We propose an improved YOLOv11-based method for detecting defects in wind turbine blades and solar panels. Wind power and photovoltaic power are currently the main forms of renewable energy development. However, harsh weather conditions such as lightning, rain, snow, and hail can damage wind turbines and solar panels. Previous studies have applied improved YOLOv8 algorithms to wind turbine blade defect detection and solar panel defect detection separately. Those methods compensate for the shortcomings of traditional inspection, but they focus on a single scenario. We integrate defect detection for two different scenes—wind turbine blades and solar panels—to achieve precise detection of both.
To address these challenges, we introduce an improved YOLOv11 network that incorporates a Convolutional Block Attention Module (CBAM) and Wavelet Transform Convolution (WTConv). The CBAM uses channel attention and spatial attention mechanisms to optimize feature maps by emphasizing important channels and spatial regions, thereby improving model performance. WTConv applies wavelet transform to decompose input images into different frequency components, performs small-size convolutions on each frequency layer, and recombines the results using inverse wavelet transform, enabling multi-scale analysis of images. Combined with real-time field data, the model completes analysis and parameter optimization, achieving typical fault detection for wind turbine equipment.
We verify our method using a self-made dataset of wind turbine blades and solar panels. The improved YOLOv11 network shows improvements in precision (P), recall (R), and mean average precision (mAP) compared to the baseline YOLOv11.
1. Improvement of YOLOv11
1.1 YOLOv11 Model
The backbone network is the core part of the YOLO architecture, mainly responsible for extracting multi-level features from input images. This process uses stacked convolutional layers and specific modules to generate feature maps at multiple scales. The YOLO backbone consists of modules such as CBS, C3K2, Spatial Pyramid Pooling – Fast (SPPF), and C2PSA. The CBS module combines convolution, batch normalization, and activation functions to efficiently extract image features. YOLOv11 retains the SPPF module from previous versions, which extracts feature information from different scales. The SPPF module fuses information from multiple scales, improving the model’s ability to detect defects of varying sizes (e.g., small cracks or large hot spots). For wind turbine blades and solar panels, defect size variations can be significant, and the SPPF module helps the model handle these changes. After the SPPF module, YOLOv11 introduces a new Cross Stage Partial Pyramid Spatial Attention Module (C2PSA). The C2PSA module enhances spatial attention in feature maps, allowing the model to focus on important regions of the image. By performing spatial pooling on features, the C2PSA module enables YOLOv11 to concentrate on specific areas of interest, thereby improving detection accuracy for objects of different sizes and positions. Moreover, it facilitates cross-stage information transfer on feature maps at different scales, enabling the model to detect defects across multiple scales—crucial for defects like small cracks or large corrosion on wind turbine blades, and local hot spots or extensive damage on solar panels.
The neck part of YOLOv11 is responsible for combining features from different scales and passing them to the head for prediction. This process typically involves upsampling and concatenation of feature maps, allowing the model to effectively capture multi-scale information. YOLOv11 replaces the C2F module in the neck with the C3K2 module. The C3K2 module has more diverse convolution kernel designs and stronger hierarchical feature extraction capabilities, better adapting to different scales. It fuses multi-scale information, balancing global large-scale features and local detail changes, thereby improving perception of tiny and fine defects. This module is suitable for detecting defects of various sizes, enabling more efficient and rapid extraction of features for small cracks, corrosion on wind turbine blades, and hot spots on solar panels, ultimately enhancing detection accuracy.
The head structure of YOLOv11 generates final predictions for object detection and classification. It processes feature maps from the neck and outputs bounding boxes and class labels of objects in the image. During training, two detection heads are jointly optimized to provide rich supervision. During inference, YOLOv10 discards the one-to-many head and uses only the one-to-one head for prediction, enabling end-to-end deployment without non-maximum suppression (NMS). YOLOv11 retains this advantage with the same head structure as YOLOv10. The detection head outputs predictions from three different feature maps, covering different scales of the image. This strategy helps detect small objects more accurately while using higher-level features to capture larger objects. The position regression branch uses two standard convolutions for feature fusion and a single convolution for position prediction. The classification branch uses depthwise separable convolution (DSC) for feature fusion and pointwise convolution for channel interaction, finally using a single convolution for classification prediction with a softmax activation function to output class probabilities.
1.2 CBAM Attention Mechanism
The CBAM principle uses channel attention and spatial attention in two sequential modules to progressively refine feature maps, making the neural network focus on key features and suppress irrelevant information. The output process is expressed as:
$$
M_c(F) = \sigma(MLP(AvgPool(F)) + MLP(MaxPool(F)))
$$
where $M_c(F)$ is the importance weight for each channel, $F$ is the input feature map, $\sigma$ is the activation function (sigmoid), $MLP$ is a multi-layer perceptron, $AvgPool(F)$ is global average pooling of $F$, and $MaxPool(F)$ is global max pooling of $F$.
The channel attention module adaptively emphasizes channels relevant to the target task. For example, in solar panel hot spot detection, the model needs to focus on channels with large temperature variations in infrared images; in wind turbine blade defect detection, the model needs to focus on channels with prominent edge or texture features. The spatial attention weight $M_s(F)$ is computed as:
$$
M_s(F) = \sigma(f^{7 \times 7}([AvgPool(F); MaxPool(F)]))
$$
where $f^{7 \times 7}$ is a convolution operation with kernel size $7 \times 7$. The spatial attention module focuses on spatial locations in the feature map. For solar panel hot spot detection, spatial attention can emphasize the hot spot region; for wind turbine blade defect detection, spatial attention enables the model to focus on areas where cracks or corrosion appear on the blade surface.
1.3 Wavelet Convolution
Wavelet transform (WT) is a mathematical transformation that decomposes a signal into components of different time scales, representing the time-frequency characteristics of the signal. Wavelet transform is widely used in signal processing, image analysis, and data compression. Wavelet convolution uses wavelet transform to provide convolutional neural networks (CNNs) with larger receptive fields while avoiding parameter increase caused by large convolution kernels. WTConv is an innovative convolution layer that uses wavelet transform to perform deep, multi-scale analysis of input data. The discrete wavelet transform calculation is:
$$
W(a, b) = \int_{-\infty}^{\infty} f(t) \psi_{a,b}(t) dt
$$
where $W(a,b)$ is the inner product of the original signal $f(t)$ with the scaled and translated wavelet basis function $\psi_{a,b}(t)$; $a$ is the scale factor; $b$ is the translation factor; $\psi_{a,b}(t)$ is the wavelet basis function; $t$ is the signal. Discrete wavelet transform uses a set of basis functions (wavelet functions) for multi-scale analysis of signals. The basic process involves decomposing the signal using low-pass and high-pass filters, then further decomposing the low-frequency and high-frequency parts. Typically, discrete wavelet transform (DWT) decomposes the signal into four parts: Low-Low (LL), Low-High (LH), High-Low (HL), and High-High (HH) frequency subbands. For our objects, wavelet transform can decompose images in horizontal and vertical directions, extracting richer information.
We adopt cross-channel wavelet convolution to improve the C3K2 module in YOLOv11, constructing a new C3K2WT module, and then replace the original C3K2 module with C3K2WT. This aims to expand the model’s receptive field. The basic operation of wavelet convolution introduces wavelet transform into the standard convolution operation. Specifically, convolution not only uses traditional convolution kernels but also performs convolution using wavelet basis functions, capturing multi-scale and multi-level features from input data. The standard convolution operation is:
$$
y(t) = (f * w)(t) = \int_{-\infty}^{\infty} f(\tau) w(t – \tau) d\tau
$$
In wavelet convolution, we replace $w(t)$ with the wavelet basis function $\psi(t)$, and the convolution operation becomes:
$$
y(t) = \int_{-\infty}^{\infty} f(\tau) \psi_{a,b}(t – \tau) d\tau
$$
By adjusting the scale factor $a$ and position factor $b$ of the wavelet basis function, wavelet convolution can extract features at different scales and positions, enabling multi-scale feature extraction in convolutional neural networks.
2. Experimental Results and Analysis
2.1 Dataset Description
To verify the accuracy and practicality of our network, we used a Matrice 350 RTK drone equipped with a Zenmuse H30 series camera to collect a total of 2000 images of wind turbine blades and solar panels. Among these, 1000 images captured hot spots on solar panels using the infrared fill light of the Zenmuse H30 camera. To prevent overfitting and improve reliability, we applied data augmentation techniques to the original images, including flipping, adding noise, cropping, rotation, and affine transformation. The data augmentation methods are summarized in Table 1. After augmentation, each dataset (wind turbine blade defects and solar panel hot spots) expanded to 3000 images. Each dataset was split into training and validation sets with an 8:2 ratio, resulting in 1200 images for training and 300 images for validation.
| Augmentation Method | Description |
|---|---|
| Flipping | Horizontal or vertical flipping of images to generate new samples |
| Cropping | Random cropping of a subregion from the original image |
| Adding Noise | Random addition of noise to images |
| Rotation | Rotation of images by random angles |
| Affine Transformation | Operations including translation, rotation, scaling, and shearing |
2.2 Evaluation Metrics
Experiments were conducted on an NVIDIA GeForce GTX 3090 GPU environment. We used the SGD optimizer with a total of 300 epochs and a batch size of 16. To evaluate the performance of the object detection model, we used three metrics: precision (P), recall (R), and mean average precision (mAP). Precision is calculated as:
$$
\text{Precision} = \frac{TP}{TP + FP}
$$
where $TP$ (true positives) is the number of samples where the detected defect class matches the ground truth, and $FP$ (false positives) is the number of samples where the detected defect class does not match the ground truth. Recall is calculated as:
$$
\text{Recall} = \frac{TP}{TP + FN}
$$
where $FN$ (false negatives) is the number of actual defect targets that were not correctly detected by the algorithm. To comprehensively evaluate model effectiveness, we also compute mAP over all defect classes. The mAP is defined as:
$$
mAP = \frac{1}{N} \sum_{n=1}^{N} AP(n)
$$
where $AP(n)$ is the average precision for class $n$, and $N$ is the number of defect classes (wind turbine blade defects and solar panel hot spots).
2.3 Experimental Results Comparison
To verify the effectiveness of our proposed model for wind turbine blade defect detection and solar panel hot spot detection, we used the same datasets and kept all experimental settings consistent. Each result is the average of 5 runs to ensure fairness. We compared our model (Ours) against baseline YOLOv11. The results are 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, our model achieves a precision of 88.6%, recall of 88.4%, and mAP of 89.5%, demonstrating significant improvements over the baseline YOLOv11. This confirms the ability of the improved network to accurately detect defects in both wind turbine blades and solar panels.
2.4 Ablation Study
To validate the effectiveness of CBAM and WTConv, we conducted an ablation study. By gradually removing or replacing key components, we systematically evaluated the contribution of each module to overall performance. The results are presented 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 (YOLOv11 + CBAM + WTConv) | 88.6 | 88.4 | 89.5 |
From Table 3, adding only CBAM or only WTConv to YOLOv11 yields slight improvements in precision, recall, and mAP. However, when both CBAM and WTConv are introduced together (our proposed method), the mAP increases by 3.3% compared to baseline YOLOv11 (from 86.2% to 89.5%). This indicates that the combination of CBAM and WTConv provides complementary benefits, effectively enhancing multi-scale feature fusion and attention to important regions, leading to more accurate defect detection for both wind turbine blades and solar panels.
3. Conclusion
In this work, we proposed an improved YOLOv11 network for detecting defects in wind turbine blades and solar panels. By integrating the CBAM attention mechanism and wavelet transform convolution (WTConv), the model achieves better multi-scale fusion capability and higher detection accuracy. Experiments on self-constructed datasets demonstrate that our method achieves a mAP of 89.5%, outperforming the baseline YOLOv11 by 3.3%. The results validate the effectiveness of the proposed approach. Nevertheless, our experiments are limited to static images. Future work will involve testing on video frames from different temporal sequences to better simulate real-world scenarios and further improve detection robustness for wind turbine blades and solar panels.