Feature Reduction Integrated with Distance Measure and SVD and Its Application in Mechanic Fault Diagnosis

Feature reduction is a key step of pattern recognition. In this paper, a new feature reduction method is presented integrated with distance measure and singular value decomposition (SVD). Firstly, feature sensitivity is put forward to distinguish the different classes, namely the different mechanic faults and defined as the ratio of within-class distance to between-class distance. Secondly, sensitivities of different features are calculated according to this definition and then selected by SVD. Thirdly, the selected features are put into the back propagation (BP) nerve network. With 144 samples as training set and 96 samples as validation set, the BP nerve network is trained fairly well. Finally, the trained BP nerve network is used to diagnose 12 diesels, of which 11 diesels are diagnosed correctly. The diagnostic correct rate is 91.67%, which shows this feature reduction method is satisfactory.