Online automatic anomaly detection for photovoltaic systems using thermography imaging and low rank matrix decomposition

Abstract

Faults occurred during the operational lifetime of photovoltaic (PV) systems can cause energy loss, system shutdown, as well as possible fire risks. Therefore, it is crucial to detect anomalies and faults to control the system’s performance and ensure its reliability. Comparing to traditional monitoring techniques based on an on-site visual inspection and/ or electrical measuring devices, the combination of drones and infrared thermography imaging evidently provides the means for faster and less expensive PV monitoring. However, the literature in this area lacks automatic and implementable algorithms for PV fault detection, particularly, using raw aerial thermography, with precise performance evaluation. The objective of this paper is, thus, to build a fully automatic online monitoring framework. We propose an analytical framework for online analysis of the raw video streams of aerial thermography. This framework integrates image processing and statistical machine learning techniques. We validate the effectiveness of the proposed framework and provide sufficient details to facilitate its implementation by practitioners. Two challenges hinder direct fault detection on raw PV images. One is that raw PV images often have non-smooth backgrounds that can impact the detection performance. This background needs to be removed before fault detection. However, this is a daunting task given the perspective of images. To deal with this challenge, we utilize the Transform Invariant Low-rank Textures (TILT) method to orthogonalize the perspective before applying edge detection to crop out the background and aligning the cropped images. The other issue is that the regular hot spots at the bottom edges of the solar panels are normal and should not be detected as anomalies. This makes the intensity-based detection method in the literature fail. These hot spots are part of the low-rank pattern of the image sequence. On the other hand, the hot spots caused by anomalies deviate from the normal low-rank pattern of the PV cells. Therefore, we propose a methodology that relies on Robust Principal Component Analysis (RPCA), which can separate sparse corrupted anomalous components from a low-rank background. The RPCA is applied to the PV images for simultaneous detection and isolation of anomalies. In addition to RPCA, we suggest a set of post-processing procedures for image denoising, and segmentation. The proposed algorithm is developed using 20 normal (with no anomalies) training samples and 100 test samples. The results showed that the algorithm successfully detects the anomalies with a recall of 0.80 and detects the significant anomalies with the maximum recall of 1. Our method outperforms two benchmark methods in terms of F1 score by 44.5% and 114.3%. The small number of false alarms is mostly due to irregular image patterns at the end of a PV array or an extreme non-orthogonal perspective. Since the number of false alarms is not large, it does not disrupt the inspection process, and they can easily be identified by an appraiser offline. The average computation time is 6.32 sec/image, which enables online automatic inspection of PV panels.

Keywords:

• Aerial thermography
• automatic anomaly detection
• robust PCA
• PV system

Acknowledgement

The authors would like to thank the editor, the editor of the Case Study section, and two anonymous referees for their constructive comments and suggestions that have considerably improved the article. The work of Paynabar was partially supported by National Science Foundation Grants CMMI-1839591. The authors are thankful to PVK S.r.l. industry (Italy) for providing the data set of the case study.

Additional information

Notes on contributors

Qian Wang

Qian Wang is currently a Ph.D. student in Machine Learning at H. Milton Stewart School of Industrial and Systems Engineering (ISyE) at Georgia Tech. His research focuses on the analysis of high-dimensional functional data including profiles, images and point cloud data using statistical machine learning tools for anomaly detection, quality control or causal analysis purposes in manufacturing.

Kamran Paynabar

Kamran Paynabar is the Fouts Family Early Career Professor and Associate Professor in the H. Milton Stewart School of Industrial and Systems Engineering at Georgia Tech. He received his Ph.D. in IOE and M.A. in Statistics from The University of Michigan. His research interests comprise both applied and methodological aspects of machine-learning and statistical modeling integrated with engineering principles. He is a recipient of the Data Mining Best Student Paper Award, the Best Application Paper Award from IIE Transactions, the Best QSR refereed paper from INFORMS, and the Best Paper Award from POMS. He has been recognized with the GT campus level 2014 Junior Faculty Teaching Excellence Award and Provost Teaching and Learning Fellowship. He served as the chair of QSR of INFORMS, and the president of QCRE of IISE. He is an Associate Editor for Technometrics and IEEE-TASE, a Department Editor for IISE-Transactions and a member of the editorial board for the Journal of Quality Technology.

Massimo Pacella

Massimo Pacella received the M.Sc. degree in computer engineering from the University of Salento (Italy) in 1998, and the Ph.D. degree in manufacturing and production systems from the Polytechnic University of Milan (Italy) in 2003. He received the Fulbright Fellowship in 2009. Currently, he is an Associate Professor with the Department of Engineering for Innovation, University of Salento. His main research interests are functional data processing, profile monitoring, design of experiments, manufacturing process control, and coordinate metrology, including machine learning techniques and methods of applied statistics. He is a member of the Italian Association for Manufacturing Technology (AITeM).