Noise Reduction Method for Nonlinear Time Series Based on Principal Manifold Learning and Its Application to Fault Diagnosis
JH Yang,JW Xu,DB Yang,M Li
DOI: https://doi.org/10.3901/jme.2006.08.154
2005-01-01
Abstract:A new noise reduction method for nonlinear time series based on principal manifold learning is proposed. After embedding the time series into a high reconstructed phase space, the principal manifold of the dynamical system attractors, which is represented in a low dimensional, global and orthogonal coordinate system, is obtained by manifold learning method such as Local Tangent Space Alignment algorithm. The final denoised result in the form of one dimensional time series is achieved after averaging of the data in phase space which are regenerated according to the principal manifold. Being different from the current noise reduction methods for nonlinear time series, for example, local projective method, the method based on principal manifold learning emphasizes more on the holistic structure of nonlinear deterministic characters hidden in the time series. Experimental results show that the proposed method has a better performance of removing Gaussian white noise in chaotic time series, compared with local projective noise reduction method. Also the method was applied to the analysis of a vibration signal sampled from a defective gear box with a broken tooth. The impact features, which are difficult to be identified from noisy data directly, can be extracted successfully after noise reduction.