This dataset provides comparative images and power output performance records of damaged photovoltaic (PV) cells, along with an extensive textual description of each degradation and power loss level. The dataset is limited to a small number of cells from a single photovoltaic technology, as it is designed for controlled multimodal correspondence rather than statistical generalization. Each PV cell (DAUERHAFT Polycrystalline AP130 × 150 MM 4.5 [Wp]), originally brand-new, was deliberately subjected to progressive damage caused by soiling and by mechanical actions such as direct impact from various blunt objects, cuts, abrasions, scrapes, bending, smashing, cracking, as well as thermally and electrically induced degradation, including direct exposure to flames and Joule effect overheating resulting in a fire. Under each condition, a set of three images was acquired, using different techniques: visible high-dynamic-range (HDR) under artificial uniform light conditions; near-infrared monochrome (NIR) under the same light conditions; electroluminescence (EL) without any external light source, but connecting the PV cell to a direct current power source current controlled. During EL, the imposed current and resulting voltage were recorded directly by the power source system. Then, after unplugging the power source, short-circuit current and open-circuit voltage were acquired under a standard test light source. Data was acquired in-door, using a laboratory electric power source, a digital multimeter, an HDR camera (for visible-HDR images), and a filtered visible/near-infrared camera (BP800 filter, for both NIR and EL images). Cameras were fixed in a binocular setup and optically calibrated with a calibration board. Using the same camera for NIR and EL resulted in an almost perfect pixel-pixel correspondence of images. The dataset highlights a significant difference between the visual information obtained by various imaging technology, showing the relevance of a multi-spectral approach as well as the valuable insight in degradation evaluation given by the EL when electrical parameters are recorded simultaneously. The dataset combines pixel-level aligned HDR, monochrome, and electroluminescence images with electrical measurements and natural-language descriptions of degradation conditions. Existing photovoltaic datasets do not appear to provide this specific combination of aligned imaging modalities, electrical data, and defect annotations. The dataset may support research on cross-modal correspondence learning and multimodal approaches that jointly exploit visual, electrical, and linguistic information. ©2026 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ )
Dataset of damaged photovoltaic cells electroluminescence, visible and near-infrared multispectral images with performance degradation data and LLM-RAG descriptors
Polenghi M.;Caldelli R.;Loconsole C.;
2026-01-01
Abstract
This dataset provides comparative images and power output performance records of damaged photovoltaic (PV) cells, along with an extensive textual description of each degradation and power loss level. The dataset is limited to a small number of cells from a single photovoltaic technology, as it is designed for controlled multimodal correspondence rather than statistical generalization. Each PV cell (DAUERHAFT Polycrystalline AP130 × 150 MM 4.5 [Wp]), originally brand-new, was deliberately subjected to progressive damage caused by soiling and by mechanical actions such as direct impact from various blunt objects, cuts, abrasions, scrapes, bending, smashing, cracking, as well as thermally and electrically induced degradation, including direct exposure to flames and Joule effect overheating resulting in a fire. Under each condition, a set of three images was acquired, using different techniques: visible high-dynamic-range (HDR) under artificial uniform light conditions; near-infrared monochrome (NIR) under the same light conditions; electroluminescence (EL) without any external light source, but connecting the PV cell to a direct current power source current controlled. During EL, the imposed current and resulting voltage were recorded directly by the power source system. Then, after unplugging the power source, short-circuit current and open-circuit voltage were acquired under a standard test light source. Data was acquired in-door, using a laboratory electric power source, a digital multimeter, an HDR camera (for visible-HDR images), and a filtered visible/near-infrared camera (BP800 filter, for both NIR and EL images). Cameras were fixed in a binocular setup and optically calibrated with a calibration board. Using the same camera for NIR and EL resulted in an almost perfect pixel-pixel correspondence of images. The dataset highlights a significant difference between the visual information obtained by various imaging technology, showing the relevance of a multi-spectral approach as well as the valuable insight in degradation evaluation given by the EL when electrical parameters are recorded simultaneously. The dataset combines pixel-level aligned HDR, monochrome, and electroluminescence images with electrical measurements and natural-language descriptions of degradation conditions. Existing photovoltaic datasets do not appear to provide this specific combination of aligned imaging modalities, electrical data, and defect annotations. The dataset may support research on cross-modal correspondence learning and multimodal approaches that jointly exploit visual, electrical, and linguistic information. ©2026 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ )I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
