基于Faster R-CNN的非侵入式负荷识别方法

Non-Intrusive Load Identification Method Based on Faster R-CNN

针对非侵入式负荷辨识终端对特征相似电器、小样本数据辨识困难等问题,提出一种利用改进的Faster R-卷积神经网络(convolutional neural networks,CNN)对小样本负荷数据进行高精度辨识的方法。在传统的Faster R-CNN目标检测算法的基础上,增大了模型输入图像尺寸,保留更多负荷图像曲线的细节,提高了对小目标图像细节的识别效果;替换模型特征提取网络VGG16为Inception V2网络,拓宽了网络宽度,减小了差异较大的图像曲线变化尺寸对模型识别造成的干扰,以满足对不同尺度的负荷特征图像曲线的特征提取。在数据集上测试Faster R-CNN对非入侵式负荷设备识别的可行性与准确性,结果表明该方法显著减小了数据处理与识别网络的计算量,使识别的准确率与召回率均有了较大提升。

Aiming at the problems that the non-intrusive load identification terminal cannot realize the identification of similar electrical appliances and the identification of small sample data in household smart meters,a method for high-precision identification of small sample load data using Faster R-CNN is proposed.Based on the traditional Faster R-CNN target detection algorithm,the input image size of the model is increased to retain more details of the load image curve,and the recognition effect of small target image details is improved.The replacement model feature extraction network VGG16 is the Inception V2 network,which widens the network width and reduces the interference caused by the large difference in image curve change size to the recognition of the model,so as to meet the feature extraction of load characteristic image curves of different scales.The feasibility and accuracy of Faster R-CNN identification for non-invasive load equipment were tested on the data set.The results show that the method significantly reduces the computation of data processing and identification network,and greatly improves the accuracy and recall rate of identification.