基于改进CNN的串联型故障电弧识别方法研究

Study on Identification Method of Series Arc Fault Based on Improved CNN

针对串联型故障电弧影响供电系统安全且不易测量的问题,提出改进卷积神经网络对串联型故障电弧的识别方法。由于SVM学习的超平面是距离各个样本最远的平面,相比于Softmax,具有更强的泛化推广能力和更高的识别准确率,故采用SVM损失函数(hingeloss)替换原有的Softmax损失函数并在CNN模型中添加三层Inception结构得到改进的CNN模型。其次,研制串联型故障电弧实验平台,在不同的实验条件下采集电流信号,利用短时傅里叶变换,得到电流的时频谱图。采用同一数据集对两个模型进行训练和测试。结果表明,改进的CNN电弧识别模型相较于CNN电弧识别模型在识别准确率和效率上都有了明显的提高。

Aiming at the problem that the series arc fault affects the safety of the power supply system and is not easy to measure, an improved convolutional neural network(CNN) is proposed to identify the series arc fault.SVM learning is the farthest plane from each sample. Compared with softmax, SVM has stronger generalization ability and higher recognition accuracy. Therefore, the SVM loss function is used to replace the original softmax loss function and three-layer Inception structure is added to the CNN model and the hyperplane ofto obtain an improved CNN model. Secondly, a series arc fault experiment platform is developed, and current signals are collected under different experimental conditions. Then, the time-frequency spectrum of current is obtained by a short-time Fourier transform. The two models are trained and tested with the same data set. The results show that the improved CNN arc recognition model can significantly improve the recognition accuracy and efficiency compared with the CNN arc recognition model.