Fault level diagnosis for planetary gearboxes using hybrid kernel feature selection and kernel Fisher discriminant analysis

This paper introduces a hybrid dimension reduction method that combines kernel feature selection and kernel Fisher discriminant analysis (KFDA). In the first stage, a kernel feature selection method is proposed to remove redundant and irrelevant features for two purposes: (1) reducing computation burden of the entire fault diagnosis system and (2) alleviating the impact of irrelevant features on KFDA. In the second stage, KFDA is used to establish a more compact feature subset by extracting a smaller number of features. We use Gaussian radial basis function as the kernel function for the two kernel stages in the proposed method. A parameter selection method for this kernel is proposed to select the optimal values for the proposed method. Experimental results on fault level diagnosis demonstrate that the proposed hybrid dimension reduction method has advantages over other approaches that use feature selection or KFDA separately.