Photovoltaic (PV) boards are a perfect way to create eco-friendly power from daylight. The defects in the PV panels are caused by various conditions; such defective PV panels need continuous monitoring. The recent development of PV panel monitoring systems provides a modest and viable approach to monitoring and managing the condition of the PV plants. In general, conventional procedures are used to identify the faulty modules earlier and to avoid declines in power generation. The existing deep learning architectures provide the required output to predict the faulty PV panels with less accuracy and a more time-consuming process. To increase the accuracy and to reduce the processing time, a new Convolutional Neural Network (CNN) architecture is required. Hence, in the present work, a new Real-time Multi Variant Deep learning Model (RMVDM) architecture is proposed, and it extracts the image features and classifies the defects in PV panels quickly with high accuracy. The defects that arise in the PV panels are identified by the CNN based RMVDM using RGB images. The biggest difference between CNN and its predecessors is that CNN automatically extracts the image features without any help from a person. The technique is quantitatively assessed and compared with existing faulty PV board identification approaches on the large real-time dataset. The results show that 98% of the accuracy and recall values in the fault detection and classification process.
Consumption of electricity is increasing day by day due to population growth and industrial growth. In this case, the world needs a cost-effective and sustainable energy source. Solar energy and other renewable energy resources are to be used to solve the energy crisis [
Defects | Nature & Severity of Damage | References |
---|---|---|
Broken PV cells | The solar cell was damaged while being handled, most likely during the soldering procedure. | [ |
Scratches on the glass | In many situations, small scratches and large scratches happen at the manufacturing unit due to mishandling of the PV module. It can degrade the yield of the PV module. | [ |
String alignment | A misaligned panel position leads to an unpleasant omission. Arching may occur if the gap between the solar cells is too small. | [ |
External particles | Another flaw that can be found is dirt, fabric, or insects on the outside of the PV cells. | [ |
Snail trails | When the solar cells are broken, they have been generating a lot of electricity along with the breaks. This creates hot spots that erode the surface of the cell. If there is water vapour in the air, snail trails of a different colour appear along the micro-cracks. This makes the solar cell produce less energy. | [ |
Hot spots | Any flaw in solar cells, such as fractures, improperly soldered junctions, and abnormalities, leads to greater resistance and hot spots. Hot spots have severe consequences, such as burned scars that destroy solar cells and back sheets if they are not controlled, eventually leading to fire. | [ |
Internal corrosion | Whenever moisture gets into the panel, it causes rust. If the lamination process is not done correctly, this might result in structural failure during operation. | [ |
Delamination | The separation of laminated parts is referred to as delamination. It also includes poorly installed module trim. | [ |
An image mosaicing technique is used to localize the PV panel’s defects based on color information. A harris corner and a hough transform detector are used to stitch the group of images together to form a larger mosaic image. That the faults are evaluated by the mosaicing technique, but it didn’t reveal any numerical results. It is used with the thermal camera, which is mounted on an Unmanned Aerial Vehicle (UAV) for inspection using the aerial thermal visual photography technique [
One of the most promising approaches developed to overcome manual inspection is CNN [
Another way to identify the faults in PV panels is based on artificial neural networks. The performance of PV modules is evaluated by attributes such as temperature, current (I), voltage (V), and evoltage (IeV). To classify the various faults, a Field Programmable Gate Array was used [
The PV panel faults are identified electrically too. The fuzzy logic control is used to monitor, identify, and detect the various PV faults based on three values such as open-circuit voltage, current, and voltage [
Both traditional image processing approaches and deep learning techniques have their pros and cons. Classic image processing algorithms are well-known, easy to understand, and optimized for performance and power efficiency. Traditional image processing algorithms are hard to come up with when there are a lot of classes to sort into or when the image isn’t very clear. Disadvantages of ANN approaches include usually hardware dependence, difficulty of conveying the issue to the network, duration of the network is uncertain. Signal processing chips are costly and it requires higher bandwidth to transmit the data to the network and skilled engineers can only work on the Signal processing devices. Deep learning, on the other hand, is more accurate and flexible, but it uses a lot of computing resources. They are especially useful for the fast implementation of high-performance systems. You don’t have to define the features and do feature engineering anymore. RMVDM was trained on GPUs. Models were trained quickly. Its 8-layer design ensures it can extract characteristics better than previous approaches. This implies little feature loss. It cancels negative gradient summation output, not the dataset. Since not all perceptrons are active, model training speed improves.
The CNN-based RMVDM defect detection and classification method is one of the best classification techniques, which is used to identify the defects in the PV module and provide exact electricity generation volumes under various conditions. Deep learning techniques consist of multiple layers to learn data. These deep learning techniques are used in speech recognition, text recognition, object recognition, object detection, object classification, pattern recognition, and many other applications to solve more complex problems. It is described as a class of machine learning algorithms that have multiple layers to perform artificial intelligence tasks. These deep learning methods are capable of learning from non-identical data without any predefined programmes [
Layers | Functions of CNN layers | References |
---|---|---|
Input Layer | The initial input has been given to this layer. Here the image pixels have a shape: the number of the input * input height * input width * input channels. | [ |
Convolution layer | The input image data is pre-processed here. Once the image is pre-processed, the image becomes abstracted into a feature map, also called an activation map. | [ |
Rectified Linear Unit (ReLU) | ReLu applies a non-linear function to the output of the preceding layer. It removes the negative values and inserts the zero value from the preceding output. | [ |
Pooling Layer | The pooling layer can reduce the size of the output data by combining the results of the layers that came before it. | [ |
Fully-connected layer | Every neuron in every layer is connected by a fully connected layer. Connecting neurons is done by all traditional neural networks. | [ |
Loss layer | It assesses the predictions of the trained model from true data labels. | [ |
In the present study, the CNN-based RMVDM framework is a technique for detecting PV panel defects that is built on a contemporary identification framework. RMVDM is a fast network that uses Deep Convolutional Neural Networks (DCNN) to conduct object detection and object classification at the same time. The RMVDM object recognition framework recognizes PV panel defects from RGB photos. It has been demonstrated to be a flexible and successful approach, with good detection accuracy and minimal computation cost; for example, it runs continuously on usually accessible GPUs. Therefore, the present work strongly recommends quantitative analysis outcomes based on an RGB image collected from real-world solar power plants. The suggested solution meets all three of the needed characteristics: accurate, quick, and capable of being executed in real-time. Finally, it is concluded that no plant-specific configuration is required for CNN.
The key benefit of employing RMVDM is that a single pass-through network provides all the information needed to discover the suspicious objects in an image without further costly processes. Conventional systems, on the other hand, require two different steps: first, they identify numerous proposal regions where an item is most likely to be discovered; and second, they categorize each region to validate or dismiss the proposal. Because running a classifier on multiple areas increases the computational cost but, the RMVDM design is simpler, yet faster and has more accuracy. The RMVDM framework was initially created to address the challenge of generic object identification, to automatically recognize and classify diverse kinds of objects. It simplifies the challenge and allows the design to be trained with much less data. RMVDM works as follows: Feature Extraction and Classification.
The collections of several PV plant image datasets are captured by a high-resolution RGB camera. The dataset comprises 6 different categories and a total of 1200 images. The PV images are taken from different lighting conditions, various altitudes, and 4 different kinds of paths (East, West, North, and South). Besides, the PV boards have a place with various PV plants and with different sizes, shapes, alignment, and colors; it is found that the dataset has enough variants to reflect many of the difficult scenarios encountered in popular applications involving the autonomous DCNN evaluation of PV installations. A deep learning technique trained and evaluated on this dataset is predicted to be resistant to changes in direction, panel makers, plant dispositions, and weather conditions and illuminations. The real-time PV images are depicted in
Normal panel | Broken | Cracked | Fade | |
---|---|---|---|---|
Training | 500 | 500 | 500 | 500 |
Testing | 1200 | 1200 | 1200 | 1200 |
The RMVDM feature extraction network has already been trained. A convolution operator, which is useful for solving complex operations, is used to get the name of these kinds of networks. The main benefit of RMVDM is that it automatically extracts features, which is the case. First, the specified input data is sent to a network for extracting features. Then, the extracted features are sent to a network for classifiers. The simplest and quickest method to extract the image features in CNN-based RMVDM is feature extraction, which has been trained on the CNN-based RMVDM architecture to extract the various features of PV images. Features of the PV images are such as color, texture, shape, position, edges, and regions. Generally, the RMVDM model has fixed input and output sizes. In this present work, the dataset image size is 3096*4128. So, all the PV images are resized to 227*227.
The architecture of feature extraction is depicted in
In single-phase recognition architecture, the entire PV image is used as input for the first layer. The network then creates a set of coordinates, each with a level of confidence. It has been tried in different kinds of deep learning settings. The RMVDM architecture consists of 201 layers. The RMVDM Blocks contain 116 layers, and the Conv Layer, Fully Connected Layer, and Transition Layers each contain 1 layer. The complete RMVDM layer details are shown in the
Layer names | RMVDM | Output |
---|---|---|
Input | 7 × 7/2 conv | 224 × 224 × 3 |
Conv layer | 7 × 7/2 conv | 112 × 112 × 64 |
Pooling layer | 3 × 3/2 max pool | 56 × 56 × 64 |
RMVDM Block 1 | 56 × 56 × 256 | |
Transition layer 1 | 1 × 1 conv, 2 × 2/2 average pool | 28 × 28 × 128 |
RMVDM Block 2 | 28 × 28 × 512 | |
Transition layer 2 | 1 × 1 conv, 2 × 2/2 average pool | 14 × 14 × 256 |
RMVDM Block 3 | 14 × 14 × 1024 | |
Transition layer 3 | 1 × 1 conv, 2 × 2/2 average pool | 7 × 7 × 512 |
RMVDM Block 4 | 7 × 7 × 1024 | |
Pooling layer | 7 × 7/7 average pool | 1 × 1 × 1024 |
Fully connected layer | 1000-D | 1 × 1 × 1000 |
The RMVDM architecture splits the input PV images into a matrix format. Each row of the matrix forecasts
To further develop the presentation when managing tiny objects, RMVDM utilizes a skip connection to exploit the fine-grained highlights when genuine discoveries are anticipated in the last layers.
Finally, Batch Normalization is used, and the backbone network design is updated by replacing every pixel value in channel C with the following rule:
It is quite significant that the RMVDM network is completely convolutional and is prepared with various sizes of images; this permits the locator to be utilized for different purposes. The subsequent precision is essentially improved, with seemingly minor impacts on the deduction time. The pseudo-code of RMVDM architecture is shown in Algorithm 2.
In this present work, a quantitative evaluation of the suggested technique in this section is presented. Also, it portrays the dataset utilized in our experimentation. Therefore, we present the work metrics used to evaluate the approaches’ performance as well as the experimental procedure utilized for the evaluation, with the accompanying equations:
The common architecture of RMVDM is depicted in
CNN based fault detection model’s results as well as performance of the proposed RMVDM fault diagnosis method are presented in this section. Techniques such as the one proposed here require a reference solar cell and camera for the module’s temperature, defect’s position and defect’s size. Analyzed results are stored in an electronic database before being transferred to an operating system on a personal computer, where the diagnostic procedure is performed.
Compared to the other conventional techniques, RMVDM results in the two primary environments, such as parameter-free configuration and fine-tuned configuration. The test images are not utilized to optimize the parameters in the event of parameter-free experimentation. This parameter-free experiment is to replicate a situation in which the person does not need to do any sort of setting on the given circumstance. In the parameter-tuned configuration, 30% of train images are utilized to fine-tune the network; the remaining 70% of images are utilized as validation data. A minor configuration in the RMVDM model boosted the overall network efficiency. Although parameter-tuning is available on the present CNN model. It is indicated that fine-tuning enables the correction of the numerous false negatives that occurred.
Using the different PV panel image datasets for training and validation also provides outstanding outcomes. The RMVDM technique needs parameter modification at each phase to adapt to the different PV panel images. One of the primary benefits of the RMVDM technique is that the installation operation is not time-consuming and this solution is perfectly suitable for usage in industrial products. Monitoring is the recurring process to identify the PV panel’s actual progress. In some cases, the same PV panel may be shown in more than one frame, making it possible for the detecting system to miss it.
The algorithm’s worst-case time complexity is represented by the function
With varied numbers of images in each class, the performance of different techniques in defect detection and classification is assessed in
Defect detection performance in % | ||||
---|---|---|---|---|
Methods | Resnet | Alexnet | Densenet | RMVDM |
1000 Images | 73 | 77 | 82 | 91 |
3000 Images | 77 | 81 | 85 | 93 |
5000 Images | 83 | 89 | 88 | 98 |
The outcomes of the RMVDM architecture with different settings are shown in
Test | Specificity | Recall | Accuracy | Precision | F1 Score |
---|---|---|---|---|---|
Test 1 | 96.00 |
97.20 |
97.24 |
97.10 |
97.09 |
Test 2 | 97.20 |
96.80 |
96.92 |
97.00 |
97.01 |
Test 3 | 97.80 |
98.20 |
98.24 |
98.00 |
98.00 |
Test 4 | 97.80 |
98.40 |
98.43 |
98.10 |
98.09 |
In this present work, the proposed methodology will give high accuracy for every type of defect. Here we used the PV images of broken, crack, fade, and normal to measure the accuracy and loss values. The above graphs consist of two-axis (x, y). The x-axis represents epoch run and the y-axis represents accuracy values. The PV panel is appropriately located in the preceding and subsequent frames; therefore there is no true loss when compared to that of others. The above graphs and tables illustrate model accuracy, recall, F1 score, precision and specificity. The trained CNN has 98% accuracy, so the estimator has a low chance of classifying a negative sample as positive. An F-1 score of 98% suggests that recall and accuracy are almost similar to the network. By recall factor, CNN can locate all positive samples 98% of the time. A CNN with a 0.001 learning rate, 32 samples as batch size, and 100 epochs of training was the best classifier of anomalies in a PV module. Due to restricted training data, ten-fold cross-validation was utilised to train and evaluate the suggested model. When compared to neural networks, the RMVDM architecture has the deepest neural structure.
One of the significant improvements in the neural network architecture is the use of anchor boxes. We can train a CNN to categorize an image as well as to produce the bounding box with four coordinates. In RMVDM, each grid cell rightly guesses the anchor box’s coordinates for which it is responsible for defects. In that way, localization becomes a simple regression issue. As an example, we might use RMVDM model which has a number of convolutional, pooling, and other layers, and reuse the fully connected layer to build a bounding box. This localization technique enables the model to perform better.
A large PV dataset consists of 10,000 different PV images in different categories, which are captured from various places and under various lighting conditions. The existing CNN framework was developed to solve the challenges of identifying PV panel faults using RGB images. Compared with previous deep learning methods, the RMVDM method offers a distinct advantage. Furthermore, it doesn’t require a lengthy process. This proposed CNN based RMVDM method is most appropriate for usage in commercial applications. When tested on solar plant characteristics that the system has never encountered before, the RMVDM detector gets a 98% accuracy rating and gets a recall value of up to 97%. The outcome is somewhat better than that obtained with other techniques.
I would like to express my gratitude to my supervisor, Dr.R.Annie Uthra and also thank Dr.J.Preetha Roslyn, who guided me throughout this research.