摘要
太阳电池的生产组装过程工艺复杂,易在多个环节产生断栅、开焊、隐裂等缺陷,该文提出一种基于卷积神经网络和随机森林相结合的太阳电池缺陷识别方法,卷积神经网络的输出层采用随机森林分类器,该网络提取到的特征用于训练随机森林分类器,从而实现太阳电池的缺陷识别。该方法首先对组装太阳电池的EL(electro luminescence)图像进行预处理,提取小块太阳电池图像,然后在卷积神经网络中添加了BN(batch normalization)层,改变卷积核的大小与输出单元的个数,该网络还使用ReLu激活函数、Dropout正则化方法、批规范层(batch normalization)和Adam优化器等方法来提升网络性能。仿真结果表明,卷积神经网络和随机森林结合能够有效提高太阳电池的缺陷识别率和识别速度,并在原图像上进行准确标注。
Due to the complexity of the process in the production and assembly of solar cells,it is easy to generate defects such as broken gates,open welds and hidden cracks. This paper proposes a method to detect the defects of solar cells based on convolutional neural network and random forest(RF),The output layer of convolutional neural network adopts(RF)classifier which is trained by the features extracted by convolutional neural network. Doing that to realize the defect recognition of solar cells. The method first preprocesses the EL image of the assembled solar cell,extracts the image of the small cell,and then adds a BN layer to change the size of convolution kernel and the number of output units. The network also uses the ReLu Activation function,Dropout,Batch Normalization layer and Adam optimizer to improve the network performance. The simulation results show that the combination of convolutional neural network and random forest can effectively improve the defect recognition rate and recognition speed,and accurately mark it on the original image.
作者
周颖
毛立
张燕
陈海永
Zhou Ying;Mao Li;Zhang Yan;Chen Haiyong(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300130,China;China Hebei Control Engineering Research Center,Tianjin 300130,China)
出处
《太阳能学报》
EI
CAS
CSCD
北大核心
2020年第12期69-76,共8页
Acta Energiae Solaris Sinica
基金
国家自然科学基金(61403119)
河北省教育厅重点项目(ZD2016071)。
关键词
太阳电池
图像处理
缺陷检测
卷积神经网络
随机森林
solar cells
image processing
defect detection
convolutional neural networks
random forests(RF)
