The Application Of Feature Fusion Technology In The State Recognition Of High-Speed Train

In the state recognition of high-speed train, one single ki nd of feature generally cannot represent the running state of the train roundly for the complexity and diversity of the cause of the fault, resulting in wrong decision easily. It is usually used to extract much more features and fuse them to one fusion feature. The fusion feature can represent the running state of the train fully and exactly and increase the identification rate of the state. Problems of the state recognition of high-speed train based on feature fusion are studied. It mainly includes three feature fusion methods:fusion method based on feature selection, feature transformation and both feature selection and transformation namely serial fusion.In the aspects of feature fusion based on feature selection, feature is selected based on single-criterion and multi-criterion according to the number of the evaluation methods of features. ReliefF criterion, relation criterion and Inter-class separability criterion are used to evaluate the same kind of features or feature sets composed by different kinds of features. The test results show that single-criterion evaluation method can evaluate the same kind of features well but not as well as feature sets. Multi-criterion evaluation method gets better result in evaluating feature sets than single-criterion. Multi-criterion evaluation method based on linear combination has to make more than one tests to decide the weight of each single-criterion evaluation method to be used. As a result it has a lower efficiency. A new nonparametric multi-criterion evaluation method based on frequency is proposed to compare with the multi-criterion evaluation method based on linear combination. The test results show that nonparametric multi-criterion evaluation method based on frequency can give a high identification rate. The new method saves a lot of time and improves the efficiency because it does not need to set suitable parameters by making a lot of tests.In the aspects of feature fusion based on feature translation,2DPCA(2 dimensional principal component analysis), a technology of feature translation, is studied. Construction method of 2 dimensional feature matrix used in 2DPCA is analyzed in detail. It adds some zeros in every single feature vector to get a two-dimensional feature matrix in the traditional construction method of 2 dimensional feature matrix. And the traditional method gets good result only at the condition of small difference in dimension of each feature vector. A new construction method of 2 dimensional feature matrix based on SVD (Singular Value Decomposition) is proposed. The new method groups all feature vectors end to end as a new one-dimensional feature vector.Then the new one-dimensional feature vector is decomposed into a two-dimensional feature matrix based on the decomposition feature of SVD with the phase of the signal unchanged. Then 2DPCA is used to translate the feature matrix into a fusion feature by decreasing dimensions of the feature matrix. The fusion feature is used to classify and identity. The test results show that the new construction method of feature matrix proposed avoids the disadvantage of adding too many zeros after every feature vector and can get a higher identification rate at the conditions of both the difference in dimension of each feature vector is small or huge.On the basis of the studies above, feature selection and feature translation are used in turn to do serial processing with the detected data of the high-speed train to get fusion feature. The relationship between the choice of feature selection and feature translation and the identification rate of the serial fusion feature are studied. The experience results show that if the method of feature selection and feature translation are the same type in the process of serial feature fusion, the fusion feature will not get higher identification rate even though its dimensions of the feature matrix are decreased. Because the feature matrix is processed by the same type method twice, which lose some important information.