New PPT
New PPT
1
er
7
M. Yamini(21341A0596)
M.Gagan Aditya(21341A05A9)
GMR Institute of Technology
M. Karishma (21341A05B3)
K. Deepthi (21341A0581)
T. Praveen(22341A0512)
compromise overall power capacity. Timely fault detection during PV module inspections is crucial
for maintaining efficiency, reliability, and safety. This paper introduces "DeepSolarGuard," a deep
learning-based system designed for multi-class fault detection in solar panels, addressing issues
like dust accumulation, bird droppings, electrical and physical damage, and snow coverage. By
using convolutional neural networks Mobilenetv3, VGG16, inception,ResNet50 and advanced
image processing techniques, DeepSolarGuard enhances the resilience of solar energy
infrastructure, enabling real-time fault detection, predictive maintenance, and optimization of
energy production. The system's effectiveness is demonstrated through a dataset encompassing
various solar panel conditions, providing a holistic approach to fault detection essential for the
sustainable operation of solar power systems in diverse environmental and operational scenarios.
OBJECTIVES:
To study the impact of pre-processing techniques, such as filters and
histogram equalization, on fault detection and demarcation accuracy in
thermal images of solar panels.
To compare the performance of different pre-processing techniques in
terms of fault detection and demarcation accuracy.
To evaluate the efficiency of fault detection and demarcation techniques
using Intersection over Union (IoU) as a measure.
Introduction
Solar energy is becoming increasingly popular as a clean and renewable source
of power for homes, businesses, and communities around the world. However,
to ensure that solar panels operate efficiently and effectively, it's crucial to detect
GMR Institute of Technology
and address any faults or issues they might encounter. This is where solar panel
fault detection comes into play.In simple terms, solar panel fault detection is like
giving a check-up to solar panels to make sure they're working as they should.
Just like how we go to the doctor to make sure we're healthy, solar panels need
regular check-ups too.The goal of solar panel fault detection is to identify any
problems or faults that might affect the performance of the panels, such as dirt,
damage, or electrical issues. By detecting these problems early on, we can
prevent them from causing bigger issues down the line and ensure that the solar
panels continue to generate electricity efficiently.There are various methods and
techniques used for solar panel fault detection, ranging from simple visual
inspections to more advanced technologies like infrared imaging and electrical
testing.
These methods help us identify different types of faults, such as dirt and snow
accumulation, cracks or physical damage, shading from nearby objects, electrical
malfunctions, and degradation over time. In this paper we are using some deep
learning techniques to find the faults in the solar panels like
MobileNetV3,InceptionV3, VGG-16, ResNet . For the images we consider we
will use the preprocessing techniques like normalization, resizing of the images ,
GMR Institute of Technology
etc. we will classify the faults by using the above mentioned models into different
class which are present in the dataset.
Literature Survey
Reference-1:-Sohail, A., et al. (2023). "Fault detection and computation of power
in PV cells under faulty conditions using deep learning." *Department of
Electrical Engineering, NUST College of Electrical and Mechanical Engineering,
GMR Institute of Technology
Rawalpindi, Pakistan.
The study aims to develop an effective deep learning method for
identifying cracks in photovoltaic (PV) cells, distinguishing between
microcracks and deep cracks, with a specific emphasis on the orientation of
microcracks. It also investigates how the cracks impact the power generated
by the PV cells. The analysis observes a clear connection, indicating that as
the size of deep cracks increases, the efficiency of power output decreases
proportionally. Four deep learning models (U-net, LinkNet, FPN, and
attention U-net) are trained and evaluated using electroluminescence images
from an online dataset sourced from the DuraMAT datahub website.
Evaluation metrics such as Intersection over Union (IoU) and F1-Score are
employed. Attention U-Net emerges as the top-performing model,
outperforming others in terms of mIoU, F1-score, precision, recall, and
accuracy.
U-Net performs well but lags behind FPN in certain metrics, while LinkNet,
though showing the least loss value during optimization, takes longer to
converge and performs less favorably in other performance metrics. Ensemble
learning further enhances segmentation accuracy, IoU, and F1-Score across
the models. In summary, the research underscores the application of diverse
GMR Institute of Technology
Reference-5-R. Al-Mashhadani et al., "Deep Learning Methods for Solar Fault Detection
and Classification: A Review," Information Sciences Letters, vol. 10, no. 2, pp. 323-331,
2021. DOI: 10.18576/isl/100213.
GMR Institute of Technology
This literature survey reviews deep learning methods applied in solar fault detection.
The findings highlight the capability of these methods in detecting visible defects, with most
achieving accuracy over 90%. Hybrid models outperform standard ones, and data augmentation,
particularly flip and rotation, improves performance. Cracks are the most commonly targeted
faults, and more than 60% of studies focus on two-class classification. Multi-class classification is
less common but exhibits lower accuracy. Some models are computationally demanding,
emphasizing the need for suitable architecture. Real-time solar fault detection shows promise but
requires further enhancement. Future work recommendations include exploring diverse models,
hybrid structures, and improving precision in real-time application.However a CNN model's
superiority, achieving 93.02% accuracy over . Data augmentation techniques, including rotation
and flipping, are also employed.
Limitations: The literature lacks explicit mention of study limitations or specific dataset
details, potentially impacting generalizability assessment.
Future Work :Suggested future work includes exploring diverse deep learning models,
investigating alternative hybrid structures, evaluating real-time dust accumulation assessment,
and expanding training datasets for enhanced model robustness.
Literature survey
Reference 6
Title: Fault Detection in Solar Energy Systems: A Deep Learning Approach
Year:2023
GMR Institute of Technology
Dataset: Infrared Images (20,000 images), consisting of 12 classes: cell, cell-multi, cracking, diode,
diode-multi, hot spot, hot spot-multi, no-anomaly, offline-module, shadowing, soiling, and
vegetation.
Methodology:
Images were initially resized to 224 x 224 and trained using Efficientb0. Then, 56 x 56 patches were
extracted and fed into fully connected layers to obtain 17,000 features. NCA feature selection was
applied to enhance classification by maximizing class separability. SVM classifier with 10-fold
cross validation was used for classification based on the selected features.
Results: These results indicate average values of 93.93% accuracy, 89.82% F1-score, 91.50%
precision, and 88.28% sensitivity, respectively.
Future work: The researchers may encompass model-specific hyperparameter optimization and the
exploration of various data augmentation techniques to further optimize the model and improve its
generalization capabilities
Literature survey
Reference 7
Title: Automatic detection of visual faults on photovoltaic modules using deep ensemble
learning network
Year:2022
Models: Depp Neural Network with Random Forest classifier
GMR Institute of Technology
Dataset: RGB images are taken by using the Unmanned Aerial Vehicle (UAV). They consider the
6 conditions like snail trail, glass breakage, delamination, discoloration, burn marks and normal
one’s. One hundred images were collected for every test condition. (80:20).
Methodology:3150 images were divided evenly into classes for a balanced dataset. Images were
resized to 512 x 512 pixels and converted to grayscale for segmentation techniques. Features were
extracted using FFT, GLDM, DWT, and GLCM, resulting in 252 features. Random forest with
100 trees was chosen for batch predictions. Evaluation metrics included accuracy and ROC curve.
The ensemble-based DNN achieved 99.68% accuracy with only two misclassifications on 630 test
samples, verified by five-fold cross validation. Random forest outperformed other classifiers with
the highest accuracy.
Result: accuracy of 99.68%
Future work:
1) multiple feature extraction strategies can be adopted, and ensemble techniques can be applied
to reduce the dimensional complexity of images.
2) we will develop the proposed model to detect other faults which can be identified by infrared
(IR) thermography and electroluminescent (EL) techniques.
Literature survey
Reference 8
Title:A deep convolutional encoder-decoder architecture for autonomous fault detection of PV
plants using multi-copters
Year:2021
GMR Institute of Technology
Dataset:A dataset of aerial imageries of PV strings affected by bird's drops was collected through
several experimental flights using multi-copters.
Model:For the purpose of bird’s drops segmentation, an improved encoder-decoder architecture
is employed and a modified VGG16 model is used as a backbone for the encoder part.
Methodology:
A dataset of aerial imageries of PV strings affected by bird's drops was collected through
experimental flights using multi-copters .An improved fully convolutional deep network, based
on a modified VGG16 model, was used for bird's drop segmentation .The encoder-decoder
architecture was employed to map low-resolution features to full-resolution ones for pixel-wise
segmentation .An image object positioning algorithm was used to find the exact position of
detected bird's drops in the local coordinate system .Detected damages were prioritized based on
parameters such as severity of shading and extent of impact on PV module's output
current .Different affected PV modules were characterized based on the output patterns of the
classification step to evaluate the effect of bird's drops on module parameters
Result:98% and 93% for training and testing
Future work: expand the fault detection method to include other types of failures and defects on
PV modules, beyond bird's drops.
Literature survey
Reference 9
Title: Intelligent Fault Pattern Recognition of Aerial Photovoltaic Module Images based on
Deep Learning Technique
Year:2020
Models: they used the CNN is adapted to the defects classifications, extracting features by the pre-
GMR Institute of Technology
trained models like vgg16 and added the fully connected layers and the svm decision layer to
classify the defects.
Dataset: this dataset consists of 2000 good conditions of PV modules, 1000 bad conditions of PV
modules containing of 100 encapsulant delamination, 100 encapsulant delamination, 100 glass
breakage,500 dust-shading and 200 snail trails.
Methodology:
Kernels of the 2nd, 4th, 7th, 10th, 13th convolutional layers are followed by a max pooling layer. In 1 st
layer it filters the 224 × 224 × 3 input image with 64 kernels of size 7 × 7 × 3 with a stride of 2
pixels and We extract the features from the hidden fully-connected layer, and each layer has 4096
neurons, which means we extracted a 4096-dimensional feature. After the feature extractions, we use
the SVM classifier to divide the deferent defects. they implemented our CNN feature extraction
strategy in python based on tensorflow, and they randomly cropped 16 patches from the original
image and their CNN model with these patch images. Max-pooling and Rectified Linear Unit
(ReLU) as non-linearity activation were also applied among these layers. they trained their models
using stochastic gradient descent with a batch size of 16 examples, momentum of 0.9, and weight
decay of 0.001.
They consider the two cases. They are case-1: good/bad conditions of the PV modules, case-2:
case-2:
Model Pre-trained Vgg16 Non pre-trained CNN
proposed method
Encapsulant delamination 68.55% 57.50%
78.91%
GMR Institute of Technology
Methodology:
Logistic Regression's optimal parameters were saga solver with 5000 max iterations and penalty 11, yielding
88.76% accuracy and 88.74% F1-score. SVM achieved best results with RBF kernel, 5000 max iterations,
resulting in 91.29% accuracy and 91.25% F1-score. Neural networks found optimal configuration with 2
hidden layers, 300 nodes, relu activation, and adam solver with 300 max iterations, reaching 95% accuracy
and 94.97% F1-score. Quantum SVM utilized 2 q-bits, 1024 shots, and state-vector simulator for 93.58%
accuracy and F1-score of .94. Quantum neural networks employed 2 q-bits, 256 shots, QASM simulator,
learning rate .03, sigmoid activation, achieving around 55-60% accuracy and F1-score around .21.
Result:
Neural networks model gave the best results among the other models which were implemented in the paper.
Future work:
1) researchers should perform tests using real solar array data as opposed to a simulated, hand-picked data set
2) future tests should be performed on real quantum computers instead of quantum simulators.
3) Researchers should also conduct future tests using Monte Carlo Simulations rather than taking a simple
average.
19
Paper 11: Literature Survey
Khan, F. A., Gautam, B. K., & Singh, S. (2020). Detection of defects in solar panels
using thermal imaging by PCA and ICA method. International Research Journal of
Engineering and Technology (IRJET), 4(06), 1700-1703.
The paper titled "Detection of Defects in Solar Panels Using Thermal
GMR Institute of Technology
Imaging by PCA and ICA Method" proposes a method for detecting defects in
solar panels using thermal imaging technology. The methodology described in
the paper involves the use of Principal Component Analysis (PCA) and
Independent Component Analysis (ICA) for feature extraction and image
reconstruction. By analyzing thermal images of both normal and defective
solar panels, the authors demonstrate the effectiveness of their approach in
identifying and locating faults accurately.
PCA and ICA are both dimensionality reduction techniques
commonly used in machine learning and data analysis. In this context, they
are employed to extract relevant features from the thermal images that can
help differentiate between normal and defective solar panels. Once these
features are extracted, a classification step is performed to determine whether
a given panel is defective or not.
Therefore, while the paper focuses on defect detection using thermal
imaging, the underlying process involves classification techniques to categorize the
panels based on their thermal characteristics and identify potential defects.
The results presented in the paper include thermal images of defected solar panels, as
GMR Institute of Technology
well as ROC curves illustrating the performance of the proposed method in terms of
false positives and true positives. Future direction mentioned in this paper are These
include the development of real-time defect detection systems, integration of machine
learning algorithms for automated defect classification.
Literature Survey
Paper12:
Pathak, S. P., & Patil, S. A. (2023). Evaluation of effect of pre-processing techniques in
solar panel fault detection. IEEE Access.
The paper introduces the context of solar power generation,
GMR Institute of Technology
The paper categorizes faults occurring in solar cells into three main
types: Micro Cracks, Snail Tracks, and Hotspots. Each type is discussed in
detail, including their causes, effects on power output, and potential solutions.
Beyond individual cell faults, the paper discusses broader module failures,
including dust and soiling, shading, and packaging material faults. These
factors can significantly impact module performance and require specific
detection and mitigation strategies.
GMR Institute of Technology
The proposed method achieves high accuracy with an accuracy rate of 98.86% in
training and 98.64% in testing, and low false detection and missed detection
The paper acknowledges that there are significant issues in terms of false
positives and missed detections in existing algorithms for defect identification
in solar cell panels, indicating room for further improvement.
GMR Institute of Technology
The study acknowledges the need for continued research and development to
address the issues of false positives and missed detections in existing
algorithms for defect identification in solar cell panels
Literature survey
Reference 16: Jumaboev, S., Jurakuziev, D., & Lee, M. (2022). Photovoltaics plant
fault detection using deep learning techniques. Remote Sensing, 14(15), 3728.
• The paper focuses on the use of deep learning techniques for photovoltaics plant
GMR Institute of Technology
fault detection using thermal images obtained from an unmanned aerial vehicle
(UAV) equipped with infrared sensors .
• Dataset: The paper utilized an open novel PV plant thermal images dataset,
which contains multiple and single damaged cells in images. The dataset was
collected during the inspection of a 66 MW power PV plant in Tombourke,
South Africa.
• Algorithms: The authors implemented three accurate segmentation models,
namely DeepLabV3+, Feature Pyramid Network (FPN), and U-Net, to detect
defective panels on large solar plantations .
• Limitations: The models were evaluated on a specific dataset and may not
generalize well to different types of PV plants or environmental conditions.
• The performance of the implemented models was evaluated using metrics such
as intersection over union (IoU), dice coefficients, and the 30
Jaccard Index (IoU
metric) .
Literature survey
• The paper highlights the importance of high-quality and timely maintenance
of solar power plants to reduce repair costs and increase the overall
generation of electricity.
GMR Institute of Technology
31
Literature survey
Reference 17: Shihavuddin, A. S. M., Rashid, M. R. A., Maruf, M. H.,
Hasan, M. A., ul Haq, M. A., Ashique, R. H., & Al Mansur, A. (2021). Image
based surface damage detection of renewable energy installations using a unified
deep learning approach. Energy Reports, 7, 4566-4576.
GMR Institute of Technology
• The paper discusses the use of deep learning techniques, specifically pretrained
convolutional neural networks (CNNs), for the diagnosis of environmental faults
in solar panels.
• Dataset : dataset comprising four specific image sets for the research
community, including wind turbine optical images, PV panel close-up optical
images, PV panel close-up infrared images, and PV panel long shot optical
images.
• Algorithms: Faster R-CNN, Inception-Resnet-V2 are some of the deep learning
object detection frameworks used for surface damage detection in renewable
energy installations.
• Limitations: The performance of the deep learning models may vary depending
on the modality and resolution of the input images.
• Future Work: To utilize a generalized CNN architecture for different
32 types of
optical image datasets in future work.
Literature survey
Reference 18 : Pierdicca, R., Paolanti, M., Felicetti, A., Piccinini, F., & Zingaretti,
(2020). Automatic faults detection of photovoltaic farms: solAIr, a deep learning-
based system for thermal images. Energies, 13(24), 6496.
• The paper discusses the latest technological improvements in digital cameras
and their combination with affordable costs, which have made PV inspection
GMR Institute of Technology
35
Literature survey
Reference 20: Pathak, S. P., Patil, S., & Patel, S. (2022). Solar panel hotspot
localization and fault classification using deep learning approach. Procedia
Computer Science, 204, 698-705.
• The paper discusses the need for advanced and robust condition monitoring
GMR Institute of Technology
• The paper provides an overview of four major PV array faults and their
causes.
• It reviews both conventional PV array protection schemes and advanced
fault detection and diagnosis (FDD) techniques for managing these faults.
• The paper conducts a performance evaluation of the advanced FDD
techniques using the FM score metric.
• Addresses the limitations of conventional protection techniques and guides
future research in this area.
37
Literature survey
• It introduces a new way of measuring the effectiveness of advanced fault
detection techniques called the Fault Management Score (FM score) and
uses it to evaluate their performance. The paper evaluates the performance
of these advanced techniques using the FM score metric. Lastly, it outlines
GMR Institute of Technology
38
Literature survey
Reference 22:Pillai, Dhanup S., Frede Blaabjerg, and Natarajan Rajasekar. "A
comparative evaluation of advanced fault detection approaches for PV
systems." IEEE Journal of Photovoltaics 9.2 (2019): 513-527.
39
Literature survey
40
Literature survey
Reference 23:Harrou, Fouzi, et al. "Reliable fault detection and diagnosis of
photovoltaic systems based on statistical monitoring approaches." Renewable
energy 116 (2018): 22-37.
• The paper suggests a statistical method for detecting and diagnosing faults
GMR Institute of Technology
42
Literature survey
Reference 24:It Coşgun, Atıl Emre, Yunus Uzun, and A. Cosgun. "Thermal
fault detection system for PV solar modules." Electr. Electron. Eng. Int.
Journal (ELELIJ) 6 (2017): 9-15
• describes various methods used for inspecting PV plants, including
GMR Institute of Technology
43
This paper explores the application of thermal monitoring with infrared
cameras for detecting faults in PV solar modules, highlighting the
benefits of using a wireless directable robotic vehicle equipped with an
RF and thermal camera for this purpose. The study mentions the
inclusion of a multimeter and an infrared camera as measurement units
GMR Institute of Technology
for the test platform, along with providing information on the electrical
characteristics of the PV panel used in the experiment, such as open
circuit voltage, optimum operating voltage, and operating temperature.
However, the paper primarily focuses on detailing the experimental
setup, methodology, and results obtained from the study, rather than
offering an extensive literature review on the topic.
44
44
Literature survey
Reference 25: Ali, Mohamed Hassan, et al. "Real time fault detection in
photovoltaic systems." Energy Procedia 111 (2017): 914-923.
45
The paper introduces a method to quickly spot problems in solar panel
systems. It compares how a faulty solar panel behaves with how it should
work normally. By looking at the differences, it can tell if something's wrong.
To do this, the paper uses a smart computer program called the Particle Swarm
GMR Institute of Technology
Optimization algorithm to help figure out how the solar panel should work.
This method works for most types of solar panels and can even mimic how
shading or damage affects them. Once the program has a good idea of how the
solar panel should perform, it checks the actual performance against this
model. If there's a big difference between what's expected and what's actually
happening, it knows there might be a problem with the solar panel. Overall,
the paper offers a simple yet effective way to detect issues in solar panel
systems by comparing how they should work with how they're actually
performing.
46
46
Comparison Table
Performance
Title year Objectives Limitations Advantages metrics Gaps
GMR Institute of Technology
ANN excels in
capturing complex
nonlinear
Solar PV’s Micro Investigating PV hot-spotting and relationships,
micro-cracks faults, the study suitable for
Crack and modeling fault In the future,
Hotspots aims to develop a fault detection patterns, while it is aimed to
Reference 1 Detection 2021 algorithm using ANN and SVM. Lack of explicit SVMs are robust implement
Technique Using The secondary goal is to compare dataset details against overfitting an average the system to
NN and SVM their performance. and algorithm and adept at accuracy of 87% function
handling high- for feed-forward under
parameters may dimensional data,
impact making them back permanent
reproducibility advantageous for propagation partial
and datasets with neural network shading
comparability. multiple input and 99% SVM conditions
parameters. technique too.
The proposed
deep learning-
based framework mIOU=58.33.accu
for detecting micro racy=97%,F1-
and deep cracks in score=48.3%
PV cells is its for attention
ability to classify Unet
micro-cracks into Limited
Fault detection dataset size
different classes and single-cell
GMR Institute of Technology
datasets for
outperform enhanced
standard models model
robustness
While the
paper The proposed Include
mentions algorithm refining the
extracting achieves a high algorithm
texture accuracy rate of 97% accurate with
information 97% in with the advanced
comparison of feature
"Fault Detection Develop a fault diagnosis system such as classifying test and training extraction
of Solar PV using technical details regarding contrast and defective and results techniques,
System using the implementation of the thermal correlation non-defective PV exploring
SVM and image processing algorithm and from thermal cells, based on alternative
2020
Reference 5 Thermal Image
the integration with Support images, it the experimental machine
Processing." Vector Machine (SVM) does not delve validation learning
into more results. This algorithms or
advanced indicates the ensemble
feature methods
effectiveness of
extraction the approach in
techniques identifying faults
accurately.
GMR Institute of Technology
Comparison Table
Title Year objectives limitations Advantages Performance Gaps
metrics
Ref-3A deep 2021 To develop an The paper focuses The use of aerial 98% and 93% to expand the fault
convolutional autonomous fault specifically on the imageries collected for training and detection method to
encoder-decoder detection method for detection of bird's through experimental testing include other types of
architecture for visually visible drops and may not flights provides a failures and defects on
autonomous fault failures and defects on address other types of comprehensive PV modules, beyond
detection of PV PV modules, with a failures or defects on dataset for training bird's drops.
plants using multi- focus on bird's drops PV modules. the deep network.
copters
GMR Institute of Technology
Ref-4Quantum Machine 2021 To compare classical Uses quantum Proposes that Logistic researchers should
Learning for Solar and quantum machine simulators instead of quantum machine Regression- perform tests using
Panel Fault learning algorithms for real quantum learning can be used88.76% real solar array data as
Detection solar panel fault computers for testing to solve complex SVM – 91.29% opposed to a
detection problems with large Neural simulated, hand-
data sets, which mayNetworks-95% picked data set, future
be useful in the area
Quantum SVM- tests should be
of fault detection 93.58% performed on real
Quantum Neural quantum computers
Networks- 55- instead of quantum
60% simulators.
Ref-5 Intelligent Fault 2020 To present and The images captured The proposed Accuracy- performance and
Pattern Recognition implement an by the UAV system intelligent UAV-based 90.23%(case-1) functionalities of the
of Aerial intelligent UAV-based may have low quality inspection system Accuracy is implemented UAV
Photovoltaic Module inspection system for with noises, which significantly improves mentioned for based inspection
Images based on asset assessment and can effect the the efficiency of each class systems needs to be
Deep Learning defect classification in accuracy of defect system monitoring individually in further validated
Technique large-scale PV systems analysis and diagnosis and fault detection in case-2 through field trails in
large-scale PV large-scale PV systems
systems located in more
complex topographies
Comparison Table
Performance
Title year Objectives Limitations Advantages metrics Gaps
Detection of 2020 Develop a solar panel defect Despite its The thermal The proposed Future
defects in solar detection system utilizing effectiveness, imaging-based method research
panels using thermal imaging to enhance the system system offers demonstrates a could focus
GMR Institute of Technology
Reference 11 thermal imaging operational reliability and may be automated high accuracy on
by PCA and ICA optimize power generation limited by the defect percentage of over enhancing
method. efficiency. initial detection, 90% in detecting real-time
investment improving defects within monitoring
required for operational solar panels, capabilities
thermal efficiency and ensuring reliable through
imaging minimizing fault integration
equipment and manual identification. with IoT and
the need for inspection AI
periodic costs. technologies
calibration. , enabling
predictive
maintenance
and further
optimizing
solar panel
performance
.
Explore the
Evaluation of 2023 Develop an ART2 NN and Reliance on Enhanced fault Achieved 100% integration
effect of pre- MNN-based solar panel fault V_OC data diagnosis recognition rate in of additional
Reference 12 processing diagnosis system using V_OC alone for fault accuracy fault diagnosis sensor data,
techniques in solar with respect to duty cycle as diagnosis may through experiments using such as
panel fault inputs to improve maintenance overlook parallel real data, temperature
detection. efficiency and reduce costs. certain types operation of validating the and
of faults not ART2 NN and effectiveness of humidity, to
reflected in MNN, leading the proposed augment
V_OC to timely neural network- fault
variations, maintenance based approach. diagnosis
potentially interventions capabilities
GMR Institute of Technology
quantifiable ve coverage.
estimates, but
real-world
conditions may
vary.
.
Reference 15 Surface Defect 2023 The study focuses on surface The study The The proposed
Recognition of defect detection of acknowledges combination of method achieves The paper
Solar Panel Based photoluminescence images the need for the improved high accuracy with acknowledge
on Percollation- collected from polycrystalline continued Serre standard an accuracy rate of s that there
Based Image solar panels. research and model, pre- 98.86% in training are
Processing and development processing and 98.64% in significant
Serre Standard to address the scheme, and testing, and low issues in
GMR Institute of Technology
55
Limitation Performanc
Title year Objectives s Advantages e metrics Gaps
To propose deep learning- The models Deep learning- Pixel
Photovoltaics Plant based techniques for detecting were based fault Accuracy
Fault Detection
2022 faults in photovoltaics (PV) evaluated on detection (PA), F1-
Using Deep plants using thermal images a specific techniques score,
Learning obtained from an unmanned dataset and enable quick Precision,
Techniques aerial vehicle (UAV) equipped may not identification of and Recall.
with infrared sensors. generalize faulty
well to components,
GMR Institute of Technology
Cost-effective Limited
Image based surface structural availability of
damage detection of Develop a cost-effective health MAP representative
renewable energy structural health monitoring training samples
installations using a monitoring for deep learning
system for large-scale power
unified deep 2021 system for models due to the
utility facilities, reducing
learning approach large-scale scarcity of
overall maintenance costs and
power utility damages on large
improving safety measures. facilities, structures during
reducing inspections.
overall
maintenance
Reference 17 costs and
improving
safety
measures.
56
Title year Objectives Limitations Advantages Performanc Gaps
e metrics
Automatic 2020 To develop an artificial The proposed Enabling the Iou, To enhance the
Faults intelligence system based on system is trained detection of F1 Score, accuracy and
Detection of deep learning for anomaly and evaluated on various types Dice efficiency of
Reference Photovoltaic
cells detection in a publicly of faults and Coefficient anomaly cell
18 Farms: photovoltaic images available dataset improving the detection
solAIr, a collected overall
Deep specifically for reliability and
GMR Institute of Technology
Environmenta 2022 Timely and accurate Potential for Potential for F1 when dealing
l Fault diagnosis of environmental improvement automated Score,recall with large image
Reference 19 Diagnosis of faults reduces damage fault detection collections, to
Solar Panels caused by faults on the make the
Using Solar panel. monitoring task
Thermal more affordable
Images in and reliable.
Multiple
Convolutional
Neural
Networks
57
Title year Objectives Limitations Advantages Performan Gaps
ce metrics
Solar panel 2022 Develop advanced and robust Enables early F1 Score,
hotspot condition monitoring systems identification MAP(Mean
localization and for photovoltaic solar power and fault avg
GMR Institute of Technology
58
Performance
Title year Objectives Limitations Advantages metrics Gaps
Address the
lack of Include case
standardization studies or
in fault examples of
GMR Institute of Technology
By comparing
multiple
Results from advanced fault
your evaluation detection
may be specific approaches,
to the systems, you provide
The fault detection techniques conditions, or valuable insights
are reviewed based on various datasets used. into their
"A comparative parameters such as types of Discuss how strengths and
evaluation of faults detected, detection time, applicable your weaknesses.
advanced fault sensor requirement, procedural findings are to This can guide
detection complexity, detection variables, other PV future research
approaches for PV and level of protection achieved. systems and and industry
Reference 22 systems." 2019 environments. practices.
59
Performance
Title year Objectives Limitations Advantages metrics Gaps
Photovoltaic
systems are
subject to
various
environmental By quickly
conditions, identifying
which might faults, the
affect the system's overall
To identify faults, the authors use performance of efficiency and
residuals of current, voltage, and statistical performance
GMR Institute of Technology
60
Performance
Title year Objectives Limitations Advantages metrics Gaps
Integrating real-
time fault
detection
systems into
existing PV
systems may
GMR Institute of Technology
61
Existing Model/Architecture of your project
GMR Institute of Technology
62
Proposed Model/Architecture of your project
GMR Institute of Technology
63
In this proposed methodology, we consider the dataset of different images of
solar panel faults like snow covering, physical damage, electrical damage, dust
accumulated, bird drop and cleaned one’s. here the dataset is divided into the
GMR Institute of Technology
training and testing part in the ratio of 80 and 20. by using the training dataset
we will do the preprocessing to the images like resizing normalization, etc. after
completing the preprocessing of the images then we will train the model with
different models like VGG-16, inceptionv3,Mobilenetv3 models in the deep
learning. From the models we will get the features extracted, identify the hidden
patterns in the images which are used for the training of the model. Then by
giving the new image we will test the model from that we will identify which
fault is that by using the classification technique from the models we considered.
64
GMR Institute of Technology
Inceptionv3
METHODOLOGY
65
GMR Institute of Technology
66
GMR Institute of Technology
67
GMR Institute of Technology
68
Inception v3 is a deep convolutional neural network architecture designed
for image classification tasks, developed by Google Research. It features
the innovative Inception module, which uses multiple convolutional filters
of varying sizes (1x1, 3x3, 5x5) and pooling operations to capture spatial
hierarchies and correlations across different scales within the same
GMR Institute of Technology
69
GMR Institute of Technology
MobileNetV3
70
In MobileNetV3 for image classification, input images are preprocessed by
resizing them to a fixed size (e.g., 224x224) and normalizing pixel values.
The network consists of lightweight depthwise separable convolution layers,
applying depthwise convolutions independently to each channel followed by
GMR Institute of Technology
71
GMR Institute of Technology
Dataset Images
72
Mobile NetV3
Summary of the model development process:
1. Base Model Definition: MobileNetV3Small with pre-trained weights from ImageNet is defined,
and its layers are frozen to keep the weights fixed during training.
GMR Institute of Technology
3. Model Compilation: In the model compilation step, we define the Adam optimizer for efficient
parameter updates, employ 'sparse_categorical_crossentropy' as the loss function for multi-class
classification, and select 'accuracy' as the metric to monitor the model's classification performance
during training.
4. Model Summary: A summary of the model is printed to inspect its architecture and parameters.
5. Model Training: The model is trained using the training data (`train`) for 10 epochs, with
validation data (`validation`), and utilizing the defined callbacks for checkpointing and early
stopping.
6.The model yields promising outcomes, boasting a commendable 94% accuracy rate on the
testing dataset and an impressive 99% accuracy rate on the validation dataset
73
GMR Institute of Technology
MobileNet-Model
74
GMR Institute of Technology
MobileNet-Model
75
GMR Institute of Technology
MobileNet-Model
76
GMR Institute of Technology
MobileNet-Model-Outputs
77
GMR Institute of Technology
MobileNet-Model
78
InceptionV3-Model
1. **Input Layer**: The model expects inputs of shape (224, 224, 3), which is typical for many
image classification tasks.
2. **Preprocessing**: The input data is preprocessed using the preprocess_input function from the
GMR Institute of Technology
InceptionV3 model. This function preprocesses the input data according to the requirements of the
InceptionV3 model.
3. **Base Model**: The preprocessed input is passed through a base model, which is likely
InceptionV3 in this case. The base model is loaded using `base_model =
tf.keras.applications.InceptionV3(weights='imagenet', include_top=False)`. This base model is a
pre-trained convolutional neural network (CNN) that has been trained on the ImageNet dataset.
4. **Global Average Pooling**: The output from the base model is passed through a Global
Average Pooling layer. This layer reduces the spatial dimensions of the feature maps to a single
value per feature map, effectively summarizing the spatial information.
5. **Dropout**: Dropout regularization with a dropout rate of 0.3 is applied to reduce overfitting
in the model.
6. **Output Layer**: Finally, the output of the dropout layer is passed through a Dense layer with
6 units ( classification task with 6 classes) and softmax activation, which gives the probability
distribution over the classes. 79
Inceptionv3-Model
7. **Model Compilation**: The model is compiled with the Adam optimizer with a learning rate of
0.001, SparseCategoricalCrossentropy loss function (as the labels are integers), and accuracy as the
metric to monitor during training.
GMR Institute of Technology
8. **Training**: The model is trained using the fit method on the training data (`train`) with validation
data (`validation`) for 15 epochs. Early stopping is employed with a patience of 3 epochs to prevent
overfitting and restore the best weights based on validation loss.
The model yields promising outcomes, boasting a commendable 86% accuracy rate on the
testing dataset and an impressive 99% accuracy rate on the validation dataset
80
GMR Institute of Technology
Inceptionv3-Model
81
GMR Institute of Technology
Inceptionv3-Model
82
GMR Institute of Technology
MobileNet-Model
83
GMR Institute of Technology
Inceptionv3-Model
84
VGG16-Model
• The project aims to build an image classification model using transfer learning
with the VGG16 architecture to classify images into six categories: Bird, Clean,
Dust, Electrical, Physical, and Snow.
• The dataset is organized based on file names containing specific keywords, which
GMR Institute of Technology
determine how images are split into training and testing sets. Images are
preprocessed by resizing to 224x224 pixels using OpenCV and labeled according
to their category.
• The VGG16 model is utilized as a feature extractor with pre-trained ImageNet
weights. Its layers are frozen to retain these weights, and a custom classification
head is added on top, consisting of dense layers with relu activations and a softmax
output layer.
• The model is compiled with an Adam optimizer and Sparse Categorical
Crossentropy loss. It is trained for 25 epochs, monitoring both training and
validation accuracies.
• The entire process is summarized with a model summary and visualized with an
architecture plot, demonstrating the use of transfer learning for efficient model
development with reduced training data requirements.
• The model yields promising outcomes, boasting a commendable 71% accuracy rate
on the testing dataset and an impressive 91% accuracy rate on the validation 85
GMR Institute of Technology
VGG16-Model
86
GMR Institute of Technology
VGG16-Model
87
GMR Institute of Technology
VGG16-Model
88
GMR Institute of Technology
VGG16-Model
89
GMR Institute of Technology
VGG16-Model
90
GMR Institute of Technology
VGG16-Model
91
RESULTS
92
Future Work
In our pursuit of advancing our computer vision models, future work will be directed
towards several key areas for improvement. Primarily, expanding our dataset to
encompass a broader spectrum of diverse and representative images will be
instrumental in enhancing the model's ability to generalize and tackle real-world
GMR Institute of Technology
93
References
1. Sohail, A., et al. (2023). "Fault detection and computation of power in PV
cells under faulty conditions using deep learning." *Department of Electrical
Engineering, NUST College of Electrical and Mechanical Engineering,
Rawalpindi, Pakistan.
GMR Institute of Technology
94
References
1. Duranay, Z. B. (2023). Fault Detection in Solar Energy Systems: A Deep
Learning Approach. Electronics, 12(21), 4397.
2. Venkatesh, S. N., Jeyavadhanam, B. R., Sizkouhi, A. M., Esmailifar, S. M.,
Aghaei, M., & Sugumaran, V. (2022). Automatic detection of visual faults on
GMR Institute of Technology
95
References
1. Khan, F. A., Gautam, B. K., & Singh, S. (2020). Detection of defects in solar
panels using thermal imaging by PCA and ICA method. International
Research Journal of Engineering and Technology (IRJET), 4(06), 1700-1703.
2. Pathak, S. P., & Patil, S. A. (2023). Evaluation of effect of pre-processing
GMR Institute of Technology
96
References
1. Jumaboev, S., Jurakuziev, D., & Lee, M. (2022). Photovoltaics plant fault
detection using deep learning techniques. Remote Sensing, 14(15), 3728.
2. Shihavuddin, A. S. M., Rashid, M. R. A., Maruf, M. H., Hasan, M. A., ul
Haq, M. A., Ashique, R. H., & Al Mansur, A. (2021). Image based surface
GMR Institute of Technology
98