InGenio Journal
Revista de Ciencias de la Ingeniería de la Universidad Técnica Estatal de Quevedo
https://revistas.uteq.edu.ec/index.php/ingenio
e-ISSN: 2697-3642 - CC BY-NC-SA 4.0
InGenio Journal
Revista de Ciencias de la Ingeniería de la Universidad Técnica Estatal de Quevedo
https://revistas.uteq.edu.ec/index.php/ingenio
e-ISSN: 2697-3642 CC BY-NC-SA 4.0
Volumen 6 | Número 1 | Pp. 1–4 | Enero 2023 Recibido (Received): 2022/mm/dd
DOI: https://doi.org/10.18779/ingenio.v6i1.nm Aceptado (Accepted): 2022/mm/dd
Experimental Study of Defect Detection in Photovoltaic
Panels Using Electroluminescence
(Estudio experimental de la detección de defectos en paneles fotovoltaicos
utilizando electroluminiscencia)
Franklin Mauricio Gómez López
, Danny Vinicio Ochoa Correa
, Isabel Dolores Cabrera
Carrera
Universidad de Cuenca, Ecuador.
franklin.gomez@ucuenca.edu.ec, danny.ochoac@ucuenca.edu.ec, isabel.cabrera@ucuenca.edu.ec
Abstract: This study presents an experimental investigation of defect detection in
monocrystalline and polycrystalline photovoltaic panels using electroluminescence imaging.
Twelve modules with varying levels of degradation, deployed in rural Andean installations,
were evaluated through visual inspection and automatic analysis based on convolutional
neural networks. The methodology included manual defect segmentation and the application
of a U-Net model for automatic classification of structural and electrical anomalies. Results
showed that fractures and inactive areas were more frequent in monocrystalline panels,
whereas potential-induced degradation (PID) and discoloration were mainly observed in
polycrystalline ones. The comparison between manual inspection and the automated model
demonstrated high performance in detecting discrete defects such as cracks and busbars,
although lower accuracy was obtained for diffuse anomalies like PID. Spatial analysis of
defect maps revealed the clustering of certain fault types within cell regions, highlighting the
importance of considering spatial correlations in future detection algorithms. These findings
support the development of targeted preventive maintenance strategies and reinforce the
potential of semantic segmentation as a diagnostic tool for photovoltaic modules.
Keywords: Electroluminescence, photovoltaic panels, preventive maintenance, semantic
segmentation, U-Net.
Resumen: Este estudio presenta una investigación experimental sobre la detección de
defectos en paneles fotovoltaicos monocristalinos y policristalinos mediante imágenes de
electroluminiscencia. Se evaluaron doce módulos con distintos niveles de degradación,
instalados en entornos rurales andinos, mediante inspección visual y análisis automático
basado en redes neuronales convolucionales. La metodología incluyó la segmentación
manual de defectos y la aplicación de un modelo U-Net para la clasificación automática de
anomalías estructurales y eléctricas. Los resultados mostraron que las fracturas y áreas
inactivas fueron más frecuentes en paneles monocristalinos, mientras que la degradación
inducida por potencial (PID) y la decoloración se observaron principalmente en paneles
policristalinos. La comparación entre la inspección manual y el modelo automático evidenció
un alto desempeño en la detección de defectos discretos como grietas y barras conductoras,
aunque con menor precisión en defectos difusos como la PID. El análisis espacial de los
mapas de defectos reveló la agrupación de ciertos tipos de fallas dentro de las celdas,
destacando la importancia de considerar correlaciones espaciales en futuros algoritmos de
detección. Estos hallazgos respaldan el desarrollo de estrategias de mantenimiento preventivo
más específicas y refuerzan el potencial de la segmentación semántica como herramienta de
diagnóstico para módulos fotovoltaicos.
Palabras clave: Electroluminiscencia, paneles fotovoltaicos, mantenimiento preventivo,
segmentación semántica, U-Net.
Experimental Study of Defect Detection in Photovoltaic
Panels Using Electroluminescence
(Estudio experimental de la detección de defectos en paneles fotovoltaicos
utilizando electroluminiscencia)
Franklin Mauricio Gómez López
, Danny Vinicio Ochoa-Correa
, Isabel Dolores Cabrera
Carrera
Volumen 9 | Número 2 | Pp. 36–45 | Julio 2026
DOI: https://doi.org/10.18779/ingenio.v9i2.1232
Recibido (Received): 2025/11/06
Aceptado (Accepted): 2026/04/10
InGenio Journal, 9(2), 36–45 37
InGenio Journal, 9(2), x–x
| 2
1. INTRODUCTION
The reliability of photovoltaic (PV) installations depends on the integrity of their components,
particularly the condition of individual solar cells within the modules. Over time, degradation
mechanisms such as micro-cracks, potential-induced degradation (PID), delamination,
discoloration, and metallization layer disruptions compromise the energy yield and operational
lifespan of PV systems [1], [2], [3]. Beyond their direct impact on energy yield, these defects also
reduce the diagnostic signal-to-noise ratio in imaging workflows and complicate large-scale
quality assurance as PV deployment accelerates worldwide [4].
Monocrystalline and polycrystalline silicon modules remain the most widely deployed PV
technologies, but they differ in structural characteristics and vulnerability to specific defect types
[5]. Several comparative studies have reported variations in performance degradation, thermal
behavior, and defect propagation across these two material classes [6]. Understanding how defect
types manifest and distribute in each case is essential for improving diagnostic accuracy and
guiding maintenance strategies. From an operations standpoint, mixed portfolios that combine
mono- and multi-Si modules introduce additional variability for condition monitoring systems;
models tuned to one substrate may generalize sub-optimally to the other, which motivates explicit,
side-by-side analyses of defect prevalence and appearance across materials.
Electroluminescence (EL) imaging has become a common diagnostic technique to inspect
defects that are not visible under standard illumination or thermal imaging. When a forward bias
is applied to a PV module in darkness, radiative recombination generates infrared emission that
can reveal hidden structural and electrical anomalies [7], [8]. The contrast and resolution of EL
images, however, depend on capture conditions, the sensitivity of the camera system, and the
severity of the defect. Manual interpretation of EL images also introduces subjectivity and may
overlook subtle features. Complementary modalities such as infrared (IR) and RGB imaging are
increasingly integrated into inspection pipelines at module and plant scale, where IR assists in
locating thermally driven anomalies and RGB highlights surface-level issues like broken glass,
discoloration, or delamination, enabling faster triage in field campaigns [4].
In recent years, machine learning approaches, particularly convolutional neural networks
(CNNs), have been employed to support automatic defect segmentation and classification.
Among these, the U-Net architecture has been applied to PV inspection tasks using
electroluminescence imaging, showing encouraging performance in pixel-level detection [9],
[10], [11]. Despite these advancements, few studies have examined whether model performance
or defect detection outcomes vary when applied to different module types. A systematic
comparison between monocrystalline and polycrystalline panels-using both visual and automated
inspection methods-remains limited in current literature.
Recent surveys emphasize a shift from handcrafted feature pipelines toward end-to-end CNNs
for EL image analysis, while underscoring persistent challenges: defect diversity, scale variation,
class imbalance, and domain shifts between laboratory and field data [12].
To mitigate data scarcity, synthetic EL datasets produced with deep convolutional generative
adversarial networks (DCGANs) have been introduced, reaching quality indicators (e.g.,
Inception Score and FID) that support their use to augment training sets and to probe the
relationship between image patterns and electrical performance via learned regressors [13].
At the detector level, one-stage architectures (e.g., YOLOv5) combined with tailored loss
functions—such as focal and efficient IoU (EIoU) variants—have shown gains in mean average
precision for multi-class EL defect detection, particularly under class imbalance and scale
heterogeneity [14].
This study presents an experimental evaluation of defect detection in monocrystalline and
polycrystalline PV panels using EL imaging. Manual and U-Net-based inspections are performed
InGenio Journal, 9(2), 36–45 38
InGenio Journal, 9(2), x–x
| 3
on a shared dataset, enabling a comparative assessment of defect type prevalence, spatial
distribution, and detection accuracy across material types. The work also analyzes how common
degradation modes manifest differently in each case, contributing to the development of more
targeted inspection and maintenance protocols for silicon-based PV modules.
2. MATERIALS AND METHODS
2.1. Panel samples and degradation overview
The experimental study was conducted on a set of 12 PV panels, including six monocrystalline
and six polycrystalline modules, representative of systems deployed in rural Andean installations.
Technical specifications of the modules are presented in Table 1. All panels exhibited varying
degrees of physical and electrical degradation, with service durations ranging from 5 to 10 years
under typical field conditions. Experimental evaluations were conducted at the Microgrid
Laboratory of the University of Cuenca, Ecuador [15].
Table 1. Specifications of tested PV modules.
Type
Model
(W)
Service(yrs)
Monocrystalline
Heckert NeMo 60 P260
260
7
Polycrystalline
Canadian Solar CS6P-250P
250
8
Visual inspections and EL imaging revealed a wide range of degradation patterns, classified
according to IEC 61215 [16]. Observed modes included interconnect ribbon corrosion, solder
bond failure, cell cracks, PID, busbar interruptions, and discoloration due to encapsulant
browning.
Figure 1 illustrates the representative degradation patterns found in the samples. Figure 1(a)
shows severe encapsulant discoloration (browning) concentrated in the active cell areas, a typical
sign of chemical degradation due to UV exposure. Figure 1(b) highlights external factors such as
accumulated dust along the frames and localized bird droppings, which induce partial shading.
Additionally, Figure 1(c) reveals oxidation. It is important to note that while these defects were
identifiable through visual inspection, structural fractures and microcracks remained invisible to
the naked eye, necessitating the subsequent use of electroluminescence imaging for their
detection.
Figure 1. Visible degradation observed during physical inspection: (a) Encapsulant
discoloration (browning) in the center of the cells; (b) Soiling accumulation and bird droppings
on the glass surface; (c) Signs of corrosion.
InGenio Journal, 9(2), 36–45 39
InGenio Journal, 9(2), x–x
| 4
2.2. Image acquisition and experimental setup
The EL image acquisition process was conducted in a darkroom to minimize ambient
interference. Each module was forward-biased using a programmable Chroma DC power supply
to induce electroluminescence, with polarization currents set to approximately 6 A (two-thirds of
). The camera used was an OWL 640 M model equipped with an InGaAs sensor, sensitive to
the 900–1700 nm wavelength range.
Images were captured using the XCAP-Std software with the following configuration:
exposure time between 30–50 ms, digital gain factor of 2.5, and three-point non-uniformity
correction (offset, gain, dark). Figure 2 shows the laboratory setup used during the inspection
process.
Figure 2. Experimental setup showing the EL imaging system in the darkroom environment.
2.3. Ground Truth generation via visual inspection
To establish a reliable reference for supervised learning, a rigorous visual inspection was
conducted solely to generate the Ground Truth dataset. Each EL image was segmented into
individual cells and independently evaluated by three expert reviewers. The classification process
labeled four main defect categories: cell breaks (visible micro-cracks disrupting current flow),
PID (darker areas with reduced luminescence), discoloration (localized dimming due to aging),
and finger interruptions (breaks in metallization lines). The manual annotations were recorded
and cross-validated among reviewers to create the final target masks used for training and
evaluating the automated system.
2.4. U-Net based semantic segmentation
Automatic defect detection was implemented using a U-Net architecture, trained on the
annotated dataset derived from the visual inspection phase. The input consisted of preprocessed
EL images, standardized in resolution and contrast.
The network consisted of four encoder-decoder levels, using ReLU activation and dropout
layers for regularization. The dataset was split into 70% training, 15% validation, and 15% testing.
The training outcomes, including F1 score, recall, and precision, are summarized in Table 2,
which presents the performance metrics obtained for the optimized U-Net_V32 model.
InGenio Journal, 9(2), 36–45 40
InGenio Journal, 9(2), x–x
| 5
Table 2. Training performance of U-Net_V32.
Model Version
F1 Score
Recall
Precision
U-Net_V32
0.9064
0.9066
0.8714
2.5. Implementation details and reproducibility
The U-Net models were implemented using the PyTorch framework (v2.3.1). Training and
inference procedures were executed on the Google Colab Pro platform, utilizing an NVIDIA T4
GPU with 16 GB of VRAM. The input EL images were resized to 256×256 pixels using bilinear
interpolation to preserve spatial details, while nearest-neighbor interpolation was applied to the
ground truth masks to maintain class integrity.
The training process employed the Adam optimizer with an initial learning rate of 0.0018 and
a batch size of 16 samples. A cross-entropy loss function was minimized over 48 epochs, with a
step-based learning rate decay (factor of 0.05 every 8 epochs) to facilitate convergence. Under
this configuration, the average inference time was approximately 32 ms per image, demonstrating
the model's feasibility for high-throughput processing in industrial scenarios.
2.6. Evaluation metrics
Model performance was evaluated using pixel-based accuracy, Dice score, and class-level
recall. These metrics were computed for each defect category on mono and polycrystalline
modules independently. Statistical comparison between module types was conducted using paired
t-tests to assess differences in segmentation accuracy.
Defect frequency maps and classification heatmaps were also generated to visualize spatial
distribution, as discussed further in the next section.
3. RESULTS
3.1. Defect frequency and type distribution
A total of 142 cells were evaluated through visual inspection using EL imagery. The
breakdown of defect types across three tested PV modules is presented in Table 3. The most
frequently detected issues were material anomalies and conductor finger interruptions. The
presence of PID was exclusive to polycrystalline modules, while contact interruptions were absent
in monocrystalline samples.
Table 3. Visually identified defects per module.
Module Type
Material
Anomalies
Contact
Interruptions
Finger
Interruptions
PID
Monocrystalline
11 cells
0
24 cells
18
Polycrystalline 1
25 cells
11 cells
0
14
Polycrystalline 2
25 cells
11 cells
0
11
3.2. Automatic detection results
The segmentation model was trained to identify three categories—conductor bars, cracks, and
inactive dark zones—using expert-generated Ground Truth masks as the reference for supervised
learning and evaluation. Table 4 presents a clear, class-by-class quantitative comparison of the
U-Net_V32 performance using precision, recall, and F1-score. In particular, the final model (U-
Net_V32) achieved an F1-score of 0.9083 for conductor bar detection, 0.6557 for crack
identification, and 0.6466 for inactive-zone segmentation. In addition to these aggregated metrics,
Figure 4 provides a qualitative, pixel-level comparison between the Ground Truth masks and the
U-Net_V32 predictions, including error maps that distinguish false positives (FP) and false
InGenio Journal, 9(2), 36–45 41
InGenio Journal, 9(2), x–x
| 6
negatives (FN), thereby highlighting where the model agrees with the manual annotations and
where discrepancies occur.
Table 4. Performance metrics for U-Net_V32.
Defect Category
Precision
Recall
F1 Score
Conductor Bars
0.9151
0.9031
0.9083
Cracks
0.6794
0.6465
0.6557
Inactive Zones
0.6643
0.6443
0.6466
3.3. Spatial patterns of defects
Figure 3 illustrates three distinct heatmaps showing the spatial distribution of predictions for
busbars, cracks, and dark zones on monocrystalline modules. The heatmap for busbars
demonstrates their consistent and localized presence within the module. In contrast, the heatmaps
for cracks and dark zones reveal that these defects had a higher occurrence in the central regions
of the cells. The co-occurrence matrix, shown in Figure 4, highlights the conditional presence of
multiple defect types in the same cell regions.
Figure 3. Heatmaps showing defect distribution across cells in monocrystalline module.
Figure 4. Co-occurrence matrix of defects in monocrystalline cells.
InGenio Journal, 9(2), 36–45 42
InGenio Journal, 9(2), x–x
| 7
3.4. Case studies
To assess model accuracy at the image level, a representative cell was manually annotated
and its predicted segmentations were compared. Figure 5 shows a visual comparison of U-
Net_V32 predictions against manual annotations for the cracks, dark zones, and busbar defects
within this single cell. The figure also includes error maps, highlighting false positives and false
negatives for each defect type. The U-Net_V32 model demonstrated proficiency in identifying
microcracks not marked by the manual classifier, suggesting that EL signal attenuation can be
captured algorithmically even when visually imperceptible. However, it occasionally
misclassified encapsulant delamination artifacts as inactive zones.
Figure 5. Qualitative comparison of segmentation performance. Each row corresponds to a
specific defect class: (a) Cracks, (b) Dark Zones, and (c) Busbars. The columns display the U-
Net prediction, the Ground Truth (manual annotation), and the spatial Error Map. In the error
maps, Cyan pixels represent False Positives (over-segmentation), while Orange pixels indicate
False Negatives (missed defects).
4. DISCUSSION
The results provide multiple layers of interpretation regarding the classification of defects in
PV modules through EL imaging. One of the central observations is the disparity in defect
prevalence between mono- and polycrystalline modules. Although both types exhibited broken
cells, inactive regions, and discoloration, fractures were especially concentrated in
monocrystalline samples. This trend is consistent with the higher structural uniformity of mono-
Si wafers, which tend to propagate mechanical stress more efficiently. Conversely, PID and snail
InGenio Journal, 9(2), 36–45 43
InGenio Journal, 9(2), x–x
| 8
trails were more commonly detected in polycrystalline modules, indicating distinct degradation
pathways influenced by material properties and manufacturing conditions.
The defect maps and co-occurrence matrices reinforce these findings. As shown in Figure 3,
defects such as cracks and dark zones appeared concurrently in multiple cells, suggesting spatial
clustering. In mono modules, 48 cells were simultaneously marked with cracks and inactive areas.
These results emphasize the importance of including spatial correlation when developing defect
detection algorithms, particularly for defect classes that emerge in proximity.
The automatic segmentation achieved through U-Net produced consistent predictions across
several categories, with high F1 scores for crack detection and busbar alignment. However,
performance varied by defect type. For example, the identification of PID was less accurate,
possibly due to its diffuse appearance and lower contrast under EL. These outcomes align with
findings from previous studies that highlight the difficulties of detecting uniform degradation
patterns compared to discrete faults such as fractures or inactive zones.
Compared with prior EL-based inspection studies that focus primarily on reporting
segmentation accuracy, the present work adds three practical elements that become relevant under
real field conditions. First, the dataset is built from mono- and polycrystalline modules with
degradation accumulated in service (rather than controlled laboratory aging), which exposes the
model to realistic variability in soiling, corrosion, and encapsulant discoloration that can interfere
with EL interpretation. Second, the Ground Truth was generated through cross-validated expert
annotations, and the evaluation is explicitly anchored to a pixel-level comparison (including
FP/FN error maps), enabling a more transparent identification of failure modes rather than relying
only on aggregate scores. Third, beyond per-class metrics, the study analyzes the spatial
organization of defects (co-occurrence and clustering), showing that cracks and inactive regions
often appear jointly within neighboring cells; information that can be integrated into inspection
workflows to prioritize modules or strings for targeted maintenance and further testing.
When comparing manual and automatic assessments, qualitative agreement was observed in
most annotated panels. Nonetheless, certain cases showed undersegmentation by the model,
especially for defects located at the edges or overlapping with frame artifacts. These discrepancies
underline the importance of training with representative samples that include peripheral and
partially obscured defects.
Limitations in the current implementation stem from the relatively small annotated dataset
and the fixed architecture of the segmentation model. Although the version trained (U-Net_V32)
was optimized with adjusted filters and skip connections, future work should examine ensemble
strategies or hybrid approaches that merge multiple model outputs. Additionally, data
augmentation strategies could be enhanced to simulate edge-based distortions and noise patterns
encountered in real field conditions.
5. CONCLUSIONS
This study evaluated the detection of structural and electrical defects in mono- and
polycrystalline PV panels through EL imaging combined with U-Net-based semantic
segmentation. A comparative analysis revealed that fractures and inactive areas were more
frequent in monocrystalline modules, whereas PID and surface discoloration appeared
predominantly in polycrystalline ones.
Manual inspection protocols were cross-validated with automated segmentation outputs.
Using expert-generated Ground Truth masks as the reference, the final U-Net version (U-
Net_V32) provided a clear class-by-class performance, achieving an F1-score of 0.9083 for
conductor bar (busbar) detection, 0.6557 for crack identification, and 0.6466 for inactive-zone
segmentation. The model achieved strong F1 performance on discrete defects such as cracks and
InGenio Journal, 9(2), 36–45 44
InGenio Journal, 9(2), x–x
| 9
disconnected busbars, but underperformed in detecting diffuse anomalies like PID. This indicates
that CNN-based segmentation can support diagnostic assessments, provided that training datasets
cover sufficient variation in defect morphology and imaging artifacts.
The spatial distribution analysis confirmed that certain defects appear in clusters, especially
in mono panels, highlighting the relevance of incorporating spatial information into diagnostic
workflows. Although the model correctly classified most defect types, limitations remain in
handling image boundaries and overlapping noise patterns. In practical terms, these results
support the use of EL-driven segmentation as a screening tool to prioritize modules for further
inspection, while emphasizing that improved labeling diversity and artifact-aware training are
needed to increase reliability for low-contrast and diffuse degradation signatures (e.g., PID).
The findings of this study also emphasize the industrial applicability of EL-based
segmentation for enhancing preventive maintenance strategies in photovoltaic systems. The
model's ability to prioritize modules for inspection based on defect severity offers significant
potential for optimizing maintenance schedules and reducing operational downtime in large-scale
PV installations.
In terms of scalability, the methodology can be expanded to handle larger datasets by
leveraging batch processing and incremental model training, ensuring that it remains feasible for
wide-scale deployment in the industry. Additionally, this framework can be adapted to different
PV technologies and expanded to include additional defect categories.
Future work should focus on extending the training dataset, improving edge-specific
augmentation strategies, and integrating temporal EL imaging for defect tracking. Further
improvements in the training process could include hybrid and ensemble approaches, as well as
better artifact handling through more diverse and realistic training data. Overall, the methodology
supports a structured framework for comparing degradation patterns across different PV
technologies and contributes to enhancing inspection strategies for preventive maintenance.
ACKNOWLEDGMENTS: The authors thank Universidad de Cuenca for easing access to the
facilities of the Microgrid Laboratory of the Faculty of Engineering, for allowing the use of its
equipment, and for authorizing its staff the provision of technical support necessary to carry out
the experiments described in this article.
REFERENCES
[1] R. Zhang et al., “Optimizing PV Panel Segmentation in Complex Environments Using Pre-
Training and Simulated Annealing Algorithm: The JSWPVI”, Land, vol. 14, no. 6, jun.
2025. [Online]. Available: https://doi.org/10.3390/land14061245
[2] H. Munawer Al-Otum, “Deep learning-based automated defect classification in
Electroluminescence images of solar panels”, Adv. Eng. Inform, vol. 58, oct. 2023. [Online].
Available: https://doi.org/10.1016/j.aei.2023.102147
[3] A. El-Tayeb, F. Li, A. Kumar, y G. Tamizhmani, “Electrical Impedance Spectroscopy: A
Complementary Approach Differentiating PID Mechanisms in Photovoltaics”, Electronics,
vol. 14, no. 5, pp. 1021, mar. 2025. [Online]. Available:
https://doi.org/10.3390/electronics14051021
[4] D. Matusz-Kalász, I. Bodnár, y M. Jobbágy, “An Overview of CNN-Based Image Analysis
in Solar Cells, Photovoltaic Modules, and Power Plants”, Appl. Sci., vol. 15, no. 10, pp.
5511, jan. 2025. [Online]. Available: https://doi.org/10.3390/app15105511
[5] A. Tihane, M. Boulaid, A. Elfanaoui, M. Nya, y A. Ihlal, “Performance analysis of mono
and poly-crystalline silicon photovoltaic modules under Agadir climatic conditions in
InGenio Journal, 9(2), 36–45 45
InGenio Journal, 9(2), x–x
| 10
Morocco”, Mater. Today Proc., vol. 24, pp. 85-90, jan. 2020. [Online]. Available:
https://doi.org/10.1016/j.matpr.2019.07.620
[6] M. Köntges, S. Kurtz, C. Packard, U. Jahn, K. A. Berger, y K. Kato, Performance and
reliability of photovoltaic systems: subtask 3.2: Review of failures of photovoltaic modules:
IEA PVPS task 13: external final report IEA-PVPS. Sankt Ursen: International Energy
Agency, Photovoltaic Power Systems Programme, 2014. [Online]. Available:
https://doi.org/10.2314/GBV:856979287
[7] T. Fuyuki, H. Kondo, T. Yamazaki, Y. Takahashi, y Y. Uraoka, “Photographic surveying
of minority carrier diffusion length in polycrystalline silicon solar cells by
electroluminescence”, Appl. Phys. Lett., vol. 86, no. 26, pp. 262108, jun. 2005. [Online].
Available: https://doi.org/10.1063/1.1978979
[8] Y. Zhang, R. Wang, F. Wang, D. Zhu, X. Gong, y X. Cheng, “Electroluminescence as a
Tool to Study the Polarization Characteristics and Generation Mechanism in Silicon PV
Panels”, Appl. Sci., vol. 13, no. 3, pp. 1591, 2023. [Online]. Available:
https://doi.org/10.3390/app13031591
[9] R. A. M. Rudro et al., “SPF-Net: Solar panel fault detection using U-Net based deep
learning image classification”, Energy Rep., vol. 12, pp. 1580-1594, dic. 2024. [Online].
Available: https://doi.org/10.1016/j.egyr.2024.07.044
[10] F. Gómez-López, D. Ochoa-Correa, y I. Cabrera-Carrera, “U-Net–based semantic
segmentation of defects in photovoltaic panels”, Ingenius, no. 35, pp. 110-121, jan. 2026.
[Online]. Available: https://doi.org/10.17163/ings.n35.2026.08
[11] A. B. Alao, O. M. Adeyanju, M. Chamana, S. Bayne, y A. Bilbao, “Photovoltaic Farm
Power Generation Forecast Using Photovoltaic Battery Model with Machine Learning
Capabilities”, Solar, vol. 5, no. 2, p. 26, 2025. [Online]. Available:
https://doi.org/10.3390/solar5020026
[12] T. Hussain, M. Hussain, H. Al-Aqrabi, T. Alsboui, y R. Hill, “A Review on Defect Detection
of Electroluminescence-Based Photovoltaic Cell Surface Images Using Computer Vision”,
Energies, vol. 16, no. 10, pp. 4012, ene. 2023. [Online]. Available:
https://doi.org/10.3390/en16104012
[13] H. F. M. Romero et al., “Synthetic Dataset of Electroluminescence Images of Photovoltaic
Cells by Deep Convolutional Generative Adversarial Networks”, Sustainability, vol. 15, no.
9, jan. 2023. [Online]. Available: https://doi.org/10.3390/su15097175
[14] S. Ding, W. Jing, H. Chen, y C. Chen, “Yolo Based Defects Detection Algorithm for EL in
PV Modules with Focal and Efficient IoU Loss”, Appl. Sci., vol. 14, no. 17, pp. 7493, jan.
2024. [Online]. Available: https://doi.org/10.3390/app14177493
[15] J. L. Espinoza, L. G. González, y R. Sempértegui, “Micro Grid Laboratory as a Tool for
Research on Non-Conventional Energy Sources in Ecuador”, en 2017 IEEE International
Autumn Meeting on Power, Electronics and Computing (ROPEC), 2017, pp. 1-7. [Online].
Available: https://doi.org/10.1109/ROPEC.2017.8261615
[16] International Electrotechnical Commission, “IEC 61215-1:2016 - Terrestrial Photovoltaic
(PV) Modules – Design Qualification and Type Approval – Part 1: Test Requirements”.
2016. [Online]. Available: https://www.iecee.org/certification/iec-standards/iec-61215-
12021
Copyright (2026) © Franklin Gómez López, Danny Ochoa Correa y Isabel Cabrera Carrera.
Este texto está protegido bajo una licencia internacional Creative Commons 4.0. Usted es libre para compartir, copiar y redistribuir el material en
cualquier medio o formato. También podrá adaptar: remezclar, transformar y construir sobre el material. Ver resumen de la licencia.
