采用多目标进化模型的无监督故障特征选择算法

高维故障特征数据易影响诊断的处理速度和识别率,而传统单目标特征选择算法易融入主观偏好,从而影响特征选择的质量。为此,提出一种无监督的多目标进化特征选择算法。采用熵度量作为相关度目标,采用相关系数的概念设计了冗余度目标,算法同时将这两个目标作为优化对象;利用样本在各个特征上的分布信息,设计了导向性的种群初始化过程和变异算子,以提高算法的优化能力;还利用集成的方法得到了所有特征的重要度序列。对5组UCI数据和3组往复式压缩机故障数据的测试结果表明,该算法比已有的几种特征选择算法更具优势。

Unsupervised feature selection algorithm utilizing multi-objective evolutionary model for fault diagnosis

Feature selection is necessary for high-dimensional features in fault diagnosis since it can improve the efficiency and accuracy. However, traditional feature selection algorithm always has a strong bias towards a single criterion, which is harmful to the quality of feature selection. An unsupervised feature selection algorithm based on multi-objective evolutionary model was proposed to solve this problem. A relevance objective based on entropy measure and a redundancy objective based on correlation coefficient were simultaneously optimized. Both initialization process and mutation operator were also designed by utilizing the distribution of samples in each feature. Besides, an ensemble method was proposed to obtain the importance order. Experiments on five UCI and three valve fault of reciprocating compressor datasets demonstrated better performance of the proposed algorithm.