一种自主核优化的二值粒子群优化–多核学习支持向量机变压器故障诊断方法

An Autonomic Kernel Optimization Method to Diagnose Transformer Faults by Multi-Kernel Learning Support Vector Classifier Based on Binary Particle Swarm Optimization

支持向量机(support vector machine,SVM)对于核函数及模型参数十分敏感,多核学习可降低模型的参数敏感性。提出了基于二值粒子群优化(binary particle swarmoptimization,BPSO)的多核学习SVM分类方法(BPSO-MKSVC)进行变压器故障诊断。多核学习支持向量机(multi-kernel support vector classifier,MKSVC)采用由多个基核线性组合的多核进行学习,其中每一个基核完成从特定样本空间提取故障特征,通过多面故障特征的线性组合,将学习分类问题转化为相应的凸规划问题进行迭代求解。采用BPSO优化算法对MKSVC中的基核数及模型参数进行优化,实现了参数的自主选择。与常用诊断算法相比,BPSO-MKSVC具有更高的诊断精度;与PSO优化的SVM方法相比,其具有更低的参数敏感性和更好的鲁棒性。

Support vector machine（SVM） is sensitive to kernel function,kernel parameter and model parameters,and the multi-kernel learning can reduce its sensitivity to parameters.To diagnose faults occurred in power transformer,a classification method using multi-kernel support vector classifier（MKSVC） based on binary particle swarm optimization（BPSO） is proposed.The learning of MKSVC is performed by multi-kernel composed of linear combination of multi basis kernels,and each basis kernel extracts partial fault characteristic within a specific sample space;then through linear combination of various partial fault characteristics,the learning and classification problem is turned into corresponding convex programming problem to perform iteration solution.The autonomic parameter selection is implemented through the optimization of basis kernel parameters and model parameters of MKSVC by BPSO.Comparing with common diagnosis algorithms,the proposed BPSO-MKSVC method can provide higher accuracy of fault diagnosis;comparing with SVM based on particle swarm optimization,the proposed BPSO-MKSVC method is not so sensitive to parameters and possesses better robustness.