针对变压器故障诊断多特征量输入时产生的冗余信息以及支持向量机核函数与惩罚因子优化对分类器产生的影响的问题,建立一种基于随机森林(random forest,RF)的特征选择,结合天牛须搜索算法(beetle antennae search,BAS)优化支持向量机的变压器油中溶解气体分析(dissolved gas analysis,DGA)故障诊断方法.该方法选择常见的5种油中溶解气体生成22维待选特征向量,通过RF算法对待选特征集进行重要度排序并消除冗余信息,得到最终的10维输入量,最后利用BAS算法对SVM中的惩罚因子及核函数进行寻优改进,对输入特征量进行故障诊断.仿真结果表明:相比PSO-SVM和ABC-SVM,BAS-SVM故障诊断模型故障诊断准确率分别高出6.83%和10.1%,诊断用时减少3.64 s和1.62 s,验证了所用方法的有效性和可行性.
Research on BAS-SVM Transformer DGA Fault Diagnosis Technology Based on RF Feature Selection
This paper investigates the impact of kernel function and penalty factor optimization of a Support Vector Machine(SVM)on a classifier,in order to resolve the issue of superfluous information in transformer fault diagnosis with multi-feature input.To this end,a feature selection technique based on Random Forest(RF)was established.A transformer DGA fault diagnosis method based on Support vector machine(SVM)was optimized by combining Beetle Antennae Search(BAS).This method takes common five kinds of gases dissolved in transformer oil to generate 22d c haracteristic vector to be choosen,and get the final 10 dimensions of input by RF importance sort algorithm selected feature set and eliminate redundant information,finally the BAS algorithm was applyed on the punishment factor and kernel function of SVM optimization improvement and fault diagnosis of the characteristic of input.Simulations show that BAS-SVM model's fault diagnosis accuracy was 6.83% and 10.1% higher than PSO-SVM's and ABC-SVM's,and the diagnosis time for the input characteristic is reduced by 3.64 seconds and 1.62 seconds,which verify the validity and feasibility of the method.
万方数据
维普
