基于随机森林的发电机定子线棒局部放电图谱特征识别方法

Recognition Method of Partial Discharge Spectrum Feature in Generator Stator Bar Based on Random Forest

发电机定子线棒局部放电的在线监测和故障诊断对于电机故障预警、故障定位和指导机组检修有重要意义。目前存在部分不同类型局部放电相似度高,模式识别方法计算时间过长的问题,需要高精度和高效的识别方法,为此提出一种基于随机森林的定子线棒局部放电识别方法。制备了6种类型定子线棒,用特高频天线检测局部放电信号,基于相位分布局部放电(PRPD)图谱比对,提出参数幅值不对称度。基于随机森林方法识别缺陷类型,计算特征重要性,选择有效特征。最后,可视化特征相似度,与传统反向传播(BP)神经网络方法对比,验证有效性。结果表明:随机森林算法可以有效识别人工缺陷定子线棒局部放电,总体识别正确率达到93.33%。筛选半数特征后,随机森林比BP神经网络的准确率提高了10.83%,特征选择对随机森林的准确率影响很小,但识别效率大幅提高。随机森林在少量特征时识别准确率、计算时间都明显优于神经网络。幅值不对称度参数的重要性排在全部特征前1/3,具有推广价值。

On-line monitoring and fault diagnosis of partial discharge(PD)in generator stator bar is of great significance for fault warning,fault locating,and helping unit maintenance.There are problems that some different types of partial discharges have high similarity and the calculation time of pattern recognition method is excessively long,which needs a high-precision and fast recognition method.Therefore,a partial discharge recognition method of stator bar based on ran-dom forest is proposed.Six types of defective stator bars are fabricated,and ultra-high frequency antenna is used to obtain partial discharge signals.The phase resolved partial discharge(PRPD)spectrum is compared to extract a feature called amplitude asymmetry.Moreover,the defect type is identified by the random forest method,and the feature importance is calculated to select effective features.Finally,from the viewpoint of visual feature similarity,the random forest is com-pared with the traditional back propagation(BP)neural network to verify the effectiveness.The results show that the random forest algorithm can effectively identify the partial discharge from artificial defects in stator bars with an overall correct identification rate of 93.33%.After halving the number of features,the accuracy of random forest is 10.83%high-er than that of BP neural network.Feature selection has a negligible effect on the accuracy of random forest,whereas the recognition efficiency is greatly improved.The recognition accuracy and calculation time of random forest are signifi-cantly better than those of neural network in a small number of features.The validity of the amplitude asymmetry is ranked in the first third of all features,which can be promoted further.