Multi-Criterion Feature Set Optimal Selection Method Based On D-S Evidence Theory

Feature set optimal selection is an important research content in pattern recognition. We may usually extract a lot of features, however not all features are conducive to classification accuracy and some features may evenly have bad effect. So it is an important and difficult point to identify which features are more effective and which features are not informative. Researchers at home and aboard proposed many feature set optimal selection methods. Firstly, special ranking methods are used to sort the features. Then we can obtain the optimal feature set according to the feature ranking. Therefore, the ranking methods directly affect the selection of the merits of the feature set. The ranking methods can be categorized into two classes:single ranking method and multi-criterion ranking method (multi-criterion ranking method is refer to the synthesize ranking method which integrates several single ranking criterions in a specific way). It is also known that the single feature ranking criterion cannot reflect characteristics of features completely. Therefore, how to fuse a collection of different ranking criteria has been another important topic which needs t o be addressed. We can select the more effective feature subsets according to the feature ranking vectors obtained by different ranking methods.This work firstly regards the measured data of common faults of high-speed train running gear as research object. Considering the characteristics of high-speed train fault data, a multi-criterion feature ranking and selection method is proposed based on Fisher’s ratio and fuzzy entropy. The proposed approach, integrating Fisher’s ratio and fuzzy entropy in parallel, can make a comprehensive evaluation of features and obtain more effective feature subsets. In the experimental section, with two sets of standard datasets to do the test first, the method is then applied to fault datasets of high-speed train. The results show the effectiveness of the proposed approach.How to combine different single evaluation criteria is the most important issue of multi-criterion feature ranking algorithm. Given a feature data, using different ranking criteria will yield different ranking orders. This implies that there are conflicts among different single-ranking algorithms, which might produce counter-intuitive results when fusing different ranking orders. The D-S based evidence conflict theory can effectively solve the uncertain and conflict existing in fusion, and eliminate combination paradox. This work presents a novel multi-criterion feature-combination selection algorithm based on D-S theory and evidence conflict theory (MCFR-DSEC), combining different criteria to improve both classification performance and stability. Compared with the existing algorithm, MCFR-DSEC shows obvious advantages. But the fusion rule of MCFR-DSEC is complex in calculation and has many parameters which affected the computing speed seriously. In response to this shortcoming, the improved D-S theory proposed by Murphy is chose as the fusion rule-MCFR-MURPHY. Compared with MCFR-DSEC, the improved algorithm is simpler in calculating process and is more effective. In addition to classification accuracy, the stability is also a most important issue in feature selection. If a feature ranking algorithm lacks robustness, it might produce unrepeatable results even only a few samples are added to or deleted from the training data set. The results show that MCFR-DSEC has the best stability. In brief, the D-S theory can not only combine different single evaluation criteria, but also has excellent stability.Finally, the multi-criteria feature ranking method MCFR-MURPHY is applied to fault diagnosis of high-speed train. MCFR-MURPHY is used to evaluate the features and get the classification accuracy of different feature space. Compared with Borda Count and single evaluation method, the new method can make evaluation more effectively based on classification accuracy. What’s more, this method is more simple, well-adapted and with good stability.This work was supported by National Natural Science Foundation of China. (No.61134002)...