A Review on Defect Detection of Electroluminescence-Based Photovoltaic Cell Surface Images Using Computer Vision

Abstract

The past two decades have seen an increase in the deployment of photovoltaic installations as nations around the world try to play their part in dampening the impacts of global warming. The manufacturing of solar cells can be defined as a rigorous process starting with silicon extraction. The increase in demand has multiple implications for manual quality inspection. With automated inspection as the ultimate goal, researchers are actively experimenting with convolutional neural network architectures. This review presents an overview of the electroluminescence image-extraction process, conventional image-processing techniques deployed for solar cell defect detection, arising challenges, the present landscape shifting towards computer vision architectures, and emerging trends.

1. Introduction

Climate change is a reality accepted universally across nations and strategies for suppressing the severity of its impact on the globe are a shared goal. A manifestation of the progress in addressing this issue can be gauged by the advent and since then continuous growth of photovoltaic (PV) installations in many developed and developing countries [1]. The government of the United Kingdom is keen on fulfilling its ambitious commitment to reduce emissions by at least 68% by the year 2030. The commitment is expected to cost up to £11.6 billion [2]. PV installations are based on the concept of electricity generation via light energy [3,4,5]. The sun provides a tremendous amount of free and abundant light energy [6,7], which photovoltaic solar cells [8,9] can collect and transform into energy without releasing harmful emissions into the environment, known as ‘green energy’ [10,11].

In order for the PV cells to provide an optimal performance, the manufacturing process involved in creating the cell surface plays an important role. The manufacturing process involves many subtle and rigorous steps from silicon extraction to wafer slicing, exposing the cell surface to various levels of stress both mechanical and thermal [12]. This warrants a stringent quality-inspection process, whereas the present manual inspection procedure has many drawbacks such as human fatigue, bias, downtime, labor costs, and low accuracy. Considering the success of deep learning, in particular, computer vision, in other domains [13,14,15,16,17,18], researchers are actively looking at the implementation of convolutional neural networks (CNN) for PV fault detection within a manufacturing setting. By doing so, the drawbacks associated with human inspection can be eliminated, for example, labor costs, fatigue, and bias.

Although many studies are available on post-deployment PV fault detection [19], this review focuses on evaluating the current research status of automated PV defect inspection within manufacturing sites, through the implementation of CNN architectures. The review looks in detail at the presented research, focusing not only on the achieved results but more importantly on the introduced frameworks, implemented methodologies, and their impact on enhancing the performance of the respective architectures.

The structure of the review paper follows with an introduction to micro-cracks and detection strategies. Then, conventional image processing techniques for PV defect detection are presented, followed by details on the key components of a CNN architecture. CNN-focused PV fault-detection work is then presented with a focus on evaluating the proposed methodologies, followed by discussion, a conclusion, and future directions.

1.1. Brief Introduction to Key CNN Components

Machine learning (a subfield of artificial intelligence) is typically utilized for projects or applications that are based on numerical data. It includes a number of algorithms that can be modified to meet the needs of a given application. Either manually or with the aid of extra algorithms, feature engineering and feature selection must be implemented accordingly. Deep learning can be helpful in this situation. Practically speaking, it is impossible to narrow down to the most useful features, especially when there is enormous amount of data with potentially thousands of features. As training advances, deep learning enables autonomous feature engineering across the buried layers via backpropagation. CNNs, the abbreviation for convolutional neural networks, are a subset of deep learning systems that use filters to process picture input. Filters are a necessary parameter for convolutional layers.

Convolutional theory is not new to CNNs or deep learning in general; instead, the technique has been employed in conventional image and signal processing for a very long period of time. For instance, there are a number of well-defined filters used in image processing that carry out specific actions on images including extracting edges, shapes, blurring, or sharpening the image [20]. The important distinction is that with deep learning, the values contained in these filters are acquired independently by the model during training, as opposed to being pre-defined by a person. This capability enables CNNs to extract more detailed and abstract features via filters that could have remained overlooked without them. A convolutional network’s kernels are initialized to a random distribution, such as Gaussian or normal. To ensure that each kernel may learn distinct features from a single image, various initialization criteria are used.

1.2. Filters

One of the most important parts of a CNN is a filter, which makes it possible to retrieve local feature representations for later processing as the internal architectural layers are added. It comprises spatial dimensions that include the width and height but are significantly smaller compared to the dimensions of the input image, yet penetrate the entire depth of the input image volume. The filter is moved over the spatial input image volume as part of the convolutional procedure, and the dot products between the filter and input values are computed at any location. The outcome of this operation is a 2D feature map that shows the filter’s reaction at every spatial position.

Activation of the specific filter is dependent on specific visual features, such as edges or shapes. This subsequently leads to multiple two-dimensional activation maps. The output volume is then produced by stacking the activation maps along the dimensional depth [21]. Following a convolutional operation, the spatial dimensions of the activation map (kernel/filter) are determined using the formula below. The formula includes a new term, ‘P’, which stands for ‘padding’. Padding is referred to as the introduction of ‘zero’ rows and columns to the picture’s borders in the effort to make up for the initial image’s reduction due to the convolutional process and to retain more data at the edges of the image.

$${\mathrm{n}}_{\mathrm{o}\mathrm{u}\mathrm{t}}=\left[\frac{{\mathrm{n}}_{\mathrm{i}\mathrm{n}}-2_{\mathrm{p}}-\mathrm{k}}{\mathrm{s}}\right]+1$$

where ${\mathrm{n}}_{\mathrm{o}\mathrm{u}\mathrm{t}}$ = feature output number, ${\mathrm{n}}_{\mathrm{i}\mathrm{n}}$ = Input feature number, $\mathrm{p}$ = border padding, $\mathrm{k}$ = kernel, and $\mathrm{s}$ = stride.

1.3. Activation Function Sigmoid

The logistic function, often known as the sigmoid function [22], was once a highly well-liked activation function. The equation above converts the input ‘x’ to a number between 0 and 1. We can see that the function is similar to a smoothed step function when we look at the visual representation of it in Figure 1. Because the function is non-linear, any non-linear mapping can be learned when it is utilized as an activation function in a neural network.

$$\mathrm{x}\to \frac{1}{1+{\mathrm{e}}^{-\mathrm{x}}}$$

When utilizing the sigmoid activation function in specific neural networks, particularly with deeper layers, there are various difficulties that can become apparent. The saturation at the two tails of the function in Figure 1 is the first problem. As we go closer to the tails, we can observe that the gradient becomes quite small. The problem is that neural networks learn by iteratively propagating gradient information from the output to the input in a backward pass until the optimum weights are found.

The vanishing gradient problem is a phenomenon where, when using a sigmoid function, the gradients may become smaller as the network depth increases so that the model is unable to learn anything valuable. In addition to this, in situations where the network is required to perform millions of equations, the function is computationally demanding, as shown by the formula. It is also to be noted that the function’s lack of a zero-centered domain is another disadvantage. Data normalization using a zero mean is a common approach throughout the data-processing stage. Therefore, it has been demonstrated in practice that a zero-centered activation function converges more quickly.

1.4. Activation Function Tanh and ReLu

The sigmoid function’s zero-centered problem is resolved by utilizing a tanh [23] activation function. The figure below (Figure 2), which is a visual depiction of the function, shows that the sigmoid function has been scaled and recentered. With an output range of −1 to 1, the ‘S’ shape denotes the non-linear character of the system. The function can converge significantly more quickly than the sigmoid function because the ‘zero-center’ problem has been solved.

$$tanh\left(x\right)=2\sigma \left(2x\right)-1$$

According to Figure 3 below, the ReLu [24] activation function is fundamentally distinct from the sigmoid and tanh functions. The equation sets zero as the value of any input less than zero; however, an input that is larger than zero is maintained at its original value.

$$\mathrm{m}\mathrm{a}\mathrm{x}(0,\mathrm{x})$$

1.5. Activation Function Sigmoid

Backpropagation [25,26] is the abbreviation for backward propagation of errors and is among the most essential parts of a neural network that enables model optimization and the decrease in a network’s error. Calculating a loss function’s gradient with respect to each network weight is the purpose of backpropagation. Once the model has made a prediction, or after the output layer, backpropagation is started. By computing the appropriate derivatives, the goal is to modify the network’s weights. It displays the output layer’s derivative. Shown is the output layer’s derivative, where ${\mathrm{Y}}_{\mathrm{Z}}$ = predicted output and ${\mathrm{t}}_{\mathrm{Z}}$ = actual output.

$$\frac{\mathsf{\Delta }\mathrm{E}}{\mathsf{\Delta }{\mathrm{Y}}_{\mathrm{Z}}}={\mathrm{Y}}_{\mathrm{Z}}-{\mathrm{t}}_{\mathrm{Z}}$$

The formula below is used to update the weights.

$$\mathrm{w}1=\mathrm{w}1-\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{n}\mathrm{i}\mathrm{n}\mathrm{g} \mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}\ast \frac{\mathsf{\Delta }\mathrm{l}}{\mathsf{\Delta }\mathrm{w}1}$$

The calculation for w1 is shown in the formula above. The updated value for w1 is obtained by multiplying the learning rate by the partial derivative of w1 and subtracting the present value of w1. In order to allow for the updating of weights, the partial derivatives of the loss derived via the difference between the expected and actual output are generated after each forward epoch.

Calculating the partial derivative of the input with regard to the loss is the objective here. We accomplish this by computing the partial derivative with reference to the loss until we reach each intermediate weight. The new value of the input weight is then calculated by multiplying all of the partial derivatives and the input derivative with respect to the loss. The chain rule of calculus is a technique for determining the derivative of an input with regard to a loss, as illustrated below:

$$\frac{\mathsf{\Delta }\mathrm{L}}{\mathsf{\Delta }\mathrm{w}1}=\frac{\mathsf{\Delta }\mathrm{L}}{\mathsf{\Delta }\mathrm{e}}\ast \frac{\mathsf{\Delta }\mathrm{e}}{\mathsf{\Delta }\mathrm{ẑ}}\ast \frac{\mathsf{\Delta }\mathrm{ẑ}}{\mathsf{\Delta }\mathrm{w}1}$$

2. Micro-Cracks

Micro-cracks are a type of PV cell degradation, sometimes referred to as cracks or micro-fractures [27]. Due to its fragility, processed silicon is susceptible to harm from minute differences at any point in the life of the cell. Imperfections are spawned by these inconsistencies and begin to spread on the cell surface. The size of the initial difference determines how long these micro-cracks will be; some may just cover a tiny portion of the cell surface, while others will cover the entire surface [28]. Micro-cracks have the potential to affect the total life of the entire PV installation as well as the output efficiency.

2.1. How Are Micro-Cracks Formed

Micro-cracks can develop at any point during the PV lifecycle, including production, shipping, installation, and post-deployment. They have the potential to develop into regional surface loss, which lowers the output of both the micro-crack-containing cells and the entire PV string that contains the defective cell. In addition, unnoticed micro-cracks can cause hot spots or diode activation, which pose a larger safety risk because they could potentially lead to fire breaking out within the cells.

The manufacturing process is where micro-cracks initially develop. The constant handling of the cells necessary for the production processes unfortunately puts them under thermal and mechanical stress [19]. When copper and silicon experience different rates of thermal expansion above 300 °C, micro-cracks may form during the manual cell-soldering process. These possible temperature differences may cause the substrate to develop micro-cracks, which would increase the cell resistance [29].

The lamination of PV modules is another production step that could lead to the development of micro-cracks [30]. Manufacturers who use low-quality ethylene vinyl acetate (EVA) with weak moisture impermeability during the panel lamination process run the danger of micro-crack formation when moisture tries to enter inside due to the non-standardization of the production process [31]. If even a slight quantity of moisture develops in this situation, its elimination through the process of evaporation is not assured. This raises the possibility of moisture penetrating the laminate’s face and oxidizing the PV contacts.

2.2. Detectecion of Micro-Cracks

Micro-cracks that are not detected have the potential to grow over the cell surface substantially over time, intensifying their negative effects on the lifespan and performance of the PV string. Its appeal stems from the fact that EL enables the discovery of concealed flaws that would otherwise be impossible to detect using techniques such as flash testing, voltage–ampere (V–A) characteristics, or infrared (IR) imaging using thermal cameras [32].

The EL procedure is effectively an X-ray of the PV cells through image analysis, exposing inactive regions (defects) that are otherwise hidden from human sight. Ideally, EL cell testing needs to be completed prior to and succeeding each significant stage of construction, during the manufacturing process, such as prior to module lamination, so that the defective cells can be replaced, and subsequent to lamination to ensure no micro-cracks result from the lamination process. However, not all manufacturers are able or willing to invest in their quality-control process output due to the cost, time, and domain expertise associated with the procedure. Inadequate quality control in the PV production environment can negatively affect the performance of the installation as a whole and, depending on the deployment’s size, cause clients to suffer significant financial losses.

EL testing is often carried out twice by Tier One manufacturers while producing solar panels to filter out and reject faulty panels. This is particularly crucial when dealing with MW/GW projects that necessitate a large number of panels to be installed. However, due to their subtle composition, even Tier-One-rated solar manufacturers occasionally misinterpret or neglect to identify micro-cracks, resulting in the shipment of faulty cells to the customer.

3. Electroluminescence (EL) Measurement Approach

The ability to assign defects to specific manufacturing processes and to identify different failure modes is a significant benefit of the EL approach. However, EL inspection can be performed at any point in the PV lifetime. For instance, flaws may appear as a result of production inconsistencies evading quality inspections, improper handling during transit, or environmental conditions and post-deployment degradation [33].

From an administrative perspective, the aforementioned factors may have a variety of effects on warranty claims, logistics, operating procedures, and maintenance activities. While EL examination of PV modules prior to delivery to the client site is desirable since manufacturing-related flaws can occur, EL inspection can also be carried out after installation if an unexpectedly low power output is discovered [34,35].

If an EL inspection is requested after deployment, IR monitoring should be performed first in order to locate any PV modules that have anomalies so that they could be taken out of the PV string before being sent for EL inspection. In addition to this, due to technical improvements such as the widespread use of drones, on-site EL inspection is now possible without disconnecting localized modules. However, a near-dark condition and particular equipment, such as an EL camera, are necessary beforehand [36]. Figure 4 presents the standard process for EL-based PV cell surface image procurement. A current is fed into the terminals of the solar cell via an external power supply. This enlightens the cell surface; however, this is not visible to the human eye [36]. Hence, an EL camera is used for capturing the illuminated cell surface image [37]. When inspecting an EL image, the illuminated surface corresponds to the active area of the cell surface, with darker regions representing busbar connections in a periodic manner. Any other darker regions appearing sporadically imply defective regions [38,39].

The length of time required to conduct a meaningful EL examination of a PV module and ongoing focus adjustment with relation to the PV cells, however, pose a significant limiting factor for post-deployment, on-site EL inspection. Due to the lack of periodicity associated with the occurrence of micro-cracks and the potential for multiple micro-cracks to reside on the surface of a single cell, it is challenging to determine the impact of each irregularity with respect to the module’s power losses established via the current–voltage ‘I–V’ curves.

Infrared vs. Electroluminescence Imaging

The primary building block of infrared imaging [40] is the derivation of temperature differences between various components of the module/strings. It has been demonstrated that inactive regions are warmer than active ones. The PV module’s temperature rises in inactive parts as a result of the energy for that region not being converted into DC power and instead remaining as excess heat [41].

Depending on the type and extent of the defect, there may be a very small temperature increase, and the variation is typically proportional to the irradiance. As a consequence, at low irradiance, it is harder or sometimes impossible to find some faults, including micro-cracks. IR is typically used to identify defects such as ‘hot-spots’ [42] that are brought on by weak electrical connections or internal module mismatches that result in significant local heat gradients.

Non-destructive measurement methods such as EL and IR produce real-time, high-resolution pictures with a two-dimensional distribution of the module feature properties [43]. IR would not be the best measurement technique for identifying micro-cracks, because the temperature difference between the normal and faulty regions is not always as great as it is with micro-cracks. On the other hand, the EL method makes use of the radiative recombination of stimulated photons inside the cell. After providing an external excitation current, an EL camera is used to collect the surface-emitted photons. The defective cell areas radiate less as a result of the external excitation, making them seem darker than the sections that are functioning well. EL inspection’s improved cell surface resolution allows for more accurate problem diagnosis than IR imaging in some cases. This study [44] shows that EL examination is a useful method for finding flaws such as micro-cracks. On the other hand, it also emphasizes that EL is incapable of providing information on the impact’s quantification with respect to the module’s power output.

Each of the aforementioned mechanisms has some advantages and disadvantages. Hence, the extent of the application and the testing goals should be taken into consideration while choosing the inspection mechanism. IR-based methods are suitable for applications that need precise measurements of the current, while bulk defect imaging applications benefit from EL, particularly micro-cracks, series resistance image data, and breakdown junction imaging [45]. The Chart below (Figure 5) gives a summary of the benefits and drawbacks of both measuring methodologies.

4. Conventional Image Processing

The identification of PV-cell-based faults has been implemented by researchers using conventional image-processing techniques. On a 550 × 550 pixel cell image, Du-Ming Tsai et al. [46] offer a ‘self-reference’ technique via Fourier image reconstruction with a 0.29 s inference speed. The investigation’s faults, however, are described by the author as appearing as ‘line or bar’-shaped objects. Due to the variety and irregular structure of these cracks, the investigation is thus restricted to a very tiny subset of them. By zeroing out the frequency elements linked to the line- or bar-shaped faults, the technique seeks to eliminate flaws. After that, the spectral image is ‘back-transformed’ into a spatial image. As a result, by comparing and evaluating the changes in grey levels between the original and converted images, the defect location can be located.

The authors further emphasize that the suggested method is appropriate for finding ‘small cracks’. This brings to light a significant flaw in traditional image processing—namely, the requirement for creating unique features to meet a limited set of clearly defined issues. We are aware that ‘cracks’ can have any size in relation to the PV cell and any simple or complex shape. The efficiency of standard image processing is limited since it may not be able to handle a wider variety of flaws due to the necessity for bespoke feature engineering and specific processing. Deep learning circumvents this limitation by concentrating on data generalization through an iterative process of weight optimization supported by backpropagation. This eliminates the need for manual feature engineering.

The automated machine vision of material surfaces uses the image-processing approach of texture analysis [47]. This technique is focused on surfaces with homogeneous textures, meaning the textural patterns have a high element of repetition or are occasionally present, together with other techniques from the image-processing domain. A set of ‘textural features’ is extracted through the use of spectral or spatial domains throughout the textural analysis procedure. A high-dimensional classifier is then used to identify any flaws that prevent the background’s texture from being uniform.

4.1. Spatial Processing

The most-often-employed spatial-domain approaches are those that generate their co-occurrence matrices from spatial grey-level co-occurrence matrices [48]. The inspection of web surfaces [49], machined surfaces [50], cotton contamination [51], and textile fabrics [52] are a few examples of the applications. The extraction of textural information from an image’s power spectrum using Fourier transforms is a commonly used method in the spectral domain [53]. Applications include the inspection of textile fibers and semiconductor wafers (see references [54,55,56]). When inspecting granite [57], textile fabrics [58,59], and steel [60], Gabor filter banks with different frequencies and phase angles are frequently employed to represent textural qualities. Wavelet transforms are yet another effective approach for multi-resolution analysis for textural feature extraction [61,62]. The wavelet energies of decomposed sub-images at various frequency ranges and channels are used in this instance to represent the discriminant qualities. The inspection of semiconductor wafers and textile fabrics have both been conducted using wavelet transforms [63,64,65].

4.2. Homogenous vs. Heterogeneous Compositions

The aforementioned methods lean closer toward conventional image processing in general. They offer local feature-extraction methods that, in terms of modelling, can be quite effective in finding defects in surfaces with uniform textures. We can take advantage of the benefits of weight optimization by employing deep learning, which eliminates the necessity for human feature-representation creation. Additionally, deep learning offers models with greater generalizability than traditional image processing, improving the performance on diverse data.

5. Convolutional Network for Photovoltaics

Artificial intelligence has been the subject of research, particularly computer vision, to eliminate the drawbacks associated with human inspection and boost solar cell production output efficiency. Convolutional neural network (CNN) architecture is suggested by M. Waqar Akram et al. [66] for the detection of faulty PV cells, as shown in Figure 6. The researchers present their findings of an ‘isolated-model’ (98.57%) that was tuned for EL-based images via transfer learning, which resulted in a 99.23% accuracy rate. The authors claim that because there are only about 800 images in the dataset, a deep architecture would lead to overfitting. There are two major phases of the research. The initial stage centers on creating a CNN using EL images. The EL-trained CNN is utilized as a pre-trained model in the second phase and is adjusted based on infrared images (IR) of faulty cells. The accuracy of the pre-trained model (EL-based) was 98.67%. Although the reported accuracy is outstanding, a closer examination of the approach, particularly the data collecting and pre-processing, reveals that faults were artificially added to the cell pictures rather than true defects. The authors select data augmentation in order to increase the size and variation of the dataset. Naturally, this results in a 6.5% improvement in the model’s accuracy. The four-block CNN network with a fully linked layer feeding into a SoftMax function was chosen by the authors, after considering the CNN architecture in detail. The choice to utilize the Adaptive Moment Estimation (ADAM) optimizer is not supported by a comparison to other optimizers such as the Stochastic Gradient Descent (SGD), but rather by the fact that it is well-recognized and widely used.

A CNN is suggested by Sachin Mehta et al. [67] for the detection of PV soiling and other faults. The authors concentrate on both the location of flaws on the PV cell as well as their detection. Traditionally, the process of classifying and localizing images is referred to as object detection. Bounding boxes are created for the locations of the points of interest as a conventional strategy to solving object-detection problems, and these boxes are subsequently used for supervised model training. The authors suggest, however, replacing bounding boxes for each image with power loss, as shown in Figure 7. Although this can be considered a novel contribution, the dataset utilized is of a considerable size containing more than 45,000 images to train the network, bringing into question the generalizability of the proposed methodology on smaller datasets. Furthermore, the hardware specifications listed by the authors, such as the use of an NVIDIA TitanX GPU, which is not commonly available for PCs, might be used to estimate the computing load of the network. However, the model can only run at 22 frames per second (FPS).

Three models are compared by Ahmad et al. [68] for the EL-image-based identification of solar cell faults. Random forest (RF) and support vector machine (SVM), the first two models, are essentially members of the machine learning (ML) family, while CNN is a deep learning (DL) model. According to the authors, a CNN with two convolutional blocks and a fully linked layer can deliver accuracy rates of more than 99%. The dataset’s augmentation, which expanded the original data by four times, is credited by the authors as a major factor in this outstanding outcome.

An SVM and a CNN network are suggested by Sergiu Deitsch et al. [69] for multiple-defect detection in EL-based solar cell images. The authors assert that SVM (82.44%) and CNN (88.42%) both offer good accuracy. Before examining the methodology used by the best performance (CNN), it is important to note that models that achieve under 90% accuracy are no longer considered to be ground-breaking due to advances in deep learning, data scaling and transforming techniques, and regularization. Investigating the CNN in detail reveals that the authors use the VGG-19 architecture to subscribe to the transfer learning domain. The fine tweaking is carried out by the writers in two stages. The fully connected layer’s weights are initialized at random in the first stage using ADAM as the weight update optimizer. The weights for all layers are refined in the second stage. The second stage of the algorithm is used with stochastic gradient descent (SGD), according to the authors, who also note that the ‘momentum’ value was adjusted at 0.9. Noting that ’momentum’ is a hyperparameter utilized in SGD-M, which differs from SGD, is crucial. The rationale for changing the ADAM optimizer in the second round of fine-tuning is also unclear.

A Mask-RCNN with a RESNET-101 backbone architecture is suggested by Yang Zhao et al. [70] for the detection of flaws in EL-based solar cell pictures. The authors’ initial goal is to classify 19 different types of faults. They attain a mean average precision (MAP) of 70.2% after model testing. This measurement is based on a 0.5 MAP localization. This means that the predicted bounding box must have at least a 50% intersection over union (IoU) between the ground truth and the forecast bounding box in order to be considered a valid prediction. The methodological strategy is subject to further examination because the authors chose to use a screening approach after achieving a modest level of accuracy. The dataset was divided into three types of defect severity. The defects that the model had the hardest time identifying were given the lowest severity ratings and were not treated as faults as a result. The MAP was dramatically raised to above 90% thanks to this strategy. We believe a more successful strategy would have been to attempt to improve the model’s generalization to these defects by attempting to enhance the variance of the damaged cells through targeted data augmentations. Additionally, to improve the generalizability of the model, hyperparameter optimization of the network itself might have been examined instead of merely eliminating the low-performing flaws using criteria-based screening.

With an accuracy of 91.58%, Ashfaq Ahmad et al. [71] suggest EL-based solar cell images with a CNN architecture for fault detection. The authors stress the value of data augmentations in scaling the dataset and introducing variance before going over the architectural considerations. The input image is used to start the chosen CNN architecture, which is then followed by four convolutional blocks with a total of 32 filters. The number of filters increases to 64 in the following two convolutional blocks, and there are 128 filters in the final two blocks. Eight convolutional blocks make up the entire designed CNN architecture, which feeds its output through a single fully connected layer. The justification for eight convolutional blocks is unclear, despite the fact that as the model becomes more sophisticated, increasing the number of kernels is a typical rule of thumb. Additionally, there is a lack of an intuitive explanation; for instance, was the substantial variation in the types of defects what drove the number of convolutional blocks? Furthermore, rather than being chosen through an experimental/optimization technique, the choice of the optimizer ADAM and the learning rate 0.001 appears to be made automatically. By changing the learning rate, for example, the authors may have found that the model could keep its accuracy with a higher learning rate, i.e., 0.02, leading to a shorter training period.

A CNN is suggested by Wuquin Tang et al. [72] for the detection of flaws in EL-based cell pictures. The implementation of a GAN network for data augmentation is, according to the authors, their first significant contribution. The justification for scaling the dataset with a GAN network and adding more variance is not quite evident. A comparison of the accuracy between the use of GAN and popular data-augmentation methods already present in deep learning frameworks such as TensorFlow, Pytorch, and Keras may have been helpful in determining the applicability of GAN. The overall accuracy was 83% after employing GAN for data augmentation, which demonstrates the technique’s inapplicability in this particular situation. Additionally, because GAN is a network in and of itself that is used to create new images, it is significantly more computationally intensive than utilizing conventional data augmentation, necessitating an excessive allocation of resources. The authors’ experiments to choose the number of kernels have yielded useful results, which is a plus. Based on their findings, the authors discuss how, up to a point, increasing the number of kernels can greatly enhance the model accuracy. After that, an increase in the kernels will not have a favorable effect on the model and can even cause overfitting.

A deep learning strategy using features is demonstrated by Christopher Dunderdale et al. [73] for the detection of solar cells with defects. We will concentrate on analyzing the approach related to the creation of the deep learning models because we have an interest in the deep learning aspect. For comparison, the authors train the PV dataset on the MobileNet [74] and VGG-16 [75] architectures. The authors’ decision to compare SGD and ADAM outcomes is commendable, rather than only using ADAM as the ‘off-the-shelf’ optimizer. When we examine the VGG-16-trained architecture in more detail, we discover that an SGD optimizer with data augmentation—horizontal and vertical flipping, and rotation—produced the best performance (85.8%). The ADAM optimizer yielded an inadequate accuracy of 27.4% for the identical conditions. By applying data augmentations such horizontal and vertical flips and using ADAM as the optimizer, however, the MobileNet architecture was able to obtain the best accuracy (89.5%). The ‘paradigm-shift’ in the findings that occurred after switching between the two optimizers and architectures is not explained by the authors. Understanding that depth-wise [76] convolutions are used rather than the usual convolutions, which can reduce the computations by up to nine-fold, makes MobileNet more computationally efficient and lightweight.

A CNN based on the VGG-16 architecture is proposed by Pierdicca R et al. [77] for the detection of faulty PV cells; the classification pipeline is presented in Figure 8. One of the explanations given by the authors for their choice of VGG-16 is its ease of implementation. However, numerous frameworks, such as Pytorch and Tensorflow created by Google and Facebook, respectively, now make it easy to use various SOTA pre-trained models. As a result, choosing an architecture should be dependent on the properties of the dataset rather than its ease of implementation. The lack of batch normalization inside the network’s convolutional layers is one of the author’s own admissions of the consequences of choosing the VGG-16 network. Based on the information we can understand, the authors do not offer any computational details; if they did, we believe the model training and convergence times would be much longer than they are for models that use batch normalization [78]. The results are considered modest, despite the authors’ claim that using data augmentation enhances the model performance, and the authors cite the large imbalance in the damaged photos as the cause. Instead of depending solely on 0.001, the model might have been further fine-tuned for greater accuracy by altering the learning rate.

A weakly supervised surface-defect-detection architecture has been suggested by Haiyong, Chen, and colleagues [79] for filtering anomalies on diverse surface textures, including solar cells. The authors describe a fused design that combines a random forest (RF) classifier with a CNN, claiming that this architecture is more resistant against complex background filtering. There are 5915 faulty cells and 15,330 normal PV cells in the dataset under examination. The five convolutional blocks that make up the CNN architecture are followed by the RF classifier and the attention network. Given the obvious imbalance in the dataset and the authors’ cross-validation results of 93.23% accuracy, it is applaudable that they used K-Fold with K = 5.

On production lines, Bolun Du et al. [80] reveal their research on detecting PV cell defects. The authors train three state-of-the-art (SOTA) models—GoogleNet, LeNet, and VGG-16—by adhering to the well-known fine-tuning method. The best performer is shown to be GoogleNet, which finished with a loss of 0.002 and 100% accuracy after 81 epochs. Because our dataset was likewise trained on this architecture and produced outstanding results, we are able to corroborate the great performance of GoogleNet. However, we went a step further and evaluated each model’s performance not just in terms of accuracy but also in terms of architectural and post-deployment criteria. Despite GoogleNet’s high level of accuracy, the model has 13 million learnable parameters, compared to 6.42 million in our suggested architecture.

A two-stage procedure is suggested by Mustafa Yusuf et al. [81] for the detection of various EL-based solar cell faults. First, the dataset is scaled using a variety of data-augmentation approaches, paying special attention to the defective classes. The authors train four well-known models, DarkNet-19, ResNet-50, VGG-16, and VGG-19, using transfer learning to extract features from the enhanced dataset, as shown in Figure 9. Following the selection of the most distinctive features, classification is carried out using several machine learning architectures, such as support vector machines and random forest. We also investigated this strategy while focusing on the first stage of the study, or transfer learning, but we believed that the pre-trained network designs were quite deep and consequently highly computationally intensive. We also demonstrate how feature extraction is directly possible with just one carefully constructed CNN architecture. With an overall F1-score of 98.0%, we employ our suggested architecture for feature extraction as well as classification.

Transfer learning is a method presented by Nawei Zhang et al. [82] for identifying flaws in poly-crystalline solar cells. The authors train a deep adoption network (DAN) using a labelled mono-crystalline dataset before testing it on an unlabeled poly-crystalline dataset, citing similarities between mono-crystalline and poly-crystalline. The 44 photos that make up the initial dataset scale to over 2000 images after augmentations. What kind of data augmentations were used is not entirely known. In order to compare the type of augmentations used with those used by others, we believe it is crucial to reveal the augmentations that have been made. Additionally, using generic augmentations without domain-specific implementation can cause model drift, which means that after deployment, despite potentially having a highly exceptional performance, the model would be unable to generalize on the true nature of the PV cells. Moreover, the authors base their architecture on ResNet-50. We demonstrate that the ResNet family of architectures, despite offering higher accuracy when measured on a variety of metrics other than accuracy, such as computational–architectural complexities, latency, and FPS, can be very demanding, and therefore is not the ideal option for a manufacturing facility deployment close to the data source.

Hussain et al. [83] present a custom lightweight architecture for automated micro-crack detection in production based on EL images of PV cell surfaces. The authors highlight the issue of representative data scaling and propose a gradient guided filter-tuning mechanism for capturing augmented images from within the developed architecture at certain feature-map checkpoints, as shown in Figure 10. The authors also highlight the importance of a lightweight architecture with respect to deployment on a production floor. Comparing the proposed architecture against other state-of-the-art architectures, the authors demonstrate their architecture as the most efficient in terms of FPS and latency while maintaining an impressive F1-score of almost 99%.

6. Discussion

This paper has presented a review of EL-based PV cell inspection for micro-cracks based on computer vision architectures, in particular focusing on the CNN-inspired architectures. From the review, it is clear that there is a shift from conventional image-processing techniques towards deep-learning-based automated detection for automated EL inspection. The main reason for this is due to the lack of generalizability of hand-crafted, feature-based algorithms to maintain a high performance on datasets with slight variations. This is critical as PV-manufacturing facilities vary in their environmental settings, such as production floor light intensities, production line configurations, etc. CNN architectures, being part of the deep learning domain, address this loophole by providing automated feature extraction via a learning process facilitated through backpropagation.

However, CNNs also have shortcomings, such as the requirement of large training samples, which can be difficult to procure from within a manufacturing facility. It is evident from

Table 1. CNN work comparison.

Reference	Year	Dataset Size	Contribution	Accuracy	Architecture
[66]	2020	~800 images	Artificial fault generation and transfer learning for better generalization	99.23%	Custom CNN
[67]	2018	45,000	Input specific feature maps learned via input reinforcement at the different levels of network applying a bidirectional input-aware fusion (BiDIAF) block	~95%	Custom CNN
[68]	2019	5400	Automated data-analysis mechanism developed for pre-processing electroluminescence module images and categorizing images to be utilized for machine learning.	99%	Custom CNN
[69]	2021	2624	Multiple defect detection in EL-based solar cell images.	88.42% (CNN)	Custom CNN VGG-19
[70]	2020	5983	Detection of flaws in EL-based solar cell pictures.	90%	Mask-RCNN with a RESNET-101
[71]	2020	2624	EL-based solar cell images with a CNN architecture for fault detection	91.58%	Custom CNN
[72]	2020	2250	Use of GAN network to detect the flaws in EL-based cell pictures compared with VGG16, ResNet50, Inception V3, and MobileNet.	83%	Custom CNN
[73]	2019	798	Detecting and classifying defective photovoltaic modules using thermal infrared images.	85.8% (VGG 16) 89.5% (MobileNet)	VGG 16 MobileNet
[77]	2018	3336	To show the degradation problem and evaluate the proposed method.	75%	VGG 16
[79]	2020	Around 24,000	Combination of random forest with CNN, claiming that they are more useful than complex background filtering.	93.23%	Custom CNN
[80]	2020	Not mentioned	Intelligent classification method for efficient defect detection.	100% (GoogleNet) 97.67% (VGG-16)	GoogleNet LeNet VGG-16
[81]	2021	2624	Lightweight CNN proposed to detect defects.	98% (DarkNet-19)	DarkNet-19 ResNet-50 VGG-16 VGG-19
[82]	2020	2000	Method to identify faults in poly-crystalline solar cells	77%	ResNet-50
[83]	2022	777	Custom lightweight architecture for automated micro-crack detection in production based on EL images of PV cell surfaces.	99%	Custom CNN
[73]	2019	798	Detecting and classifying defective photovoltaic modules using thermal infrared images.	85.8% (VGG 16) 89.5% (MobileNet)	VGG 16 MobileNet
[77]	2018	3336	To show the degradation problem and evaluate the proposed method.	75%	VGG 16
[79]	2020	Around 24,000	Combination of random forest with CNN, claiming that they are more useful than complex background filtering.	93.23%	Custom CNN
[80]	2020	Not mentioned	Intelligent classification method for efficient defect detection.	100% (GoogleNet) 97.67% (VGG-16)	GoogleNet LeNet VGG-16
[81]	2021	2624	Lightweight CNN proposed to detect defects.	98% (DarkNet-19)	DarkNet-19 ResNet-50 VGG-16 VGG-19
[82]	2020	2000	Method to identify faults in poly-crystalline solar cells	77%	ResNet-50
[83]	2022	777	Custom lightweight architecture for automated micro-crack detection in production based on EL images of PV cell surfaces.	99%	Custom CNN

that the vast majority of the reviewed articles focused on CNN-inspired PV defect detection are based on datasets composed of less than 10,000 images, with the exception of [67,79].

Researchers are actively investigating into addressing the issue of data scaling not only through generic data augmentations but also through generating representative data samples by modeling the EL-extraction process and capturing the resultant feature maps at various internal architectural checkpoints, as recently presented by [84] and tested with various types of CNN architectures, regularization, and optimization techniques.

Another emerging trend found within the literature was the comparison of developed or proposed architectures against other state-of-the-art architectures, as presented in

Table 2. Multiple architecture comparison.

Ahmad Maroof Karimi et al.	[68]	RF, SVM, Custom CNN
Sergiu Deitsch et al.	[69]	SVM, VGG-19
Christopher Dunderdale et al.	[73]	VGG-16, MobileNet
Bolun Du et al.	[80]	LeNet, GoogleNet, VGG-16
Mustafa Yusuf et al.	[81]	DarkNet-19, ResNet-50, VGG-16, VGG-19
Muhammad Hussain et al.	[83]	Custom CNN, ResNet-18, AlexNet, GoogleNet, MobileNetV2

. This is appreciated as it provides a manifestation of the efficiency of a proposed architecture and justifies its selection. In particular, for custom-developed CNN architectures, it is important that researchers evaluate them against state-of-the-art architectures in order to substantiate the need for a custom network. For example, Hussain et al. [83] demonstrated the effectiveness of the custom-designed CNN architecture by comparing it against various architectures on a wide range of metrics including accuracy, architectural, computational, and post-deployment performance.

7. Conclusions

This review primarily focused on the implementation of CNN architectures for automated defect detection within EL-based PV cells. An overview of the EL-extraction process along with conventional image-processing techniques was presented prior to the CNN literature in order to present a complete and compact review of the domain. It is clear from the number of research papers and the recent publication dates that more researchers are taking an active interest in CNN-based defect detection of PV-cell defects due to its numerous benefits, such as automated feature extraction, leading to better generalizability, reduced labor cost, reduced bias and downtime, and improved accuracy as a result of automation.

However, to accelerate the advancement of CNN-inspired visual PV defect detection in a more reliable and robust manner, there is a need for PV-focused benchmark datasets. The advent of benchmark PV datasets would enable developers to rigorously test their CNN architectures in a universal manner, leading to progressive advancements that can be benchmarked and lead to deployable solutions within manufacturing facilities. Interestingly, the majority of the reviewed studies presented in Table 1 utilized limited data samples for training and reported impressive accuracies, such as [83] achieving 99% based on a dataset consisting of only 777 images. Although this is impressive, there is a question mark on whether the trained architecture has truly generalized on the underlying distribution of PV visual faults. The ‘true generalization’ question reinforces the case for benchmark datasets containing a larger sample and wider distribution statistics for validating the performance, i.e., the generalization capacity of the developed architecture.

In addition to the need for benchmark datasets, we anticipate more research focus on custom CNN development aimed at providing architectures that are not only highly performing with respect to accuracy but also computationally lightweight and hence deployable on standard hardware without requiring additional graphical processing units (GPUs). As shown in Table 2, various studies have gone beyond evaluating only their own architectures but have focused on a comparison with other architectures on a broader range of metrics in order to prove the high efficacy of their architectures. For example, [83] goes beyond evaluating its custom architecture in conventional terms, i.e., validation accuracy, by being inclusive of other metrics such as the computational complexity, number of parameters, etc.