Feature extraction based on supervised kernel locality preserving projection algorithm and application to fault diagnosis

Fault redundancy information can increase computation complexity and reduce the precision of fault diagnosis. Feature extraction becomes very important to improve the performance of fault diagnosis. A supervised kernel learning algorithm based on manifold is presented to carry out feature extraction. The proposed algorithm firstly implements locality preserving projection in reproducing kernel Hilbert space. Using the quotient of between-class scatter matrix dividing within-class scatter matrix as discriminant criterion, it constructs feature space by selecting discriminant vector that reflects difference among classes. Discriminant vector that mainly reflects difference within classes is discarded. The proposed method is applied to fault diagnosis of switch open-circuit fault in brushless dc motor power converter, using proximal support vector machine classifier. Experimental result shows that the proposed algorithm has high diagnosis accuracy.