A Review on Multimodal-Based Fault Detection and Diagnosis in Photovoltaic Systems: Fusion and Attention Mechanisms

The reliability and efficiency of photovoltaic (PV) systems are critical to ensuring the stability and sustainability of modern power grids, particularly with the increasing integration of renewable energy sources. However, PV installations are exposed to diverse environmental and operational conditions that can lead to various faults, reducing power output and system lifespan. Traditional fault detection and diagnosis (FDD) methods often rely on single-source measurements and handcrafted features, which limit their adaptability and diagnostic precision in complex, real-world conditions. In this context, multimodal learning has emerged as a promising paradigm that leverages heterogeneous data sources such as electrical, thermal, visual, and environmental information to enhance fault detection accuracy and robustness. This paper surveys recent advancements in multimodal-based FDD for PV systems, emphasizing fusion strategies and attention mechanisms that enable effective cross-domain feature integration. We review key approaches at the data, feature, and decision levels, along with hybrid architectures that exploit attention to adaptively weight informative modalities. Critical research challenges, including data heterogeneity, sensor synchronization, imbalance in fault samples, and real-time implementation, are thoroughly discussed. Furthermore, emerging directions such as transformer-based architectures, self-supervised representation learning, edge-intelligent diagnosis, and privacy-preserving federated learning are explored as enablers for scalable and interpretable PV fault diagnosis. This review provides a comprehensive roadmap toward the development of intelligent, adaptive, and resilient FDD frameworks for next-generation photovoltaic systems.