Abstract: As the core component of photovoltaic power generation systems, solar panels are prone to various defects during the power generation process, leading to a reduction in power generation efficiency and affecting the overall system’s power generation capacity and safety. This thesis introduces object detection algorithms based on deep learning into the field of solar panel surface defect detection, utilizing deep learning-based image augmentation algorithms to enrich the dataset of solar panel surface defects. Additionally, transfer learning is employed to construct a power generation efficiency loss model, which can accurately detect defect types and grades while estimating the corresponding power generation efficiency loss. The following is a detailed introduction to this research.
1. Introduction
The installed market for photovoltaic power generation is exhibiting strong growth trends. In this context, ensuring the safety and stability of photovoltaic systems during power generation has become increasingly important. Solar panels, as the most crucial component in photovoltaic power generation systems, directly determine the overall system’s power generation efficiency based on their energy conversion efficiency. Exposed to harsh weather conditions such as wind, rain, temperature variations, sandstorms, and hail, solar panels are susceptible to various defects on their surfaces, such as cracking, dust accumulation, and foreign object coverage. These defects, if not addressed promptly based on their severity, can consume energy in the form of heat, leading to localized temperature increases on the solar panels, the formation of hot spots, and ultimately impacting the power generation efficiency and safety of the entire system. Therefore, accurately detecting various defects on the surface of solar panels and assessing their grades play a vital role in the safe and efficient operation and maintenance of photovoltaic power generation systems.
2. Image Augmentation for Solar Panel Surface Defects
2.1 Overview of Traditional Image Augmentation
Traditional image augmentation algorithms expand datasets primarily through data transformations of existing images, which can be divided into single-sample-based image augmentation and multi-sample-based image augmentation. However, these methods often struggle to generate diverse and high-quality augmented images, limiting their effectiveness in improving the generalization performance of deep learning models.
2.2 Introduction to DCGAN Algorithm
Deep Convolutional Generative Adversarial Networks (DCGANs) have shown great potential in image generation tasks. By pitting a generator against a discriminator in a zero-sum game, DCGANs can learn the data distribution of the training set and generate new images that are similar to the real ones.
2.3 Improved DCGAN Algorithm for Solar Panel Defect Image Augmentation
To address the issues of insufficient image quantity and category imbalance in the solar panel defect dataset, an improved DCGAN algorithm was proposed. This algorithm combines attention mechanisms, ELU activation functions, and Wasserstein functions to build a model. Table 2-1 below shows the experimental configuration for validating the effectiveness of this improved algorithm.
| Experimental Configuration | Details |
|---|---|
| Dataset | Self-made solar panel surface defect dataset (D_initial), consisting of 1800 defect images |
| Experimental Part | Ablation experiment, comparison with other deep learning image augmentation algorithms, and verification of detection performance improvement |
| Evaluation Metrics | mAP (mean Average Precision) |
Experimental results demonstrate that the images generated by the proposed algorithm outperform those from other image augmentation algorithms in quality. Moreover, the detection model trained on the dataset augmented by this algorithm exhibits a significant improvement in detection accuracy, with an mAP increase of 6% compared to the SSD algorithm.
Table 2-3: Distribution of Different Defects in D_initial Dataset
| Defect Type | Number of Images |
|---|---|
| Cracking | 600 |
| Dust Accumulation | 600 |
| Foreign Object Coverage | 600 |
3. Solar Panel Surface Defect Detection Based on Improved SSD Model
3.1 Introduction to SSD Algorithm
Single Shot MultiBox Detector (SSD) is an object detection algorithm that combines the advantages of Faster R-CNN and YOLO, achieving high detection accuracy while maintaining fast detection speed.
3.2 Improvements to the SSD Model
To further enhance the detection performance of the SSD model on solar panel surface defects, the following improvements were made:
- Backbone Network Improvement: Replacing the VGG-16 network in the SSD algorithm with ResNet50, which reduces the number of parameters and computational cost while improving model accuracy due to the introduction of residual modules.
- Introduction of Feature Fusion Pyramid Network: Optimizing the network structure by introducing a feature fusion pyramid network to enhance the model’s ability to detect targets of various sizes by capturing context information more effectively.
- Loss Function Improvement: Modifying the loss function to better balance the detection of small targets and improve the model’s overall detection performance.
3.3 Experimental Results and Analysis
The dataset used for training various algorithm models in this chapter is the deep learning image augmentation dataset (D_deep) obtained using the improved DCGAN algorithm for augmentation in Chapter 2, consisting of 3600 images. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio. After referring to the training parameter settings of other common detection algorithms and conducting multiple experiments, the number of training epochs was set to 30. The mAP value was used as the evaluation metric to assess the detection accuracy of different models. The larger the mAP, the higher the detection accuracy.
4. Design and Implementation of the Solar Panel Surface Defect Detection System
4.1 Estimation of Solar Panel Power Generation Efficiency Loss
To investigate the impact of solar panel surface defects on power generation efficiency, a method for estimating power generation efficiency loss based on transfer learning was proposed. By collecting and analyzing solar panel defect images, the power generation efficiency loss corresponding to each defect grade was calculated, and a dataset of power generation efficiency losses for different defect types and grades was constructed.
4.2 Construction of the Power Generation Efficiency Loss Model
Based on the work in Chapters 2 and 3, the solar panel surface defect dataset was augmented, and the SSD target detection model was improved to achieve accurate identification of solar panel surface defects. However, in practical work, it is often necessary to assess the grades of solar panel surface defects and the corresponding power generation efficiency losses. To address this, a power generation efficiency loss model was trained using transfer learning strategies.
4.2.1 Basic Principles of Transfer Learning
Transfer learning is a technique that allows a model trained in one domain to accelerate the learning process by adapting to a new task. It leverages existing knowledge to assist in learning new knowledge, similar to humans’ abilities of analogy and reasoning.
4.2.2 Implementation Process
After constructing the SSD network, the data weights of the classification model were first trained using the D_deep dataset, with the Conv4_x and Conv5_x convolutional layers frozen. Subsequently, the original fully connected layer was replaced with a new fully connected layer with 12 output nodes. Finally, the fully connected layer of the classification model was trained using the power generation efficiency loss dataset to obtain the power generation efficiency loss model.
4.2.3 Comparison of Models Generated by Different Transfer Learning Methods
Experimental results show that the model trained using the method of gradually thawing high-level convolutional layers can accurately detect the types and grades of solar panel surface defects as well as the corresponding power generation efficiency losses.
4.3 Construction of the Software Analysis Platform
To comprehensively demonstrate the research results of this thesis, a software platform for solar panel surface defect detection was designed and implemented. This software can basically meet the daily inspection needs of photovoltaic power stations.
4.3.1 Software Architecture Design
The software architecture mainly includes data preprocessing, image augmentation, defect detection, result display, and power generation efficiency loss estimation modules.
4.3.2 Software Function Implementation
The software realizes functions such as image import, defect detection, result display, and power generation efficiency loss estimation through the combination of front-end and back-end technologies.
4.3.3 Software Interface Display
(Insert an image of the software interface here)
5. Conclusion and Future Work
This thesis applies deep learning-based image augmentation and target detection algorithms to solar panel surface defect detection, addressing the difficulties in solar panel surface defect detection through algorithmic and theoretical innovations, combined with practical engineering applications, to develop a solar panel surface defect detection scheme based on deep learning. The main research findings are summarized below:
- An improved DCGAN image augmentation algorithm was proposed to address the issues of insufficient image quantity and category imbalance in the collected solar panel surface defect images.
- The SSD target detection model was improved by introducing a feature fusion pyramid network and optimizing the loss function, significantly improving the detection accuracy of solar panel surface defects.
- A power generation efficiency loss model was constructed using transfer learning strategies, enabling accurate detection of solar panel surface defect types and grades while estimating the corresponding power generation efficiency losses.