摘要
针对复杂环境下光伏电池板热斑故障的多尺度目标导致检测困难的问题,提出一种融合知识蒸馏和注意力机制的检测算法。为实现故障特征信息高效提取与保留,设计一种融合高阶空间交互和通道注意力的模块以提升网络对于故障特征信息的表达能力;为增强复杂背景下目标信息表达能力,构建一种结合通道和位置信息的注意力模块来提高网络对于故障位置信息的识别准确率;采用知识蒸馏思想将教师网络的参数迁移至学生网络,在不增加任何复杂度的前提下提升学生网络的检测精度。为进一步精确定位热斑目标,引入Focal-CIoU损失函数加速网络收敛,从而提升检测性能。为验证算法有效性,与8种经典算法进行比较,实验结果表明,本文算法的检测精度最高,精度达84.8%,对于分辨率为640×512的图像检测速度可达142 FPS。
A detection algorithm combining knowledge distillation and attention mechanism is proposed to solve the problem that multi-scale target of the hot spot fault of photovoltaic panel in a complex environ⁃ment leads to difficult detection.To efficiently extract and retain fault feature information,a module that integrates higher-order spatial interaction and channel attention was designed to improve the expression ability of fault feature information.To further enhance the ability of expressing target information in a com⁃plex background,an attention module combining channel and location information was constructed to im⁃prove the recognition accuracy of fault location information.The parameters of teacher network were trans⁃ferred to student network by knowledge distillation,and the detection accuracy of student network was im⁃proved without adding any complexity.A focal-CIoU loss function was introduced to accelerate network convergence and improve detection performance.In verifying the effectiveness of the proposed algorithm against eight classical algorithms,the experimental results show that the proposed algorithm has the high⁃est detection accuracy(84.8%),and the detection speed can reach 142 FPS for images with a resolution of 640×512.
作者
郝帅
吴瑛琦
马旭
李彤
王海莹
HAO Shuai;WU Yingqi;MA Xu;LI Tong;WANG Haiying(College of Electrical and Control Engineering,Xi'an University of Science and Technology,Xi'an 710054,China)
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2023年第24期3640-3650,共11页
Optics and Precision Engineering
基金
国家自然科学基金资助项目(No.51804250)
中国博士后科学基金资助项目(No.2019M653874XB,No.2020M683522)
陕西省科技计划资助项目(No.2021JQ-572,No.2020JQ-757)
陕西省教育厅科研计划资助项目(No.18JK0512,No.21JK0769)。
关键词
深度学习
光伏热斑检测
知识蒸馏
注意力机制
deep learning
photovoltaic hot spot detection
knowledge distillation
attention mechanism
