为了检测太阳能电池的缺陷, 建立了太阳能电池板的电致发光(EL)图像与其缺陷类型间的神经网络预测模型, 可以对太阳能电池板不同类型缺陷进行自适应检测。首先, 采用主成分分量分析(PCA)算法对电致发光(EL)图像训练样本集降维; 然后, 将降维后得到的数据输入神经网络预测模型进行学习, 对模型的参数进行优化选取; 最后, 将训练好的网络对测试样本集进行仿真。仿真结果表明: 在采用相同的训练样本集和测试样本集条件下, 与反向传播神经网络(BPNN)相比, 径向基神经网络(RBFNN)具有全局最优特性, 结构简单, 最高识别率达96.25%, 计算时间较短, 能满足在线检测的要求。

In order to detect the defect on solar panels and improve the conversion efficiency, two neural network models were established between solar panels electroluminescence (EL) images and defect types, which can detect different types of defects on solar panels adaptively. Firstly, the dimensions of EL images training samples set were reduced by using principal component analysis (PCA). Then, EL images training samples set after dimension reduction was put into the neural networks for training. Finally, the testing samples set was simulated by the trained network through choosing the best parameters. Compared with BPNN, RBFNN has the advantages of global optimization characteristics and simple structure, which leads to the highest accuracy rate of 96.25% and shorter computational time. The experiment results show that RBFNN can meet the requirements of online detection.