As solar energy continues to establish itself as one of the most commercially viable green energy sources, the efficiency of photovoltaic systems becomes paramount. Solar panels, being the core components for photoelectric conversion, are susceptible to dust accumulation during operation. This dust obstructs solar radiation, leading to a decline in conversion efficiency and output power. Existing methods for dust detection often rely on manual inspections or traditional image processing techniques based on threshold segmentation. While these approaches can identify dust to some extent, their accuracy and efficiency are constrained by environmental factors such as lighting conditions and weather changes, making them inadequate for complex real-world scenarios. Recent advancements in deep learning have introduced new solutions for object detection and image segmentation. However, the performance of these models heavily depends on the richness and quality of annotated data, which is challenging to obtain in practical applications due to the variability in dust distribution and environmental conditions on solar panels.
To address these limitations, I propose a novel dust detection method that integrates deep learning with traditional image processing techniques. This approach leverages the YOLO11 algorithm for robust feature extraction and generalization, combined with histogram analysis and dual-light image fusion. The methodology involves several key steps: first, using the YOLO11-OBB rotating object detection algorithm to accurately extract photovoltaic strings and individual solar panels from both visible light and thermal imaging images; second, applying the YOLO11-SEGMENT algorithm to detect dust areas in visible light images and thermal spots in thermal images, followed by temperature analysis of these spots; and finally, employing traditional image processing methods to analyze visible light images and predict dust coverage ratios based on histogram features. By combining the strengths of deep learning and image processing, this method introduces a dual-validation mechanism that enhances detection accuracy in complex environments. Additionally, the integration of thermal imaging data allows for severity classification of dust accumulation based on abnormal temperature readings, providing a comprehensive solution for maintaining the efficiency of solar panels.
The core of this method lies in the improved YOLO11 architecture, which incorporates several enhancements to boost performance in detecting solar panels and associated defects. The YOLO11 model builds upon previous versions by optimizing the balance between accuracy and efficiency. Key improvements include the integration of the Triple Attention mechanism into the C3K2 module, forming the C3K2-T module, which enhances feature sensitivity through channel, spatial, and contextual attention. The attention mechanisms are combined as follows: channel attention adjusts weights for each channel to emphasize critical features, spatial attention focuses on key regions by weighting pixel locations, and contextual attention captures long-range dependencies by integrating global information. The combined attention feature is computed as:
$$ F_{out} = X \cdot C \cdot S \cdot S_{context} $$
where ( C ) represents the channel attention weights, ( S ) the spatial attention map, and ( S_{context} ) the contextual attention map. The channel attention is calculated as:
$$ C = \sigma(W_c \cdot X) $$
and the contextual attention as:
$$ S_{context} = \text{SoftMax}(W_{context} \cdot X) $$
Furthermore, the C2PSA module is enhanced by embedding the SEAM (Squeeze-and-Excitation Attention Module) to form the C2PSA-S module, which improves detection under occlusions and varying light conditions by optimizing convolution structures and feature representations. The SEAM loss function comprises three components: classification loss, equivariant regularization loss, and equivariant cross-regularization loss, expressed as:
$$ L = L_{cls} + L_{ER} + L_{ECR} $$
where the classification loss ( L_{cls} ) is given by:
$$ L_{cls} = \frac{1}{2} \left( l_{cls}(Z^0, l) + l_{cls}(Z^t, l) \right) $$
the equivariant regularization loss ( L_{ER} ) by:
$$ L_{ER} = \left| y^0 \cdot A – y^t \right| $$
and the equivariant cross-regularization loss ( L_{ECR} ) by:
$$ L_{ECR} = \left| y^0 \cdot A – y^t \right| + \left| y^t \cdot A – y^0 \right| $$
These enhancements enable the model to handle the intricate patterns of dust on solar panels more effectively, reducing false positives and missed detections.
The overall system workflow for dust detection on solar panels involves multiple stages. Initially, drones equipped with RTK positioning and dual-light cameras capture visible and thermal images of photovoltaic stations at fixed intervals or positions. The images undergo preprocessing, such as distortion correction, to ensure quality. The YOLO11-OBB algorithm then extracts photovoltaic strings and individual solar panels from both image types. For each panel, dust regions are detected in visible light images using YOLO11-SEGMENT, while thermal spots are identified in thermal images. A confidence threshold of 0.65 is applied to deep learning detections; results below this threshold are re-evaluated using traditional histogram-based methods, which assign a dust score. If the score exceeds 0.7, the region is confirmed as dusty. Similarly, panels not detected by the model are processed through the histogram pipeline, and those with scores above 0.8 are classified as dusty. Finally, dust regions are matched with thermal spots via pixel-to-coordinate conversion, and severity is graded based on thermal data: spots with average temperatures between 40°C and 60°C indicate mild dust accumulation, while those above 60°C indicate severe accumulation. This integrated approach ensures high accuracy and reliability in diverse conditions for solar panels.
The histogram-based dust assessment process plays a critical role in validating detections on solar panels. After extracting panel regions using YOLO11, the images are converted to grayscale to simplify processing while retaining structural information. The grayscale conversion uses the formula:
$$ I_{gray} = 0.299 \times R + 0.587 \times G + 0.114 \times B $$
where ( R ), ( G ), and ( B ) are the red, green, and blue channel values, respectively. Gaussian filtering is then applied to reduce noise, with the kernel function defined as:
$$ G(x, y) = \frac{1}{2\pi\sigma^2} e^{-\frac{x^2 + y^2}{2\sigma^2}} $$
where ( \sigma ) controls the smoothing degree. Morphological operations, including erosion and dilation, are used to remove small interferences like white stripes and enhance dust regions. Erosion is expressed as ( A \ominus B = { z | (B)_z \subseteq A } ), and dilation as ( A \oplus B = { z | (B^*)z \cap A \neq \emptyset } ), where ( A ) is the image and ( B ) the structuring element. Histogram analysis follows, where the grayscale histogram is computed and smoothed with Gaussian filtering to identify peaks. The minimum gray value between two peaks in the 100–150 range is used as a threshold for binarization. Pixels with values above 235 are set to 100 (representing panel edges), those below the threshold to 100 (clean areas), and those above to 200 (dusty areas). Regions with fewer than 8 rows or columns are removed to avoid edge effects. This process yields a binary image that clearly separates dusty and clean areas on solar panels, enabling accurate dust coverage calculation.
To evaluate the method, experiments were conducted on a dataset comprising approximately 100,000 visible and thermal images from various photovoltaic stations, including centralized and mountainous setups. The dataset encompasses different lighting and weather conditions to test robustness. The real dust subset contains about 5,000 images. The experimental setup uses an Ubuntu OS with an Intel Xeon Gold 5218 processor, 45 GB RAM, and NVIDIA Quadro RTX 8000 GPUs. The code is implemented in Python 3.8 with PyTorch 1.12.1, CUDA 10.2, and cuDNN 7.6.5. Performance metrics include mean Average Precision (mAP), precision, recall, parameters, and GFLOPs. The mAP is calculated as:
$$ \text{mAP} = \frac{1}{n} \sum_{i=0}^{n} \text{AP}_i $$
with Average Precision (AP) given by:
$$ \text{AP} = \int_0^1 P(y) dy $$
Precision ( P ) and recall ( R ) are defined as:
$$ P = \frac{TP}{TP + FP} $$
$$ R = \frac{TP}{TP + FN} $$
where ( TP ) is true positives, ( FP ) false positives, and ( FN ) false negatives.
Comparative experiments with mainstream algorithms demonstrate the superiority of the proposed method for detecting solar panels and dust. The improved YOLO11-OBB model achieves higher mAP scores compared to Faster R-CNN, SSD, YOLOv5, and YOLOv8, as shown in the table below. Ablation studies confirm that integrating the C2PSA-S and C3K2-T modules enhances performance without significantly increasing computational cost.
| Algorithm | Recall | mAP@0.5 | mAP@0.5:0.95 | Param (M) | GFLOPs |
|---|---|---|---|---|---|
| Faster R-CNN | 0.76 | 0.77 | 0.78 | 370.0 | 220 |
| SSD | 0.73 | 0.76 | 0.72 | 12.3 | 62 |
| YOLOv5s | 0.73 | 0.75 | 0.75 | 7.2 | 17 |
| YOLOv8s | 0.76 | 0.78 | 0.77 | 11.2 | 28 |
| YOLO11s-OBB | 0.80 | 0.82 | 0.79 | 13.5 | 20 |
| Ours | 0.85 | 0.88 | 0.83 | 7.8 | 20 |
For dust detection on solar panels, the method shows improved recall and mAP over alternatives, as illustrated in the following table. The two-stage approach—first detecting photovoltaic strings and then segmenting components—reduces background noise and improves accuracy despite limited dust data.
| Algorithm | Recall | mAP@0.5 | mAP@0.5:0.95 | Param (M) | GFLOPs |
|---|---|---|---|---|---|
| Faster R-CNN | 0.53 | 0.55 | 0.55 | 370.0 | 220 |
| SSD | 0.51 | 0.52 | 0.52 | 12.3 | 62 |
| YOLOv5s | 0.55 | 0.56 | 0.58 | 7.2 | 17 |
| YOLOv8s | 0.56 | 0.58 | 0.58 | 11.2 | 28 |
| YOLO11s-OBB | 0.56 | 0.59 | 0.61 | 13.5 | 20 |
| Ours | 0.62 | 0.65 | 0.63 | 7.8 | 20 |
Thermal spot detection on solar panels, which correlates with dust-induced hotspots, also benefits from the proposed enhancements. The table below compares thermal detection performance, where the improved YOLO11-OBB model achieves higher recall and mAP, facilitating accurate severity assessment based on temperature.
| Algorithm | Recall | mAP@0.5 | mAP@0.5:0.95 | Param (M) | GFLOPs |
|---|---|---|---|---|---|
| Faster R-CNN | 0.64 | 0.61 | 0.59 | 370.0 | 220 |
| SSD | 0.55 | 0.57 | 0.53 | 12.3 | 62 |
| YOLOv5s | 0.66 | 0.62 | 0.57 | 7.2 | 17 |
| YOLOv8s | 0.65 | 0.67 | 0.62 | 11.2 | 28 |
| YOLO11s-OBB | 0.66 | 0.69 | 0.64 | 13.5 | 20 |
| Ours | 0.73 | 0.75 | 0.71 | 7.8 | 20 |
Ablation studies further validate the contributions of individual components. The table below shows that integrating the C2PSA-S and C3K2-T modules into YOLO11s-OBB progressively improves recall and mAP, with the full model achieving a 7.3% increase in mAP@0.5 over the baseline.
| Method | C2PSA-S | C3K2-T | Recall | mAP@0.5 | Param (M) | GFLOPs |
|---|---|---|---|---|---|---|
| YOLO11s-OBB | 0.80 | 0.82 | 13.5 | 20 | ||
| ① | √ | 0.81 | 0.83 | 15.5 | 22 | |
| ② | √ | 0.83 | 0.85 | 13.5 | 20 | |
| ③ | √ | √ | 0.85 | 0.88 | 15.0 | 21 |
In comprehensive testing on 1,000 real-world images of solar panels, the method demonstrates high accuracy and robustness. The histogram-based dust assessment reliably computes coverage ratios, with results showing clear separation between dusty and clean regions. For instance, in sample outputs, dust coverage percentages are accurately predicted, such as 31.30% or 16.86%, based on the binarization and histogram analysis. The integration of thermal data allows for practical severity grading, enabling prioritized cleaning schedules for solar panels with severe dust accumulation.
In conclusion, this research presents an effective solution for dust detection on solar panels by combining YOLO11 deep learning with traditional image processing and dual-light analysis. The method overcomes the limitations of existing approaches, achieving high precision and adaptability in complex environments. Future work will focus on enhancing model generalization and exploring dynamic monitoring systems for automated cleaning of solar panels, thereby supporting the sustainable operation of photovoltaic power stations.