基于非线性尺度空间与极坐标分布熵的GIS局放特征提取方法

Feature Extraction Method for Partial Discharge in GIS Based on KAZE and Distribution Entropy in Polar Coordinates

气体绝缘组合电器(gas insulated switchgear,GIS)局部放电(partial discharge,PD)缺陷类型的准确识别对电力设备的状态评估与故障诊断至关重要,为了解决PD模式识别中特征提取准确性不足的问题,该文提出一种基于极坐标分布熵优化的非线性尺度空间特征(KAZE)提取方法。首先,对光、电单参量图谱进行非下采样轮廓波变换(non-subsampled contourlet transform,NSCT),得到包含信息更丰富的光电融合图谱;然后,利用KAZE算法提取图谱典型特征点,并依据相位、幅值、尺度信息将特征点发散至极坐标,提取子区域分布熵构成特征向量;最后,将特征信息带入经自适应增强算法(Adaboost)优化的长短期记忆网络(long short-term memory,LSTM)进行模式识别验证,并与KAZE法、统计参数法、卷积神经网络(convolutional neural networks,CNN)进行对比。结果表明,该文提出的特征提取方法在不同训练集分布下均可达到较高的识别率,最高可达91%,相较于统计参数法、CNN分别提升8.8%和4.4%,可以为提高GIS局部放电特征提取准确性提供参考。

Accurate identification of partial discharge(PD)defect of gas-insulated switchgear(GIS)is very important for the state assessment and fault diagnosis of power equipment,to solve the problem of insufficient accuracy of feature extraction in PD pattern recognition,a feature extraction method based on KAZE and distribution entropy in polar coordinates is proposed.First,non-subsampled contourlet transform(NSCT)is used to obtain the photoelectric fusion images containing more information.Then,the KAZE method is used to extract the typical feature points,and the feature points are diverged to the extreme coordinates according to phase,amplitude,and scale information.The subregion distribution entropy is extracted to form the feature vector.Finally,the feature information is brought into the long short-term memory network(LSTM)optimized by an adaptive enhancement algorithm(Adaboost)for pattern recognition verification,which will be compared by the KAZE method,statistical parameter method,and convolutional neural networks(CNN).The results show that the feature extraction method proposed in this paper can achieve a high recognition rate of up to 91%under different training set distributions,which is 8.8%and 4.4%higher than the statistical parameter method and CNN.This method can provide a reference for improving the accuracy of GIS PD feature extraction.