Evolving kernel principal component analysis for fault diagnosis

Abstract

Feature extraction is the core of a fault diagnosis system. This paper presents a novel approach, called evolving kernel principal component analysis (EKPCA), to transform the original features to a more effective nonlinear combination in fault classification. EKPCA is based on the integration of kernel principal component analysis (KPCA) and an improved evolutionary optimization algorithm. As a coordinate transformation technique, KPCA is a superset of principal component analysis (PCA), which is utilized to project the original data space to a nonlinear feature space via the appropriate kernel function, and then PCA is performed in the projected feature space. Compared with PCA, KPCA is more flexible in extracting a group of new nonlinear features. However, the efficiency of KPCA in real-world applications depends mainly on the kernel function chosen a priori. It remains an issue of how to select the kernel function from the viewpoint of optimization. This paper addresses this issue using the techniques from evolutionary computation (EC). An improved evolutionary algorithm incorporated with a Gaussian mutation operator that is inspired from evolutionary strategies (ES) and evolutionary programming (EP) can enhance both the global and the local search performances without substantially increasing the computational effort. The application in fault diagnosis to a large-scale rotating machine shows that EKPCA is effective and efficient in discovering the optimal nonlinear features corresponding to real-world operational data. Thus, this method can improve the recognition power of a fault diagnosis system.