SDHS-RLDNet: A real-time lightweight detection network for small-dense photovoltaic hot-spots

To ensure the safe and stable operation of photovoltaic power plants, it is crucial to conduct regular fault inspections on solar arrays. In a complex inspection environment, it is difficult to ensure the accuracy of detecting small and densely distributed photovoltaic hot-spot faults using traditio...

Full description

Saved in:
Bibliographic Details
Published in:Journal of real-time image processing 2025-01, Vol.22 (1), p.34, Article 34
Main Authors: Hao, Shuai, Li, Tong, Ma, Xu, Li, Tianqi, Chang, Chi, Qi, Tianrui
Format: Article
Language:English
Subjects:
Accuracy
Algorithms
Computer Graphics
Computer Science
Computer vision
Deep learning
Design
Fault detection
Faults
Image Processing and Computer Vision
Lightweight
Modules
Multimedia Information Systems
Parameters
Pattern Recognition
Power plants
Real time
Signal,Image and Speech Processing
Solar arrays
Solar collectors
Target detection
Weight reduction
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:To ensure the safe and stable operation of photovoltaic power plants, it is crucial to conduct regular fault inspections on solar arrays. In a complex inspection environment, it is difficult to ensure the accuracy of detecting small and densely distributed photovoltaic hot-spot faults using traditional algorithms, and real-time detection is also challenging. To solve the above issues, a real-time lightweight detection network for small and dense photovoltaic hot-spots is proposed. Firstly, a multi-scale target extraction module is designed to enhance the feature extraction capability of the backbone network, which can effectively detect faults at different scales. Additionally, a small target prediction head is added to improve the detection performance of small hot-spot faults. Secondly, a dense object detection module is designed to enhance the positional information of hot-spot faults and effectively suppress background interference caused by complex backgrounds. Furthermore, to achieve network lightweighting, the method of knowledge distillation is adopted. By transferring the parameters of the teacher network to the student network, it simplifies the network parameters, improves model inference speed, and ensures real-time detection performance of the network. Finally, to verify the superiority of the proposed network, seven classical algorithms are selected for comparison experiments. The experimental results demonstrate that SDHS-RLDNet can accurately detect multi-scale hot-spot faults under various conditions with an accuracy rate reaching 86.6%.
ISSN:1861-8200
1861-8219
DOI:10.1007/s11554-024-01600-y