Automatic Fault Detection in Photovoltaic Modules Using Infrared Thermography and Deep Learning: A Comprehensive Review

Abstract

The rapid expansion of photovoltaic (PV) installations worldwide has created a growing demand for reliable and scalable inspection techniques capable of identifying module faults that reduce system efficiency and pose potential safety risks. Infrared (IR) thermography has become an important noncontact diagnostic tool for detecting thermal irregularities such as hotspots, cracked cells, and bypass diode failures, which typically appear as localized temperature variations when modules operate under similar environmental conditions [1], [8], [11].

At the same time, recent developments in deep learning have improved the analysis of thermographic data by enabling automated extraction of meaningful visual features from thermal images. Compared with traditional image processing methods that rely on manually designed features or rule-based detection, deep learning approaches are better suited for capturing complex thermal patterns associated with PV defects [3–5].

This paper presents a comprehensive literature survey of recent research that combines IR thermography, UAV-based imaging, and deep learning techniques for automated photovoltaic module fault detection. The review examines commonly used imaging platforms, available datasets, labeling methodologies, and machine learning architectures, including convolutional neural networks, transfer learning strategies, segmentation–classification frameworks, and hybrid machine learning–deep learning models [7], [10].

Findings from the surveyed studies indicate that CNN-based models trained on curated module-level datasets can achieve high detection accuracy, while UAV-based thermographic inspection enables efficient monitoring of large-scale PV installations [6], [12]. Techniques such as transfer learning and hybrid modeling have also shown improved performance in situations where labeled data is limited [7]. Nevertheless, several challenges remain, including dataset diversity, domain shifts between different PV plants, class imbalance, annotation complexity, and limitations related to real-time deployment on edge devices. This review summarizes the current progress in thermography-based deep learning systems for PV fault detection and highlights important research directions for improving their reliability and practical adoption.