基于方向梯度直方图属性空间的局部放电模式识别改进算法

Improved Algorithm for Partial Discharge Pattern Recognition Based on Histogram of Oriented Gradient Attribute Space

该文提出一种基于方向梯度直方图(HOG)属性空间的局部放电模式识别改进算法,旨在提高特征对样本信息的概括能力,并克服分类器对高维特征的识别局限。首先,构造局部放电脉冲序列分布(PRPS)模式作为识别依据,利用局部细分叠加窗口滑移的迭代算法自动化构建PRPS图像的HOG属性空间;然后,通过线性变换协方差矩阵的方式重构HOG属性空间,使其满足相互独立性;接着,根据属性重要性重新排列空间后,依次增加输入朴素贝叶斯分类器的属性个数,基于分类精度搜索最佳属性子集;最后,按照归约属性的相对重要性进行加权,最终设计出HOG属性选择加权朴素贝叶斯分类器。大量样本测试结果证明,此算法能够达到很高的识别精度,对传统识别算法的优化效果明显,有较好的应用价值。

This paper proposed an improved algorithm for partial discharge pattern recognition based on the attribute space of histogram of oriented gradient(HOG),aiming to enhance the summary ability of features and overcome the limitation of traditional classifier for the high-dimensional features.Firstly,the partial discharge phase resolved pulse sequence(PRPS)patterns were constructed as the recognition basis.And the HOG attribute space of PRPS images was formed automatically by the iterative algorithm of local cell superimposed slipped window.Secondly,in order to satisfy the mutual independence,the HOG attribute space was reconstructed by linear transformation of covariance matrix and was rearranged according to attribute importance.Thirdly,the number of attributes input to the Naïve Bayesian classifier was sequentially increased,and then the best attribute subset was obtained based on the classification accuracy.After weighting the reduced attribute according to their relative importance,the HOG attribute selective weighted naïve bayes classifier was finally designed.The test results of a large number of samples prove that the improved algorithm can achieve high recognition accuracy,and has an obvious optimization effect and good application value.