Intelligent fault detection in photovoltaic systems: sensing modalities, AI methods, generalization and future trends

Abstract

Reliable fault detection is essential for maintaining the safety, energy yield and long-term reliability of photovoltaic (PV) systems as deployment expands worldwide. Artificial intelligence (AI) has substantially advanced PV fault diagnosis through automated analysis of electrical measurements, operational data and multimodal imagery. This Review critically examines intelligent PV fault-detection methods across I–V characteristics, supervisory control and data acquisition (SCADA), visible-light imaging, infrared thermography (IRT), electroluminescence (EL), unmanned aerial vehicle (UAV) inspection and multimodal sensing. In contrast to previous reviews that predominantly address individual sensing modalities, specific fault categories or AI techniques, we establish a decision-oriented framework that links fault observability and sensing physics with AI methodology, dataset provenance, localization capability, cross-site generalization and deployment requirements. Representative evidence illustrates both the potential and the limitations of current approaches: electrical-data models have reported accuracies above 98%, including 98.3% average accuracy on measured string-level I–V data; on the public ELPV benchmark, CNN-based EL classification achieved 88.42% average accuracy compared with 82.44% for SVM; and a field RGB–IRT framework reported 91.7% mAP@0.5, 95% defect-detection accuracy and 36 frames s -1 . However, these results are study-specific and are not directly comparable because datasets, fault taxonomies and validation protocols differ substantially. Across modalities, the principal barriers to practical deployment are the scarcity of representative and synchronized multimodal datasets, class imbalance, heterogeneous acquisition conditions and limited cross-site validation. Multimodal diagnosis is further differentiated into data-, feature- and decision-level fusion, together with emerging attention- and transformer-based approaches. We identify standardized multi-site benchmarks, self-supervised and foundation models, physics-informed learning, missing-modality robustness and lightweight edge intelligence as priorities for translating high-performing algorithms into scalable PV monitoring and predictive-maintenance systems.