In addition to our parallel event on September 15th, SOLARIS will share additional technical insights though poster presentations.

Make sure to join us at EU PVSEC!

Augmentation of Computer Vision-Analysis using Tandem Visual and Thermal Field-Imaging of Photovoltaic Modules (15/09/2026 at 10:30 (4BO.7))

Speaker: Dr. Thøger Kari (DTU University)

Tandem RGB and infrared thermography (IRT) imaging is widely used for drone-based monitoring of photovoltaic (PV) plants, particularly for detecting faulty or damaged modules. While both modalities are well established individually, their combined potential for automated fault detection remains underexplored. RGB provides high spatial resolution and rich visual information, but limited insight into non-visible faults. IRT reveals thermal anomalies associated with electrical faults, but its lower spatial resolution can lead to ambiguous patterns, particularly at cell level. This research will evaluate whether combining RGB and IRT can improve the reliability, robustness, and interpretability of automated PV inspection, with a particular focus on reducing false positives. It builds on the in-house AI model CartesiaNet, currently used to detect PV modules and their corners in SWIR images with high accuracy. As module features are similar in RGB and SWIR, this capability is expected to be transferable to RGB imagery. Detected module corners can then be mapped onto corresponding IRT images, enabling accurate localization, extraction, and rectification of individual modules in the thermal domain without requiring a dedicated IRT detection model. The aligned RGB and IRT data will support complementary fault-detection strategies, particularly for faults such as glass cracks and soiling that may be difficult to distinguish using IRT alone. For example, RGB information could help distinguish soiling-related thermal irregularities from actual module defects and identify modules requiring cleaning. The research will also investigate camera lens-distortion calibration, improvements to CartesiaNet for RGB-based corner detection, and methods to refine corner localization. As CartesiaNet relies primarily on synthetic training data, the transfer of its data-synthesis approach to RGB will be explored to minimise the need for extensive manual labelling. Finally, the study will assess whether improved localization and access to precise cell positions can enhance fault localization and overall inspection accuracy.

Investigation of Low-Energy Glass Crack Detection in Photovoltaic Modules During Daylight Electroluminescence Inspections (16/09/2026 at 17:00 (4CO.9))

Speaker: Mr. Rodrigo del Prado Santamaría (DTU University)

Broken glass in photovoltaic (PV) modules is a critical degradation mode, posing significant safety risks and potentially compromising module reliability. Cracks can result from mechanical stress during installation and transport, external impacts such as hail or debris, or long-term environmental loads. Even minor fractures can propagate over time, causing moisture ingress, corrosion, delamination, inverter tripping, and increased safety hazards. Early and reliable detection is therefore essential for safe and efficient PV plant operation. Current inspection methods rely mainly on visual inspection and UAV-based infrared thermography (IRT) and RGB imaging. However, these techniques may fail to detect low-energy cracks, particularly in bifacial modules where fractures can originate on the rear side without producing clear thermal or visual signatures. This research will investigate a motion-based daylight electroluminescence (EL) imaging technique designed to reveal low-energy glass cracks that are difficult to detect with conventional IRT and RGB inspections. The approach exploits changes in reflected and scattered sunlight caused by fractured glass surfaces during camera motion. These variations are accumulated during image reconstruction, enhancing crack visibility while preserving the EL emission from the underlying cells, thereby enabling the simultaneous detection of electrical defects and glass cracks. The research will investigate the physical mechanisms governing crack visibility and assess the feasibility of implementing the technique with lower-cost NIR or RGB cameras. Particular attention will be given to detecting cracks on the rear side of bifacial modules. Controlled indoor experiments using halogen illumination will determine the minimum illumination levels required for reliable detection, while spectral analysis will identify the optimal wavelength range for crack visibility and help explain the advantages of SWIR imaging. The findings will be validated through field inspections, including rear-side daylight imaging of bifacial modules, to assess crack detectability under realistic operating conditions. The objective is to establish a robust and scalable inspection methodology for the early detection of glass cracks, strengthening diagnostic capabilities for one of the most safety-critical degradation modes in PV plants.

A First-Principles Algorithm for Solar-Cell Feature-Segmentation, for Benchmarking, Labeling, and AI-Model Enhancement (17/09/2026 at 8:30 (3DO.16))

Speaker: Dr. Thøger Kari (DTU University)

Numerous methods have been developed for detecting and quantifying faults in photovoltaic (PV) cells from electroluminescence (EL) images, including physical models, regression approaches, deep learning, and threshold-based techniques. Each has advantages and limitations: physical models can be complex or lack precise crack localization, regression models are fast but provide limited spatial awareness, while AI-based approaches can be powerful but require large datasets and may suffer from generalisation issues. Thresholding methods also struggle with variations in crack intensity. In addition to accuracy, robust benchmarking and reliable image labelling remain important challenges. This research proposes a first-principles algorithm for automatically segmenting solar cell EL images into areas of interest and converting greyscale images into multi-channel representations containing statistical and structural features such as mean intensity, standard deviation, and gradient strength. A positional mask identifying busbar boundaries and non-cell areas provides spatial awareness and enables heuristic detection of line cracks and disconnected-area cracks without model training. The approach can support both automated labelling and benchmarking of more complex fault-detection and power-prediction methods. Its feature-based representation may also provide cleaner and more distinct inputs for AI models, potentially reducing data requirements and improving generalisation. Initial results show advantages over conventional thresholding and line-crack detection applied directly to the original images. Further research will assess the method’s potential as an industrial benchmarking and labelling tool and investigate its ability to enhance both AI and non-AI models. Their performance will be compared using both original EL images and the proposed feature-flattened representations.

Data-Driven Power-Loss Modelling of Photovoltaic Cells Using Physically Based Syntehtic EL-IV Data (17/09/2026 at 15:15(4DV.4))

Speaker: Rodrigo del Prado Santamaría (DTU University)

Accurately estimating the power output of photovoltaic (PV) cells and modules is important for the operation and assessment of large PV plants. While current performance assessments typically rely on electrical measurements, recent research has shown that machine learning (ML) models can estimate power output directly from electroluminescence (EL) images, which reveal cracks and other defects affecting module performance. However, training these models requires large datasets combining EL images with electrical measurements, which are costly and time-consuming to obtain. Rare degradation mechanisms are particularly difficult to represent in real-world datasets. This research investigates the use of synthetic data generated with Griddler, a solar cell simulation tool capable of producing both EL images and current-voltage (I-V) curves. The approach enables the creation of large datasets covering different defect types and severity levels. Two power-loss simulation models, developed by DTU and the University of Genova (UNIGE), will be evaluated using real cracked-cell samples with measured I-V curves. Synthetic EL/I-V datasets will then be combined with real measurements to determine the amount of synthetic data required and assess its impact on model performance. A further objective is to investigate whether synthetic data can accelerate the development of power-loss models for degradation mechanisms beyond cell cracks. To demonstrate this potential, synthetic datasets of potential-induced degradation (PID) will be generated and used to retrain the models. The research aims to reduce the experimental effort required to develop data-driven PV diagnostic tools and enable more scalable prediction of performance losses from EL images alone.