A new acoustic emission (AE) signal processing and feature extraction approach to bearing fault diagnosis is presented in this paper. The presented approach uses time-frequency manifold analysis to extract time-frequency manifold feature (TFMFs) from AE signals. It reconstructs a manifold by embedding AE signals into a highdimensional phase space. The tangent direction of the neighborhood for each point is then used to approximate its local geometry. The variation of the manifolds representing different condition states of the bearing can be revealed by performing multi-way principal component analysis. The AE signals acquired from a bearing test rig are used to validate the presented approach. The test results have shown that the presented approach can interpret different bearing conditions and is effective for bearing fault diagnosis.