Abstract: The vibration signals of planetary gearboxes are composed of complex frequency components and interfering noises, and their spectra have intricate sidebands, which cause difficulty in and even misleading fault identification. In different fault cases, the vibration signatures in multiple domains typically differ from normal states with different discrepancies. Based on this hypothesis, time and frequency domain features are extracted for the purposes of fault identification. The vibration signal is adaptively decomposed into a set of mono-components, and the instantaneous energy of each mono-component is calculated in time-frequency domain by exploiting the merits of local mean decomposition, including its better robustness to noise and freedom from pseudo-mode and negative frequency problems. Manifold learning is utilized to tackle the high-dimensionality and non-linearity aspects of multiple-domain feature space construction. A new method is proposed for estimating the intrinsic dimension and selecting the k-nearest neighborhood based on the improved pseudo-nearest neighbor. In addition, isometric feature mapping (ISOMAP) is utilized to reduce the dimensions of the multiple-domain feature space. The proposed method is validated by analyzing the planetary gearbox lab experimental dataset. Based on the clustering analysis results of the extracted manifold features, the localized faults on the sun, planet, and ring gears are successfully identified.