关键词:

互信息, 特征选取, 评价准则, 故障诊断

Abstract:

Focusing on the problem that information-theoretical measure, namely, mutual information (MI) in a high dimensional space is difficult to be calculated for feature selection, an equation of estimation for the conditional mutual information between the candidate feature and output class given the subset of selected features was derived under the assumption that the information of features was distributed uniformly in the high dimensional space.When the information of features did not violate the uniform distribution severely, this equation could be used to rank features effectively, and based on the estimation equation, a feature selection filter algorithm was proposed on the basis of the second order mutual information of selected features.Unlike MI-based feature selection (MIFS) in which parameters relating to the redundancy of selected features were needed to be preset manually, this algorithm could estimate the redundancy of selected features adaptively. Hence the performance of the method was improved largely.The experimental results of Tennessee-Eastman process (TEP) showed that the proposed algorithm could provide more accurate and more effective feature evaluation measures, and have higher accuracy of fault diagnosis.

Key words:

互信息, 特征选取, 评价准则, 故障诊断