Time-frequency analysis can reveal an intrinsic signature for representing nonstationary signals for machine health diagnosis. This paper proposes a novel time-frequency signature, called time-frequency manifold (TFM), by addressing manifold learning on generated time-frequency distributions (TFDs). The TFM is produced in three steps. First, the phase space reconstruction (PSR) is employed to reconstruct the inherent dynamic manifold embedded in an analyzed signal. Second, the TFDs are calculated to represent the nonstationary information in the phase space. Third, manifold learning is conducted on the TFDs to discover the intrinsic time-frequency structure of the manifold. The TFM combines nonstationary information and nonlinear information and may thus provide a better representation of machine health pattern. By evaluating the characteristics of top two TFMs, a synthetic TFM signature is further proposed to improve the time-frequency structure. The effectiveness of the TFM signature is verified by means of simulation studies and applications to diagnosis of gear fault and bearing defects. Results indicate the excellent merits of the new signature in noise suppression and resolution enhancement for machine fault signature analysis and health diagnosis.