Abstract
Photovoltaic (PV) power stations, predominantly constructed in harsh environments, are susceptible to hot spot failures due to natural erosion on solar panels. Such failures can impact the lifespan of solar panels and, in severe cases, lead to fires. Therefore, timely detection of hot spots on solar panels has become a crucial research direction. To address the issues of low accuracy and slow detection speed in traditional unmanned aerial vehicle (UAV) inspections for hot spot faults in photovoltaic power stations, as well as to eliminate misidentification caused by background noise in infrared images, this study leverages convolutional neural networks (CNNs) to design a method for rapid and accurate hot spot detection. This article details the research content and findings.
Keywords: solar panels, hot spot detection, unmanned aerial vehicles, infrared imaging, convolutional neural networks
1. Introduction
1.1 Research Background and Significance
The majority of photovoltaic power stations in China are built in harsh areas, where solar panels are prone to hot spot failures due to environmental factors. These hot spots can significantly affect the operational efficiency and lifespan of solar panels, and in extreme cases, may even trigger fires. Traditional methods of manual inspection are not only time-consuming and labor-intensive but also face challenges such as low detection accuracy and high risk. Therefore, developing an efficient and accurate hot spot detection technology has become a pressing need.
1.2 Current Research Status at Home and Abroad
1.2.1 UAV Inspection Status
With the rapid development of UAV technology, UAVs have gradually become an important tool for inspecting photovoltaic power stations. UAVs can quickly collect large amounts of data on solar panels, significantly improving inspection efficiency.
1.2.2 Research Status of Hot Spot Faults in Solar Panels
Previous studies on hot spot faults in solar panels have mainly focused on theoretical analysis and simulation experiments, with limited practical applications.
1.2.3 Research Status of Object Detection Models Based on CNNs
In recent years, CNNs have shown remarkable performance in image recognition and object detection tasks, providing new ideas for hot spot detection in solar panels.
1.3 Research Content and Technical Route
1.3.1 Research Content
This study focuses on developing a hot spot detection model for aerial infrared images of solar panels using deep learning techniques. The specific research content includes: constructing an infrared image dataset of solar panels, designing and optimizing a hot spot detection model, and evaluating the model’s performance.
1.3.2 Technical Route
The technical route of this study is as follows: (1) Collect infrared images of solar panels using UAVs; (2) Construct an infrared image dataset of solar panels and perform data enhancement; (3) Design and optimize a hot spot detection model based on CNNs; (4) Train and test the model, and evaluate its performance.
Table 1: Technical Route Overview
| Step | Description |
|---|---|
| 1 | UAV Image Collection |
| 2 | Dataset Construction |
| 3 | Model Design |
| 4 | Model Training & Testing |
2. Analysis of Infrared Image Characteristics of Hot Spot Faults in Solar Panels and Dataset Construction
2.1 Causes and Infrared Image Characteristics of Hot Spot Faults in Solar Panels
2.1.1 Causes of Hot Spot Faults in Solar Panels
Hot spot faults in solar panels are mainly caused by shading, dust accumulation, bird droppings, and other factors, which lead to local heating of the solar panel and the formation of hot spots.
2.1.2 Analysis of Infrared Image Characteristics of Hot Spot Faults
Infrared images of hot spot faults in solar panels exhibit the following characteristics: (1) Hot spots are usually small, occupying a very small area in the entire photovoltaic array image; (2) Hot spots often appear at the edges of photovoltaic modules; and (3) The edges of hot spots are distinct.
2.2 Collection and Screening of Infrared Images of Solar Panels
In this study, infrared images of solar panels were collected using a DJI M300 UAV equipped with a Zenmuse XT2 thermal imaging camera. A total of 2,188 infrared images were collected, and after screening, 1,557 images containing solar panels were obtained for dataset construction.
Table 2: Image Collection and Screening Results
| Total Images Collected | Images Containing Solar Panels |
|---|---|
| 2,188 | 1,557 |
2.3 Construction of Infrared Image Dataset of Solar Panels
2.3.1 Data Enhancement
To address issues such as data diversity and overfitting, data enhancement techniques were employed to expand the dataset. Specifically, methods such as image rotation, compression, contrast adjustment, and other transformations were used to generate diverse infrared images.
Table 3: Data Enhancement Methods
| Data Enhancement Method | Description |
|---|---|
| Image Rotation | Rotation of images at arbitrary angles |
| Image Compression | Compression of images to simulate different shooting heights |
| Contrast Adjustment | Adjustment of image contrast to simulate different lighting conditions |
After data enhancement, the number of effective images increased to 7,785.
2.3.2 Image Annotation
The LabelImg software was used to manually annotate the images, obtaining position information for the solar panels. After annotation, XML files were generated to store the position information of the solar panels.
3. Design of Solar Panel Recognition Model Based on YOLOv4
To address the issue of misidentification caused by background noise in infrared images, it is necessary to extract the solar panels from the images before performing hot spot segmentation. Therefore, an improved YOLOv4 solar panel recognition model was proposed in this study.
3.1 Model Selection and Optimization
The YOLOv4 model was selected as the basis for constructing the improved model due to its high detection accuracy and speed. To achieve lightweight design, the MobileNet-V2 network was used to replace the backbone feature extraction network of YOLOv4. Additionally, depthwise separable convolutions were used to modify the standard convolutions in the PANet network.
Table 4: Model Optimization Methods
| Optimization Method | Description |
|---|---|
| Backbone Replacement | Replacement of YOLOv4’s backbone network with MobileNet-V2 |
| Convolution Modification | Modification of standard convolutions in PANet with depthwise separable convolutions |
3.2 Model Training and Testing
Transfer learning was used to train the model on the constructed infrared image dataset of solar panels. Experimental results showed that the improved model achieved high recognition accuracy and speed for solar panels.
Table 5: Model Performance Comparison
| Model | Accuracy (%) | FPS | Model Size (MB) |
|---|---|---|---|
| Original YOLOv4 | – | – | – |
| Improved YOLOv4 (MobileNet-V2) | 99.56 | 22.1 | Reduced to 1/5 of original |
4. Training and Testing of Hot Spot Segmentation Model
4.1 Model Selection and Optimization
Based on the results of the solar panel recognition model, a hot spot segmentation model was designed and trained using the constructed infrared image dataset. Multiple semantic segmentation models, including FCN, Dilated FCN, and DeeplabV3+, were compared and analyzed.
4.2 Model Training and Testing
Transfer learning was also used to train the hot spot segmentation model. Experimental results showed that the improved DeeplabV3+ model achieved high accuracy and speed for hot spot detection.
Table 6: Hot Spot Segmentation Model Performance
| Model | Mean Average Precision (mAP) (%) | Mean Intersection over Union (mIoU) (%) | FPS | Model Size (MB) |
|---|---|---|---|---|
| Improved DeeplabV3+ | 95.99 | 85.58 | 24.5 | Reduced to 1/10 of original |
5. Conclusion
This study proposes a method for rapid and accurate detection of hot spots in aerial infrared images of solar panels based on deep learning. By constructing an infrared image dataset of solar panels, designing and optimizing a solar panel recognition model and a hot spot segmentation model, and evaluating their performance.