View in EDS

A layer-2 solution for inspecting large-scale photovoltaic arrays through aerial LWIR multiview photogrammetry and deep learning: A hybrid data-centric and model-centric approach.

Saved in:
Bibliographic Details
Title: A layer-2 solution for inspecting large-scale photovoltaic arrays through aerial LWIR multiview photogrammetry and deep learning: A hybrid data-centric and model-centric approach.
Authors: Zefri, Yahya1 (AUTHOR) y.zefri@iav.ac.ma, Sebari, Imane1 (AUTHOR), Hajji, Hicham1 (AUTHOR), Aniba, Ghassane2 (AUTHOR), Aghaei, Mohammadreza3,4 (AUTHOR)
Source: Expert Systems with Applications. Aug2023, Vol. 223, pN.PAG-N.PAG. 1p.
Subjects: Deep learning, Blended learning, Photogrammetry, Workflow, Tiles
Abstract: Defective components within solar photovoltaic (PV) arrays overheat, resulting in particular temperature patterns under the long-wave thermal infrared (LWIR) spectrum. The detection and on-field localization of these patterns is of paramount aid to the operations and maintenance of PV installations. In this context, we develop a two-layer end-to-end inspection solution for the detection, quantification and on-field localization of overheated regions on PV arrays from LWIR UAV imagery. Layer 1 generates a georeferenced orthomosaic of the inspected site via a Structure from Motion-MultiView Stereo (SfM-MVS) photogrammetric acquisition/post-processing workflow. Layer 2 is a tile-based deep semantic segmentation stage that extracts and quantifies the affected regions from the generated orthomosaic. We collect aerial images from 103 PV sites, comprising approximately 342 000 modules. After a SfM-MVS workflow, we produce and annotate 7910 orthorectified unique affected tiles, posteriorly augmented to prepare the state-of-the-art dataset in terms of size and representativeness. Through a training/cross-validation and test process, we investigate the implementation of 9 models in the segmentation process: FCN, U-Net, FPN, DeepLab, LinkNet, DANet, CFNet, ACFNet and TransU-Net, each of which experimented with 2 backbones: ResNet50 and DenseNet121. The models feature efficient encoder-to-decoder feature map transfers, pyramidal feature recognition, spatial and channel attention, feature co-occurrence, class center as well as vision transformers. The best performance is achieved by FPN-DenseNet121, with a mean mIoU of 93.44% and an F1-score of 96.39% on our test set. The two-layer solution takes the best of the data-centric and model-centric paradigms, alongside addressing the limitations of conventional inspection procedures. It is put into a concrete application framework, where it provides a pixel-based and a tile-based quantification of the affected regions within a PV plant. The results are promising, and the selected model can be deployed efficiently for extensive aerial monitoring of large-scale PV plants. [ABSTRACT FROM AUTHOR]
Copyright of Expert Systems with Applications is the property of Pergamon Press - An Imprint of Elsevier Science and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:Defective components within solar photovoltaic (PV) arrays overheat, resulting in particular temperature patterns under the long-wave thermal infrared (LWIR) spectrum. The detection and on-field localization of these patterns is of paramount aid to the operations and maintenance of PV installations. In this context, we develop a two-layer end-to-end inspection solution for the detection, quantification and on-field localization of overheated regions on PV arrays from LWIR UAV imagery. Layer 1 generates a georeferenced orthomosaic of the inspected site via a Structure from Motion-MultiView Stereo (SfM-MVS) photogrammetric acquisition/post-processing workflow. Layer 2 is a tile-based deep semantic segmentation stage that extracts and quantifies the affected regions from the generated orthomosaic. We collect aerial images from 103 PV sites, comprising approximately 342 000 modules. After a SfM-MVS workflow, we produce and annotate 7910 orthorectified unique affected tiles, posteriorly augmented to prepare the state-of-the-art dataset in terms of size and representativeness. Through a training/cross-validation and test process, we investigate the implementation of 9 models in the segmentation process: FCN, U-Net, FPN, DeepLab, LinkNet, DANet, CFNet, ACFNet and TransU-Net, each of which experimented with 2 backbones: ResNet50 and DenseNet121. The models feature efficient encoder-to-decoder feature map transfers, pyramidal feature recognition, spatial and channel attention, feature co-occurrence, class center as well as vision transformers. The best performance is achieved by FPN-DenseNet121, with a mean mIoU of 93.44% and an F1-score of 96.39% on our test set. The two-layer solution takes the best of the data-centric and model-centric paradigms, alongside addressing the limitations of conventional inspection procedures. It is put into a concrete application framework, where it provides a pixel-based and a tile-based quantification of the affected regions within a PV plant. The results are promising, and the selected model can be deployed efficiently for extensive aerial monitoring of large-scale PV plants. [ABSTRACT FROM AUTHOR]
ISSN:09574174
DOI:10.1016/j.eswa.2023.119950