Innovative high-speed method for detecting hotspots in high-density solar panels by machine vision

Defects in production

Photovoltaic cells with defects might have a non-uniform density of current. This may occur because of defects in the material itself or errors during manufacturing (Deng et al. 2017). This causes a mismatch between currents in cells that results in decreased output power. Because the cells are in series, the total current is limited to the current of the cell producing the least amount. During the operation, the cells with defects heat and are prone to irreversible damage (Waqar Akram et al. 2020).

Degradation

With the increasing tendency towards thinner panels, silicon wafers are more prone to crack development. These cracks may develop at any time during the panel’s lifetime (Akram et al. 2019). They mostly start during the water production and develop further during the operation of the photovoltaic panel. These cracks cause a mismatch in the produced current of the cells and in return cause dissipation of the power produced by healthy cells (Li et al. 2018).

Shading

Hot spots resulting from external agents such as shading of trees or birds, etc., and soiling are usually temporary, because the source of the hot spots may be removed after a while. For example, the shadow of a tree moves at different hours of the day. Shading is one of the most common causes of hot spots (Hanifi, Schneider, and Bagdahn 2015).

Junction failures and bad connections

Junction boxes connect the strings of photovoltaic cells and the output terminals. Also, junction boxes contain a diode for each string of photovoltaic cells. Bad connections and soldering can cause high resistance between the terminals, diodes, and cells. This high resistance can cause heat and, in turn, Hot spots.

Bypass diodes

Most photovoltaic cell brands use a bypass diode with opposite polarity for the cells and parallel to the cell modules. Hence if a cell is shaded, it will be cut off from the circuit.

In Figure 1 the reverse current of a photovoltaic cell is shown. When the saturation current of the diode is increased, the R _sh is reduced (Yang et al. 2010). Normally the photovoltaic cells are forward biased, and the bypass diode is reverse biased and no current flows through the diode. When a cell has an issue and its output current is reduced, it will become reverse biased, which in turn causes the bypass diode to be forward biased. The current produced by healthy modules flows through the bypass, diode and the hot spotted region is cut off (Figure 2).

Figure 1:

Reveres bias in a photovoltaic cell (Yang, Wang, and Wang 2012).

Figure 2:

Function of the bypass diode during the occurrence of hot spot (https://www.alternative-energytutorials.com/photovoltaics/bypass-diode.html/, 2021, validation data3).

The heat effect

Photovoltaic panels have a better performance at lower temperatures. Their output power is much higher in lower temperatures rather than in higher degrees. When a photovoltaic panel is subjected to sun rays, it heats up. Most of this heat is due to the absorption of infrared radiation. This increase in the panel’s temperature is significant. We know from experience that the surface temperature of photovoltaic panels can reach up to 100°C. Most manufacturers of photovoltaic panels accommodate users with information about the heat effect on the performance of the photo-voltaic panel. In this information, a factor is introduced that is known as the temperature coefficient of the nominal output power of a photovoltaic cell. It is reported as a percentage of power reduction per each centigrade degree of temperature increase. In Figure 3, the effect of the increase of ambient temperature on the V–I characteristics of a panel is shown. As demonstrated, with temperature increase, the current increases as well but by such a small amount that it can be neglected. However, with the temperature increase, the output voltage is dropped significantly. The output power is the product of voltage and current, and as seen in Figure 3, with an increase in temperature, the output power is decreased (Moshtagh Dezfoli and Roshandel 2015). Evaluation of the output power can also be a useful method for the identifying of hot spots.

Figure 3:

The effect of temperature increase on panel performance (Moshtagh Dezfoli and Roshandel 2015).

There are several methods in the identification of hot spots, such as identification by AC signal injection, light projection, and by infrared imaging with thermo-vision. Currently, the best method for identifying hot spots is infrared thermal imaging which is usually supervised by operators to detect possible defects in the panels.

YOLO^[1] algorithm

First, let us start with the word YOLO. YOLO means that humans can easily understand what objects are in a picture, where they are, and how they relate to each other, but this process is very difficult for a computer (Redmon et al. 2016). In this research, after receiving the pictures and applying the required graphical processing, it is revealed that with the help of the YOLO algorithm, unified and automated monitoring can be conducted. Because in power plants with a high density of photovoltaic panels, high accuracy and low error are required, it is hard for a human operator to detect and distinguish errors quickly.

For the task of image processing, a few basic functions were first selected that are commonly used in graphic programs on PC and sometimes in features of commercial electronics (Kotyza and Kasik 2016).

YOLO employs a 24-layer Convolutional Neural Network for feature extraction and two fully connected layers for predicting items and their coordinates on the image. Figure 4 shows the architecture of the YOLO algorithm.

Figure 4:

Overview of the layers of YOLO architecture (Redmon et al. 2016).

The use of convolutional neural networks (CNN, CONVnet) instead of standard classification algorithms is a trend in the research area of human detection, regardless of whether the task is nighttime detection using thermal imaging cameras or daytime detection using standard optical cameras. Thermal images are mainly used to detect the presence of people at night or in bad lighting conditions. However, they can perform poorly in the daytime when there is insufficient thermal contrast between the people and their surroundings (Ghose et al. 2019).

The development of CNNs helped move surveillance systems to embedded devices like the Raspberry Pi. Khalifa, Badr, and Elmahdy (2019).

Super-pixel refers to an irregular pixel block with a certain visual representation which consists of adjacent pixels with similar characteristics such as texture color, brightness, and so on (Achanta et al. 2012).

Loss function of YOLO algorithm

With the increase in panel density in photovoltaic power plants, the Mean Squared Error (MSE) Function of the YOLO algorithm can be utilized. Because optimizing this loss function is easy and compatible with the regression problem in the YOLO algorithm. However, some changes need to be made to the MSE before it can be used in object detection according to the requirements, see equation (1);

In this function x ˆ i , y ˆ i , w ˆ i , and h ˆ i are the coordinates of the object located in the image that is already given to the algorithm, and the algorithm must be trained accordingly; x ˆ i and y ˆ i are the coordinates of the center of the object in the cell i.

w ˆ i and h ˆ i are the width and height of the bounding box that contains the object. C ˆ i is the confidence of an object existing in cell i. p ˆ i ( c ) is the existence probability of the object belonging to class c in cell i. The values without (˄) are the values extracted by the algorithm during each repetition in the training phase that is compared to the values given to the algorithm (also known as ground truths) and result in the loss function.

1 _i ^obj shows if there is an object in cell i, and 1 _ij ^obj shows that the predictor box j is responsible for the prediction. However, there are some problems with this function. In the MSE loss function, the loss from localization and classification error has the same weight, which reduces the accuracy. Also, in every picture, most cells do not contain any objects. This causes the confidence score of these boxes to be zero and overwhelm the cells that contain an object and leave them neglected. This causes model instability. This also means that instead of converging the amount of error, which suggests the end of the training, the error values diverge and disrupt the learning of the neural network. To solve this issue, we increase the loss of localization while reducing the loss of confidence rating, and to realize this, we set (λ _coord = 5) and (λ _noobj = 0.5).

Also, in MSE, the weight of the error is the same in large and small boxes. At the same time, small errors in larger boxes should be reflected in less than the same number of errors in small boxes. In other words, a minor error in a large box should be less punishable than the same error in a smaller box. That is why we used the square root for the width and height of the bounding boxes, as seen in Figure 5.

Figure 5:

Comparison of a variable and the square root of a variable (https://howsam.org/yolo-algorithm/, 2021, validation data4).

YOLO network is trained with Train and Validation data from PASCAL VOC 2007 and 2012 database (Figure 6). While testing on PASCAL VOC 2012 database, PASCAL VOC 2007 test data is also used for training (http://host.robots.ox.ac.uk/pascal/VOC/voc2007, 2007, Validation data1).

Figure 6:

Overview of the detection process of R-CNN algorithm (Girshick et al. 2014).

R-CNN^[2] algorithm

R-CNN is a relatively older method of object detection. In this algorithm, around 2000 regions are proposed. Each of these regions passes the convolutional neural network and is classified. This means that to detect the objects in a picture, the picture has to pass through the neural network 2000 times (Girshick et al. 2014).

IOU^[3] benchmark review

This criterion measures the degree of overlap between two “regions”; This value is equal to the area resulting from the overlap of these two areas, divided by the area resulting from the union of these two areas. This criterion shows the quality of the predictions produced by the object detection system compared to the correct answers of the ground truth. It compares them with each other, which is extracted according to relationship number (3).

Benchmark relationship IOU

The accuracy measure calculates the percentage of positive predictions that are correctly predicted as positive. The accuracy measure also calculates the ratio of “True Positives” to all possible outputs. The values of the precision and recall criteria are included in equations number (4) and (5), which are based on the predicted classes of false positives and true positives. It is that in Figure 9, the accuracy criterion function is drawn in a classified manner, which includes the final result of the performance accuracy for us, based on the final F1 relationship (equation number (6)), which is embedded in the Yolo processing algorithm. These two criteria have an inverse relationship with each other, which is shown in Figure 9.

Figure 9:

Anticipated classes.

Validation and validation of accuracy: data related to checking the correctness of the performance of the YOLO-based machine vision algorithm with Train and Validation data from the PASCAL VOC 2007 and 2012 database and thermo-vision images of real defective panels for simulation in MATLAB to implement object detector evaluation functions to measure common criteria. Such as the average precision (valuateDetectionPrecision) and the average error rates (evaluateDetectionMissRate) of the presentation are used. Average accuracy provides a single number that combines the detector’s ability to classify (precision) correctly and the detector’s ability to find all relevant objects (recall). Figure 10 is the average precision simulated function.

Figure 10:

The average precision simulated function with YOLO version 3.

In this study, we compared two object detection methods in detect hotspots accurately. The first method is Faster-R-CNN, which is a slower method but, according to the articles, is more accurate at low density. The second method is the YOLO algorithm, which is faster but slightly less accurate at low densities.

In this study, we investigate the speed and accuracy of these two algorithms with the data set related to solar panel hot spots. It is worth noting that the same data set was used to train both algorithms, and the image file format was JPEG in the training and testing phases.

To utilize these algorithms, we must first train the convolutional neural network and obtain certain weight and configuration files that contain the information required by the network to detect a certain object. Then by employing these files, we use the neural networks to detect hot spots in pictures the networks have never seen before.

Comparison of YOLO and faster R-CNN

This algorithm also needs high processing power for its training, and the convolutional neural network of this algorithm is known as the darknet. In the Google Colab environment, we implemented the code related to learning this algorithm and started training the algorithm. The algorithm reached proper accuracy in the training phase after 2220 repetitions, which took about 6 h.

It can be seen in Figure 14 that the input picture includes 10 hot spots. While YOLO did not detect one hot spot, it reported nine. However, Faster-R-CNN has reported eight hot spots because it counted three hot spots as one.

However, the Faster-R-CNN has serious problems in this regard. It is observed in Figure 15 that the Faster-R-CNN has performed badly in detecting small hot spots while the YOLO has detected them very well. Furthermore, the Faster-R-CNN’s error in detecting the frame instead of the hot spot is very obvious. However, the YOLO has done it flawlessly. Figure 11 shows a comparison of these algorithms in accuracy.

Figure 11:

Comparison of the accuracy of the algorithms.

In the training phase, using Google Colab’s Tesla T4 graphics card, Faster-R-CNN took less time to train, about 3 h. However, YOLO needed 6 h for it to be trained. The speed results are shown in Figure 12. It is worth noting that the detection speed of the algorithms is not dependent on the input image. It means that it takes the same amount of time to perform detection on an image with four hot spots as it takes on an image with 10 hot spots or even no hot spots. This is why the speed is reported in a column chart.

Figure 12:

Comparing the detection speed of the algorithms.