Evaluating deep learning detectors for thermal photovoltaic fault localization using independent source families

Abstract Unmanned aerial vehicle (UAV)-based automated infrared thermography (IRT) offers a scalable approach to condition monitoring in utility-scale photovoltaic (PV) plants, but published deep-learning evaluations exhibit persistent methodological weaknesses: frame-to-frame data leakage, unaddressed class imbalance, inconsistent input resolutions, and unstandardized detector benchmarking. This work presents a controlled, source-family-independent benchmark comparing single-stage (YOLOv8s, YOLO11s) and two-stage (Faster R-CNN with ResNet-50-FPN) object detectors on 7,339 cleaned thermal images with 20,995 annotated defect instances from the Thermo-Solar dataset. To reduce evaluation bias, dataset partitioning enforces zero leakage across 2,431 unique source families defined by physical PV module identity, flight pass sequence, and camera telemetry. The single-stage anchor-free detectors clearly dominate the two-stage baseline in a Pareto sense. YOLOv8s and YOLO11s attain similar localization accuracy (0.6958 and 0.6876 [email protected], respectively), with YOLO11s offering a 15.3% smaller parameter footprint (9.43M vs. 11.14M) and lower computational complexity (21.7 GFLOPs). Faster R-CNN achieved lower localization accuracy (0.6542 [email protected]), 7.3× slower forward-pass inference latency on an NVIDIA Tesla T4 GPU (75.86 ms/image vs. 10.20–10.40 ms/image), and an 8× higher computational training runtime (7.86 hours for 20 epochs vs. ~1.1 hours for 50 epochs). Despite a raw 119.6:1 class imbalance, the single-stage models retained stronger robustness on minority fault classes (String Reversed Polarity: 0.5817 [email protected]) than the Region Proposal Network baseline (0.2781 [email protected]), while spatially dense Multi Hotspots caused severe bounding-box suppression problems (0.4342–0.4781 [email protected]). This work provides a rigorous and reproducible baseline for operational UAV thermography.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-11340775/v1
Primary Topic
Industrial Vision Systems and Defect Detection
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Evaluating deep learning detectors for thermal photovoltaic fault localization using independent source families

Alberto Ernesto Coboi, Nam Pham Van, Trinh Trong Chuong, Diarra Dramane
5 links
Research Square
Industrial Vision Systems and Defect Detection
preprint

Evaluating deep learning detectors for thermal photovoltaic fault localization using independent source families

Alberto Ernesto Coboi, Nam Pham Van, Trinh Trong Chuong, Diarra Dramane
preprint en

Abstract

Abstract Unmanned aerial vehicle (UAV)-based automated infrared thermography (IRT) offers a scalable approach to condition monitoring in utility-scale photovoltaic (PV) plants, but published deep-learning evaluations exhibit persistent methodological weaknesses: frame-to-frame data leakage, unaddressed class imbalance, inconsistent input resolutions, and unstandardized detector benchmarking. This work presents a controlled, source-family-independent benchmark comparing single-stage (YOLOv8s, YOLO11s) and two-stage (Faster R-CNN with ResNet-50-FPN) object detectors on 7,339 cleaned thermal images with 20,995 annotated defect instances from the Thermo-Solar dataset. To reduce evaluation bias, dataset partitioning enforces zero leakage across 2,431 unique source families defined by physical PV module identity, flight pass sequence, and camera telemetry. The single-stage anchor-free detectors clearly dominate the two-stage baseline in a Pareto sense. YOLOv8s and YOLO11s attain similar localization accuracy (0.6958 and 0.6876 [email protected], respectively), with YOLO11s offering a 15.3% smaller parameter footprint (9.43M vs. 11.14M) and lower computational complexity (21.7 GFLOPs). Faster R-CNN achieved lower localization accuracy (0.6542 [email protected]), 7.3× slower forward-pass inference latency on an NVIDIA Tesla T4 GPU (75.86 ms/image vs. 10.20–10.40 ms/image), and an 8× higher computational training runtime (7.86 hours for 20 epochs vs. ~1.1 hours for 50 epochs). Despite a raw 119.6:1 class imbalance, the single-stage models retained stronger robustness on minority fault classes (String Reversed Polarity: 0.5817 [email protected]) than the Region Proposal Network baseline (0.2781 [email protected]), while spatially dense Multi Hotspots caused severe bounding-box suppression problems (0.4342–0.4781 [email protected]). This work provides a rigorous and reproducible baseline for operational UAV thermography.

Research Square
Pan African University Institute for Basic Sciences, Technology and Innovation (KE), Hanoi University of Industry (VN)
Đại học Công nghiệp Hà Nội
Affordable and clean energy, Industry, innovation and infrastructure
Industrial Vision Systems and Defect Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon