Improved kernel fisher discriminant analysis for fault diagnosis

This paper improves kernel fisher discriminant analysis (KFDA) for fault diagnosis from three aspects. Firstly, a feature vector selection (FVS) scheme based on a geometrical consideration is given to reduce the computational complexity of KFDA when the number of samples becomes large. Secondly, a new kernel function, called the Cosine kernel, is proposed to increase the discriminating capability of the original polynomial kernel function. Thirdly, nearest feature line (NFL) classifier is employed to further enhance the fault diagnosis performance when the sample number is very small. Experimental results show the effectiveness of our methods.