Gearbox incipient fault diagnosis using feature sample selection and principal component analysis

Abstract

Monitoring and diagnostic systems have played an important role in modern industry. Many intelligent or signal processing methods have been successfully applied in the manufacturing process. Inspired by the research of kernel function approximation (KFA), a novel kernel principal component analysis (KPCA) method is proposed in this paper and applied in gearbox incipient fault detection. The proposed KPCA is realised by feature sample selection and principal component analysis, which can also be called FSS-PCA. The key issues studied in this paper are non-linear feature extraction, optimal feature sample selection and diagnostic performance assessment. Firstly, the integral operator Gaussian kernel function is used to realise the non-linear map from the raw input space of gearbox vibration features to a high dimensional space, where appropriate feature samples are selected to construct the feature subspace. Then PCA is used to classify two kinds of gearbox running conditions: normal and tooth crack. The quantity of selected samples is much fewer than that of whole sample sets, which has quickly expedited the computation process. Experiment results indicate the effectiveness of FSS-PCA for gearbox incipient fault diagnosis.