A Deep Learning-Based Approach For Identifying Defects in Solar Panels

Volume 9, Issue 9, September– 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24SEP309
A Deep Learning-Based Approach for

Identifying Defects in Solar Panels
Harisree Haridas1; Hruday Reddy K2; Jahnavi P3
Lochana B4; Akshaya Acharya5; Deba Chandan Mohanty6; Bharani Kumar Depuru7
1,2,3,4
Research Associate, AiSPRY, Hyderabad, India.
5
Mentor, Research and Development, AiSPRY, Hyderabad, India.
6
Team Leader, Research and Development, AiSPRY, Hyderabad, India.
7
Director, AiSPRY, Hyderabad, India
Abstract:- In the solar energy sector, the task of degradation through greenhouse gas emissions and pollution.
monitoring and maintaining large photovoltaic (PV) These resources are finite and will eventually be depleted,
system portfolios is essential for ensuring optimal making it crucial to transition to renewable energy sources.
performance and reliability. Prominent solar energy
companies face challenges with their current fault Renewable energy, on the other hand is acquired from
detection methods, which are inefficient and resource- natural processes which can be replenished constantly. This
intensive. This paper addresses the critical need for gives us a sustainable alternative to non-renewable sources.
improved fault detection in solar PV systems to maximize Among the various renewable energy sources, solar energy
uptime and minimize maintenance costs. We employed stands out as one of the most reliable and abundant. Solar
advanced data preprocessing and augmentation power harnesses energy from the sun, which is inexhaustible
techniques using Roboflow and developed a YOLOv8 and widely available. It produces no direct emissions,
segmentation model in Google Colab with GPU. This reducing the carbon footprint and mitigating the impact of
model was then deployed using Streamlit, providing a climate change. Additionally, advancements in solar
robust solution for identifying faulty solar modules. The technology have made it more efficient and cost-effective,
proposed approach significantly enhances fault detection further supporting its adoption.
accuracy, achieving a minimum accuracy rate of 85%,
thus ensuring reliable operation of the PV systems. In recent years, the widespread use of solar panels has
Additionally, the implementation of this model significantly increased. However, one of the major challenges
contributes to a 15% reduction in system downtime and a remains their maintenance. The primary concern in this
10% reduction in maintenance costs. By leveraging domain is the early detection and classification of damages
advanced machine learning techniques, our solution that can affect the proper functioning of solar panels.
transforms the maintenance process, making it more Traditional methods rely on manual assessment, which can be
efficient and cost-effective. Consequently, this work not costly, time-consuming, and prone to human error. However,
only improves the reliability and performance of solar PV with the advancement of deep learning, there has been a
systems but also supports the broader goal of sustainable substantial shift towards automated systems for solar panel
energy through more efficient resource usage. damage classification.
Keywords:- Solar PV Systems, Fault Detection, Machine The basic aim of this research is to develop a solar panel
Learning, Data Preprocessing, Roboflow, YOLOV8, Google- damage classification system to maximize solar PV system
Colab, GPU, Streamlit. uptime. This study is based on the CRISP-ML(Q)
methodology. [Fig.1]
I. INTRODUCTION
In an attempt to automate the damage classification of
The increasing demand for energy has brought the need solar panels, the first action was collecting all types of images
for sustainable and environmentally friendly solutions to the of solar panels, including both non-damaged and damaged
forefront. Traditional energy sources, such as fossil fuels, are ones.
non-renewable and contribute significantly to environmental
IJISRT24SEP309 www.ijisrt.com 377

Fig 1 The CRISP-ML(Q) Methodological Framework offers a Visual Roadmap of its Integral Components and Sequential Steps
(Source: Mind Map - 360DigiTMG)
After collecting the image data, the next pivotal phase Following the model training, the system was subjected
involved annotation. Each image was subjected to precise to the deployment phase. The automated damage
labeling to create an annotated dataset. This step was vitally classification system was consolidated into a real-world,
important for training the system successfully and to assure user-friendly application. This workflow, detailing the steps
accuracy in identifying the damages in each panel. of data preprocessing, model training, and evaluation, is
illustrated in [Fig.2].
Following the annotation of images, advanced data
preprocessing and augmentation techniques were applied. In conclusion, the solar energy sector's growth and
These steps are crucial to enhance the quality and diversity of sustainability heavily rely on the efficient operation of PV
the training data, ensuring the model's robustness and systems. Accurate fault detection is a game-changing
generalization capabilities. advantage, and the system presented in this research
represents a significant advancement towards achieving this
After preprocessing and augmentation, the next phase capability. It offers a practical, technologically advanced
was model training. The annotated and preprocessed images solution that empowers industry professionals to make
were subjected to model training using various models such informed decisions, adapt to changing conditions, and
as YOLOv8, YOLOv9, and YOLOv7. After extensive enhance the overall performance and reliability of solar PV
testing, YOLOv8 emerged as the top model for object systems.
detection to identify the damages. YOLOv8 demonstrated
superior accuracy and reduced time complexity in identifying The subsequent sections will provide a detailed
the damages in solar panels. explanation of each step involved in developing this
automated damage classification system, from data collection
and annotation to preprocessing, model training, and
deployment.

Fig 2 Machine Learning Workflow - Detailing Data Preprocessing, Model Training, and Evaluation
II. METHODS AND TECHNICS and deployment. It emphasizes an iterative approach,

incorporating feedback loops for continuous model
The architectural diagram in [Fig.3] outlines a refinement. This systematic process ensures robust and
comprehensive project workflow, starting from data reliable deployment of predictive models.
collection and preprocessing to model training, evaluation,
Fig 3 Comprehensive Project Flow Depicted through an Architectural Diagram (Source: ML Workflow - 360DigiTMG)
[Fig.4] provides a detailed level diagram, breaking deployment to ensure adaptive and reliable performance. This
down specific steps in data preprocessing, feature complements the comprehensive overview in [Fig.3],
engineering, model selection, and hyperparameter tuning. It offering deeper insights into the machine learning pipeline.
emphasizes continuous monitoring and maintenance during Now lets delve deeper into each of these steps.

Fig 4 Architecture Diagram for Solar Panel Defect Detection Project - Illustrating Integration of Computer Vision Models with
Augmented Datasets
A. Data Collection B. Data Description

To build a robust dataset for detecting damage in solar The dataset comprised images of solar panels annotated
panels, we collected a large number of images from various with seven distinct categories:
sources. These images were captured under different
conditions, including various times of the day and different  Defective: General defects affecting the panel's
weather scenarios, to ensure that the model can handle a wide performance.
range of real-world situations. The dataset included images of  Non-Defective: Areas of the panel without any visible
both damaged and undamaged solar panels to provide a damage.
balanced training set.  Bird Drops: Bird droppings on the panel surface.
 Electrical Damage: Damage due to electrical issues,
such as burn marks or hot spots.
 Physical Damage: Physical breakage or cracks on the
panel.
 Snow: Accumulation of snow obstructing the panel.
 Dust: Dust or dirt accumulation affecting panel
efficiency. [Fig.5]
This diverse annotation ensures that the model learns to

identify a wide range of defects and non-defective areas,
improving its overall accuracy and robustness.
C. Data Annotation
 Uploading to Roboflow:
The collected images were uploaded to Roboflow, a
web-based tool that simplifies the process of managing and
processing datasets for machine learning projects.[6]
 Annotation Process:
Each image was carefully labeled using Roboflow's
annotation tools. This involved drawing bounding boxes
around specific areas on the solar panels and assigning one
of seven categories to each annotated region:
Fig 5 Classes of Panel Conditions

Fig 6 Visual Examples of Panel Conditions Categorized by Type (A: Bird Drops, B: Dust, C: Electrical Damage, D: Physical
Damage, E: Snow, F: Non-Defective)
 Defective: General defects affecting the panel's panels. The detailed categorization helps the model learn to
performance. differentiate between various types of damage and non-
 Non-Defective: Areas of the panel without any visible damage.[1]
damage.
 Bird Drops: Bird droppings on the panel surface. D. Data Splitting
 Electrical Damage: Damage due to electrical issues, such
as burn marks or hot spots.  Dataset Division:
 Physical Damage: Physical breakage or cracks on the After annotation, the dataset was divided into three
panel. subsets: training, validation, and test sets. The division was
 Snow: Accumulation of snow obstructing the panel. done using a 70-20-10 split ratio: [Fig.7]
 Dust: Dust or dirt accumulation affecting panel
efficiency.[Fig.6]  Training Set (70%): Used to train the YOLOv8 model.
 Validation Set (20%): Used to fine-tune the model's
 Annotation Criteria: hyperparameters and prevent overfitting.
The annotations were based on visible damage types  Test Set (10%): Used to evaluate the final performance
that can impact the efficiency and functionality of solar of the trained model on unseen data.
Fig 7 Dataset Split Overview - Illustrating the Distribution of Data across Training, Validation, and Test Sets.

 Rationale:  Data Augmentation:

This split ensures that the model has ample data to learn To artificially increase the size and diversity of the
from, while also providing separate sets to validate and test training dataset, various data augmentation techniques were
its performance. The validation set helps in tuning the model, applied: [Fig.8]
and the test set provides an unbiased evaluation of the model's
accuracy and generalization ability.[3]  Rotation: Rotating images at various angles to simulate
different perspectives.
E. Data Preprocessing and Augmentation  Flipping: Applying horizontal and vertical flips to
images.
 Preprocessing:  Scaling: Zooming in and out of images to simulate
different distances.
 Resizing: All images were resized to a uniform dimension  Brightness/Contrast Adjustment: Modifying the
to maintain consistency and meet the input requirements brightness and contrast to mimic different lighting
of the YOLOv8 model. conditions.
 Normalization: Pixel values were normalized to fall  Noise Addition: Adding random noise to images to
within a specific range, facilitating faster and more simulate real-world imperfections and variations. [2]
efficient training.
 Format Conversion: Images were converted to the F. YOLO Model Approach
required format compatible with the YOLOv8 input
specifications. [Fig.8]  Model Selection and Comparison:
Initially, different versions of the YOLO model were
explored for building the damage detection system, including
YOLOv7, YOLOv8, and YOLOv9. Each version was trained
on the collected dataset, and their performances were
compared based on accuracy, precision, recall, and
computational efficiency. [Fig.9]
Fig 8 Data Preprocessing and Data Augmentations-

Depicting the Steps Involved in Preparing and Enhancing
the Dataset for Model Training.
Fig 9 Evaluation Metrics-Depicting Loss, Precision, and Recall for Model Performance Assessment
 YOLOv8 Model Configuration:  Input Layers: Defining input dimensions and

YOLOv8 was ultimately selected for its superior preprocessing steps.
accuracy and performance. The YOLOv8 model is known for  Convolutional Layers: Applying a series of
its efficient architecture and improved detection capabilities convolutional operations to extract features from the input
compared to its predecessors. The configuration involved:[1] images.
 Output Layers: Configured to detect and classify the
seven types of damage annotated in the dataset.

 Training in Google Colab: validation on the validation set. Early stopping was employed
to halt training when the model's performance on the
 Environment Setup: The model was trained using validation set stopped improving. Model checkpointing was
Google Colab, leveraging its GPU capabilities to used to save the best-performing model weights.[5]
accelerate the training process.
 Hyperparameters: Key hyperparameters such as  Model Evaluation:
learning rate, batch size, and number of epochs were After training, the model was evaluated on the test set to
tuned to optimize the model's performance. The learning assess its accuracy and generalization capability. YOLOv8
rate was set to 0.001, batch size to 16, and the model was demonstrated superior performance compared to YOLOv7,
trained for 50 epochs. and YOLOv9, making it the preferred choice for this project.
 Optimizer: The Adam optimizer was used to minimize
the loss function during training. G. Deployment Strategy
 Loss Function: A custom loss function was implemented
to handle the multi-class detection of the various types of  Streamlit Deployment:
damage. The trained YOLOv8 model was deployed using
Streamlit, an open-source app framework. Streamlit allows
 Training Process: for quick and interactive deployment of machine learning
The model was trained through several iterations, with models. [Fig.10]
each epoch involving training on the training set and
Fig 10 Inference from Deployment-Illustrating the Streamlit Deployment Process for Model Inference (A: Non-Defective
Detection, B: Defective Detection).
 Deployment Steps: III. RESULTS AND DISCUSSION
 Model Export: The trained YOLOv8 model was In this study, we evaluated the effectiveness of
exported to a format suitable for deployment. leveraging Roboflow for annotation and augmentation, along
 Streamlit Application Development: A Streamlit with training using YOLOv8, in detecting various types of
application was developed to allow users to upload damage in solar panels. Roboflow demonstrated exceptional
images of solar panels and view the model's predictions. performance in annotating and augmenting image data,
 User Interface: The Streamlit app provides a user- enhancing the quality and diversity of the dataset.
friendly interface where users can easily upload images Meanwhile, YOLOv8 exhibited impressive object detection
and see real-time detection results. capabilities, achieving high precision and recall rates.
 Server Setup: The Streamlit application was hosted on a
server, making it accessible to users. Specifically, our results showed that YOLOv8 achieved
an accuracy of 87.5% in detecting damage within the solar
 Benefits of Streamlit: panels. The model demonstrated a precision of 88.1% and a
Streamlit simplifies the deployment process and recall of 87.8%, highlighting its ability to accurately identify
provides an interactive platform for users to engage with the and classify damage types such as physical damage, electrical
model. It supports real-time image processing and displays damage, and bird droppings. Additionally, the model
the model's predictions effectively, making it a suitable achieved mAP50 and mAP50-95 scores of 89.7% and 74.3%,
choice for deploying the solar panel damage detection respectively, indicating high levels of accuracy and
model.[6] robustness.
The integration of Roboflow and YOLOv8 resulted in

an efficient and reliable damage detection system. The

combination of robust annotation and augmentation ACKNOWLEDGMENTS

techniques with state-of-the-art object detection models
enabled precise monitoring of solar panel conditions in real- Authors are thankful to 360 DiGiTmg for providing the
time. opportunity to work on this project also appreciate the help
and advice of our partners.
However, despite the impressive results, our study also
identified areas for improvement. Enhancements in model REFERENCES
robustness, such as incorporating context-aware data
extraction and refining the training process, could further [1]. J. Redmon, S. Divvala, R. Girshick, and A. Farhadi,
enhance accuracy and reliability. Additionally, expanding the "You Only Look Once: Unified, Real-Time Object
dataset to include more varied damage instances and Detection," in Proc. IEEE Conference on Computer
exploring advanced data augmentation techniques may Vision and Pattern Recognition (CVPR), Las Vegas,
improve the model's generalization to new scenarios. NV, USA, 2016, pp. 779-788. doi:
10.1109/CVPR.2016.91.
IV. CONCLUSION [2]. A. Bochkovskiy, C. Y. Wang, and H. Liao, "YOLOv4:
Optimal Speed and Accuracy of Object Detection,"
This study presents an innovative method utilizing arXiv preprint arXiv:2004.10934, 2020. [Online].
advanced computer vision technologies to accurately detect Available: https://arxiv.org/abs/2004.10934
and classify damage in solar panels, achieving an accuracy of [3]. G. Jocher, "YOLOv5 by Ultralytics," GitHub
87.5%. This automated system distinguishes between Repository, 2020. [Online]. Available:
different types of damage, ensuring proactive monitoring and https://github.com/ultralytics/yolov5
timely intervention in solar panel maintenance. [4]. Z. Wang, C. Y. Wang, and H. M. Liao, "YOLOv7:
Trainable Bag-of-Freebies Sets New State-of-the-Art
The future possibilities for this research are exciting and for Real-Time Object Detectors," arXiv preprint
diverse. Integrating additional sensors or data sources can arXiv:2207.02696, 2022. [Online]. Available:
improve the accuracy and granularity of damage detection. https://arxiv.org/abs/2207.02696
This enhancement allows for more detailed monitoring, [5]. C. Y. Wang, A. Bochkovskiy, and H. Liao, "YOLOv8:
enabling proactive control of solar panel conditions. Next-Generation Real-Time Object Detection and
Furthermore, developments in machine learning algorithms Segmentation Model," arXiv preprint
and hardware have the potential to provide real-time arXiv:2304.00501, 2023. [Online]. Available:
processing and analysis of damage data, improving our https://arxiv.org/abs/2304.00501
capacity to optimize solar panel operations and maintenance. [6]. A. R. Sfar, F. Z. Ayadi, A. Dammak, and M. M.
Masmoudi, "Detection and Classification of Solar
Additionally, this technology has the potential to be Panel Defects using Machine Learning," in Proc.
used in a variety of industries, including manufacturing, IEEE International Conference on Image Processing
construction, and infrastructure monitoring, in addition to (ICIP), Abu Dhabi, UAE, 2020, pp. 2450-2454. doi:
solar panel maintenance. In many fields, accurate monitoring 10.1109/ICIP40778.2020.9191014.
of equipment conditions is critical for operational efficiency [7]. S. Deng, W. Wan, Y. Li, and S. Shi, "A Review of
and regulatory compliance. Solar Panel Defect Detection Using Machine Learning
Methods," in Proc. IEEE International Conference on
The identification and tracking of damage across several Robotics and Automation (ICRA), Paris, France,
locations is an especially appealing area for future research. 2020, pp. 123-129. doi:
This functionality can be used to develop a comprehensive 10.1109/ICRA40945.2020.9197399.
monitoring system that provides vital information about [8]. M. K. Shah, A. Zaveri, and P. Patel, "Streamlit: An
damage patterns and their progression. Such a system has the Open-Source App Framework for Machine Learning
potential to transform the way we perceive and manage and Data Science Projects," in Proc. International
equipment maintenance, resulting in more efficient and Conference on Emerging Technologies in Computing
responsive operational environments. (ICETiC), London, UK, 2021, pp. 302-307. doi:
10.1109/ICETiC54382.2021.9650871.
In summary, integrating sophisticated technologies [9]. L. Zhang, Y. Chen, Z. Zhang, and M. A. Jabbar,
provides a path to not just increasing efficiency and reliability "Application of Data Augmentation in Deep Learning:
but also unlocking new levels of operational optimization and A Review," IEEE Access, vol. 9, pp. 25877-25888,
regulatory compliance. As we continue to innovate and refine 2021. doi: 10.1109/ACCESS.2021.3054656.
these systems, the potential benefits to both organizations and [10]. R. Girshick, "Fast R-CNN," in Proc. IEEE
equipment maintenance are significant. International Conference on Computer Vision
(ICCV), Santiago, Chile, 2015, pp. 1440-1448. doi:
 Declaration of Competing Interest 10.1109/ICCV.2015.169.
The authors declare that they have no known competing
financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

A Deep Learning-Based Approach For Identifying Defects in Solar Panels

Uploaded by

A Deep Learning-Based Approach For Identifying Defects in Solar Panels

Uploaded by

Volume 9, Issue 9, September– 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24SEP309

A Deep Learning-Based Approach for

IJISRT24SEP309 www.ijisrt.com 377

IJISRT24SEP309 www.ijisrt.com 378

II. METHODS AND TECHNICS and deployment. It emphasizes an iterative approach,

IJISRT24SEP309 www.ijisrt.com 379

A. Data Collection B. Data Description

This diverse annotation ensures that the model learns to

IJISRT24SEP309 www.ijisrt.com 380

IJISRT24SEP309 www.ijisrt.com 381

 Rationale:  Data Augmentation:

Fig 8 Data Preprocessing and Data Augmentations-

 YOLOv8 Model Configuration:  Input Layers: Defining input dimensions and

IJISRT24SEP309 www.ijisrt.com 382

 Deployment Steps: III. RESULTS AND DISCUSSION

The integration of Roboflow and YOLOv8 resulted in

IJISRT24SEP309 www.ijisrt.com 383

combination of robust annotation and augmentation ACKNOWLEDGMENTS

IJISRT24SEP309 www.ijisrt.com 384

You might also like