作者：孙海蓉1, 李帅1,2
作者单位：1. 华北电力大学控制与计算机工程学院，河北 保定 071003;
2. 华北电力大学河北省发电过程仿真与优化控制工程技术研究中心，河北 保定 071003
关键词：图像识别;热斑检测;胶囊网络;迁移学习
摘要：
为了提高小样本条件下光伏电池片热斑分类模型的特征提取能力，该文提出一种基于迁移学习的注意力胶囊网络。该网络不同于传统胶囊网络单一卷积层的特征提取，它是利用迁移学习的方法通过VGG-19网络和注意力机制进行特征提取，得到关键数据特征。然后构造向量神经元输入胶囊网络，利用动态路由算法，得到用以分类的数字胶囊层，实现光伏电池片的分类。实验结果表明：当样本数量较小时，该网络对电池片红外图像的识别准确率分别高于传统卷积神经网络和胶囊网络13.15%、7.91%，泛化能力更强，运算速度更快。
Improved capsule network method for small sample photovoltaic hot spot recognition
SUN Hairong1, LI Shuai1,2
1. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China;
2. Hebei Engineering Research Center of Simulation Optimized Control for Power Generation, North China Electric Power University, Baoding 071003, China
Abstract: In order to improve the feature extraction ability of the photovoltaic cell hot spot classification model under the condition of small samples, this paper proposes an attention capsule network based on transfer learning This network is different from the feature extraction of a single convolutional layer of the traditional capsule network. It uses the method of migration learning to extract the features through the VGG-19 network and the attention mechanism to obtain key data features. Then construct the vector neuron input capsule network, and use the dynamic routing algorithm to obtain the digital capsule layer for classification to realize the classification of photovoltaic cells. The experimental results show that when the number of samples is small, the recognition accuracy of the infrared image of the battery is higher than that of the traditional convolutional neural network and the capsule network by 13.15% and 7.91%, with stronger generalization ability and faster calculation speed.
Keywords: image recognition;hot spot detection;capsule network;migration learning
2023, 49(2):106-112 收稿日期: 2021-07-20;收到修改稿日期: 2021-09-15
基金项目: 河北省自然科学基金（E2018502111）
作者简介: 孙海蓉(1972-),女,河北保定市人,副教授,研究方向为深度学习在发电领域的应用
参考文献
[1] 孙邦伍. 光伏电站优化运维技术研究[D]. 南京: 南京理工大学, 2019.
[2] 毛峡, 石天朋. 光伏热斑图像有效区域分割算法研究[J]. 太阳能学报, 2018, 39(5): 1270-1276
[3] 严勤, 邓高峰, 胡涛, 等. 基于深度循环神经网络的异常用电检测方法[J/OL]. 中国测试: 1-6[2021-07-20]. http://kns.cnki.net/kcms/detail/51.1714.TB.20201104.1600.012.html.
[4] 周林春, 陈春俊. 复数卷积神经网络滚动轴承故障诊断研究[J]. 中国测试, 2020, 46(11): 109-115
[5] 王飞龙, 刘萍, 张玲, 等. 改进胶囊网络的小样本图像分类算法[J/OL]. 计算机科学与探索: 1-9[2021-07-20]. http://kns.cnki.net/kcms/detail/11.5602.TP.20210426.1002.002.html.
[6] 王昕, 赵飞, 蒋佐富, 等. 迁移学习和卷积神经网络电力设备图像识别方法[J]. 中国测试, 2020, 46(5): 108-113
[7] XUAN G R, ZHANG W and CHAI P Q. EM algorithms of Gaussian mixture model and hidden Markov model[C]//Proceedings 2001 International Conference on Image Processing , 2001.
[8] SABOUR S, FROSST N, HINTON G E. Dynamic routing between capsules[C]//Proceeding of the Neural Information Processing Systems, 2017.
[9] WANG D, LIU Q. An optimization view on dynamic routing between capsules[C]//6th International Conference on learning Representations (ICIR 2018), 2018.
[10] 姚红革, 董泽浩, 喻钧, 等. 深度EM胶囊网络全重叠手写数字识别与分离[J/OL]. 自动化学报: 1-15[2021-07-20]. http://kns.cnki.net/kcms/detail/11.2109.TP.20200608.1049.002.html.
[11] 谢军, 肖朝轩, 张思刚, 等. 基于迁移学习和特征融合的复合绝缘子憎水性等级判别方法别方法[J/OL]. 电网技术: 1-9[2021-07-20]. https://doi.org/10.13335/j.1000-3673.pst.2021.0710.
[12] 朱佳丽, 宋燕. 基于迁移学习的注意力胶囊网络[J]. 智能计算机与应用, 2021, 11(2): 44-49
[13] 马欣睿. 基于基元特征分析的图像识别算法研究[D]. 西安: 西安电子科技大学, 2019.
[14] 郭梦浩, 徐红伟. 基于Faster RCNN的红外热图像热斑缺陷检测研究[J]. 计算机系统应用, 2019, 28(11): 265-270
