Feature extraction method of mechanical impulse based on nonlinear manifold learning

Feature extraction is of great importance in condition monitoring of machinery. Manifold learning theories brought a new idea for recognizing and predicting the underlying nonlinear behavior. In this paper, in order to extract the key feature of mechanical signal, a principle manifold feature extraction method based on the local tangent space alignment is proposed. Integrated with reconstruction of phase space, the method can extract the manifold feature which provides a more truthful low dimensional representation. During the searching of embedding manifold, integrated with the advantage of the kurtosis and skewness indexes, the selection of local neighborhood parameters is introduced to evaluate the feature. The industrial measurements show that this approach, compared with the wavelet soft-threshold and the orthogonal matching pursuit methods, is more effective to extract the weak periodic features from mechanical signals.