VibroNet: A YOLO-Based Framework with Frequency Enhancement and Attention for Vibratory Position Detection

Abstract

Accurate detection of vibratory positions in thermal solar panels is essential for maintaining photovoltaic system efficiency, ensuring structural stability, and enabling early fault diagnosis. However, subtle vibration signatures are often obscured by thermal noise and low-contrast temperature gradients, limiting the effectiveness of conventional computer vision and deep learning approaches. To address this challenge, this study proposes a vibration-aware detection framework based on an enhanced YOLO architecture. The proposed framework integrates three main components: (i) a Vibration Attention Module (VAM) that combines spatial and channel attention to highlight weak vibration-related features, (ii) Frequency Domain Enhancement using multi-scale high-pass filtering and Laplacian pyramid processing to strengthen vibration-sensitive signals while suppressing background noise, and (iii) a Progressive Multi-Resolution Learning strategy that improves training stability and robustness under varying thermal conditions. Experiments conducted on the TRSAI.v2i thermal solar panel dataset demonstrate that the proposed method achieves 0.982 mAP@50, 0.847 mAP@50–95, 0.968 precision, 0.982 recall, and an F1-score of 0.975. The results confirm that the proposed framework significantly improves vibration detection in thermal imagery, supporting reliable condition monitoring in solar energy systems.