SPLTRAK Abstract Submission
A Deep Learning Approach for PV Failure Mode Detection in Infrared Images: First Insights
Daniel Rocha1,2,3, Miguel Lopes7, Jennifer P. Teixeira1, Paulo A. Fernandes1,4,5, Modesto Morais7, Pedro M. P. Salomé1,6
1INL-International Iberian Nanotechnology Laboratory, Braga, Portugal
/2Algoritmi R&D, University of Minho, Guimarães, Portugal
/32Ai, School of Technology, Polytechnic Institute of Cávado and Ave, Barcelos, Portugal
/4I3N, Universidade de Aveiro, Campus Universitário de Santiago, Aveiro, Portugal
/5CIETI, Departamento de Física, Instituto Superior de Engenharia do Porto, Porto, Portugal
/6Departamento de Física, Universidade de Aveiro, Campus Universitário Santiago, Aveiro, Portugal
/7IEP - Instituto Electrotécnico Português, Custóias, Portugal

Large-scale solar power plants require cheap and quick inspections, for this unmanned aerial vehicle (UAV's) for high resolution optical and infrared imaging were introduced in the past years. While using UAV’s is fast for image acquisition, image is a time-consuming process where the best of practice today is still for an expert to individually analyze each image. As such, in this work we use computer vision to accelerate this process. We performed an instance segmentation assessment using a pre-trained mask R-CNN for the segmentation of defective modules, and cells, as well as for segmentation and classification of failures. This method was chosen due its good past performance. In this work we created a database from a solar power plant consisting of 42048 modules and an expert analyzed the images. Later on, our computer algorithm results were benchmarked against the expert. Our algorithm achieved a mean average precision (mAP) in defective module segmentation mask of 72.1 % and 47.9 % in segmentation mask of failure type with an intersection over union threshold (IoU) of 0.50, without human interference. The presented preliminary results allow to assess the methodology advantages and drawbacks to increase performance and pave the way to a large-scale study.