RP-CFANet: An Adaptive Photovoltaic Hot-Spot Fault Detection Network Based on Region Perception and Cross-Channel Feature Aggregation

• Published in: IEEE transactions on instrumentation and measurement Vol. 74; pp. 1 - 14
• Main Authors: Hao, Shuai, Qi, Tianrui, Ma, Xu, Li, Jiahao, Li, Tianqi
• Format: Journal Article
• Language: English
• Published: New York IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Attention mechanism; Attention mechanisms; Autonomous aerial vehicles; Classification algorithms; cross-channel feature aggregation; Data mining; Fault detection; Fault diagnosis; Faults; Feature extraction; Heuristic algorithms; hot-spot fault detection; Infrared detectors; infrared sensors; Modules; Perception; Photovoltaic cells; Shape; U-Net
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2025.3555674

In the process of utilizing thermal infrared sensors to identify photovoltaic (PV) hot-spot faults, due to the diverse shapes of hot-spot faults and environmental interferences, fault characteristics cannot be effectively expressed, which poses a challenge for traditional detection networks in achieving accurate detection. Consequently, an adaptive PV hot-spot fault detection network based on region perception and cross-channel feature aggregation is proposed. First, to reduce the interference of pseudo-highlight features in complex backgrounds, a lite-UNet segmentation network is designed to remove background redundant information and enable the detection network to concentrate on the region of the PV panel. Second, to fully capture the geometric deformations and weak edge characteristics of hot-spot faults, a C2f_DCN module is designed, which enhances the feature extraction capabilities by adaptively adjusting the receptive field. Subsequently, to address the problem of feature coupling often encountered in extracting dense hot-spot fault features, a dense object visual enhancement (DOVE) module is proposed. Through the dynamic aggregation of cross-space features, the feature information of different channels is integrated to improve the detection accuracy of hot-spot faults. Additionally, to further enhance the effective fusion of multiscale features, an adaptive scale converter module (ASCM) detection head is designed. Finally, seven traditional detection methods are chosen for comparison to confirm the benefits of the proposed method. According to the experimental results, the suggested method can detect hot-spot faults with an accuracy of up to 86.5% in complicated inspection environments, which is an increase of 5.6% compared with the basis of the original network.