基于深度自编码网络的电弧故障检测与负载类型识别

Arc Fault Detection and Load Type Identification Based on Deep Auto-encoding Network

针对有监督学习的电弧故障检测方法需要大量带标签数据且大多只检测电弧故障而未对负载类型进行识别的问题,提出一种基于深度自编码网络的电弧故障检测与负载类型识别方法;采用小波包分解的节点系数作为自编码网络的无标签输入特征量,并运用逐层训练方法对自编码网络进行预训练;为了使所提出方法的权重系数达到全局最优,采用少量有标签数据对所得权重进行微调,通过Softmax多分类器输出电弧故障检测结果,并根据负载类别最大概率识别电弧故障可能的负载类型。结果表明,所提出的方法对电弧故障检测与负载类型识别准确率达到98.56%,高于相同层数和参数规模的有监督学习网络的准确率。

To solve the problem that arc fault detection approaches based on supervised learning required a large amount of labeled data and most of them only detected arc faults without identifying load types,a method of arc fault detection and load type identification based on deep auto-encoding network was proposed.Node coefficients of wavelet packet decomposition were used as the unlabeled input characteristic quantity of deep auto-encoding network,and layer-by-layer training method was employed to pre-train the deep auto-encoding network.In order to make weight coefficients of the proposed method reach the global optimum,a small amount of labeled data were used to fine-tune the obtained weight.The arc fault detection results were output by using a Softmax multi-classifier.The approximated load type for each arc fault was determined according to the maximum probability of the load category.The results show that the accuracy of the proposed method for arc fault detection and load type identification reaches 98.56%,which is higher than that of supervised learning network with the same number of layers and parameter scale.