In the current era of renewable energy development, wind power and photovoltaic power generation are primary forms, with solar panels playing a crucial role in harnessing solar energy. However, harsh weather conditions such as lightning, rain, snow, and hail can cause damage to wind turbine blades and solar panels, leading to efficiency losses and safety risks. Traditional inspection methods are often labor-intensive and error-prone, prompting the need for automated defect detection systems. Previous studies have explored improved YOLOv8 algorithms for defect detection in wind turbine blades and solar panels separately, but these approaches are limited to single-scenario applications. In this work, I propose a unified framework based on an enhanced YOLOv11 model to achieve precise defect detection for both wind turbine blades and solar panels, addressing the shortcomings of prior methods.
To tackle these challenges, I integrate the Convolutional Block Attention Module (CBAM) and Wavelet Transform Convolution (WTConv) into the YOLOv11 architecture. CBAM leverages channel and spatial attention mechanisms to refine feature maps by emphasizing important channels and spatial regions, thereby enhancing model performance. Meanwhile, WTConv employs wavelet transforms to decompose input images into different frequency components, performs multi-scale analysis through small-sized convolutions on each frequency layer, and reconstructs the results via inverse wavelet transforms. This combination improves multi-scale fusion capabilities and detection accuracy. By incorporating real-time field data, the model undergoes analysis and parameter optimization to facilitate typical fault detection in wind turbine equipment and solar panel arrays.
For validation, I utilize a custom dataset comprising images of wind turbine blades and solar panels. The improved YOLOv11 network demonstrates enhancements in Precision (P), Recall (R), and Mean Average Precision (mAP), indicating its effectiveness in detecting defects such as cracks, corrosion, and hotspots. This article details the methodological improvements, experimental setup, and results, with a focus on leveraging advanced techniques to boost renewable energy infrastructure monitoring.
Enhancements to YOLOv11
The YOLOv11 model serves as the backbone for our defect detection system. Its core consists of a backbone network that extracts multi-level features from input images using stacked convolutional layers and specialized modules. Key components include CBS modules (combining Convolution, BatchNorm, and Activation), C3K2 modules, Spatial Pyramid Pooling – Fast (SPPF), and the newly introduced Cross Stage Partial Pyramid Spatial Attention Module (C2PSA). The SPPF module aggregates features at different scales, improving the detection of variably sized defects, such as small cracks in wind turbine blades or large hotspots in solar panels. The C2PSA module enhances spatial attention within feature maps, allowing the model to concentrate on critical areas, which is vital for identifying defects of diverse sizes and locations. For instance, in solar panels, attention can be directed towards regions with temperature anomalies indicating hotspots.
The neck of YOLOv11 integrates multi-scale features through upsampling and concatenation. Specifically, the C2F module is replaced with the C3K2 module, which offers diverse convolutional kernel designs and stronger hierarchical feature extraction capabilities. This improves scale adaptability, enabling the model to capture both global and local details, thus enhancing the perception of subtle defects like fine cracks or corrosion in wind turbine blades and localized hotspots in solar panels. The C3K2 structure facilitates efficient feature extraction across varying defect morphologies. The head structure of YOLOv11 generates final predictions for object detection and classification, utilizing a one-to-one detection head during inference to enable end-to-end deployment without Non-Maximum Suppression (NMS). It outputs predictions from three different feature maps to cover multiple scales, with a regression branch for bounding box prediction and a classification branch using Depthwise Separable Convolution (DSC) and Softmax activation for category probability output.
Incorporation of CBAM Attention Mechanism
The CBAM attention mechanism is sequentially applied to feature maps to refine them by focusing on relevant channels and spatial regions. This is particularly beneficial for defect detection in solar panels, where specific channels may highlight thermal variations, and spatial attention can pinpoint hotspot areas. The channel attention module computes weights for each channel using global average and max pooling, followed by a multilayer perceptron (MLP) and a sigmoid activation function. The process is defined as:
$$M_c(F) = \sigma(MLP(AvgPool(F)) + MLP(MaxPool(F)))$$
where \(M_c(F)\) represents the channel attention weights, \(F\) is the input feature map, \(\sigma\) denotes the sigmoid function, \(MLP\) is a multilayer perceptron, \(AvgPool(F)\) and \(MaxPool(F)\) are global average and max pooling operations, respectively. For solar panels, this module amplifies channels that correspond to thermal signatures or texture patterns indicative of defects.
The spatial attention module emphasizes important spatial locations by concatenating average and max pooled features and applying a convolutional layer. It is expressed as:
$$M_s(F) = \sigma(f^{7*7}([AvgPool(F); MaxPool(F)]))$$
where \(M_s(F)\) denotes the spatial attention weights, \(f^{7*7}\) is a \(7 \times 7\) convolutional operation, and \([ \cdot ]\) indicates concatenation. In the context of solar panels, this allows the model to concentrate on regions with anomalies, such as hotspots or physical damage, thereby improving detection precision.
Integration of Wavelet Transform Convolution
Wavelet Transform Convolution (WTConv) is employed to enhance multi-scale analysis within the convolutional neural network. Wavelet transforms decompose signals into time-frequency components, enabling detailed examination of image features across different scales. The discrete wavelet transform (DWT) is given by:
$$W(a,b) = \int_{-\infty}^{\infty} f(t) \psi_{a,b}(t) dt$$
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, and \(b\) is the translation factor. For images, DWT typically decomposes into four subbands: LL (low-low frequency), LH (low-high frequency), HL (high-low frequency), and HH (high-high frequency), extracting rich horizontal and vertical details. This is especially useful for analyzing solar panels, where defects may manifest at various frequencies.
WTConv integrates wavelet transforms into standard convolution operations, allowing for multi-scale feature extraction without increasing parameters significantly. The standard convolution is defined as:
$$y(t) = (f * w)(t) = \int_{-\infty}^{\infty} f(\tau) w(t – \tau) d\tau$$
where \(y(t)\) is the output, \(f\) is the input signal, and \(w\) is the kernel. In WTConv, the kernel is replaced with a wavelet basis function:
$$y(t) = \int_{-\infty}^{\infty} f(\tau) \psi_{a,b}(t – \tau) d\tau$$
By adjusting the scale factor \(a\) and translation factor \(b\), WTConv captures features at different scales and positions. In this work, I implement cross-channel wavelet convolution to modify the C3K2 module, creating a C3K2WT module that replaces the original C3K2 in the YOLOv11 network. This expands the model’s receptive field and enhances its ability to detect defects in solar panels and wind turbine blades across multiple scales.
Experimental Framework and Results
To validate the proposed model, I conducted experiments using a custom dataset. The dataset includes images captured by a Matrice 350 RTK drone equipped with a Zenmuse H30 series camera, comprising 2,000 original images of wind turbine blades and solar panels. Specifically, 1,000 images of solar panels were obtained using infrared supplementary lighting to highlight hotspots. To prevent overfitting and ensure robustness, data augmentation techniques were applied, increasing the total number of images to 3,000. The augmentation methods are summarized in Table 1.
| Augmentation Method | Description |
|---|---|
| Flipping | Horizontal or vertical flipping to generate new samples. |
| Cropping | Random cropping of sub-regions from original images. |
| Noise Addition | Adding random noise to images. |
| Rotation | Rotating images by random angles. |
| Affine Transformation | Applying translation, rotation, scaling, and skewing operations. |
The augmented dataset was split into training and validation sets in an 8:2 ratio, resulting in 2,400 images for training and 600 for validation. All experiments were performed in an NVIDIA GeForce GTX 3090 GPU environment, using the SGD optimizer over 300 iterations with a batch size of 16. Evaluation metrics included Precision, Recall, and mAP, calculated as follows:
$$Precision = \frac{TP}{TP + FP}$$
$$Recall = \frac{TP}{TP + FN}$$
$$mAP = \frac{1}{N} \sum_{n=1}^{N} AP(n)$$
where \(TP\) denotes true positives, \(FP\) false positives, \(FN\) false negatives, \(AP(n)\) the average precision for each defect class, and \(N\) the number of defect types. These metrics provide a comprehensive assessment of model performance, particularly for solar panels where accurate hotspot detection is critical.
The experimental results, averaged over five runs, are presented in Table 2, comparing the proposed model with the baseline YOLOv11. The improved YOLOv11 with CBAM and WTConv demonstrates superior performance across all metrics.
| Method | Precision (%) | Recall (%) | mAP (%) |
|---|---|---|---|
| YOLOv11 | 83.7 | 85.0 | 86.2 |
| Proposed Model | 88.6 | 88.4 | 89.5 |
As shown, the proposed model achieves a Precision of 88.6%, Recall of 88.4%, and mAP of 89.5%, representing a significant improvement over the baseline. This highlights its efficacy in detecting defects in both wind turbine blades and solar panels, with enhanced accuracy for solar panel hotspots.
To further analyze the contributions of individual components, I conducted an ablation study, summarized in Table 3. This study systematically evaluates the impact of CBAM and WTConv by incrementally adding them to the baseline YOLOv11.
| 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 |
| Proposed Model (YOLOv11 + CBAM + WTConv) | 88.6 | 88.4 | 89.5 |
The results indicate that while adding CBAM or WTConv individually yields modest improvements, their combination boosts mAP by 3.3%, underscoring the synergistic effect of attention mechanisms and multi-scale convolution for defect detection. Specifically, for solar panels, this combination enhances the model’s ability to identify subtle hotspots and physical damages across varying scales.
Discussion and Future Directions
The proposed model demonstrates robust performance in detecting defects in wind turbine blades and solar panels, leveraging advanced techniques to address the limitations of previous approaches. The integration of CBAM and WTConv enhances feature refinement and multi-scale analysis, leading to higher precision and recall. However, the study is based on a relatively limited dataset focused on static images, which may not fully capture real-world dynamic scenarios. In future work, I plan to extend the evaluation to video data by extracting frames from different sequences, enabling more accurate and real-time defect detection for solar panels and wind turbine blades in operational environments. This could involve adapting the model for temporal analysis and incorporating additional sensor data to further improve reliability. Moreover, expanding the dataset to include more diverse weather conditions and defect types will enhance the model’s generalizability, ensuring its applicability in monitoring renewable energy infrastructure like solar panels under varying operational stresses.
Conclusion
In this article, I presented an improved YOLOv11-based framework for defect detection in wind turbine blades and solar panels. By incorporating CBAM attention mechanisms and Wavelet Transform Convolution, the model achieves enhanced multi-scale fusion and detection accuracy. Experimental results on a custom dataset show that the proposed method attains an mAP of 89.5%, outperforming the baseline YOLOv11 by 3.3%. This underscores its potential for practical applications in renewable energy monitoring, particularly for maintaining the integrity of solar panels. Future efforts will focus on leveraging dynamic data and refining the model for real-world deployment, contributing to the sustainable management of wind and solar power systems.