Globality-based uncorrelated linear extension of graph embedding for fault feature extraction

作者：Tang Diyin^*; Yu Jinsong; Chen Xiongzi; Wang Honglun

来源：Journal of Beijing University of Aeronautics and Astronautics, 2013, 39(3): 411-415.

• Fault feature extractions
• Feature extraction methods
• Global distribution
• Graph embeddings
• Information redundancies
• Linear extensions
• Locality-preserving
• Statistic correlations

Systematic approach to extract the most effective information from original features is of great importance and efficiency for fault detection where physical modeling is highly difficult and the original features are highly dimensional and nonlinear. An algorithm named globality-based uncorrelated linear extension of graph embedding for fault feature extraction was therefore proposed. Supervised learning was used to establish the relationship between original features, and the linear extension of graph embedding was adopted as the feature extraction framework. Great efforts were taken to combine the locality-preserving properties inside the classes and global distribution between different classes, in order to discover both the local and global structure of original features. Information redundancy was greatly reduced by eliminating the statistic correlation between extracted features. Experimental results on standard dataset demonstrate the superiority of this proposed algorithm to many classical feature extraction methods. Thus, a better efficiency in the convergence of training network and in the fault detection can be achieved.

• 出版日期2013
• 单位北京大学