A non-parametric complementary ensemble empirical mode decomposition (NPCEEMD) is proposed for identifying bearing defects using weak features. NPCEEMD is non-parametric because, unlike existing decomposition methods such as ensemble empirical mode decomposition, it does not require defining the ideal SNR of noise and the number of ensembles, every time while processing the signals. The simulation results show that mode mixing in NPCEEMD is less than the existing decomposition methods. After conducting in-depth simulation analysis, the proposed method is applied to experimental data. The proposed NPCEEMD method works in following steps. First raw signal is obtained. Second, the obtained signal is decomposed. Then, the mutual information (MI) of the raw signal with NPCEEMD-generated IMFs is computed. Further IMFs with MI above 0.1 are selected and combined to form a resulting signal. Finally, envelope spectrum of resulting signal is computed to confirm the presence of defect.