基于S变换和深度置信网络的变压器局部放电模式识别  

Pattern Recognition on Partial Discharge Signals of Transformers Based on S-transform and Deep Belief Network

作　　者：贾亚飞[1] 朱永利[1] 兰志堃 王刘旺[1] 

机构地区：[1]新能源电力系统国家重点实验室(华北电力大学),河北保定071003 [2]国网河北省电力公司保定供电公司,河北保定071000

出　　处：《广东电力》2017年第1期108-115,共8页Guangdong Electric Power

基　　金：国家自然科学基金(51677072)资助项目;中央高校基本科研业务费专项资金资助(2016XS101)

摘　　要：针对目前常用的浅层模式识别方法无法有效处理高维特征量的问题,提出一种基于深度置信网络(deep belief network,DBN)的局部放电模式识别方法,并提出在DBN学习过程中采用自适应学习率来控制其在全局和局部的寻优能力。该方法首先对局部放电信号进行S变换得到二维时频矩阵;其次考虑时频矩阵中特征量之间的相关性和计算复杂度,对二维时频矩阵采用双向二维主成分分析(two-directional two-dimensional principal component analysis,(2D)2PCA)进行降维处理。最后,将降维得到的特征量输入DBN,从低层到高层逐层训练,并将训练好的DBN用于测试样本的模式识别。用上述方法对实验室条件下的四种不同放电模型产生的放电信号进行特征提取和模式识别,并与反向传播网络得到的识别结果进行比较,结果表明该方法对于高维特征量具有更高的正确识别率和更快的运行速度,更适用于高维度特征量的模式识别。In allusion to the problem of common pattern recognition method in shallow architecture being unable to effective- ly handle with high dimensional feature, this paper presents a kind of partial discharge （PD） pattern recognition method based on deep belief network （DBN） and proposes to use self-adaptive learning rate to control global and partial optimizing ability in the DBN learning process. This method firstly uses S-transform on PD signals to obtain two-dimensional time-fre- quency matrix, then adopts two-directional two-dimensional principal component analysis （（2D）2 PCA） to handle with two- dimensional time-frequency matrix for dimensionality reduction by considering correlation of feature among elements of time-frequency matrix and calculation complexity. Finally, the feature is input into DBN for training from the low layer to the high layer, and the trained DBN is applied in pattern recognition for testing samples. This method is used for feature ex- traction and pattern recognition on discharge signals of four different discharge models and the acquired results are compared with recognition result from backing propagation network. Comparison result indicates that this method has higher correct recognition rate and faster running speed which is more suitable for pattern recognition on high dimensional feature.

关 键 词：电力变压器 局部放电 模式识别 深度置信网络 S变换 

分 类 号：TM85[电气工程—高电压与绝缘技术]