In the current landscape of renewable energy development, wind power and photovoltaic power are the two dominant forms. However, extreme weather conditions such as lightning, rain, snow, and hail frequently cause damage to wind turbine blades and solar panels. Traditional detection methods are often inefficient and limited in scope. To address these challenges, I propose an improved YOLOv11-based defect detection framework that simultaneously handles both wind turbine blade defects and solar panel defects. My approach integrates a Convolutional Block Attention Module (CBAM) and a Wavelet Transform Convolution (WTConv) into the YOLOv11 architecture, thereby enhancing multi-scale feature fusion and detection accuracy. I validate the proposed model using a self-constructed dataset comprising images of wind turbine blades and solar panels. The experimental results demonstrate that the improved model achieves a mean Average Precision (mAP) of 89.5%, which is 3.3% higher than the baseline YOLOv11 model. This work provides a unified and efficient solution for defect detection across two critical components of renewable energy infrastructure.
The rest of this paper is organized as follows. Section 1 introduces the improvements made to YOLOv11, including the CBAM mechanism and the WTConv module. Section 2 presents the experimental setup, dataset, evaluation metrics, and results. Section 3 concludes the work and discusses future directions.
1. Improvements to YOLOv11
1.1 Baseline YOLOv11 Model
YOLOv11 is a state-of-the-art object detection architecture that inherits the advantages of previous YOLO versions while introducing several key enhancements. The backbone network of YOLOv11 consists of CBS (Convolution + BatchNorm + SiLU), C3K2, SPPF (Spatial Pyramid Pooling – Fast), and C2PSA (Cross Stage Partial Pyramid Spatial Attention) modules. The CBS module efficiently extracts image features by combining convolution, batch normalization, and activation. The SPPF module aggregates multi-scale contextual information, which is particularly important for detecting defects of varying sizes—such as small cracks on wind turbine blades and large hot spots on solar panels. The C2PSA module further enhances spatial attention, enabling the model to focus on critical regions. The neck network of YOLOv11 replaces the traditional C2F module with C3K2, which offers more diverse convolutional kernel designs and better scale adaptability. The detection head adopts a dual-head structure (one-to-many and one-to-one) to provide rich supervision during training and efficient end-to-end inference without non-maximum suppression. The overall architecture is depicted conceptually in the original paper.
1.2 Convolutional Block Attention Module (CBAM)
CBAM is a lightweight attention mechanism that sequentially applies channel attention and spatial attention to refine feature maps. The channel attention module computes importance weights for each channel using both average pooling and max pooling, followed by a shared multi-layer perceptron. The spatial attention module identifies important spatial regions by applying a 7×7 convolution to the concatenated average-pooled and max-pooled features. The CBAM operation can be expressed as follows.
Channel attention output:
$$M_c(F) = \sigma(MLP(AvgPool(F)) + MLP(MaxPool(F)))$$
where $$F$$ is the input feature map, $$\sigma$$ denotes the sigmoid activation function, and $$MLP$$ is a multi-layer perceptron with a hidden layer.
Spatial attention output:
$$M_s(F) = \sigma(f^{7\times7}([AvgPool(F); MaxPool(F)]))$$
where $$f^{7\times7}$$ is a convolution with a 7×7 kernel. By integrating CBAM into YOLOv11, the model can adaptively emphasize important channels and spatial locations. For solar panel defect detection, CBAM helps focus on temperature-variant channels in thermal images; for wind turbine blade defect detection, it enhances edge and texture features associated with cracks or corrosion.
1.3 Wavelet Transform Convolution (WTConv)
WTConv is an innovative convolution layer that leverages the discrete wavelet transform (DWT) to perform multi-scale analysis without significantly increasing the number of parameters. The DWT decomposes an input signal into four frequency subbands: low-low (LL), low-high (LH), high-low (HL), and high-high (HH). In the context of image processing, this decomposition captures both approximate and detail information at different scales. The wavelet convolution operation can be defined as follows.
Given an input signal $$f(t)$$ and a wavelet basis function $$\psi_{a,b}(t)$$, where $$a$$ is the scale factor and $$b$$ is the translation factor, the wavelet transform is:
$$W(a,b) = \int_{-\infty}^{\infty} f(t) \psi_{a,b}(t) dt$$
The convolution operation with a wavelet kernel is:
$$y(t) = \int_{-\infty}^{\infty} f(\tau) \psi_{a,b}(t-\tau) d\tau$$
By replacing standard convolutions with WTConv in the C3K2 module, I construct a new module called C3K2WT, which is substituted into the YOLOv11 neck network. This modification enlarges the receptive field and enables the model to capture defects at multiple scales—critical for detecting both tiny cracks on wind turbine blades and large-area hot spots on solar panels.
2. Experimental Results and Analysis
2.1 Dataset Construction
To validate the proposed model, I constructed a dataset using a Matrice 350 RTK drone equipped with a Zenmuse H30 series camera. A total of 2,000 images were collected, including 1,000 images of wind turbine blades and 1,000 images of solar panels. For the solar panel subset, infrared illumination was used to highlight hot spots. To prevent overfitting and improve generalization, I applied data augmentation techniques as listed in Table 1. After augmentation, the dataset was expanded to 3,000 images per category. The dataset was split into training and validation sets in an 8:2 ratio, resulting in 1,200 training images and 300 validation images for each defect type.
| Augmentation Method | Description |
|---|---|
| Flip | Horizontal or vertical flipping to generate new samples |
| Crop | Random cropping of a sub-region from the original image |
| Noise | Adding random noise to images |
| Rotation | Random rotation by a certain angle |
| Affine Transform | Translation, rotation, scaling, and shearing operations |
2.2 Experimental Setup and Evaluation Metrics
All experiments were conducted on an NVIDIA GeForce RTX 3090 GPU. The SGD optimizer was used with a batch size of 16 and 300 epochs. The initial learning rate was set to 0.01, with cosine annealing scheduling. The evaluation metrics include Precision (P), Recall (R), and mean Average Precision (mAP). These are computed as follows.
Precision:
$$Precision = \frac{TP}{TP + FP}$$
Recall:
$$Recall = \frac{TP}{TP + FN}$$
mAP:
$$mAP = \frac{1}{N} \sum_{n=1}^{N} AP(n)$$
where $$TP$$ is the number of true positives, $$FP$$ is false positives, $$FN$$ is false negatives, $$N$$ is the number of defect classes (crack, corrosion for wind turbine blades; hot spot for solar panels), and $$AP(n)$$ is the average precision for class $$n$$.
2.3 Main Results
I compared the improved model (Ours) with the baseline YOLOv11 under identical experimental settings. Each result is the average of five independent runs. The comparison is shown in Table 2.
| Method | Precision (%) | Recall (%) | mAP (%) |
|---|---|---|---|
| YOLOv11 | 83.7 | 85.0 | 86.2 |
| Ours | 88.6 | 88.4 | 89.5 |
As shown in Table 2, the proposed model achieves significant improvements across all metrics. Precision increases by 4.9%, Recall by 3.4%, and mAP by 3.3%. This demonstrates that the integration of CBAM and WTConv effectively enhances the model’s ability to detect both wind turbine blade defects and solar panel hot spots.
2.4 Ablation Study
To verify the individual contributions of CBAM and WTConv, I conducted an ablation study. The results are summarized in Table 3.
| Method | Precision (%) | Recall (%) | mAP (%) |
|---|---|---|---|
| YOLOv11 | 83.7 | 85.0 | 86.2 |
| YOLOv11 + CBAM | 85.3 | 85.2 | 86.8 |
| YOLOv11 + WTConv | 84.5 | 85.1 | 86.3 |
| Ours (CBAM + WTConv) | 88.6 | 88.4 | 89.5 |
From Table 3, it can be observed that adding either CBAM or WTConv alone yields only modest improvements. However, when both modules are combined, the performance gains become substantial. The mAP improves by 3.3% over the baseline, confirming the synergistic effect of the two modules. The CBAM mechanism helps the model focus on critical channels and spatial regions, which is especially beneficial for detecting subtle defects such as small cracks on wind turbine blades and hot spots on solar panels. The WTConv module provides multi-scale analysis, enabling the model to capture defects of various sizes without a significant increase in computational cost.
2.5 Qualitative Analysis
Figure 1 shows an example of a solar panel image from our dataset. The improved model accurately detects hot spots, which appear as bright regions in the thermal image. In contrast, the baseline YOLOv11 sometimes misses small hot spots or produces false positives. Similarly, for wind turbine blades, the proposed model successfully identifies cracks and corrosion even when the defects are small and located at the blade edges.
The above figure illustrates a typical solar panel image containing hot spots. The improved model’s detection results are overlaid, demonstrating high precision and recall.
3. Conclusion
In this work, I proposed an improved YOLOv11-based defect detection method for both wind turbine blades and solar panels. By incorporating the CBAM attention mechanism and the WTConv module, the model achieves enhanced multi-scale feature fusion and better focus on critical regions. Experiments on a self-constructed dataset show that the proposed method raises the mAP from 86.2% to 89.5%, a 3.3% improvement over the baseline. The ablation study confirms the complementary benefits of CBAM and WTConv. This unified framework can effectively detect a variety of defects—including cracks, corrosion, and hot spots—in two major renewable energy components. Future work will involve extending the method to video-based real-time detection and testing on larger, more diverse datasets collected under varying environmental conditions. Additionally, integrating temporal information from consecutive frames could further improve detection stability for dynamic scenes.