Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Photovoltaic panel inspection using deep learning systems on unmanned aerial vehicle based images

Dimopoulos Konstantinos

Full record

---

Summary

It is apparent to everyone that the world that we live in is slowly but steadilytransitioning from fossil fuels to renewable and clean energy. This is quite beneficial forthe environment, the humans that inhabit this planet and the fauna and flora as well.One of the most rapidly growing sources of green energy is solar anergy. Constructionsof huge solar farms are taking place each year. With each farm containing hundreds ofthousands of individual photovoltaic panels (PV), the need for a faster, safer, and moreefficient way of condition monitoring for each panel has risen. Surveillance monitoringvia the use of unmanned aerial vehicles (UAVs) has been proposed and applied, showingexceptional results. In this thesis, the focus is on investigating the use of novel deeplearning models, that take as an input infrared spectrum images of individualphotovoltaic panels with non-annotated multiple damages. The goal is first to create anannotated dataset for training, validation and testing purposed and a general dataaugmented dataset for further testing purposes. The second goal is to achieve anoptimal training of 2 individually trained models. The first of the models excels in binarydetection of surface of solar panels, meaning that it can detect efficiently whether aphotovoltaic panel is damaged or not. The other model excels at multi-classclassification, meaning that it can detect, identify and classify the specific error on adamaged panel, and then proceeds to put a bounding box over the error, pinpointingits exact location on the image. Throughout this this sensitivity investigation, theYOLOv8 and YOLONAS models are used. For YOLOv8, the detection performance on thetraining phase achieved for the binary model an F1 score of 93.8% and mAP50-90 scoreof 97.7%. Our multi-class classification model can correctly identify 8 distinct categoriesof solar panel surface faults, with different percentages of success depending on thefault class in question (e.g. F1 score: 64.4% and mAP50-90 score: 45.3%). The specs ofthese trained models are further analyzed in this thesis, together with all the processthat was required to reach this point. These models were tested both on part of theinitial database as well as a smaller database that was created using data augmentation.For first time, two cutting-edge DNN models were used for multiclass detection ofdamages in PV panels paving the way for future more efficient and massive detection ofdamages in PV farms.

Available Files

Services