A lightweight network for photovoltaic cell defect detection in electroluminescence images based on neural architecture search and knowledge distillation
Abstract
Nowadays, the rapid development of photovoltaic(PV) power stations requires increasingly reliable maintenance and fault diagnosis of PV modules in the field. Due to the effectiveness, convolutional neural network (CNN) has been widely used in the existing automatic defect detection of PV cells. However, the parameters of these CNN-based models are very large, which require stringent hardware resources and it is difficult to be applied in actual industrial projects. To solve these problems, we propose a novel lightweight high-performance model for automatic defect detection of PV cells in electroluminescence(EL) images based on neural architecture search and knowledge distillation. To auto-design an effective lightweight model, we introduce neural architecture search to the field of PV cell defect classification for the first time. Since the defect can be any size, we design a proper search structure of network to better exploit the multi-scale characteristic. To improve the overall performance of the searched lightweight model, we further transfer the knowledge learned by the existing pre-trained large-scale model based on knowledge distillation. Different kinds of knowledge are exploited and transferred, including attention information, feature information, logit information and task-oriented information. Experiments have demonstrated that the proposed model achieves the state-of-the-art performance on the public PV cell dataset of EL images under online data augmentation with accuracy of 91.74% and the parameters of 1.85M. The proposed lightweight high-performance model can be easily deployed to the end devices of the actual industrial projects and retain the accuracy.
keywords:
Defect detection, Photovoltaic cells, Electroluminescence, Deep learning, Neural architecture search, Knowledge distillation1 Introduction
The lifetime of photovoltaic(PV) modules is essential for power supply and sustainable development of solar technology. However, the PV cells are easily affected by various external factors. During the manufacturing process, minor operational errors may result in module damages. In addition, vibration and shock during transportation and installation may also cause module breakage. Defects such as cracks, solder corrosion, cell interconnect breakage can make PV modules unusable, and microcracks that are hard to observe will potentially affect future output power and lifetime[NDIAYE2013140, li2019thermo]. The above defects in PV cells may cause module failure during operation, which can lead to power reduction and even safety problems for the whole system[kontges2014review].
The current-voltage(I-V) curve is used for detection of defective PV modules. Changes in I-V characteristics can reflect those heavily degraded modules. However, tiny cracks can hardly affect I-V characteristics, thus are difficult to identify. These microcracks have the potential possibility of separation and degradation, which can seriously affect the future use[kajari2012criticality]. As described in some research, microcracks can cause power attenuation, the loss of which varies from 0.9% to 42.8%, and may cause hot spot effect[dhimish2020micro, abdelhamid2013review]. Besides I-V curve, infrared thermal(IRT) imaging[TSANAKAS2016695] is another technology which can be used to detect defects. The temperature of PV cells with defects is significantly higher than other cells around them. However, the hot spots of the PV modules are not necessarily caused by the defects. Other factors like object occlusion can also lead to the abnormal detection results. Also, microcracks which have not yet affecting power efficiency can not be recognized by IRT images with a relatively low resolution.
Due to the high resolution of imaging, electroluminescence (EL) imaging[fuyuki2009photographic] has become one of the most commonly used methods for defect detection of PV modules. EL imaging system is a non-destructive technology with high imaging resolution which can be used to detect microcracks[breitenstein2011can]. In EL images, cracks and other defects in defective PV cells appear as dark gray lines and areas. In early stage, traditional methods based on manual features are proposed to detect defects in EL images. These methods depend on large amounts of manual design experiments and the performances are limited.
Because of the strong feature capturing ability of convolutional neural network(CNN), the methods using deep learning have gradually become the mainstream to detect defects in EL images. However, while CNN has greatly improved the detection accuracy, it also requires more time and hardware resources, making it hard to be deployed in end devices of practical applications. In order to meet the requirements of both the accuracy and speed of defect detection in the industrial field, a lightweight and efficient detection network is required.
There are a few works[karimi2019automated, tang2020deep, akram2019cnn] which have proposed
lightweight CNN-based methods to detect defective PV cells in EL images. These lightweight CNN-based methods are all based on manual design, which require a lot of experiments to find a suitable network structure.
To obtain a lightweight structure for practical application with much less manual work, we introduce neural architecture search (NAS) into the defective PV cell classification task, which is the first method using NAS to automatically design networks in the field of PV defect detection.
Aiming at the automatic design of network architecture, NAS can reduce manual intervention and make better use of computing resources in an automated manner.
Since the defects can be any size, we propose a search space which can enhance features at different scales, obtaining a lightweight network architecture that can better extract multi-scale features.
To make better use of the prior knowledge, knowledge distillation is introduced to learn the priors obtained by the existing pre-trained large-scale model to improve the performance of the searched lightweight network. Different kinds of knowledge are transferred, including attention information, feature information, logit information and task-oriented information. The obtained lightweight network has a high performance, which even outperforms the existing large-scale teacher model.
The contributions of the proposed method can be summarized as follows:
-
1.
We propose a lightweight network structure for detection of defective PV cells with high accuracy of 91.74% and size of 1.85M parameters, achieving the state-of-the-art performance on public PV cell dataset[ELPV] of EL images under online data augmentation. The proposed model also has high accuracy on defective PV cells up to 94.26% on our private dataset.
-
2.
We introduce NAS to the field of PV cell defect detection for automatic lightweight network design, which reduces the workload of manual design. To detect defects with any size, the search space is designed by considering multi-scale characteristic into the network architecture.
-
3.
To make full use of the priors already learned by the existing large-scale network, we utilize knowledge distillation to transfer various prior knowledge into our model. We consider attention information, feature information, logit information and task-oriented information into the knowledge transfer process and the experiments prove the effectiveness of knowledge distillation to enhance the ability of recognizing defective PV cells.
2 Related work
2.1 Traditional methods
Some research uses traditional image processing methods to detect defects in EL images. These methods usually rely on the manually selected features.
Dhimish et al.[dhimish2019novel, dhimish2019solar] used the bit-by-bit OR gate method to process EL images and enhance crack images, but the detection accuracy and other results were not given. Tsai et al.[tsai2012defect] presented an independent component analysis technique, but finger cracks that have little effect on crack detection were identified as other cracks. Anwar et al. [anwar2014micro] proposed an improved image segmentation method based on anisotropic diffusion filter and support vector machine(SVM) was employed to detect micro-crack defects based on medium-sized datasets, but this method requires a higher level of pre-processing. Su et al.[su2019classification] improved a new feature descriptor, which combines central pixel gradient information with central symmetric local binary mode to obtain more recognizable defect features under uneven background interference.
In traditional image processing methods, edge gradient information is often used to describe the features. Due to the similarity between the change of edge gradient of defects and the grain under complex background, it is easy to be disturbed when distinguishing defects from background grains. At the same time, these methods tend to be applied only on small datasets, and their generalization ability is not strong.
2.2 Deep learning based methods
Due to the popularity of deep learning, surface defect detection of PV cells based on deep learning has become a research hotspot in this field. CNN is becoming a widely used detection method because of its strong feature extraction ability.
Sun et al.[sun2017defect] proposed a crack classification network based on LeNet5[LeNet5], which can classify four kinds of crack defects. Bartler et al.[bartler2018automated] designed an improved classification network based on VGG16[vgg] structure and explored the effects of a few oversampling and data expansion methods on performance improvement. Deitsch et al.[deitsch2019automatic] conducted two defect classification methods based on VGG19 and SVM, and contributed a PV cell dataset of EL images. Shou et al.[shou2020defect] presented an unsupervised defect detection method based on generative adversarial networks(GAN), but the stability of the network needs to be further discussed. Both studies by Liu et al.[liu2019surface] and Su et al.[su2020deep] improved the region proposal network on the basis of Faster-RCNN, and realized the detection of small cracks in PV cells.
These studies are based on existing networks by transfer learning or improvement on some layers and parameters. Compared with traditional methods using EL images, deep learning methods have better generalization ability and higher accuracy.
2.3 Lightweight methods
Most of the existing deep learning models are large, requiring high hardware deployment in field for PV cell defect detection.
To solve this problem, a few researchers proposed lightweight networks by manual design. Karimi et al.[karimi2019automated] designed a 4-layer CNN structure for classification of 3 kinds of defects. Tang et al.[tang2020deep] designed a 9-layer CNN structure and improved the performance with a mixture of GAN generation and traditional data augmentation. Inspired by VGG11, a 9-layer CNN structure was designed by Akram et al.[akram2019cnn] and validated on the public PV cell dataset[ELPV]. Wang et al.[WANG2022119203] utilized octave convolution to build a lightweight network with high inference speed. All these studies are based on manual network structure design, which are difficult and require a large amount of experiments. Besides, manual structure design depends a lot on the existing data and is less universal.
To reduce the manual workload in model design, we introduce neural architecture search(NAS) into the task of PV cell classification for effective automatic architecture design. For better model training, we transfer different prior knowledge already learned by large-scale model based on knowledge distillation. In this process, attention information, feature information, logit information and task-oriented information are exploited and transferred to enhance the performance of the searched lightweight model.
3 Methodology
An effective lightweight network is proposed in this section for detection of defective PV cell by NAS and knowledge transfer. To automatically design lightweight network, NAS is introduced to the field of PV cell defect detection for the first time. To detect defects with any size, the network architecture search space is designed by adding multi-scale characteristic. Then a variety of prior knowledge is transferred by knowledge distillation to make full use of the priors already learned by the large-scale network. The illustration of our method is depicted in Figure 1.
