面向高维特征故障数据的进化软子空间聚类算法  

Evolutionary Soft Subspace Clustering Method for Fault Diagnosis of High-Dimensional Features

作　　者：夏虎 庄健[1] 于德弘[1] 

机构地区：[1]西安交通大学机械工程学院,西安710049

出　　处：《西安交通大学学报》2013年第5期115-120,共6页Journal of Xi'an Jiaotong University

基　　金：中央高校基本科研业务费专项资金资助项目(XJJ20100142);广东省战略性新兴产业核心技术攻关项目(2011A091101004);陕西省科技厅工业攻关项目(2011K09-45)

摘　　要：针对复杂机械设备故障诊断中特征量众多且对各种故障敏感程度不同的现象,提出了采用软子空间聚类算法来实现故障的识别方法。同时,针对传统软子空间聚类易陷入局部最优,目标函数设计受限制的缺点,又提出了采用进化计算实现聚类的方法。利用同类样本在相关特征维上方差小的假设,新的目标函数能更好地评价聚类结果的质量。在该算法中,通过设计类中心和权重值的混合编码以及聚类导向搜索算子,使算法更适于聚类问题的优化,而且设计的修复算子可有效地去除不合理的聚类结果。采用5组UCI数据集、2组轴承滚珠故障数据集和3组往复式压缩机气阀故障数据集对算法进行了测试,结果表明:该算法明显好于几种的软子空间聚类算法,在Rand指标上最多可高出0.226 6,并且对2组不同工况下一级缸气阀故障可实现100%的故障识别。Many features can be collected in a complex machine, however, using all these features for fault diagnosis is inappropriate due to different sensitivity of each feature. A soft subspace clustering schedule is proposed. A novel evolutionary algorithm is suggested to avoid the shortages in the traditional soft subspace clustering, such as local optimum and constraint in designing objective function. Under the hypothesis that the samples in the same cluster have small variance in relevant dimensions, a new objective function is constructed in the proposed algorithm to optimize. Considering the special meaning of each gene, a new encoding frame and a cluster-oriented search operator are designed. A repair operator enables to eliminate some meaningless individuals in the evolution process. The experiments on five UCI, two bearing fault and three valve fault of reciprocating compressor datasets demonstrate the better performance of the new algorithm. It is at most 0. 226 6 higher than other competitive soft subspace clustering methods on Rand index （RI）, especially for two of the valves, fault diagnosis rate gets 100%

关 键 词：故障诊断 软子空间聚类 进化算法 相关特征维 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]