Time-frequency manifold for gear fault signature analysis

Abstract

Time-frequency analysis can reveal intrinsic feature of representing non-stationary signal for machine health diagnosis. This paper proposes a novel time-frequency feature, called time-frequency manifold, by addressing manifold learning on the time-frequency distributions (TFDs). The new feature is produced from an analyzed signal in three steps. First, a high-dimensional phase space is reconstructed as a preparation for manifold analysis. Second, the TFDs are calculated to represent the non-stationary information in the reconstructed space. Third, the manifold learning is conducted on the TFDs to produce the nonlinear manifold structure. The time-frequency manifold combines non-stationary information and nonlinear information, and may thus provide better representation of machine health pattern. The new feature is exactly suited for machine health diagnosis, which is verified by an application to gear fault signature analysis.