Hierarchical Modeling for Low-Resolution and Ambiguous Thermal Fault Classification

¹ FPT School of Business & Technology, Vietnam
² University of Science and Technology - The University of Danang, Vietnam

^* e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
^** e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Thermal fault classification in photovoltaic modules remains challenging due to low-resolution infrared images, class imbalance, and strong visual ambiguity between fault types. Conventional approaches treat this task as a flat multi-class problem, which requires a single model to simultaneously separate heterogeneous classes and often leads to unstable decision boundaries. This paper presents a structured modeling approach that decomposes the classification process into a hierarchy of conditional decisions, including anomaly detection, family-level routing, and subtype classification. This formulation aligns the learning process with the intrinsic organization of fault categories and simplifies the underlying decision structure. Experiments conducted under a controlled setting show that the hierarchical formulation improves classification accuracy by up to 8.8% compared to a flat baseline using the same backbone and training conditions. Ablation results further demonstrate that each stage contributes to performance gains, while a collapsed-label analysis indicates that remaining errors are primarily concentrated in fine-grained subtype distinctions. These findings suggest that restructuring the classification problem is more effective than increasing model complexity for low-resolution and ambiguous thermal fault recognition.