1. Introduction
Solar energy can be used widely in the world—e.g., for hot water, heating, and power generation,. Solar energy is a clean alternative energy that causes no air or water emissions and poses no threats to people’s health [
1]. According to the International Energy Agency, despite rising prices, solar energy still dominates renewable energy capacity additions. Despite the surging commodity prices and increasing manufacturing costs for solar energy, its capacity additions were forecast to grow by 17% in 2021. This set a new annual record of almost 160 GW. Solar photovoltaic cells account for 60% of all renewable capacity additions. Approximately 1100 GW will become operational over the forecast period in our main case, which is double the rate of the previous five years [
2]. This indicates that solar energy will become increasingly important in the future. Currently, people value utilizing solar-cell panels to obtain solar energy [
3]. Hence, many researchers have recently focused on solar panels. For example, take the solution—processed organic solar cells (OSCs), tandem solar cells (TSCs), and the nonfullerene-blend solar cells system. Solution-processed OSCs provide low cost and convenience, and impressive power conversion efficiencies (PCEs) [
4]. TSCs made of multiple junction devices have high open-circuit voltages, tunable optical bandgaps, and low energy losses [
5]. Nonfullerene-blend solar cell systems have broad absorption and semitransparency features [
6]. Additionally, to assess the energy and environment impacts, lifecycle assessment studies related to these systems are constantly increasing [
7]. Moreover, the absorptivity of solar-cell panels has been a critical issue recently because the lower absorption rate means the less energy created and stored. However, many factors decrease the absorptivity of solar-cell panels. In this study, we refer to these causes as solar-cell panel defects. Examples are cracking or dust on the solar panel, and hotspots on a solar panel’s diode. Hence, it necessary to design an efficient technique for detecting defects in the solar-cell panels [
8].
Various approaches for detecting defects in solar-cell panels have been proposed over time. Some are classed as sensor-based defect detection methods, and others are classed as image-based defect detection methods. Approaches such as solar panel shadow detection and solar panel electrical diagnostics are sensor-based defect detection methods [
9,
10]. Deep learning image defect detection is a form of image-based defect detection [
11]. In this study, we used a deep learning image-based method to detect defects in solar-cell panels. The biggest difference between sensor-based and image-based methods is that sensor-based detection requires a lot of both time and money to detect defects in solar panels. This is because the sensor hardware is more expensive than image-based software, and the sensor hardware needs to be implemented for longer than the deep learning software. Therefore, using deep learning techniques to recognize defects in solar-cell panels is crucial. Several approaches for detecting defects in solar-cell panels using deep learning have been proposed. The approaches include utilizing transfer learning techniques with VGG16 [
12], VGG19 [
13], GoogLeNet [
14], ResNet18 [
15], Unet [
16], FPN [
17], LinkNet [
18], and EfficientNet [
19] to detect the defects on solar-cell panels [
20].
In this study, we used neural network models such as VGG16, VGG19, GoogLeNet, ResNet, and EfficientNet to detect defects in solar-cell panels. To improve the accuracy and speed performance, we designed a new neural network, the lightweight inception residual network (LIRNet), to recognize defects in solar panels. LIRNet is a low-overload convolutional neural network with a residual block and an inception module. It is a robust model. It is based on using hierarchical classification concepts to detect defects in solar panels. The main ideas have been divided into two parts, regarding the hierarchical classification concepts. The first part is the data preprocessing stage. Here, we use the clustering algorithm to combine similar clusters and guide the neural network model for better classification. The second part involves training the LIRNet model. Here, we designed a LIRNet to be trained on defects of solar-cell panels. Due to the properties of solar-cell panels, the images of solar panels in datasets are usually thermal images, and people use unmanned aerial vehicles (UAVs) to collect such datasets. As solar panels are located on the tops of buildings, using a UAV to take pictures of them is reasonable and convenient. In addition, solar panels need to be covered by sunlight to extract solar energy, so there can be severe light reflection when attempting to find defects on solar panels in normal pictures. Hence, using thermal images for solar panels is suitable for detecting defects. For the experiment, the solar-cell-panel images were all thermal images, and we used 20,000 pieces of pictures [
8].
Herein, we used deep learning techniques to detect defects in solar-cell panels, such as cracking and hot spots, and then returned the detected results to the solar panels’ maintenance engineers so that they could repair them. Therefore, maintenance engineers need to directly and in real-time know the issues of solar panels through deep learning techniques. In the experiment, we focused on designing a more efficient neural network model and methods to enhance performance (i.e., accuracy and running speed). The contributions of this study are summarized below.
We explain the detection of defects in solar panels in simple terms and focus on deep learning techniques that can be used to detect the defects. Regarding deep learning methods, we implemented the neural network models of STOA to recognize the faults and compared them in a performance evaluation. In addition, we propose a new effective neural network model, LIRNet. LIRNet’s accuracy was 89%, and its average inference time was 1.39 ms. Its throughput was 27,841 input/s, which is higher than those of other neural network models, such as EfficientNet, GoogLeNet, and ResNet.
The primary value of LIRNet is in the data preprocessing stage, where the K-means clustering algorithm is used to improve the imbalance problem of the dataset. As similar categories will have similar characteristics, the distances between similar categories and the center of the group will be similar. After using K-means grouping, similar categories are merged. Finally, we can merge the unbalanced data of some original categories with other categories to alleviate the problem of unbalanced data. Hence, we adopted the unsupervised learning approach to modify our supervised learning model.
This paper is organized as follows: The introduction and an overview of solar energy and solar-cell panels are presented in
Section 1. We have also described many approaches for detecting defects and the core methods used herein.
Section 2 presents the related works on deep learning for the classification of defects and some techniques. The core methods we propose are discussed in
Section 3. Here, we introduce the method’s architecture and the idea used. The experiments are presented in
Section 4. We show many experiment results for our proposed method and various comparison experiments. Finally, the conclusions of the study are presented in
Section 5.
5. Conclusions
The use of deep learning methods to identify the defects in solar panels is relatively common. This study discussed the problem of the identification of defects in solar panels. A neural network model was proposed in this study, LIRNet, and it was compared with the current common neural network models, such as VGG-16, VGG-19, GoogLeNet, ResNet, and EfficientNet, through identification accuracy and general performance evaluation. The experimental results indicate that LIRNet is better than most. In LIRNet, it is crucial to use the two-phase deep learning classification. It increased the training stability and slightly improved the recognition accuracy. The LIRNet model also makes some progress in the learning curve of training; that is, it is balanced compared to other common neural network models. LIRNet has better recognition accuracy than other neural network models; however, it needs higher recognition accuracy to be practically applied in real scenarios. The accuracy rate should be significantly improved to 95%. The learning curve of the model will not be severely overfitted, which is one of the research directions for the future. The K-means clustering of LIRNet can be proved from the experimental results to have specific utility for the identification problem of solar panels, but it may only be effective on some datasets. For example, this method of reducing the number of learning categories to improve the recognition accuracy could also be useful if applied to the recognition of a fruit. Therefore, a future task to do the method proposed in this study for different situations to verify the broad practicability of the experimental method.