5%, LSCP by 6.2%, and MBRL by 5.2%. The complete model integrating all modules (DEML-DEIM) achieved the best performance: precision of 90.7%, recall of 78.6%, and mAP of 86.7%, while the number of parameters (from 3.71M to 3.99M) and computational cost (from 6.90G to 9.08G) increased only slightly.

[0067] These results demonstrate that each module makes a significant contribution to improving detection accuracy while maintaining computational efficiency.

[0068] 3) Comparative experiment: This invention compares the proposed DEML-DEIM model with several state-of-the-art detectors, including CNN-based models (SSD, Faster R-CNN, YOLOv5s / v8s / v11s / v13s) and Transformer-based models (RT-DETR-r18, r34, r50).

[0069]

[0070] As shown in Table 2, DEML-DEIM outperforms all the comparison methods, achieving a precision of 90.7% and a mAP of 86.7%, which is significantly higher than the baseline DEIM (precision of 82.5%, mAP of 84.5%) and YOLOv13s (precision of 86.4%, mAP of 84.1%).

[0071] Although DEML-DEIM's recall (78.6%) is slightly lower than some models (e.g., YOLOv11s at 84.8%), it demonstrates a good precision-recall trade-off, which helps reduce false positives in industrial applications.

[0072] 4) Generalization experiment: This invention evaluates the generalization ability of DEML-DEIM on a publicly available wind turbine blade defect dataset without fine-tuning or domain adaptation; the generalization results are presented in Table 3:

[0073] The model achieved a precision of 85.9% and a median AP of 81.1%, outperforming the baseline DEIM (79.5% mAP) and other detectors. Despite a lower recall (74.6%), DEML-DEIM maintained a good precision-recall balance, particularly excelling in the pollution category (76.7% AP). These results demonstrate that DEML-DEIM generalizes well to unseen datasets while maintaining computational efficiency.

[0074] 5) Visual analysis: This invention provides a qualitative analysis of the detection results for dirt, damage, and crack defects, and compares DEIM and DEML-DEIM. The qualitative visualization results are as follows: Figure 7 As shown, DEML-DEIM exhibits stronger robustness to background interference, better detection of fine cracks, enhanced sensitivity to low-contrast stains, and more accurate localization of structural damage. The improved model concentrates feature responses in defect areas, effectively suppressing irrelevant background features. These improvements demonstrate the significant advantages of DEML-DEIM in practical wind turbine blade inspection.

[0075] 6) Conclusion: This invention proposes DEML-DEIM, a lightweight defect detection model for wind turbine blades that improves multi-scale feature representation while maintaining computational efficiency. By integrating DAMS, MBRL, EGEA, and LSCP modules, the model enhances feature extraction and alignment capabilities, thereby achieving more accurate defect localization and improving robustness against background interference. Experimental results show that compared to the baseline DEIM, DEML-DEIM achieves an 8.2% improvement in precision and a 2.2% improvement in mAP@0.5, while increasing the number of parameters and computation by only 0.28M and 2.18 GFLOPs, respectively, while maintaining a real-time detection performance of 47.96 FPS. Compared to the RT-DETR and YOLO series models, it achieves a better accuracy-efficiency tradeoff (3.99M parameters, 9.08 GFLOPs computation). Furthermore, zero-shot experiments achieved 85.9% precision and 81.1% mAP, demonstrating its strong generalization ability across datasets. The model still has certain limitations under extreme weather conditions (such as dense fog and heavy rain). Future work will focus on: (1) multimodal fusion (infrared + visible light) to further improve robustness; and (2) semi-supervised / unsupervised learning to reduce dependence on data labeling and enhance the generalization ability of the model.

Claims

1. A method for detecting surface defects on wind turbine blades suitable for edge computing devices, characterized in that, Includes the following steps: Step 1: Construct a surface defect image dataset for the physical scenario of wind turbine blades and perform data processing to form the original dataset; Step 2: Perform image enhancement and spatial mapping on the original dataset to simulate a real complex outdoor industrial environment, expand the sample size to an effective training scale, and divide it into training set, validation set and test set according to a preset ratio; Step 3: Construct a lightweight edge-guided multi-scale target detection network DEML-DEIM based on the DEIM paradigm; Step 4: Input the training set obtained in Step 2 into the DEML-DEIM network constructed in Step 3 for iterative training; Step 5: Deploy the trained model to an edge computing device to detect surface defects on wind turbine blades.

2. The method according to claim 1, characterized in that, In step 1, the specific steps are as follows: Acquire a dataset of wind turbine blade surface images taken by drones from multiple angles and distances, containing real outdoor physical scenes; clean the image data to meet the actual application requirements of defect detection in complex industrial environments, remove neglected areas that are prone to background physical noise interference, such as severely non-uniform lighting or non-defect texture areas, and retain only defect labels with clear physical geometric contours, including three types of defects: stains, damage, and cracks, to adapt to high-precision wind turbine blade defect detection tasks.

3. The method according to claim 1, characterized in that, In step 2, the specific steps are as follows: the cleaned wind turbine blade image is converted into a digital matrix format, and image enhancement and preprocessing operations are performed to match the interference of a real outdoor complex industrial environment. Image enhancement and preprocessing operations include: random cropping, segmentation, and size normalization to simulate the drastic scale changes caused by different physical distances and defect morphologies during UAV inspection; color correction and dynamic brightness adjustment to simulate the interference of natural lighting changes; and expanding the spatial distribution characteristics of defects through geometric transformations such as random flipping and rotation. After completing the above spatial and color feature transformations, the model is divided into training, validation, and test sets according to a preset ratio to improve its robustness in open physical scenarios.

4. The method according to claim 1, characterized in that, In step 4, the specific steps are as follows: Input the wind turbine blade image matrix from the training set into the lightweight edge-guided multi-scale target detection network DEML-DEIM constructed in step 3 for iterative training; During training, the optimal matching relationship between the set of bounding boxes predicted by the model and the set of real defect annotation boxes is calculated based on the bipartite graph matching strategy. Based on the optimal matching result, the classification loss and the bounding box regression loss reflecting the degree of overlap in physical space are jointly calculated. The AdamW optimizer is used, the initial learning rate and weight decay are configured, and the network weights are continuously updated through the backpropagation algorithm until the loss function converges, thus obtaining a lightweight real-time defect detection model that has been trained.

5. The method according to claim 1, characterized in that, In step 5, the specific steps are as follows: Deploy the trained lightweight defect detection model to an edge computing device; acquire real-time images of the surface of the wind turbine blade to be detected and input them into the model; the model directly outputs the category confidence of each defect target in the image and the bounding box coordinates that accurately represent the actual physical spatial location of the defect through a single forward inference; finally, filter the prediction results according to the preset confidence threshold and output the final defect location and classification information to support subsequent downstream actions such as automated alarms and maintenance decisions.

6. The method according to claim 1, characterized in that, In step 3, the constructed lightweight edge-guided multi-scale target detection network DEML-DEIM based on the DEIM paradigm includes a backbone network, an encoder network connected to the output of the backbone network, and a decoder network connected to the output of the encoder network; the backbone network is specifically as follows: The input image is fed into the Defect Adaptive Multi-Scale Backbone Module (DAMS). The output of DAMS is connected to the input of the first-stage module Ino_Stage1. The output of Ino_Stage1 is connected to the input of the second-stage module Ino_Stage2. The output of Ino_Stage2 is connected to the input of the third-stage module Ino_Stage3 and the input of the first convolution module in the encoder. The output of Ino_Stage3 is connected to the input of the fourth-stage module Ino_Stage4 and the input of the second convolution module in the encoder. The output of Ino_Stage4 is connected to the input of the third convolution module in the encoder.

7. The method according to claim 6, characterized in that, The internal connection structure of the first stage module Ino_Stage1, the second stage module Ino_Stage2, the third stage module Ino_Stage3, and the fourth stage module Ino_Stage4 is as follows: Each stage module includes a pre-convolution module and multiple concatenated internal blocks Ino_Block; In each internal block Ino_Block, the input features are passed to one input of the multi-branch reparameterized lightweight module MBRL and the addition module Add, respectively; the multiple parallel branches of the multi-branch reparameterized lightweight module MBRL are fused by the concatenation module Concat and then connected to the input of the edge-guided enhanced attention module EGEA. The edge-guided enhanced attention module EGEA uses edge prior knowledge to guide the network to focus on high-frequency defect regions, and its output is connected to another input of the addition module Add. The output of the Add module is used as the output of the internal block Ino_Block.

8. The method according to claim 6, characterized in that, In the encoder, specifically: The output of the third convolutional module in the encoder is connected to the input of the attention interaction feature module AIFI. The output of the attention interaction feature module AIFI is connected to the input of the fourth convolutional module in the encoder. The output of the fourth convolutional module in the encoder is connected to the input of the first upsampling module. The outputs of the first upsampling module and the second convolutional module in the encoder are connected to the input of the first lightweight spatial channel pixel collaborative fusion module LSCP. Meanwhile, the output of the fourth convolution module in the encoder is connected to the input of the second lightweight spatial channel pixel collaborative fusion module LSCP, and the output of the second lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the first high-efficiency layer aggregation network module ELAN. The output of the first upsampling module is connected to the input of the second lightweight spatial channel pixel co-fusion module LSCP. The output of the second lightweight spatial channel pixel co-fusion module LSCP is connected to the input of the second high-efficiency layer aggregation network module ELAN. The output of the second high-efficiency layer aggregation network module ELAN is connected to the input of the fifth convolution module in the encoder. The output of the fifth convolution module is connected to the input of the second upsampling module. The output of the second upsampling module and the output of the first convolution module in the encoder are connected to the input of the third lightweight spatial channel pixel co-fusion module LSCP. The output of the third lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the third high-efficiency layer aggregation network module ELAN. The output of the third high-efficiency layer aggregation network module ELAN is connected to the input of the first downsampling module. The outputs of the first downsampling module and the fourth convolution module in the encoder are connected to the input of the fourth lightweight spatial channel pixel collaborative fusion module LSCP. The output of the fourth lightweight spatial channel pixel collaborative fusion module LSCP is connected to the input of the fourth high-efficiency layer aggregation network module ELAN. The output of the fourth high-efficiency layer aggregation network module ELAN is connected to the input of the second downsampling module. The output of the second downsampling module is connected to the input of the second lightweight spatial channel pixel collaborative fusion module LSCP. The outputs of the first high-efficiency layer aggregation network module ELAN, the second high-efficiency layer aggregation network module ELAN, and the third high-efficiency layer aggregation network module ELAN are connected to the input of the flatten module in the decoder.

9. The method according to any one of claims 6 to 8, characterized in that, In the decoder, the output of the Flatten module is connected to the input of the initial feature sequence layer; the output of the initial feature sequence layer is divided into two flow branches: The first main branch is connected to the main input of the first decoder layer, Decoder_Layer1; The second branch serves as the object query branch, and after being led out, it is connected to the query input of the first decoder layer Decoder_Layer1, the query input of the second decoder layer Decoder_Layer2, the query input of the third decoder layer Decoder_Layer3, and so on, until it is finally connected to the query input of the Nth decoder layer Decoder_LayerN. The output of the first decoder layer Decoder_Layer1 is connected to the main input of the second decoder layer Decoder_Layer2, the output of the second decoder layer Decoder_Layer2 is connected to the main input of the third decoder layer Decoder_Layer3, and so on, until the output of the (N-1)th decoder layer Decoder_LayerN-1 is connected to the main input of the Nth decoder layer Decoder_LayerN at the end. The decoder utilizes the aforementioned hierarchical structure and object query mechanism to achieve end-to-end defect target prediction, completely eliminating the cumbersome post-processing steps of traditional nonmaximum suppression and solving the problem of overly bloated models. This significantly improves the real-time inference efficiency of blade inspection on edge computing devices with extremely limited computing resources. The output of the Nth decoder layer, Decoder_LayerN, is connected to the prediction result output layer and finally outputs the surface defect detection results of wind turbine blades suitable for edge computing devices.

10. The method according to claim 6, characterized in that, The Defect Adaptive Multi-Scale Backbone Module (DAMS) is specifically as follows: It includes an initial downsampling layer, a multi-scale feature extraction layer, a defect-aware attention fusion layer, and a final downsampling layer; The input image is fed into the input of the initial downsampling layer; the initial downsampling layer uses a 3×3 convolution with a stride of 2, batch normalization, and ReLU activation function for spatial dimensionality reduction, and its output is connected to the input of the four parallel branches of the multi-scale feature extraction layer. The four parallel branches are as follows: the first branch includes 3×3 convolution, batch normalization, and ReLU; the second branch includes 5×5 convolution, batch normalization, and ReLU; the third branch includes depthwise separable convolution consisting of 3×3 channel-wise convolution and 1×1 pointwise convolution, followed by batch normalization and ReLU; and the fourth branch includes adaptive average pooling, batch normalization, ReLU, bilinear interpolation upsampling, and 1×1 convolution. The number of output channels of each branch is configured to be divisible by the number of subsequent convolutional groups. The outputs of the four branches are connected to the input of the defect-aware attention fusion layer. This fusion layer concatenates the input features along the channel dimension, compresses the spatial dimension through global average pooling, and then generates channel attention weights by sequentially passing two 1×1 convolutions, ReLU, and Sigmoid activation functions. These weights are then multiplied with the concatenated features channel by channel to enhance the defect region. The output of the defect-aware attention fusion layer is connected to the input of the final downsampling layer; the final downsampling layer uses a 3×3 convolution with a stride of 2 for secondary spatial downsampling, and its output is the final output of the DAMS module.

Images