Measurement 232 (2024) 114683

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Comparative investigation of imaging techniques, pre-processing and visual

fault diagnosis using artificial intelligence models for solar photovoltaic
system – A comprehensive review
Gurukarthik Babu Balachandran a, *, M. Devisridhivyadharshini a,
Muthu Eshwaran Ramachandran b, R. Santhiya a
a
Kamaraj College of Engineering and Technology (Autonomous), K. Vellakulam-625701, Near Virudhunagar, Madurai District, Tamil Nadu, India
b
AAA College of Engineering & Technology, Kamarajar Educational Road, Amathur, Sivakasi 626123, Tamil Nadu, India

A R T I C L E I N F O A B S T R A C T

Keywords: Photovoltaic systems provide an eco-friendly key to meet our increasing energy demand while mitigating the
Photovoltaic Systems adverse impacts of conventional fossil fuel-based energy generation. The effective power generation from PV
Machine learning systems can be obtained from fault-free systems. Detecting and correcting faults in solar photovoltaic (PV)
Artificial Intelligence
systems is vital to ensure their best performance, safety and durability. In the existing literature, some have
Inspection Techniques
concentrated only on possible faults in PV, some on inspection techniques for monitoring, and some on Machine
learning models applied in fault detection. In this work, a combined review of the types of possible PV system
failures, image acquisition methodologies, preprocessing techniques, and artificial intelligence (AI) models could
accurately localise and distinguish the faults in PV systems that are presented. This work discusses different
literature on automatic fault detection methodologies. This enables readers to focus on critical aspects while
developing a practical fault detection technique for Solar PV systems.

1. Introduction accomplishing sustainable development by fulfilling energy demands

and then protecting the environment [4]. Solar energy, being the most
The world is experiencing an extraordinary challenge at the inter­ significant energy source, presently experiences noteworthy evolution
section of rising population growth and surging energy demand. As with a capacity of more than 940 GW installed globally in 2021, whereas
global population statistics continue to increase, the stress on conven­ the capacity of 70 GW only in 2011, which is represented in Fig. 1. Solar
tional energy sources and the environment strengthens. The authors [1] technology has experienced noteworthy evolution through innovations
have stated that an increase of 1 % in the population density surges the in materials and manufacturing processes.These technological pro­
electricity demand by 0.77 % and fuel consumption by 0.25 %. Meeting gressions have not only expanded the range and applicability of solar
this swelling energy demand while mitigating the adverse effects of energy but also have some complexities that require vigilant mainte­
fossil fuels on our planet has become a chief concern. In response, the nance. Various faults and performance issues can be caused by the lack
implementation and improvements of renewable technologies have of maintenance that is in accordance with the effectiveness of solar
been evolving as a promising solution to address these crucial issues. systems.
Sustainable development seeks clean and accessible energy entirely, Every year, solar panels struggle from the efficiency loss of 0.5 % − 1
which can only be achieved with renewable energy as they are distrib­ % which results in the reduction of power generation. This loss arises
uted largely across the globe [2]. Solar power remains a principal from electrical and environmental faults [5]. [6] has analysed the
candidate as the finest among renewables for a brighter, cleaner, and mismatch faults of the PV system by considering the electrical param­
more sustainable energy future. [3]It would support the energy prices to eters of voltage, resistance and temperature. Arduino controller is used
get balanced and provide abundant environmental and socio-economic for the analysis. Regardless of the material technology or its type,
profits. This could be indicated by the contribution of solar energy in modules are also subjected to varied environmental conditions related to

* Corresponding author.
E-mail addresses: [email protected], [email protected] (G.B. Balachandran).

https://doi.org/10.1016/j.measurement.2024.114683
Received 21 January 2024; Received in revised form 22 March 2024; Accepted 6 April 2024
Available online 15 April 2024
0263-2241/© 2024 Elsevier Ltd. All rights reserved.
G.B. Balachandran et al. Measurement 232 (2024) 114683

Nomenclature

PV Photovoltaics DT Decision Tree

AI Artificial Intelligence RF Random Forest
I-V curve Current-Voltage Curve LSTM Long Short-term Memory
VI Visual Inspection RGB Red Green Blue
IR-T Infrared thermography AC Alternative Current
UV-F Ultra-violet Fluorescence DC Direct current
EL Electroluminescence LLF Line-Line Fault
PL Photoluminescence GF Ground Fault
FDD Fault Detection and Diagnosis RCAG-Net Residual channel wise attention gate network
MLT Machine Learning Techniques GAN Generative Adversarial Network
CNN Convolutional Neural Network CCD Charge coupled device
PID Potential Induced Degradation ReLU Rectified Linear Units
LID Light Induced Degradation PPV Positive Predictive Value
DCNN Deep Convolutional Neural Network FDR False Discovery Rate
DC Discriminant Classifier ROC Receiver-Operating Characteristic
ANN Artificial Neural Network TP True Positive
SVM Support Vector Machine TN True Negative
FL Fuzzy Logic FP False Positive
kNN k-nearest neighbours FN False Negative

the outdoor arrangement [7]. Those outdoor environmental conditions honeycomb and bridge configurations, proposed by Ganesan et al.,
are harsh for modules, leading them to extreme photochemical or 2023, is capable of detecting line-to-line and open-circuit faults. Imaging
thermo-mechanical stress. In addition to the manufacturing defects, the methods serve as faster diagnostic tools for defect detection in com­
environmental surroundings pay immensely to the PV’s ageing rate, parison. Fig. 2 indicates the number of literatures, identified defects
degradation and defects. Environmental faults include shading, soiling, through five different imaging techniques, including visual inspection
snowing and varying climatic conditions like temperature increase (VI), infrared thermography (IR-T), electroluminescence (EL), photo­
incline to cause substantial loss of power in the PV panels[8]. Physical luminescence (PL) and ultraviolet fluorescence (UV-F). The images are
causes like partial shading, glass breakage, soiling, corrosion, hotspots, acquired using Unmanned aerial vehicles, hand-held cameras, and vi­
junction box failure, bypass diode failure and short circuits, which sual inspection using manpower or electrical characterisation. After the
reduce the reliability, performance and lifetime of PV Modules, make it Image acquisition, The images undergo preprocessing techniques, which
mandatory for the detection of faults at the correct instant. D. P. Win­ include feature selection, Data cleansing, Labelling, and Normalization
ston, 2020 has analysed the modification of string connections with the techniques [10]. Machine learning techniques (MLT) are capable to
junction boxes in the solar panel during faulty and hotspot conditions. work with complex, non-linear problems. MLTs are applied in Fault
Various characterisation methodologies have developed along with Detection and Diagnosis (FDD) which consist of methods with certain
different capabilities for identifying PV defects and have experienced principles and distinct architectures. The common methods include
different complexities in execution. To select a suitable diagnostic Convolutional neural network (CNN) [11]and Artificial Neural Network
method, the existing number of defects needs to be examined first. (ANN) from deep learning, Fuzzy Logic (FL) [12], Support Vector Ma­
Typical diagnosis strategies for PV faults can be broadly divided into two chine (SVM), k-Nearest Neighbor (kNN), Decision Tree (DT) and random
categories: visual inspection and automatic fault analysis. forest (RF) based machine learning techniques. There are different op­
The most popular methodologies for automatic analysis include data- timisations involved in the CNNs to enhance the efficiency of the system
driven methods and model-based residual analysis. For the data-driven in defect classification. The techniques based on Artificial intelligence
methods, Electrical measurements, environmental data or Panel images (AI), including neural networks, are exceptional optimisers for non-
are used. These analyses can be accomplished by using different tech­ linear functions. However, there are some common challenges
niques like statistical methods or by using ML technology. The electrical involved, like configuration of the model, database collection, or
properties can be monitored by recording its current–voltage (I–V) availability to increase the model’s efficiency and scalability in detect­
characteristics [9]. While a single panel’s I–V curve can reflect the ing different kinds of faults. Sakthivel et al., 2023 have analysed
performance largely and it is highly time-consuming to measure each different AI techniques, including wavelet, Fuzzy inference, SVM, and k-
module or panel in the PV plant. A novel fault detection scheme using NN, to enhance fault detection accuracy. Fig. 3 shows the number of
literatures that have employed various types of images. Fig. 4 showcases
the number of literatures that used the different types of AI models.

2. Organization of the review

The paper is interested in carrying out the review of previous studies

majorly to explain the following topics:

• The inspection techniques involved in fault diagnosis

• Different types of imaging techniques
• The causes and effects of faults involved in the solar PV
• Image preprocessing techniques
• Different types and configurations of AI and ML models
• Common challenges and prospects of the models employed

The remaining sections of the article explain the Inspection systems

Fig. 1. Adapted from [44], Evolution of photovoltaic installations globally over
past decade. in the section 2; Identification of solar faults in section 3; Review on

2
G.B. Balachandran et al. Measurement 232 (2024) 114683

Fig. 2. Number of studies included different types of faults calculated since 2010.

3. Stage 1: Inspection techniques in solar PV

Visual inspections become the first option for detecting PV modules’

failure modes and signatures that can be observed through human eyes.
Visual inspection of PV modules should be conducted before and after
the modules have been subjected to the environmental or electrical
factors. Table 3 presents the literature works reviewed majorly the types
of faults and their inspection techniques. Fig. 6 represents the various
types of Inspection techniques employed for the detection of faults in
Solar PV.
Visual inspection easily allows to detect failures during installation
or due to environmental influences and ageing. It majorly focuses on
symptoms due to the failures such as texture changes, damage to back
sheets, discoloration, haziness, bubbles, breakages, etc., than “diag­
Fig. 3. Number of studies employed different types of imaging technology
nosis” (e.g., PID, hotspots, etc.). [13]discussed relation between the
since 2010. visual inspection and the result of electrical test from 608 degraded-
modules when operated in dry and hot climates in Algeria. Likewise,
authors like [14] has utilised visual inspection for diagnosis of various
images required for the study in Section 4; Preprocessing techniques
failure signatures that consist of the colour change in cells or browning,
involved in the section 5; AI models contribution in fault diagnosis of PV
snail trails and failure in junction box while outdoor installation of sil­
in section 6; Finally, the validation metrics for the models in the last
icon modules. Results from Visual inspection are experimental in nature.
section. Fig. 5 represents the organization or layout of the article.
So, they cannot give qualitative explanations about the module failure
reasons since they can only expose the apparent damages.

Fig. 4. Number of studies employed different AI models since 2010.

3
G.B. Balachandran et al. Measurement 232 (2024) 114683

Fig. 5. Organization of the Article.

4
G.B. Balachandran et al. Measurement 232 (2024) 114683

Fig. 6. Inspection techniques for fault detection in solar PV.

3.1. Electrical characterisation employed for the PV fault detection analysis. Those techniques are
capable of identifying various defect modes that cannot be detected by
The electrical (I-V) characteristics comprise a large amount of in­ visual inspections, thereby leading to a severe reduction in PV module
formation about the health condition of panels. V-I curve analysis for performance and also creating safety problems.
different operating conditions involves plotting the characteristics of
output current and voltage from the PV module or PV array. Deviations 3.2.1. EL imaging
from the typical V-I curve can indicate damages like shading, soiling, Electroluminescence is the process in which a material emits light
cell damage, or electrical component faults. Abnormalities in the curve, when an electric current is passed through it. In the context of PV, EL
like the reduction in open-circuit voltage (Voc) or the reduction in short- imaging relies on the principle that defects or imperfections within the
circuit current (Isc), can generate problems within the PV system. The solar cells can emit light when subjected to an electrical bias. A solar cell
Table 2 showcases the electrical characterisation for different types of or module is subjected to a reverse-bias voltage whether it has some
faults observed from reviewed literature. defects. It causes to emit light from defects and areas with reduced
electron-hole recombination efficiency.
3.2. Imaging techniques When a forward bias current is applied, the infrared light is emitted
as a result of radiative recombination. It is obtained using the camera
Imaging techniques of PV arrays are sensible to the degradation with the sensor features of a silicon Charge-coupled Device (CCD). The
modes and defects while manufacturing and installation, and even based sensor records every pixel individually and denotes the spatially
on ageing from the deployment of panels in the field. Various defect resolvable datapoint of photon emission. The captured images are in
modes could be solved if they are identified and distinguished promptly. grayscale. The defective areas are represented as the darkened parts as
In the current works of literature, imaging methodologies like photo­ the faulted parts do not irradiate and are disconnected electrically,
luminescence (PL), electroluminescence (EL), and ultraviolet (UV) which in turn reduces the power production.
fluorescence techniques are popular characterisation techniques that are

5
G.B. Balachandran et al. Measurement 232 (2024) 114683

3.2.2. PL imaging thin layer which covers the solar array’s surface. The particles of the
PL imaging is a contactless characterisation technique similar to EL dust would have less than 10 mm in diameter.[20]. However, it also
imaging, for analysing PV panel conditions by giving information for depends upon location and the environment. The dust can be settled
determining the performance, spatially resolved quality and defects. The from various sources such as pollution by wind, vehicular movements
both techniques EL and PL depends on the acquisition of luminescence and pedestrian volcanic eruptions. The dust accumulation over time
signal released by the material of solar cell but the only difference is the heightens the effect of soiling[21]. The amount of dust accumulated on
way that the signal generated[9]. EL could detect several failures like the PV surface badly disturbs the overall power generated from the PV
PID, cell cracks, electrical mismatch etc., but the PL is able to detect the on a daily, monthly, annual and seasonal basis. This fault is one of the
series resistance and the minority carrier lifetime. most commonly occurring types which can be addressed through
PL could be efficiently carried out without the electrical connections cleaning and maintaining periodically.
to the cells. It does not need any modifications in wiring. When the
charge carriers recombine with those excited by solar irradiation, the
solar cell emits radiative signals that PL can capture. The bright portions 4.2. Shading Issues
of the images indicate the healthy cells, and the darker portions of weak
light emission indicate the defective areas of the solar panel [15]. In The shading effect is caused by impediments like buildings, clouds,
current literatures, authors have employed PL as an imaging tool for trees or nearby erections, which can cast shadows on solar panels and
analysing various failure modes. lead to uneven energy generation. It causes the shaded cells to work as
load and bypass diode to conduct[22]. It results the internal temperature
3.2.3. UV fluorescence imaging to increase for both shaded cell and the bypass diode. If some portion of
UV fluorescence imaging is a non-destructive option and is capable of the PV array is shaded and the remaining regions are entirely exposed to
the examination of PV faults, especially the discoloration of the encap­ the irradiance, the output current considerably decreases. The impact of
sulant of the panel. The UV fluorescence (FL) of EVA was utilised for the Shading Faults on the PV system increases when the number of panels in
first time to analyse the discolouration of PV modules in 1997. UV-FL a row of an array increases.
imaging could be utilised as the alternative technique for EL and PL The partial shading causes a reduction in solar irradiance, which
imaging. thereby reduces the performance of the panel because output current is
UV fluorescence spectroscopy works on the principle of fluorescence, the function of irradiance. The system is disturbed with the non-linearity
which occurs when molecules absorb photons of light and subsequently between voltage and current under partial shaded conditions. According
emit photons of lower energy (longer wavelength) in response.[16]In to [21], the losses of power can vary from 10 % − 70 % due to shading.
UV fluorescence spectroscopy, the sample is exposed to UV or visible Effective planning and designing of the panels by implementing bypass
light of a specific wavelength (excitation wavelength), causing mole­ diodes in solar panels can mitigate shading issues.
cules in the sample to become excited. When the excited molecules re­
turn to their ground state, they release fluorescent light at longer
wavelengths (emission wavelength). Then, the emitted fluorescence is 4.3. Hotspots
measured. It is analysed to obtain information about the sample’s
composition, concentration, and characteristics. The major cause for the occurrence of hotspots is instability of the
solar irradiance and ambient temperature. [23] have experimented and
4. Stage 2: Identification of solar PV faults showcased the ambient temperature while the experimentation is found
to be 26.1 ◦ C. The highest temperature observed in the hotspot area is
The power generated from every photovoltaic (PV) module should be found to be 29.3 ◦ C and the difference in temperature is 3.2 ◦ C for one
summed up to get the whole array power. And the power of the array module. Hotspots occur when a specific cell or portion of a solar panel
should be very much nearer to the power predicted for the normal becomes extremely hot due to mismatched cells, partial shading, or
operation of PV. Though, in practical conditions, there are several fac­ electrical faults. Hotspots are also caused due to degradation, cracks,
tors which lead to the reduced power output from the PV array[17]. contamination, broken diodes, etc….
Faults occur in PV systems due to various factors, including environ­ While reverse bias happens in abundance because of shading or any
mental conditions, manufacturing defects, installation issues and other kind of faults, high power is dissipated in the unhealthy cell, which
equipment wear and tear. Detecting and addressing these faults leads to overheating or hotspot. The hotspot then creates cracks,
promptly is essential to maintain the reliability of a solar installation. breakage, melting of solder, degradation of the solar cells, or even other
Fig. 7 gives the complete chart for the types of faults based on three destructions. (M. M. [24]Prolonged hotspots can cause cell damage or
aspects: factors inducing the faults; DC/AC side of PV; the array level, even fires, making their detection and mitigation critical.
module level, conditioning unit and cable faults. Here are the common
fault classifications occurring in PV systems explained in this section.
Table 1 shows the different literature papers in which various ML 4.4. Encapsulant Discoloration
models applied for specific fault detection, their preprocessing tech­
niques and the evaluation criteria are discussed. The change in color or appearance of the encapsulant material causes
the encapsulant discoloration where the encapsulant material is used to
4.1. Dust Accumulation protect and insulate the electronic components or photovoltaic cells.
This encapsulant material is often a polymer or resin [25]. Continued
The settlement of dust, pollen, and debris on the solar panel’s surface exposure to UV radiation from the sun can cause degradation and
could block the sunlight and diminish the efficiency of the panel. change in color of encapsulant material to degrade. Then, the color will
Temperature, irradiation, air pressure, polluted air, Wind speed, direc­ be changed. The molecular structure of the material will be broken by
tion of wind, dust storm and humidity are the environmental factors of the UV radiation. High temperatures can accelerate the ageing of
dust accumulation[18]. The daily average energy reduction is about 4.4 encapsulant materials and that would lead to discoloration [26]. Expo­
% due to the dust settlement on PV panels for 1 year. It may also raise up sure to severe chemicals or environmental pollutants can react with the
to 20 % on without rainfall condition for long time[19]. encapsulant, affecting appearance. Discoloration may lead to loss of
Solar panel temperature from 1 − 10 degrees Celsius could be functionality and reduce performance characteristics of encapsulated
credited to the external factors like soiling or dust. Actually, dust is a components.

6
G.B. Balachandran et al. Measurement 232 (2024) 114683

Fig. 7. The chart for different types of possible solar PV faults.

7
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 1
Inference from different literature papers reviewed.
Author Type of work Proposed technique Dataset Inference Evaluation
details techniques

[37] Early diagnosis of the hot-spots • SVM Not mentioned DC performed well with • Confusion matrix
• KNN accuracy = 98 % • ROC
• Decision tree
• Discriminant classifiers (DC)
[35] Automatic fault diagnosis by • Normalization • UAVs • 1000 images set CNN 99 % Not mentioned
thermographic images tool accuracy
• Homogenization • ground-based operators – • computational time- 30
Dataset of 1000 images min using Low Mid-Range
CPU −
• Discrete wavelet transforms • 200 images set 90 %
• Grey scaling accuracy using MLP,
• Thresholding • 100 % accuracy using CNN
• Box blur filtering,
• Sobel Feldman
[45] Surface damage detection Deep learning-based inspection Four specific image sets accuracy = 0.79 Not mentioned
image analysis methods. from high- to low-resolution
images
from thermal images to visual
images
from PV panels to wind
turbines
[32] Electroluminescence (EL) images Channel attention, spatial Not mentioned Mean intersection over-union 5-fold cross
attention into Multi attention U- (m-IOU) = 0.699, F-measure validation
net = 0.799.
[40] IR thermographic images gamma correction function Not mentioned High detection accuracy and Confusion matrix
(preprocess) improved + (CNN); low time consumption
CNN model using IR
temperatures;
eXtreme Gradient Boosting
(XGBoost) algorithm replaced
with CNN.
[46] PV modules with micro-cracks, Feed Forward and Back Not mentioned • average accuracy − 87 % %Accuracy
hotspots Propagation technique and SVM for feed-forward back
propagation neural
network
• 99 % for SVM
[47] Multiple large-scale images Deep joint learning model Not mentioned • Outperforms the Precision
transformed to detect the overheated benchmarked approaches Recall
region precisely. by its effectiveness and
efficiency
[48] Kirsch Operator Based Image Deep convolutional neural Comprehensive dataset • Compared with Three-Fold Cross
Segmentation − multiple visible network-three-phase ML collected from real-world conventional pattern Validation
defects approach. solar PV plants recognition algorithm,
(Encapsulant delamination, Dust • Pretrained Vgg16
shading, Glass breakage, Gridline
Corrosion, Yellowing and Snail trails) −
RGB colour images from UAV
• Standard CNN-Based
solution
[39] IR camera mounted on UAV RCAG-Net • 700 defect-free images • Performance improved to • precision
• 312 defective images 92.61 % of F-measure, and • recall
recall for the hot spot • F-measure
defect to 97.06 %
[49] Planetary gearbox, gear, and bearing- Fault diagnosis method based on First 1000 samples for Accuracy = 99.5 % for single- 99.5 %. Accuracy of
gear mixture, including the fault time–frequency representation training and last 1000 samples speed load cases, also the model-
diagnosis of bearing. (TFR) and deep reinforcement for testing where each sample outperforms in multi-work Planetary gearbox
learning (DRL) with 1280 data points conditions
[50] IR camera mounted on UAV rgSIFT descriptor with k-NN • 300 images (80 % of total • Naive Bayes Accuracy = 98.7 %
375) for training.
• Remaining 20 % images (75 • decision tree
images for testing • Stochastic gradient
descent,
• Random forest
• k-NN
[51] Not mentioned • Review on 9732 images with different Not mentioned The average
• GAN rotation angles of 90◦, 180◦ accuracy ranged
• Deep CNN and 270◦ for training. between 95 % and
• Deep convolutional GAN 98.4 %

4.5. Encapsulant Delamination away from each other or from the substrate they meant to protect, it
becomes encapsulant delamination. When the bonding between the
Encapsulant delamination is a significant and one of the common encapsulant layers or between the encapsulant and the substrate is
field failures. When the layers of encapsulant material detach or peel inadequate, it can lead to delamination [27]. Frequent temperature

8
G.B. Balachandran et al. Measurement 232 (2024) 114683

fluctuations can cause differential expansion and contraction that leads loss from the shaded or malfunctioning cells. When bypass diodes fail, it
to delamination over time. Moisture can infiltrate the encapsulant and can cause a drop in overall panel efficacy, which might need replace­
weaken the adhesion that causes delamination. Excessive mechanical ment. Bypass diodes are one of the important components in PV panels
stress can also result in delamination. Delamination can expose sensitive and arrays, exclusively in situations where multiple solar cells are con­
electronic components or the PV cells to environmental factors which nected in series[22]. Bypass diodes are used to prevent power losses due
cause damage and reduce performance. Voids and air gaps can be to shading, soiling, or cell mismatch. When solar cells in the panels are
created because of the delamination, which dissipates heat and affects partially shaded or even not functioning appropriately, the bypass diode
the overall system integrity (M. C. C. [28]. permits the current to bypass the shaded or faulty cell, ensuring that the
Addressing discoloration and delamination of encapsulant material other cells produce electricity.
is essential for making the system reliable, and that leads to the long life
of various encapsulated systems, from electronic devices to solar panels. 4.10. Wiring and connection Problems
Selecting proper materials, manufacturing processes, and environmental
controls are essential for preventing these issues. Consistent inspection Faults in the wiring, loose connections, or corroded terminals could
and maintenance could help one to identify and address the problems as cause electrical losses, increase resistance, and cause overheating.[31]
they arise. Periodical inspection and maintenance are vital to make sure that all
electrical connections and components are safe and functional. Proper
4.6. PID (Potential-Induced Degradation) wiring and connections are essential to guarantee the effectual and
secure operation of a PV system. Fig. 8 presents the types of faults that
PID arises when leakage currents are produced due to the voltage occur in the solar panel components.
potential between the frame and the solar cell, which reduces the per­
formance efficiency of the solar panels. Ground fault protection systems 5. Stage 3: image preprocessing techniques
and PID-resistant modules can aid in avoiding the degradation problem.
PID is the phenomenon that has the potential to negatively impact solar Image preprocessing is a critical step in developing an automatic
PV panel performance and, over time, reduce power output. When fault detection system using AI algorithms. Table 4 shows the papers
various parts of a solar PV system establish an unexpected electrical that reviewed the detailed procedure of automatic fault detection,
potential, often by voltage difference, it causes the solar cells to degrade. including image acquisition, preprocessing, AI models deployed, and
These solar complications must be found and solved through regular their evaluation metrics.
inspections, monitoring systems, and timely maintenance. A positive
approach to problem management ensures that a solar installation runs 5.1. Noise Reduction
effectively, maintains its durability, and maximises the return on in­
vestment and contributing to sustainable energy production. As real-world environments often contain noise, preprocessing
techniques are employed on raw solar panel images to improve quality,
4.7. Light-Induced degradation (LID) reduce noise, and prepare them for effective analysis by AI algorithms.
[32] Noise Reduction includes techniques like Gaussian smoothing
LID is a phenomenon that occurs particularly in monocrystalline and Median Filtering. Applying a Gaussian filter helps to reduce high-
silicon, where performance efficiency temporarily reduces, immediately frequency noise and smoothens the image. Median filtering is an effec­
after installation. LID can lead to efficiency reduction in the PV panel by tive technique that removes the salt-and-pepper noise which can occur
up to 2–3 % during the initial hours or days of exposure. This effect leads in images captured under difficult conditions [33].
to defects in the wafer itself, which is produced by continued exposure to
light [29]. This can be mitigated with specific materials and 5.2. Image Enhancement
manufacturing processes. LID is a phenomenon that can disturb the
panels’ initial performance shortly after they are exposed to sunlight. Image enhancement techniques, such as contrast adjustment, histo­
LID generally occurs in the initial hours or days of sunlight exposure by gram equalization, and sharpening, can be implemented to improve
the panel and can lead to a temporary reduction in solar cell efficiency image quality. These techniques help make images visually appealing
[30]. LID is chiefly associated with the interaction between boron and and help to highlight some special features. If multiple images are
oxygen within the silicon material utilised in solar cells. When the solar employed, they might need to be registered to ensure that they align
cells are exposed to sunlight, this defect is activated, which leads to a correctly.
reduction in the efficiency of cells. Solar panel images are obtained using cameras or other imaging
devices placed within or near the solar PV installation. The images
4.8. Inverter Failures contain information about the complete solar array or may even focus on
individual solar panels or sections. Contrast adjustment and brightness
Inverters are fundamental components of PV systems which convert correction are involved to enhance the Images acquired. They help
DC power produced from the panels into usable AC power. Inverter improve the visibility of information and potential faults.
faults or failures can upset the entire system and may result from issues Techniques like histogram equalization or contrast stretching can be
such as component wear, overheating, or electrical faults. Inverter applied to images [34]. Adjusting the brightness levels warrants the
failure is a common issue in solar power systems, which can significantly image is well-exposed, and that avoids underexposure or overexposure
influence the efficiency of solar PV installation. The efficiency of the issues.
inverters declines at a minimal rate once it attains the peak level with
incident energy of around 400–700 W/m2. It happens because of the 5.3. Standardization
temperature rise within the inverter while it supplies a heavy load[29].
Inverters are essential as they convert DC power from solar PV to AC Standardizing image size, color, or intensity can be important for
power that can be used in homes and also fed into the grid. consistency analysis. While training the machine learning model, fixed-
size images are needed for the essential training of ML models, which
4.9. Bypass diode Failures could improve the computational efficiency during processing.
If necessary, color conversion can be done, where Solar panel images
Bypass diodes are employed in solar panels in order to prevent power are often in RGB colors, but for fault detection, they may be converted to

9
G.B. Balachandran et al. Measurement 232 (2024) 114683

Fig. 8. Faults that occur in the panel components.

grayscale to simplify processing and reduce computational load. Resiz­ 5.6. Normalizing feature Data
ing the image to a standard size can also help guarantee reliable input for
AI algorithms. It can also reduce computational complexity. Extracting relevant features from the preprocessed image is essential
for fault detection. These features may include texture, shape, and in­
5.4. Isolating regions of Interest tensity information. Then, Normalizing the feature data to ensure that
all features have the same scale. Common normalization techniques
[33]Segmentation techniques can be used to isolate regions of in­ include mean normalization and min–max scaling. Then, the pre­
terest (ROI) within the images and make it easier to focus on the objects processed image dataset will be separated into training, validation, and
or areas where defects might occur. test sets for training and then for evaluating the efficiency of AI model.
Identifying the region of interest within the image, which actually
includes the solar panels and surrounding areas. This step reduces the
5.7. Image Annotation
computational complexity and focuses the analysis on the appropriate
portions of the image. Image Segmentation is the vital part that includes
The preprocessed data is then saved along with labels which in­
Thresholding and Edge detection. [35] Thresholding techniques can be
dicates the presence and absence of the faults and then used for training
employed to separate the solar panels and their surroundings from the
and testing AI algorithms. The process of labelling or adding metadata to
background. Edge detection algorithms like Canny or Sobel can be
images is to provide additional information about the objects or regions
employed to outline the panels and distinguish them from rest of the
within the image.
image. Morphological operations, such as dilation and erosion, can be
The preprocessed images are then fed into AI algorithms for training
applied to clean up the segmented image and remove minor artifacts or
and detection of panel faults. Proper image preprocessing is critical as it
gaps in the panels.
helps the AI model focus on relevant information, reduce noise, and
enhance its ability to detect faults precisely and proficiently.
5.5. Augmentation Techniques

To improve the robustness of an AI model, augmentation techniques 6. Stage: 4 Artificial intelligence models
like rotation, flipping, and cropping can be employed to generate
additional training samples, which are optional in the process [35]. ML is the subsection from AI. ML is the one of fresh divisions which
has arisen recently in engineering after the IInd World War to learn and
develop intelligent units. ML flourishes on mining the meaningful details
from plenty amount of data. Hence, ML discusses technology that an­
swers questions by using the current availability of data. Artificial

10
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 2
Electrical characterisation of corresponding types of fault.
Fault condition Corresponding electrical characterization

Shading condition with and without bypass diode

Degradation, partial shading, open circuit and

short circuit.

(continued on next page)

11
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 2 (continued )
Fault condition Corresponding electrical characterization

• CONTACT FAILURE
• LID FAILURE
• PID FAILURE
• CIRCUIT FAILURE
• GLASS ANTI REFLECTIVE COATING
DEGRADATION
• DIODE FAILURE
• CELL INTERCONNECT FAILURE
• DISCOLORATION
• DELAMINATIONCORROSION OF PANEL

FRACTURED SOLAR CELL

POTENTIAL-INDUCED DEGRADATION (PID)

QUICK CONNECTOR FAILURE

DELAMINATION

(continued on next page)

12
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 2 (continued )
Fault condition Corresponding electrical characterization

INTERNAL CIRCUITRY DISCOLORATION

ENCAPSULANT DISCOLORATION

JUNCTION BOX/BYPASS DIODE

LID/LeTID

intelligence, Statistical analysis, Information theory, computational defect labels. In order to minimise the computational complexity, the
complexity, biology, control theory, Philosophy and cognitive science 2nd and 5th convolution layers are followed by the max-pooling layer.
are some fields from which ML attracts on theories and proceedings. Training parameters used are given by Batch size, Momentum, Learning
ML and AI can be seen in our day-to-day activities taking from rate, Weight decay and epochs. High accuracy is obtained with the
transportation-self-driving and self-parking cars, banking-fraud detec­ model in classifying the fault types.
tion and in finance sector, google searches, social media applications, In the work of [34], for classifying thermal images of PV, The
image recognition, recommendation systems, prescription and in the Comparative analysis is conducted using different ML algorithms based
medical field also. In Solar PV applications, it is used in tracking the on their performance in training the features. Using 3 different datasets
duration of sunshine and its clearness index, the solar irradiation, mean (dataset I, dataset II, and dataset III), three experiments are conducted
temperature, diffusion fractions and the insolation. In addition to that, on training those datasets and proposed a hybrid feature dataset III for
they are employed for modelling, simulation, sizing, configuring and study. For the first dataset, the optimised SVM hyperparameters for D
control of systems, forecasting of output power obtained from stand­ and γ are found to be 1 and 0.004. For second dataset, D and γ obtained
alone PV systems and grid-connected PV systems and in fault diagnosis. are as 0.46 and 0.1. For third dataset, 0.2 and 0.46 are obtained for D
Table 5 shows the literature papers that reviewed different artificial and γ. The SVM model proposed, predicts 305 out of 315 thermal im­
intelligence models for different classifications of solar faults. ages. Confusion matrix, positive predictive value (PPV), false discovery
The paper [36] has proposed the CNN architecture which has 5 rate (FDR), receiver-operating characteristic (ROC), F-Measures, and
convolution layers and 3 fully connected (FC) layers. By employing training time for all 3 training datasets are discussed. The model trained
different kernels, Convolution layers are there to extract the features has shown the highest accuracy of 92 % while testing with the new
from EL images. The extracted features are then sent to FC layers. dataset.
Finally, the output of FC layer is made available to 4-way classifier. The The contribution of M. R. U. [32] is towards Multi attention U-net
classifier adopted non-linear ReLU, and the features are classified as 4 (MAU-net). The proposed model contains two layers namely channel

13
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 3 Table 3 (continued )

Papers which reviewed about the classification of faults and their inspection Ref Inspection/Diagnosis Fault types
techniques.
[27] • Scanning electron microscope • Failure modes discussed in this
Ref Inspection/Diagnosis Fault types (SEM) review:
[52] • Scanning electron microscopic • Behaviour of dust: deposition, • Energy-dispersive X-ray • Electromigration
images of dust rebound, and resuspension spectroscopy (EDS)
[53] • EL imaging camera − • Yellowing/browning of • Images − silver fingers at • Delamination
microcracked cell encapsulants and back sheets delaminated areas
• FLIR thermal imaging camera – • Bubble formation • Corrosion
Hotspot Solar cell • Delamination of encapsulants and • Potential Induced Degradation
back sheet [58] NA • Frame defects
• Discoloration of busbars • Junction box defects
• Oxidation formation • Milky pattern
• Hotspots • Darkening
• Corrosion of connection • Cell cracks
• Cracks in back sheet • Hot spots
• Cell breakage • Front grid and AR layer oxidation
• Microcracks • Bubbles
[54] • IR − imaging • Whole-cell part • Back-sheet delamination
• Main aspect – Authors – Year – • Part of hot cell [18] • Thermal imaging • Dust properties
Citation • Bird droppings
• IR-images showing thermal • Single hot point = hot spot • Fig. – Hail storm affected PV
anomalies module
• Thermal signature category – • Uniform hot substring patchwork • Multiple direction breakages
Description – Reason pattern of hot cells [23] • NA • Failures in:
• Hot junction box • Cell gridlines
[9] • I-V Curve • Cell cracks • Wires/Connectors
• Visual Inspection • Hotspots • Back sheet
• Ultraviolet fluorescence • Delamination • Delamination
• Infrared thermography • Internal circuitry discolouration • Frames
• Luminescence techniques • Potential induced degradation • Between cell gaps
• Electroluminescence • Light and elevated temperature- • Metallization
induced degradation • Interconnect Ribbons
• Daylight Photoluminescence • Light-induced degradation • Junction Box
• Infrared thermography • Failures in bypass diode, junction • Cell crack
box, quick connector • Front Glass
• Electroluminescence • Resistive solder bonds
[55] • Visual Inspection • Typical degradation modes with • Bypass diode
moisture ingress: [59] • Thermography images • Hot spot
• Electrical Parameter • Encapsulant discolouration • EL images • PID
measurement • Visual, Imaging, Data analytic • Crack at cells (snail track)
• Imaging • Cracks and breakages using IV curve, Signal analysis,
• EL Imaging • PID …
• PL Imaging • LID • Browned cell
• UV Fluorescence Imaging • LeTID [60] • Photobleaching effect • Edges and through cracks
• Adhesion loss • Delamination of front
• Corrosion encapsulant
• Delamination • Bubbles
[56] NA • Shading • Burn marks
• Soiling • Discoloration
• Dust Accumulation • Back-sheet delamination
• Weather • Corrosion
• Delamination • PID
• Discolouration • Snail trails
[57] NA • Manufacturing process induced: [44] • Infrared thermography • Resistive solder bonds hotspots
• Finger Interruptions • Thermography • Types of defects and IR pattern
• Finger cracks (M. C. NA • Discolouration of EVA
• Finger shunts C. encapsulant
• Cell and Interconnect warpage [28] • Corrosion in metallic contact
• Thermo-mechanical fatigue • Delamination
induced degradation: • Bubbles
• Solder joint breakages [61] NA • Classification of faults
• Intermetallic compound cracks • Arc faults
• Finger breakages • Shade faults
• Interconnect breakages • Other faults
• Chemical modes of degradation: [62] NA • Encapsulation failures
• Silver oxide depositions • Backsheet adhesion loss
• Glass frit corrosion and finger • Cell cracking
delamination • Broken interconnection
• Solder material corrosion • Shading and soiling
• Snail trails • Hotspots
• Interconnect corrosion • Module corrosion
• System voltage assisted • PID
degradation: • LID
• PID assisted delamination at • Inverter failure modes and other
fingers failure modes (Junction box,
• Interconnect burnouts Bypass diode, mismatch faults,
Ground fault, LL fault, Arc faults)
(continued on next page)

14
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 3 (continued ) Table 3 (continued )

Ref Inspection/Diagnosis Fault types Ref Inspection/Diagnosis Fault types

[31] • Infrared equipment • Fires • Electroluminescence imaging • Broken interconnect

• Hotspot (EL imaging)
• Diode fault • Ultrasonic inspection • Junction box failure
• Shading • Infrared imaging (IR imaging) • Solder Bond Failures
• Soiling • Open circuits leading to Arcing
• Dust accumulation • Delamination of encapsulant
• Oxidation and discoloration • Corrosion
• Broken glass • Encapsulant loss of adhesion and
• delamination and bubbles elasticity
[29] NA • Dust accumulation • Hot spots
• Shading • Encapsulant discoloration
• Soiling of PV panels • Shunts at the scribe lines
• Degradation-Glass breakage, • Ground fault
Hotspots • Electrochemical corrosion of TCO
[22] • Electroluminescence test in • Extreme Dirt and cell cracks • Bypass diode failures
dark room [69] NA • Dust on glass
• detection of faults (IV by • Failures in bypass diodes • Transmissive degradation −
capacitor method, Transparent materials
[63] • c-Si based technologies • PID mechanisms in PV [70] NA • Degradation of PV:
• Thin film technologies • Module level: Glass material, • Discoloration
• Test methods for PID Encapsulation • Corrosion
susceptibility: • Delamination
• PID testing on module level • Breakage and cracks in cells
• PID testing on cell level • PID
[17] • Fault detection methodology: • Classification of faults • Hot spots
• IV characteristics • Side (AC/DC) – Type of fault- • Bubbles
• Visual inspection Description-Ref. [14] • Visual Inspection, • Degradation modes:
• IR thermal imaging Measurements:
• Ultrasonic inspection • Dark I-V measurements • Optical degradation
• EL imaging • Individual cell shunt resistance • Front surface soiling
• Lock in thermography measurements
[21] NA • Shading by soiling on PV • Light Intensity measurements • Cell degradation
performance • Worst-Case Cell Determination • LID
• Dust accumulating • Determination of Temperature • Temperature-Induced
[64] NA • Hot spot Coefficients Degradation
• Glass breakage • Monitoring Isc and Voc • Mismatched cells
• Ribbon discoloration
• Encapsulant discoloration
• PID attention layer and spatial attention network to highlight defective area
• Cell breakage especially. The model proposed is evaluated using 5-fold cross-
• Corrosion in cells
[30] NA • Boron-oxygen degradation
validation while implemented on the real PV − EL image datasets.
• Copper related degradation The results have shown the network which is able to segment and
• Quasi-mono and multi-crystalline identify complex defects exactly. The mean intersection over-union (m-
silicon IOU) obtained as 0.699 and F-measure as 0.799 that performs well when
[65] • I-V pattern • Optical degradation (glass
compared with previous models.
breakage and delamination)
• Infrared thermography • Electrical mismatch and The quadratic discriminant classifier is studied for fault detection in
diagnosis degradation the study of[37]. MATLAB is used for the analysis. The minimum ac­
• Thermal image processing • PID, Defective bypass diode curacy, 25 % is obtained using DT for 1st dataset, and 98 % of maximum
• Thermal orthophotoplan • Cell cracks, consequent snail trails accuracy obtained using 3rd dataset. The classifications considered in
technique
• Degraded soldering
the study are healthy PV module, one hot-spotted solar cell, two hot-
• Broken interconnecting ribbons spotted solar cells, and three or more hot-spotted solar cells. Thus, the
(disconnected cells) ML tool’s accuracy helps predict early-stage hotspots.
[66] • Crack Statistics • Crack types: The paper[38]has used ensemble-based DNN and RF classifier to
• Monte Carlo Simulations of • Dentritic
classify glass breakage, snail trail, delamination, discoloration, burn
Cracked Modules • Several directions
• +45 degrees marks and healthy panels. The performance of ensemble-based DNN
• − 45 degrees model is evaluated by comparing it with the pre-trained models.
• Parallel busbars Transfer learning is also used where trained weights of other networks
• Perpendicular busbars are shifted to designed model with less changes in final layers. In the
• Crosscrack
[67] NA • Catastrophe faults:
field of computer vision, various pre-trained models are applied for the
• Ground faults classification tasks.
• LL faults The study by [33] employs encoder–decoder neural network. These
• arc faults networks are built of contraction part and symmetric expansion part.
[68] NA • Degradation in terrestrial PV
The former compresses the information of high-dimensional image into
• Statistical analysis
• Degradation rates due to feature sets. Then the latter gradually updates the encoded features back
accelerated aging to original resolution. Encoders and decoders combination provides the
• Indoor PID testing and Outdoor segmentation networks. There are tuning parameters which allow to
PID testing generate multiple segmentation models. For encoder part, Mobile-net,
[26] • Electrical characterization • Degradation
• Visual inspection • Broken cell
ResNet, VGG-net, and U-net are used. For decoder part, U-net, FCN-
net, PSP-net, and SegNet are used.
[39] have proposed a residual channel-wise attention gate network

15
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 4
Papers which reviewed about the detailed procedure of automatic fault detection.
Ref Inspection Fault types Preprocessing Model developed Performance indices

[71] • Infrared module dataset class • Damage in diode • Data augmentation • CNN • K-fold cross validation
• Damaged cells due to Shading or • Non-linear activation • Z-score normalization
severe soiling functions: RelU and Softmax • Loss function
• Shadowing • Confusion matrix
• Hot-Spot
Cracking
• Soiling
• Vegetation
• No-Anomaly
[72] • EL images • Cracking orientation: • Data augmentation • Deep learning approach NA
• CDCR images • Parallel to busbar
• RGB images • Dendritic/Branched
• Deep cracks isolate the cell parts
• Perpendicular to busbar
• − 45 degrees
• +45 degrees
• Other faults are:
• Contact forming failure
• Finger failure along cracks
• Silicon material
• Finger failure
[10] • External quantum efficiency plans • PV System failures: NA • Classification of ML NA
• X-ray photos − normal and • Shading Approaches
unhealthy condition of bypass diode • Degradation • Conventional ML methods for
Open-circuit Condition monitoring
• Bypass Diode • Advanced DL methods for
• Line-to-Line Condition monitoring.
• Bridging • Knowledge-guided methods
• Illustration of Common degradation for Condition monitoring.
types:
• Hotspots
• Ribbon discolouration
• Glass breakage
• Encapsulant discoloration
• Delamination
• Bubbles
[73] • Electrical parameters taken at • degradation assessment on ground NA NA • Final Yield
Standard Test Conditions (STC) mount PV plants and rooftop PV • Reference yield
plants. • Performance ratio
(PR)
• Levelized cost of
electricity (LCOE)
• Capacity Utilization
Factor (CUF)
• PV System efficiency
• Annual Degradation
rate
[74] NA NA NA • ANN • RMSE, MBE, MAE, R2
• FUZZY LOGIC
• Genetic Algorithm (GA)
• Hybrid models:
• ANN + GA and FL + GA
• Adaptive neuro-fuzzy inter­
face system (ANFIS)
[75] NA • Line to line faults • K-fold cross • Ensemble technique voting • Confusion matrix
validation K = 10 based probability
• Feature extraction • Classifier − (NB, SVM, KNN)
• Data normalization
• Feature Selection
algorithm
[76] • Visual inspection • Faults due to exte Ensemble NA • Wavelet – Multi layer • Computational time
technique voting based rnal causes Perceptron neural network • Accuracy (%)
• Thermography • Snail trails (MLPNN) • Confidence intervals
• Electroluminescence • Defective bypass diode
• Photoluminescence • Broken interconnects
• Microplasma luminescence • PID

(RCAG-Net) to attain a fusion of multiscale features to suppress the identification.

complex and unwanted backgrounds and then highlight the major fea­ Three different analyses are done in this paper [40]. Improved
tures of defects mainly to identify small hot spot defects. In this paper by gamma correction function is combined with CNN for preprocessing in
[11], Different CNN architectures are used for semantic segmentation of the first method. IR images of PV modules are employed for the study.
background elimination and segmenting the faults from the panels. Secondly, Threshold function is used as preprocessing tool and given to
Then, classification is also done using CNN architecture for fault CNN model for classification. Then, XGBoost is used for classification

16
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 5 Table 6
Papers which reviewed only the different types of models and the fault Papers which reviewed only different types of field inspection techniques for
inspection. fault classification.
Ref Inspection Model Reviewed Ref Inspection

[77] • Endogenous data • Recursive autoregressive and moving [84] • Infrared thermography (IRT)
average (RARMA) • Aerial Infrared Thermography
• Meteorological • Coupled autoregressive and • Unmanned Aerial Vehicles (UAV)
measurements dynamical system (CARDS) [85] • Thermographic cameras for PV inspection
• Numerical weather • ANN with only the past irradiance • Thermographic images captured and RGB image captured
prediction data in global horizontal irradiance • Thermographic image and Light weight aerial thermographic camera
forecasting image captured
• Satellite data • AI based forecasting • Aerial thermographic inspection
• Sky images • KNN-based forecasting [86] • Electroluminescence
• Spatio-correlated • ANN-based solar forecasting • Photoluminescence
hierarchical information • Electron beam induced current
• DL-based solar forecasting • Laser beam induced current
• Ensemble learning-based solar • High critical temperature superconductor superconducting quantum
forecasting interference device (HTS-SQUID)
(M. M. • Visual Inspection • Clustering based computation • Scanning electron acoustic microscopy
[24] • I-V Characteristics • Lamb wave air coupled ultrasonic testing (LAC-UT)
• EL Imaging • Resonance ultrasonic vibration
• IR Imaging • Visible optical NDT
• UV–FL Imaging • Dark lock in thermography
• SEM Images • Induction thermography
• EBIC Imaging • Illuminated lock-in thermography
(B. • Visual images • Shallow neural network (SNN) • Quantitative lock-in carrierographic (LIC)
[78] • Electroluminescence • DNN models • Machine vision
images • Spectral and spatially resolved photoluminescence
• Infrared images • Hybrid ANN for Fault detection • Mechanical test
• 3 scenarios of ANN (SNN, DNN,
Hybrid application)
[79] NA • Artificial Intelligence and Machine
Learning types Table 7
• Distribution of AI-based fault detec­ Papers which reviewed about the influence of economics in the Solar panel.
tion methods for PV systems found in
Ref Faults considered Inference
the available literature.
• Neural network-based methods in [9] • Cell cracks • Ranked based on Cost Priority
existing literature Number calculations
• Regression-based methods in existing • PID • Strongest financial impact lead by
literature: Cell cracks, PID short-circuited
• Decision tree-based methods in exist­ Bypass Diodes
ing literature: • Short circuit bypass diode • PID are the energy losses
• SVM-based methods in existing • Cell cracks constitute fixing costs
literature: [87] • Failure or degradation of • Ranked based on risk priority
• Neuro-fuzzy-based methods in components and subcomponents number
existing literature: of PV module • In first place the inverter,
• Wavelet-based methods in existing grounding/ lightning protection
literature: system.
• Other methods in existing literature • In a second position, the modules,
[80] NA • DL-based methods for fault diagnosis cells and contacts.
of PV systems [88] • Mono-string mono-module fault • Influence of normal interest rate
[81] • Infrared Thermography • VLAD and CNN • Multi-string multi-module fault on NPV and payback period
(IRT) • Convolutional feature maps • Mono-string multi-module fault • Influence of energy selling price
• Deep feature pooling on NPV and payback period
[82] NA • Physical analysis [25] • Broken cells • Summarized based on risk priority
• Fast Fourier transform • Encapsulant discolouration number
• Time Domain analysis • Encapsulant delamination • Using 3 factors: Occurrence
• Wavelet analysis • Backsheet warping or detaching rating, Detection rating, Severity
• AI method • Metallization or Busbar rating
[83] NA • Genetic algorithm and its variants discoloration
• Particle Swarm Optimization • Burn through backsheet
• Chaos Optimization algorithm • Solder bond failure
• Artificial Bee Swarm Algorithm • Diode failure
Optimization • Hotspots
• Artificial Bee Colony Optimization [89] • Aging of components • Cash flow analysis of two
• Biogeography Based Optimization locations (Northern and southern
(BBO) regions of Italy)
• Flower Pollination Algorithm (FPA)
• Bacterial Foraging Algorithm (BFA)
• Bird Mating Optimization Algorithm instead of CNN in the third analysis. To conclude, Hybrid models
(BMO)
perform well when compared to single models.
Simulated Annealing Algorithm (SA)
• Harmony Search Algorithm (HS) Statistical assessment and AI detection of defects in PV only allow
• Pattern Search algorithm (PS) some things to be evaluated as their occurrence and appearance. Then,
• Teaching Learning Based evaluating the time frames of repairing or its downtime, it is not enough
Optimization (TLBO)
for the economical assessments because of their impact. Different
methodologies are applied for the economic analysis. Table 7 shows the
various works that conducted the research on the economical influence

17
G.B. Balachandran et al. Measurement 232 (2024) 114683

in solar photovoltaics. cameras, sensors, or specialised imaging equipment. It’s essential to set
up cameras with appropriate parameters like resolution, focus, exposure
7. Stage: 5 evaluation methodology time, and lighting conditions to capture high-quality images. Controlled
and consistent lighting is crucial to ensure uniform image quality. The
The models employed for detection of the faults are employed using different techniques in the detection of faults and their diagnosis are
metrics. They are accuracy, precision, recall, F1 score and the confusion explained in this literature. The inspection techniques use unmanned
matrix. The metrics are helpful in calculating errors between the actual aerial vehicles, Infrared thermography, electroluminescence, photo­
and the predicted values by the machine learning models[41]. luminescence, and Lock-in thermography. Table 6 shows the papers that
Confusion matrices are precious tools for understanding the perfor­ only discussed the inspection techniques for fault diagnosis.
mance of classification models, especially when dealing with imbal­
anced datasets or when different types of errors have different 8.2. Summary of AI techniques
consequences in applications. It helps the readers and machine learning
experts to make informed decisions about the model selection and fine- Artificial techniques include Deep learning models, object detection
tuning. A confusion matrix is a basic tool in field of ML and statistics techniques, Classification models, Transfer learning models, and ML
specifically while evaluating performance of classification models. It models like SVM, DT, RF, etc.… for anomaly detection of different faults.
provides a clear and brief way of understanding the model’s perfor­ CNN is generally employed for visual fault detection tasks. It is partic­
mance by comparing the predictions with the actual results [42]. ularly well-suited for image-based classification and object detection.
A confusion matrix is classically used for binary classification prob­ For identifying and locating defects on objects, models like Faster R-
lems where there are two possible classes or outcomes: positive and CNN, YOLO or SSD are used. If Goal is to classify the entire object as
negative. However, it can also be adapted for multi-class classification either faulty or not, CNN-based image classification models can be used.
problems by extending it to cover various multiple classes. Pre-trained models (e.g., from ImageNet) can be fine-tuned on the spe­
Confusion matrix helps one to make well-versed decisions about the cific defect detection dataset to leverage the knowledge learned from a
improvements of model employed. It provides detailed perceptions broader range of images.
about the strengths and weaknesses of numerous models. This allows Unsupervised techniques like autoencoders or one-class SVMs can be
one to gain better decision on which particular model is the best fit for used for anomaly or defect detection when there are no labelled exam­
the corresponding task. The behaviour of the model can be interpreted ples of defects. Then, Training process is carried out for a dataset of
clearly by understanding the confusion matrix. This provides more in­ labelled images including both defective and non-defective items which
formation about the misclassifications, which require further concen­ is crucial. To enhance the ability of the model to generalise, Augmen­
tration or feature engineering. tation techniques like rotation, flipping, cropping are often implemented
Confusion matrix is a square matrix that consists of 4 main elements: to rise the diversity of the training samples. The general metrics for
evaluating visual fault detection models include accuracy, precision,
1. True Positives (TP): When the model predicts the positive class recall, F1-score, and area under the receiver operating characteristic
correctly, then it becomes true positive. In other words, the model curve (AUC-ROC).
predicts the positive result where actual outcome is positive.
2. True Negatives (TN): When the model predicts the negative class 8.3. Summary of evaluation methodologies
correctly, where the actual class is also negative, it becomes true
negative. To assess model generalisation, techniques like k-fold cross-
3. False Positives (FP): When the model wrongly predicts the positive validation can be used to divide the dataset to train and validation
class as negative, then it becomes false positive. subsets. Depending on the specific application, it may need to set a de­
4. False Negatives (FN): When the model wrongly predicts the negative cision threshold to balance precision and recall based on the cost related
class as positive, then it becomes false negative. with the false positives and false negatives. Analysing the confusion
matrix provides insights into model performance, especially in terms of
Once these components are obtained, Various performance metrics false positives and false negatives.
can be evaluated, including Accuracy, Precision, Recall and F1 score Table 8 presents the summary of literatures conducted based on all
[43]. Accuracy can be calculated by the proportion of right predictions the reviewed aspects including inspection techniques, types of solar
(TP + TN) out of the total predictions. Proportion of the true positives faults taken under research, Image processing techniques, AI models
out of all positive predictions (TP / (TP + FP)) is precision. Precision employed and their evaluation metrics.
measures how good the model performs when it predicts a positive
outcome. Recall is defined as the proportion of true positives out of all 9. Recommendations and insights
actual positives (TP / (TP + FN)). Recall measures how well the model
captures all the positive cases. The F1 Score is harmonic mean to the A comprehensive review on imaging techniques, preprocessing
precision and recall, which strengthens the trade-off between the two methods, and visual fault diagnosis utilizing artificial intelligence (AI)
metrics. The prediction or computational time by the models employed models in solar PV systems is conducted. In this review, various Imaging
can also be calculated, and the computational memory usage is calcu­ technologies includes Infrared thermography, electroluminescence im­
lated for the models and can be compared. aging, and visible light imaging are assessed for the applicability in
detecting diverse types of Solar PV faults. It is also explored that the
8. Discussion advancements in imaging technologies, such as high-resolution cameras,
drones, and satellite imagery, create an impact on fault detection. While
The digest of the topics, which involved the techniques for acquisi­ getting into the preprocessing domain, this review examines different
tion of images, preprocessing and the AI explained in the literature, is techniques such as noise reduction, contrast enhancement, and image
discussed in brief in the below section. normalization for enhancing the quality of PV images.
The fault diagnosis accuracy can be further improved by exploring
8.1. Summary of inspection techniques and preprocessing the integration of external data sources using weather data and histor­
ical performance data. This visual fault diagnosis is conducted with the
Visual fault detection typically starts with acquiring images of ob­ help of AI models like CNNs and RNNs that effectively classify faults
jects or products. These images can come from various sources, such as based on visual information. The review will also investigate

18
G.B. Balachandran et al. Measurement 232 (2024) 114683

Table 8
Summary table of the papers which are reviewed based on the five aspects: (PV inspection, PV Faults, Image processing techniques, AI models and Evaluation metrics).
Reference Panel Inspection Faults analysed Image processing techniques AI model employed Evaluation metrics

[48] UAV-based: Visible and thermal Dust-shading, Encapsulant Kirsch operation-based Image Multiple Three-fold cross validation,
images delamination, Glass Segmentation (Edge detection, classification support Accuracy
breakage, Gridline Background elimination, vector machine (MC-
corrosion, Snail trails and Interference Elimination), Deep SVM) algorithm
Yellowing defect Feature Extraction
[38] UAV equipped with a high- Snail trail, Glass breakage, Discrete wavelet transform (DWT), Ensemble-based 5-fold cross validation,
resolution digital camera Delamination, Discoloration fast Fourier transform (FFT), texture, Deep neural network Accuracy, Receiver
and Burn marks grey level co-occurrence matrix with Random Forest operating characteristic
(GLCM) and grey level difference classifier curve
method (GLDM).
[32] Real industrial dataset: Crack defect, Finger Multi attention network with spatial Multi attention U-net 5-fold cross validation,
Electroluminescence images interruption and defect-free and channel attention architecture mean intersection over-
union (m-IOU), F-measure,
accuracy, recall, precision.
[43] Infrared Imaging camera and Normal and defective panels Data augmentation: cropped, resized, Isolated learning 5-fold cross validation,
images collected from internet rotation and flipping operations from scratch and confusion matrix,
Develop-model Computational cost
transfer learning
[36] Obtained from public domain and Micro-crack, Break, Finger- GAN based augmentation CNN architecture Accuracy, loss function and
private dataset provided by interruption and Defect-free (Generative adversarial network) computational time.
JinkoPower Company (China)
Electroluminescence images
[35] UAV-based and ground-based Normal and Hotspot Homogenization of pixels, CNN architecture Confusion matrix,
operators: Thermographic images Normalization, Thresholding, Accuracy, loss function,
Grayscaling, Discrete wavelet Receiver operating
transform, Box blur filtering and characteristic curve
Sobel Feldman filtering.
[37] Not mentioned Early-stage PV hotspots: 1 Normalization of data Discriminant Accuracy, Confusion
hot-spotted solar cells, 2 classifiers matrix, Receiver operating
hot-spotted solar cells, more characteristics curve
than 3 hot-spotted solar cells
[90] FLIR thermal camera Defective, Non-defective Grayscale conversion, Histogram nBayes-based multi- Accuracy
without hotspots, Non- equalization, Gray Level Co- class density-based
defective with hotspots Occurrence Matrix (GLCM), classifier
Histogram of gradient, Principal
component analysis
[42] FLIR A310 IR camera: eddy current Broken edge, Surface Independent Component Analysis LeNet-5, VGG-16, Accuracy
lock-in thermography (ECLT) and impurity, scratch crack, (ICA), Principal Component Analysis GoogleNet
eddy current pulsed thermography Hotspot and Large area (PCA) and Non-negative Matrix
(ECPT) damage Factorization (NMF)
[11] Obtained from different internet Defect-free, cracks, shadows Semantic segmentation using CNN CNN architecture for Accuracy
search engines and Dusty architecture classification

explainable AI techniques to boost interpretability. Prominence will be with conventional transfer learning models and ML models like SVM and
placed on a thorough literature review, finding research gaps and dis­ DT. The evaluation techniques, like the Classification report metrics and
cussions on real-time applications, challenges, and potential future di­ Confusion matrix for the models, are discussed elaborately. Finally, this
rections, ensuring the exploration of the present state and future review of fault detection in solar and diagnosis methodologies utilising
possibilities in the field of AI-based fault diagnosis for solar PV systems. ML models and imaging techniques could provide essential insights into
optimising the maintenance and reliability of PV systems, eventually
10. Conclusion contributing to the advancement of solar energy technology and its
widespread adoption.
In this paper, several articles have been reviewed, and research has
been made based on 5 aspects: (i) Inspection of the PV panels for faults, CRediT authorship contribution statement
(ii) The possible faults that can occur in PV panels, (iii) Different types of
imaging techniques used for the research, (iv) Role of artificial intelli­ Gurukarthik Babu Balachandran: Supervision, Methodology,
gence in fault detection of the PV panels and (v) Evaluation metrics for Conceptualization. M. Devisridhivyadharshini: Validation, Method­
the AI techniques used. ology, Investigation, Formal analysis, Conceptualization, Visualization,
Various possible visual faults and their causes, such as degradation, Writing – original draft. Muthu Eshwaran Ramachandran: Concep­
corrosion, delamination, partial shading fault, interruptions, cracks and tualization, Methodology. R. Santhiya: Formal analysis, Methodology,
delamination, faults occur in the bypass diodes and blocking diodes are Validation.
elaborated. Images acquired from the solar panels are processed before
being given to the models. The various types of images, which include Declaration of competing interest
electroluminescence, photoluminescence, thermal imaging and ultravi­
olet fluorescence imaging techniques, are discussed in detail. The pre­ The authors declare that they have no known competing financial
processing techniques are discussed and summarised from different interests or personal relationships that could have appeared to influence
literatures. From the fault detection methodologies reviewed from the the work reported in this paper.
literature studies, AI models are evolving in the domain. The neural
networks and machine learning models used in fault detection are
reviewed. The neural networks are performing well when compared

19
G.B. Balachandran et al. Measurement 232 (2024) 114683

Data availability [18] V. Gupta, M. Sharma, R.K. Pachauri, K.N. Dinesh Babu, Comprehensive review on
effect of dust on solar photovoltaic system and mitigation techniques, Sol. Energy
191 (May) (2019) 596–622, https://doi.org/10.1016/j.solener.2019.08.079.
No data was used for the research described in the article. [19] Zorrilla-Casanova, J., Piliougine, M., Carretero, J., Bernaola, P., Carpena, P., Mora-
Lopez, L., & Sidrach-de-Cardona, M. (2011). Analysis of Dust Losses in
Acknowledgement Photovoltaic Modules. Proceedings of the World Renewable Energy Congress – Sweden,
8–13 May, 2011, Linköping, Sweden, 57(July 2014), 2985–2992. doi: 10.3384/
ecp110572985.
Authors would like to thank the Department of Electrical and Elec­ [20] G. Acciani, O. Falcone, S. Vergura, Typical defects of PV-cells, IEEE International
tronics Engineering at Kamaraj College of Engineering and Technology Symposium on Industrial Electronics 2745–2749 (2010), https://doi.org/10.1109/
ISIE.2010.5636901.
for providing continuous support to complete this research study. [21] M.R. Maghami, H. Hizam, C. Gomes, M.A. Radzi, M.I. Rezadad, S. Hajighorbani,
Power loss due to soiling on solar panel: A review, Renew. Sustain. Energy Rev. 59
Funding (2016) 1307–1316, https://doi.org/10.1016/j.rser.2016.01.044.
[22] F. Spertino, J. Ahmad, A. Ciocia, P. Di Leo, Techniques and Experimental Results
for Performance Analysis of Photovoltaic Modules Installed in Buildings, Energy
This research did not receive any specific grant from funding Procedia 111 (September 2016) (2017) 944–953, https://doi.org/10.1016/j.
agencies in the public, commercial, or not-for-profit sectors. egypro.2017.03.257.
[23] M. Islam, G. Hasan, I. Ahmed, M. Amin, S. Dewan, M.M. Rahman, Infrared
Thermography Based Performance Analysis of Photovoltaic Modules, International
Ethics approval statement: Conference on Energy and Power Engineering: Power for Progress, ICEPE 2019 (1)
(2019) 1–5, https://doi.org/10.1109/CEPE.2019.8726565.
[24] M.M. Rahman, I. Khan, K. Alameh, Potential measurement techniques for
Not Applicable. photovoltaic module failure diagnosis: A review, Renew. Sustain. Energy Rev. 151
(August) (2021) 111532, https://doi.org/10.1016/j.rser.2021.111532.
References [25] J.M. Kuitche, R. Pan, G. Tamizhmani, Investigation of dominant failure mode(s) for
field-aged crystalline silicon PV modules under desert climatic conditions, IEEE J.
Photovoltaics 4 (3) (2014) 814–826, https://doi.org/10.1109/
[1] I.F.U. Muzayanah, H.H. Lean, D. Hartono, K.D. Indraswari, R. Partama, Population
JPHOTOV.2014.2308720.
density and energy consumption: A study in Indonesian provinces, Heliyon 8 (9)
[26] V. Sharma, S.S. Chandel, Performance and degradation analysis for long term
(2022) e10634.
reliability of solar photovoltaic systems: A review, Renew. Sustain. Energy Rev. 27
[2] P.A. Owusu, S. Asumadu-Sarkodie, A review of renewable energy sources,
(2013) 753–767, https://doi.org/10.1016/j.rser.2013.07.046.
sustainability issues and climate change mitigation. Cogent, Engineering 3 (1)
[27] Hasan, A. A. Q., Alkahtani, A. A., Shahahmadi, S. A., Alam, M. N. E., Islam, M. A.,
(2016), https://doi.org/10.1080/23311916.2016.1167990.
& Amin, N. (2021). Delamination-and Electromigration-Related Failures in Solar
[3] G.B. Balachandran, P.W. David, R.K. Mariappan, A.E. Kabeel, M.M. Athikesavan,
Panels — A Review.
R. Sathyamurthy, Improvising the efficiency of single-sloped solar still using
[28] M.C.C. de Oliveira, A.S.A. Diniz Cardoso, M.M. Viana, V. de Lins, F. c., The causes
thermally conductive nano-ferric oxide, Environ. Sci. Pollut. Res. 27 (26) (2020)
and effects of degradation of encapsulant ethylene vinyl acetate copolymer (EVA)
32191–32204, https://doi.org/10.1007/s11356-019-06661-2.
in crystalline silicon photovoltaic modules: A review, Renew. Sustain. Energy Rev.
[4] A.O.M. Maka, J.M. Alabid, Solar energy technology and its roles in sustainable
81 (June) (2018) 2299–2317, https://doi.org/10.1016/j.rser.2017.06.039.
development, Clean Energy 6 (3) (2022) 476–483, https://doi.org/10.1093/ce/
[29] M.M. Fouad, L.A. Shihata, E.S.I. Morgan, An integrated review of factors
zkac023.
influencing the performance of photovoltaic panels, Renew. Sustain. Energy Rev.
[5] P. Jenitha, A. Immanuel Selvakumar, Fault detection in PV systems, Applied Solar
80 (June) (2017) 1499–1511, https://doi.org/10.1016/j.rser.2017.05.141.
Energy (english Translation of Geliotekhnika) 53 (3) (2017) 229–237, https://doi.
[30] J. Lindroos, H. Savin, Review of light-induced degradation in crystalline silicon
org/10.3103/S0003701X17030069.
solar cells, Sol. Energy Mater. Sol. Cells 147 (2016) 115–126, https://doi.org/
[6] M., m., v., a., k., v., & d., p. w., Permanent Mismatch Fault Identification of
10.1016/j.solmat.2015.11.047.
Photovoltaic Cells Using Arduino, ICTACT Journal on Microelectronics 01 (02)
[31] A. Mellit, G.M. Tina, S.A. Kalogirou, Fault detection and diagnosis methods for
(2015) 79–82, https://doi.org/10.21917/ijme.2015.0014.
photovoltaic systems: A review, Renew. Sustain. Energy Rev. 91 (March) (2018)
[7] I. Kaaya, M. Koehl, A.P. Mehilli, S. De Cardona Mariano, K.A. Weiss, Modeling
1–17, https://doi.org/10.1016/j.rser.2018.03.062.
Outdoor Service Lifetime Prediction of PV Modules: Effects of Combined Climatic
[32] M.R.U. Rahman, H. Chen, Defects Inspection in Polycrystalline Solar Cells
Stressors on PV Module Power Degradation, IEEE J. Photovoltaics 9 (4) (2019)
Electroluminescence Images Using Deep Learning, IEEE Access 8 (2020)
1105–1112, https://doi.org/10.1109/JPHOTOV.2019.2916197.
40547–40558, https://doi.org/10.1109/ACCESS.2020.2976843.
[8] T. Selvaraj, R. Rengaraj, G. Venkatakrishnan, S. Soundararajan, K. Natarajan,
[33] E. Sovetkin, E.J. Achterberg, T. Weber, B.E. Pieters, Encoder-Decoder Semantic
P. Balachandran, P. David, S. Selvarajan, Environmental Fault Diagnosis of Solar
Segmentation Models for Electroluminescence Images of Thin-Film Photovoltaic
Panels Using Solar Thermal Images in Multiple Convolutional Neural Networks,
Modules, IEEE J. Photovoltaics 11 (2) (2021) 444–452, https://doi.org/10.1109/
International Transactions on Electrical Energy Systems 2022 (2022), https://doi.
JPHOTOV.2020.3041240.
org/10.1155/2022/2872925.
[34] M.U. Ali, H.F. Khan, M. Masud, K.D. Kallu, A. Zafar, A machine learning framework
[9] L. Koester, S. Lindig, A. Louwen, A. Astigarraga, G. Manzolini, D. Moser, Review of
to identify the hotspot in photovoltaic module using infrared thermography, Sol.
photovoltaic module degradation, field inspection techniques and techno-
Energy 208 (May) (2020) 643–651, https://doi.org/10.1016/j.
economic assessment, Renew. Sustain. Energy Rev. 165 (May) (2022) 112616,
solener.2020.08.027.
https://doi.org/10.1016/j.rser.2022.112616.
[35] D. Manno, G. Cipriani, G. Ciulla, V. Di Dio, S. Guarino, V. Lo Brano, Deep learning
[10] T. Berghout, M. Benbouzid, T. Bentrcia, X. Ma, S. Djurović, L.H. Mouss, Machine
strategies for automatic fault diagnosis in photovoltaic systems by thermographic
learning-based condition monitoring for pv systems: State of the art and future
images, Energ. Conver. Manage. 241 (May) (2021) 114315, https://doi.org/
prospects, Energies 14 (19) (2021) 1–24, https://doi.org/10.3390/en14196316.
10.1016/j.enconman.2021.114315.
[11] A. Rico Espinosa, M. Bressan, L.F. Giraldo, Failure signature classification in solar
[36] W. Tang, Q. Yang, K. Xiong, W. Yan, Deep learning based automatic defect
photovoltaic plants using RGB images and convolutional neural networks, Renew.
identification of photovoltaic module using electroluminescence images, Sol.
Energy 162 (2020) 249–256, https://doi.org/10.1016/j.renene.2020.07.154.
Energy 201 (November 2019) (2020) 453–460, https://doi.org/10.1016/j.
[12] O. Kisi, Modeling solar radiation of Mediterranean region in Turkey by using fuzzy
solener.2020.03.049.
genetic approach, Energy 64 (2014) 429–436, https://doi.org/10.1016/j.
[37] M. Dhimish, Defining the best-fit machine learning classifier to early diagnose
energy.2013.10.009.
photovoltaic solar cells hot-spots, Case Studies in Thermal Engineering 25 (April)
[13] A. Bouraiou, M. Hamouda, A. Chaker, A. Neçaibia, M. Mostefaoui, N. Boutasseta,
(2021) 100980, https://doi.org/10.1016/j.csite.2021.100980.
A. Ziane, R. Dabou, N. Sahouane, S. Lachtar, Experimental investigation of
[38] S. Naveen Venkatesh, B. Rebecca Jeyavadhanam, A.M. Moradi Sizkouhi, S.
observed defects in crystalline silicon PV modules under outdoor hot dry climatic
M. Esmailifar, M. Aghaei, V. Sugumaran, Automatic detection of visual faults on
conditions in Algeria, Sol. Energy 159 (November 2017) (2018) 475–487, https://
photovoltaic modules using deep ensemble learning network, Energy Rep. 8 (2022)
doi.org/10.1016/j.solener.2017.11.018.
14382–14395, https://doi.org/10.1016/j.egyr.2022.10.427.
[14] E.L. Meyer, E.E. Van Dyk, Assessing the reliability and degradation of photovoltaic
[39] B. Su, H. Chen, K. Liu, W. Liu, RCAG-Net: Residual Channelwise Attention Gate
module performance parameters, IEEE Trans. Reliab. 53 (1) (2004) 83–92, https://
Network for Hot Spot Defect Detection of Photovoltaic Farms, IEEE Trans. Instrum.
doi.org/10.1109/TR.2004.824831.
Meas. 70 (2021), https://doi.org/10.1109/TIM.2021.3054415.
[15] L. Zhang, P. Liang, H.S. Zhu, P.D. Han, Detection of finger interruptions in silicon
[40] H.P.C. Hwang, C.C.Y. Ku, J.C.C. Chan, Detection of malfunctioning photovoltaic
solar cells using photoluminescence imaging, Chin. Phys. B 27 (6) (2018)
modules based on machine learning algorithms, IEEE Access 9 (2021)
1496–1502, https://doi.org/10.1088/1674-1056/27/6/068801.
37210–37219, https://doi.org/10.1109/ACCESS.2021.3063461.
[16] Köntges, M., Kurtz, S., Packard, C. E., Jahn, U., Berger, K., Kato, K., Friesen, T., Liu,
[41] Ü. Ağbulut, A.E. Gürel, Y. Biçen, Prediction of daily global solar radiation using
H., & Van Iseghem, M. (2014). Performance and reliability of photovoltaic systems
different machine learning algorithms: Evaluation and comparison, Renew.
subtask 3.2: Review of failures of photovoltaic modules: IEA PVPS task 13: external
Sustain. Energy Rev. 135 (August 2020) (2021), https://doi.org/10.1016/j.
final report IEA-PVPS. In IEA-Photovoltaic Power Systems Programme.
rser.2020.110114.
[17] S.R. Madeti, S.N. Singh, A comprehensive study on different types of faults and
[42] B. Du, Y. He, Y. He, J. Duan, Y. Zhang, Intelligent Classification of Silicon
detection techniques for solar photovoltaic system, Sol. Energy 158 (August)
Photovoltaic Cell Defects Based on Eddy Current Thermography and Convolution
(2017) 161–185, https://doi.org/10.1016/j.solener.2017.08.069.

20
G.B. Balachandran et al. Measurement 232 (2024) 114683

Neural Network, IEEE Trans. Ind. Inf. 16 (10) (2020) 6242–6251, https://doi.org/ [66] A. Morlier, F. Haase, M. Kontges, Impact of cracks in multicrystalline silicon solar
10.1109/TII.2019.2952261. cells on PV module power - A simulation study based on field data. 2015 IEEE 42nd
[43] M.W. Akram, G. Li, Y. Jin, X. Chen, C. Zhu, A. Ahmad, Automatic detection of Photovoltaic Specialist Conference, PVSC 2015 (2015) 1–7, https://doi.org/10.1109/
photovoltaic module defects in infrared images with isolated and develop-model PVSC.2015.7355664.
transfer deep learning, Sol. Energy 198 (November 2019) (2020) 175–186, https:// [67] M.K. Alam, F. Khan, J. Johnson, J. Flicker, A Comprehensive Review of
doi.org/10.1016/j.solener.2020.01.055. Catastrophic Faults in Mitigation Techniques, IEEE J. Photovoltaics 5 (3) (2015)
[44] Á.H. Herraiz, A.P. Marugán, F.P.G. Márquez, A review on condition monitoring 982–997.
system for solar plants based on thermography, in: Non-Destructive Testing and [68] A. Phinikarides, N. Kindyni, G. Makrides, G.E. Georghiou, Review of photovoltaic
Condition Monitoring Techniques for Renewable Energy Industrial Assets, Elsevier degradation rate methodologies, Renew. Sustain. Energy Rev. 40 (2014) 143–152,
Ltd., 2019, https://doi.org/10.1016/B978-0-08-101094-5.00007-1. https://doi.org/10.1016/j.rser.2014.07.155.
[45] A.S.M. Shihavuddin, M.R.A. Rashid, M.H. Maruf, M.A. Hasan, M.A. Haq, ul, [69] T. Sarver, A. Al-Qaraghuli, L.L. Kazmerski, A comprehensive review of the impact
Ashique, R. H., & Mansur, A. Al., Image based surface damage detection of of dust on the use of solar energy: History, investigations, results, literature, and
renewable energy installations using a unified deep learning approach, Energy mitigation approaches, Renew. Sustain. Energy Rev. 22 (2013) 698–733, https://
Rep. 7 (2021) 4566–4576, https://doi.org/10.1016/j.egyr.2021.07.045. doi.org/10.1016/j.rser.2012.12.065.
[46] D.P. Winston, M.S. Murugan, R.M. Elavarasan, R. Pugazhendhi, O.J. Singh, [70] A. Ndiaye, A. Charki, A. Kobi, C.M.F. Kébé, P.A. Ndiaye, V. Sambou, Degradations
P. Murugesan, M. Gurudhachanamoorthy, E. Hossain, Solar PV’s Micro Crack and of silicon photovoltaic modules: A literature review, Sol. Energy 96 (2013)
Hotspots Detection Technique Using NN and SVM, IEEE Access 9 (2021) 140–151, https://doi.org/10.1016/j.solener.2013.07.005.
127259–127269, https://doi.org/10.1109/ACCESS.2021.3111904. [71] R.H. Fonseca Alves, G.A. de Deus Júnior, E.G. Marra, R.P. Lemos, Automatic fault
[47] Y. Su, F. Tao, J. Jin, C. Zhang, Automated Overheated Region Object Detection of classification in photovoltaic modules using Convolutional Neural Networks,
Photovoltaic Module with Thermography Image, IEEE J. Photovoltaics 11 (2) Renew. Energy 179 (2021) 502–516, https://doi.org/10.1016/j.
(2021) 535–544, https://doi.org/10.1109/JPHOTOV.2020.3045680. renene.2021.07.070.
[48] X. Li, W. Li, Q. Yang, W. Yan, A.Y. Zomaya, An Unmanned Inspection System for [72] R. Al-Mashhadani, G. Alkawsi, Y. Baashar, A.A. Alkahtani, F.H. Nordin, W. Hashim,
Multiple Defects Detection in Photovoltaic Plants, IEEE J. Photovoltaics 10 (2) T.S. Kiong, Deep learning methods for solar fault detection and classification: A
(2020) 568–576, https://doi.org/10.1109/JPHOTOV.2019.2955183. review, Information Sciences Letters 10 (2) (2021) 323–331, https://doi.org/
[49] H. Wang, J. Xu, C. Sun, R. Yan, X. Chen, Intelligent Fault Diagnosis for Planetary 10.18576/isl/100213.
Gearbox Using Time-Frequency Representation and Deep Reinforcement Learning, [73] N. Bansal, S.P. Jaiswal, G. Singh, Comparative investigation of performance
IEEE/ASME Trans. Mechatron. 27 (2) (2022) 985–998, https://doi.org/10.1109/ evaluation, degradation causes, impact and corrective measures for ground mount
TMECH.2021.3076775. and rooftop solar PV plants – A review, Sustainable Energy Technol. Assess. 47
[50] M.U. Ali, S. Saleem, H. Masood, K.D. Kallu, M. Masud, M.J. Alvi, A. Zafar, Early (July) (2021) 101526, https://doi.org/10.1016/j.seta.2021.101526.
hotspot detection in photovoltaic modules using color image descriptors: An [74] K.S. Garud, S. Jayaraj, M.Y. Lee, A review on modeling of solar photovoltaic
infrared thermography study, Int. J. Energy Res. 46 (2) (2022) 774–785, https:// systems using artificial neural networks, fuzzy logic, genetic algorithm and hybrid
doi.org/10.1002/er.7201. models, Int. J. Energy Res. 45 (1) (2021) 6–35, https://doi.org/10.1002/er.5608.
[51] A. Mellit, S. Kalogirou, Artificial intelligence and internet of things to improve [75] A. Eskandari, J. Milimonfared, M. Aghaei, Line-line fault detection and
efficacy of diagnosis and remote sensing of solar photovoltaic systems: Challenges, classification for photovoltaic systems using ensemble learning model based on I-V
recommendations and future directions, Renew. Sustain. Energy Rev. 143 characteristics, Sol. Energy 211 (September) (2020) 354–365, https://doi.org/
(February) (2021), https://doi.org/10.1016/j.rser.2021.110889. 10.1016/j.solener.2020.09.071.
[52] Z. Quan, H. Lu, W. Zhao, C. Zheng, Z. Zhu, J. Qin, M. Yue, A Review of Dust [76] A. Haque, K.V.S. Bharath, M.A. Khan, I. Khan, Z.A. Jaffery, Fault diagnosis of
Deposition Mechanism and Self-Cleaning Methods for Solar Photovoltaic Modules, Photovoltaic Modules, Energy Sci. Eng. 7 (3) (2019) 622–644, https://doi.org/
Coatings 13 (1) (2023) 1–20, https://doi.org/10.3390/coatings13010049. 10.1002/ese3.255.
[53] G. Goudelis, P.I. Lazaridis, M. Dhimish, A Review of Models for Photovoltaic Crack [77] C. Feng, Y. Liu, J. Zhang, A taxonomical review on recent artificial intelligence
and Hotspot Prediction, Energies 15 (12) (2022), https://doi.org/10.3390/ applications to PV integration into power grids, Int. J. Electr. Power Energy Syst.
en15124303. 132 (September 2020) (2021) 107176, https://doi.org/10.1016/j.
[54] C. Buerhop, L. Bommes, J. Schlipf, T. Pickel, A. Fladung, I.M. Peters, Infrared ijepes.2021.107176.
imaging of photovoltaic modules: a review of the state of the art and future [78] B. Li, C. Delpha, D. Diallo, A. Migan-Dubois, Application of Artificial Neural
challenges facing gigawatt photovoltaic power stations. Progress, Energy 4 (4) Networks to photovoltaic fault detection and diagnosis: A review, Renew. Sustain.
(2022), https://doi.org/10.1088/2516-1083/ac890b. Energy Rev. 138 (October) (2021), https://doi.org/10.1016/j.rser.2020.110512.
[55] O.A. Alimi, E.L. Meyer, O.I. Olayiwola, Solar Photovoltaic Modules’, Performance [79] A. Abubakar, C.F.M. Almeida, M. Gemignani, Review of artificial intelligence-
Reliability and Degradation Analysis—A Review. Energies 15 (16) (2022) https:// based failure detection and diagnosis methods for solar photovoltaic systems,
doi.org/10.3390/en15165964. Machines 9 (12) (2021), https://doi.org/10.3390/machines9120328.
[56] J.A. Dhanraj, A. Mostafaeipour, K. Velmurugan, K. Techato, P.K. Chaurasiya, J. [80] M. Mansouri, M. Trabelsi, H. Nounou, M. Nounou, Deep Learning-Based Fault
M. Solomon, A. Gopalan, K. Phoungthong, An effective evaluation on fault Diagnosis of Photovoltaic Systems: A Comprehensive Review and Enhancement
detection in solar panels, Energies 14 (22) (2021) 1–14, https://doi.org/10.3390/ Prospects, IEEE Access 9 (2021) 126286–126306, https://doi.org/10.1109/
en14227770. ACCESS.2021.3110947.
[57] S. Kumar, R. Meena, R. Gupta, Finger and interconnect degradations in crystalline [81] B. Du, Y. He, Y. He, C. Zhang, Progress and trends in fault diagnosis for renewable
silicon photovoltaic modules: A review, Sol. Energy Mater. Sol. Cells 230 (January) and sustainable energy system based on infrared thermography: A review, Infrared
(2021) 111296, https://doi.org/10.1016/j.solmat.2021.111296. Phys. Technol. 109 (May) (2020) 103383, https://doi.org/10.1016/j.
[58] L. Lillo-Sánchez, G. López-Lara, J. Vera-Medina, E. Pérez-Aparicio, I. Lillo-Bravo, infrared.2020.103383.
Degradation analysis of photovoltaic modules after operating for 22 years. A case [82] Z. Wu, Y. Hu, J.X. Wen, F. Zhou, X. Ye, A Review for Solar Panel Fire Accident
study with comparisons, Sol. Energy 222 (May) (2021) 84–94, https://doi.org/ Prevention in Large-Scale PV Applications, IEEE Access 8 (2020) 132466–132480,
10.1016/j.solener.2021.04.026. https://doi.org/10.1109/ACCESS.2020.3010212.
[59] Livera, A., Theristis, M., Makrides, G., & Georghiou, G. E. (2019). Recent advances [83] D.S. Pillai, N. Rajasekar, Metaheuristic algorithms for PV parameter identification:
in failure diagnosis techniques based on performance data analysis for grid- A comprehensive review with an application to threshold setting for fault detection
connected photovoltaic systems. In Renewable Energy (Vol. 133, pp. 126–143). in PV systems, Renew. Sustain. Energy Rev. 82 (January) (2018) 3503–3525,
Elsevier B.V. doi: 10.1016/j.renene.2018.09.101. https://doi.org/10.1016/j.rser.2017.10.107.
[60] A. Omazic, G. Oreski, M. Halwachs, G.C. Eder, C. Hirschl, L. Neumaier, G. Pinter, [84] A.K.V. de Oliveira, M. Aghaei, R. Rüther, Automatic Inspection of Photovoltaic
M. Erceg, Relation between degradation of polymeric components in crystalline Power Plants Using Aerial Infrared Thermography, A Review. Energies 15 (6)
silicon PV module and climatic conditions: A literature review, Sol. Energy Mater. (2022), https://doi.org/10.3390/en15062055.
Sol. Cells 192 (September 2018) (2019) 123–133, https://doi.org/10.1016/j. [85] S. Gallardo-Saavedra, L. Hernández-Callejo, O. Duque-Perez, Technological review
solmat.2018.12.027. of the instrumentation used in aerial thermographic inspection of photovoltaic
[61] D.S. Pillai, N. Rajasekar, A comprehensive review on protection challenges and plants, Renew. Sustain. Energy Rev. 93 (April) (2018) 566–579, https://doi.org/
fault diagnosis in PV systems, Renew. Sustain. Energy Rev. 91 (July 2017) (2018) 10.1016/j.rser.2018.05.027.
18–40, https://doi.org/10.1016/j.rser.2018.03.082. [86] B. Du, R. Yang, Y. He, F. Wang, S. Huang, Nondestructive inspection, testing and
[62] A. Triki-Lahiani, A. Bennani-Ben Abdelghani, I. Slama-Belkhodja, Fault detection evaluation for Si-based, thin film and multi-junction solar cells: an overview,
and monitoring systems for photovoltaic installations: A review, Renew. Sustain. Renew. Sustain. Energy Rev. 78 (January) (2017) 1117–1151, https://doi.org/
Energy Rev. 82 (September) (2018) 2680–2692, https://doi.org/10.1016/j. 10.1016/j.rser.2017.05.017.
rser.2017.09.101. [87] A. Colli, Failure mode and effect analysis for photovoltaic systems, Renew. Sustain.
[63] W. Luo, Y.S. Khoo, P. Hacke, V. Naumann, D. Lausch, S.P. Harvey, J.P. Singh, Energy Rev. 50 (2015) 804–809, https://doi.org/10.1016/j.rser.2015.05.056.
J. Chai, Y. Wang, A.G. Aberle, S. Ramakrishna, Potential-induced degradation in [88] S.R. Madeti, S.N. Singh, Online fault detection and the economic analysis of grid-
photovoltaic modules: A critical review, Energ. Environ. Sci. 10 (1) (2017) 43–68, connected photovoltaic systems, Energy 134 (2017) 121–135, https://doi.org/
https://doi.org/10.1039/c6ee02271e. 10.1016/j.energy.2017.06.005.
[64] R.L. Achey, Life Expectancy, J. Palliat. Med. 19 (4) (2016) 468, https://doi.org/ [89] F. Chiacchio, F. Famoso, L. Cedola, Grid-Connected Photovoltaic Power Plant with
10.1089/jpm.2015.0452. a and the South of Italy, Energies, Lv, 2019.
[65] J.A. Tsanakas, L. Ha, C. Buerhop, Faults and infrared thermographic diagnosis in [90] K.A.K. Niazi, W. Akhtar, H.A. Khan, Y. Yang, S. Athar, Hotspot diagnosis for solar
operating c-Si photovoltaic modules: A review of research and future challenges, photovoltaic modules using a Naive Bayes classifier, Sol. Energy 190 (February)
Renew. Sustain. Energy Rev. 62 (2016) 695–709, https://doi.org/10.1016/j. (2019) 34–43, https://doi.org/10.1016/j.solener.2019.07.063.
rser.2016.04.079.

21