基于注意力机制的光伏组件热斑检测算法

樊涛; 孙涛; 刘虎

doi:10.13700/j.bh.1001-5965.2021.0457

基于注意力机制的光伏组件热斑检测算法

doi: 10.13700/j.bh.1001-5965.2021.0457

樊涛¹,
孙涛^2, ,,
刘虎¹

1.
国家电网有限公司, 北京 100031
2.
国网数字科技控股有限公司, 北京 100053

基金项目:

国家重点研发计划 2018YFB1500800

国家电网有限公司科技项目 SGTJDK00DYJS2000148

详细信息

通讯作者:
孙涛, E-mail: suntao@sgec.sgcc.com.cn

中图分类号: TP391
计量
- 文章访问数: 480
- HTML全文浏览量: 141
- PDF下载量: 37
- 被引次数: 0
出版历程
- 收稿日期: 2021-08-11
- 录用日期: 2022-01-03
- 网络出版日期: 2022-01-25
- 整期出版日期: 2022-07-20

Hot spot detection algorithm of photovoltaic module based on attention mechanism

FAN Tao¹,
SUN Tao^{2
, ,},
LIU Hu¹

1.
State Grid Corporation of China, Beijing 100031, China
2.
State Grid Digital Technology Holding Co., Ltd., Beijing 100053, China

Funds:

National Key R & D Program of China 2018YFB1500800

Technology Project of State Grid Corporation of China SGTJDK00DYJS2000148

More Information

Corresponding author: SUN Tao, E-mail: suntao@sgec.sgcc.com.cn

摘要

摘要:
热斑现象是造成光伏组件发电能力下降的重要原因之一，热斑检测是光伏电站运维必不可少的工作。然而分布式光伏电站的规模普遍较小、选址分散、环境复杂多样，使用传统的热斑检测算法需要投入大量的人力资源。基于此，提出了一种基于注意力机制的热斑检测算法HSNet。通过图像分割消除反光影响，结合通道注意力机制，学习通道间的特征信息，增强目标区域的重要性，采用自定义锚点的方法提高检测速度，使用焦点损失激活函数和基于物体先验概率的类别预测方式改善训练目标样本不均衡导致的分类准确性低的问题，通过回归方法获取准确的目标位置。实验表明：设计的目标检测算法在窗体回归精度和分类准确性方面均有明显的优势，边界框平均精度和准确率分别提升了3.18%和2.42%。
- 热斑检测 /
- 目标检测 /
- 残差网络 /
- 注意力机制 /
- 分布式光伏电站
Abstract:
The hot spot phenomenon is one of the important reasons for the reduction of power generation capacity of photovoltaic panels, and the detection of hot spots is an essential task for operation and maintenance personnel. The scale of distributed photovoltaic power plants is generally small, the site is scattered, the environment is complex and diverse, and the operation and maintenance personnel need to invest a lot of human resources to detect hot spots using traditional hot spot detection methods. In this paper, we propose a new hot spot detection algorithm HSNet. Firstly, the influence of reflection is eliminated through image segmentation. Secondly, the feature information between channels is learned in combination with the channel attention mechanism to enhance the importance of the target area. The method of user-defined anchor points is used to improve the detection speed. Then, the focus loss activation function and the category prediction method based on the prior probability of objects are used to improve the problem of low classification accuracy caused by the imbalance of training target samples, Finally, the accurate target position is obtained by regression method. Experiments show that the target detection algorithm designed in this paper has significant advantages over other algorithms in terms of window regression accuracy and classification accuracy, and the mean accuracy and accuracy of the bounding box are improved by 3.18% and 2.42%, respectively.
- hot spot detection /
- object detection /
- residual networks /
- attention mechanism /
- distributed photovoltaic plants

HTML全文

图 1 无人机拍摄的光伏阵列

Figure 1. Photovoltaic arrays photographed by a UAV

下载: 全尺寸图片幻灯片

图 2 反光消除网络结构

Figure 2. Network structure of reflection elimination

下载: 全尺寸图片幻灯片

图 3 热斑检测模型处理框架HSNet

Figure 3. Hot spot detection framework of HSNet

下载: 全尺寸图片幻灯片

图 4 Residual单元结构

Figure 4. Structure of Residual unit

下载: 全尺寸图片幻灯片

图 5 SE模块结构

Figure 5. Structure of SE block

下载: 全尺寸图片幻灯片

图 6 SE-Residual模块结构

Figure 6. Structure of SE-Residual block

下载: 全尺寸图片幻灯片

图 7 反光消除网络效果

Figure 7. Qualitative results of reflection elimination

下载: 全尺寸图片幻灯片

图 8 不同模型的P-R曲线比较

Figure 8. Comparison of P-R curves of different models

下载: 全尺寸图片幻灯片

图 9 不同检测算法的结果比较

Figure 9. Comparison of detection results with different detection algorithms

下载: 全尺寸图片幻灯片

表 1 符号表示

Table 1. Symbolic representation

符号	含义
D	训练集
I_i	第i张图像
y_ij	第i张图像的第j个光伏组件的类别
x	层运算输入
F	卷积等层计算公式
L_cls	分类损失函数
PL(pt)	概率为pt的Focal Loss损失函数
bbox_cls	回归损失函数
SE	SE处理模块
GT	Ground Truth人工标注框
M	训练模型
ReLU	ReLU激活函数
θ_F	初始化网络模型参数
batch_size	训练图像批量大小
FM	提取的特征图
(O_x, O_y), O_w, O_h	输出目标的位置及目标的宽度和高度

下载: 导出CSV

表 2 不同标注框的数量

Table 2. Number of different kinds of labeled boxes

种类	图像数量	标注框数量	占比/%
normal	440	24 846	89.32
dirty	286	2 736	9.84
hotspot	28	109	0.39
hotline	98	125	0.45

下载: 导出CSV

表 3 不同模型的结果对比

Table 3. Comparison of results with different models

算法	bbox_mAP/%	mAP/%	准确率/%
HSNet	55.97	70. 95	98.22
Faster R-CNN	50.64	67.77	95.80
RetinaNet	20.98	38.17	43.92
SSD300	66.57	67.96	96.50
YOLOv3	48.78	59.48	91.32

下载: 导出CSV

表 4 消融实验结果

Table 4. Results of ablation experiments

算法	bbox_mAP/%	mAP/%	准确率/%
F-RCNN	50.64	67.77	95.80
F-RCNN+UN	51.72	68.86	96.92
F-RCNN+UN+RN	53.68	68.92	96.99
F-RCNN+UN+SE	55.38	69.18	96.43
F-RCNN+UN+FL	52.21	69.30	96.95
F-RCNN+UN+RN+FL	54.23	69.58	97.34
F-RCNN+UN+SE+FL	55.85	70.50	98.01
F-RCNN+UN+SE+FL+PRI	55.97	70.95	98.22

下载: 导出CSV

参考文献(28)

[1]	云平, 张志祥, 刘恒. 基于I-V特性的多晶硅光伏组件故障及失效研究[J]. 太阳能, 2020(1): 61-69. doi: 10.3969/j.issn.1003-0417.2020.01.008 YUN P, ZHANG Z X, LIU H. Study on failure mechanism of polycrystalline silicon PV module base on I-V characteristics[J]. Solar Energy, 2020(1): 61-69(in Chinese). doi: 10.3969/j.issn.1003-0417.2020.01.008
[2]	温浚铎, 郭珂, 王瀚笠, 等. 光伏阵列热斑故障在线检测方法研究[J]. 电子制作, 2019(15): 55-58. doi: 10.3969/j.issn.1006-5059.2019.15.021 WEN J D, GUO K, WANG H L, et al. Photovoltaic array hot spot fault online detection method research[J]. Practical Electronics, 2019(15): 55-58(in Chinese). doi: 10.3969/j.issn.1006-5059.2019.15.021
[3]	陈斯, 戴新宇, 杨增汪, 等. 光伏旁路二极管在线监测系统[J]. 电测与仪表, 2012, 49(10): 96-100. doi: 10.3969/j.issn.1001-1390.2012.10.021 CHEN S, DAI X Y, YANG Z W, et al. A bypass diode of photovoltaic on-line monitoring system[J]. Electrical Measurement & Instrumentation, 2012, 49(10): 96-100(in Chinese). doi: 10.3969/j.issn.1001-1390.2012.10.021
[4]	谢彭宇, 徐新. 基于多尺度联合学习的行人重识别[J]. 北京航空航天大学学报, 2021, 47(3): 613-622. doi: 10.13700/j.bh.1001-5965.2020.0445 XIE P Y, XU X. Multi-scale joint learning for person re-identification[J]. Journal of Beijing University of Aeronautics and Astronautics, 2021, 47(3): 613-622(in Chinese). doi: 10.13700/j.bh.1001-5965.2020.0445
[5]	李广帅, 苏娟, 李义红. 基于改进Faster R-CNN的SAR图像飞机检测算法[J]. 北京航空航天大学学报, 2021, 47(1): 159-168. doi: 10.13700/j.bh.1001-5965.2020.0004 LI G S, SU J, LI Y H. An aircraft detection algorithm in SAR image based on improved Faster R-CNN[J]. Journal of Beijing University of Aeronautics and Astronautics, 2021, 47(1): 159-168(in Chinese). doi: 10.13700/j.bh.1001-5965.2020.0004
[6]	廖华年, 徐新. 基于注意力机制的跨分辨率行人重识别[J]. 北京航空航天大学学报, 2021, 47(3): 605-612. doi: 10.13700/j.bh.1001-5965.2020.0471 LIAO H N, XU X. Cross-resolution person re-identification based on attention mechanism[J]. Journal of Beijing University of Aeronautics and Astronautics, 2021, 47(3): 605-612(in Chinese). doi: 10.13700/j.bh.1001-5965.2020.0471
[7]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems, 2017: 5998-6008.
[8]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]//2017 IEEE International Conference on Computer Vision. Piscataway: IEEE Press, 2017: 2999-3007.
[9]	LECUN Y, BOTTOU L, BENGIO Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324. doi: 10.1109/5.726791
[10]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90. doi: 10.1145/3065386
[11]	BENGIO Y, SIMARD P, FRASCONI P. Learning long-term dependencies with gradient descent is difficult[J]. IEEE Transactions on Neural Networks, 1994, 5(2): 157-166. doi: 10.1109/72.279181
[12]	DUVENAUD D, RIPPEL O, ADAMS R P, et al. Avoiding pathologies in very deep networks[EB/OL]. (2016-07-08)[2021-08-01]. https://arxiv.org/abs/1402.5836.
[13]	ORHAN A E, PITKOW X. Skip connections eliminate singularities[EB/OL]. (2018-03-04)[2021-08-01]. https://arxiv.org/abs/1701.09175.
[14]	HINTON G E, SALAKHUTDINOV R R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786): 504-507.
[15]	ZHANG J Z, HE T X, SRA S, et al. Why gradient clipping accelerates training: A theoretical justification for adaptivity[EB/OL]. (2020-02-10)[2021-08-01]. https://arxiv.org/abs/1905.11881.
[16]	GLOROT X, BORDES A, BENGIO Y. Deep sparse rectifier neural networks[C]//Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, 2011: 315-323.
[17]	MAAS A L. Rectifier nonlinearities improve neural network acoustic models[C]//Proceedings of the 30th International Conference on Machine Learning, 2013.
[18]	CLEVERT D A, UNTERTHINER T, HOCHREITER S. Fast and accurate deep network learning by exponential linear units (ELUs)[C]//Proceedings of the 4th International Conference on Learning Representations, 2015.
[19]	HE K M, ZHANG X Y, REN S Q, et al. Delving deep into rectifiers: surpassing human-level performance on ImageNet classification[C]//2015 IEEE International Conference on Computer Vision. Piscataway: IEEE Press, 2015: 1026-1034.
[20]	IOFFE S, SZEGEDY C. Batch normalization: Accelerating deep network training by reducing internal covariate shift[C]//Proceedings of the 32nd International Conference on Machine Learning, 2015: 448-456.
[21]	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2016: 770-778.
[22]	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2017: 2261-2269.
[23]	XIE S N, GIRSHICK R, DOLLÁR P, et al. Aggregated residual transformations for deep neural networks[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2017: 5987-5995.
[24]	DAI J F, QI H Z, XIONG Y W, et al. Deformable convolutional networks[C]//2017 IEEE International Conference on Computer Vision. Piscataway: IEEE Press, 2017: 764-773.
[25]	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//2014 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2014: 580-587.
[26]	GIRSHICK R. Fast R-CNN[C]//2015 IEEE International Conference on Computer Vision. Piscataway: IEEE Press, 2015: 1440-1448.
[27]	REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[28]	HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2018: 7132-7141.