混合聚类新算法及其在故障诊断中的应用  

NOVEL HYBRID CLUSTERING ALGORITHM AND ITS APPLICATION TO FAULT DIAGNOSIS

作　　者：雷亚国[1] 何正嘉[1] 訾艳阳[1] 胡桥[1] 丁锋[1] 

机构地区：[1]西安交通大学机械工程学院

出　　处：《机械工程学报》2006年第12期116-121,共6页Journal of Mechanical Engineering

基　　金：国家自然科学基金重点(50335030);国家自然科学基金(50575171;50505033);国家重点基础研究发展计划(2005CB724106);西安市科技攻关计划(GG050410)资助项目。

摘　　要：针对模糊C-均值(FCM)聚类算法假设各维特征和每个样本对聚类贡献相同,同时需要预先设定聚类数的不足,利用3层前馈神经网络、点密度函数算法和聚类有效性指标对其进行改进,提出一种新的混合聚类算法。该算法考虑到不同特征和不同样本对聚类结果有不同程度的影响,并根据聚类有效性指标的变化自适应确定聚类数来实现聚类。利用基于梯度下降的3层前馈神经网络通过无监督训练来自适应学习特征权值,使用基于点密度函数的算法获取样本权值,给不同特征和不同样本赋予权重,突出敏感特征和典型样本的主导作用,抑制其他特征和样本对聚类的干扰,以提高聚类性能。研究结果表明,对于国际标准测试数据和某机车轴承的早期故障诊断,该混合聚类算法不但能自动确定聚类数,而且聚类的准确性明显比FCM高。Aiming at the fuzzy C-means （FCM） clustering algorithm supposing the uniform influence to clustering by different features and samples, and setting the cluster number beforehand, a novel hybrid clustering algorithm based on 3 layer forward neural networks（FNN）, an algorithm of distribution density function of data point and the cluster validity index is proposed. Feature weighting and sample weighting are considered in this hybrid clustering algorithm and the cluster number is automatically set by using the cluster validity index to finish clustering. Feature weights are adaptively learned via FNN with the gradient descent technique under the unsupervised mode of training,and sample weights are computed through the algorithm of distribution density function of data point. Then, the feature weights and the sample weights are assigned to the correspond- ing features and samples to emphasize the leading effect of sensitive features and typical samples, and weaken the interference of other features and samples. The proposed algorithm is employed to analyze the benchmark data and the practical data from locomotive hearings, and the results show that the algorithm enables to automatically and correctly set cluster number and its clustering performance is better than that of the FCM.

关 键 词：样本权值 特征权值 聚类有效性指标 混合聚类 故障诊断 

分 类 号：TH17[机械工程—机械制造及自动化] TP18[自动化与计算机技术—控制理论与控制工程]