It is difficult to extract the weak information of rolling bearing fault under complex and changeable operating environments. To address this issue, an optimized resonance sparse decomposition (RSD) based on maximizing the cyclic pulse index (CPI) is proposed, which takes full advantage of pulse characteristics and cycle period characteristics of fault impacts. Firstly, the variation coefficient of the short-time pulse peak moment is used as the CPI to comprehensively characterize the pulse and periodicity of the bearing fault signal. Then, the quality factors of RSD are optimized by the multi-scale simplified particle swarm optimization algorithm with the objective of maximizing the CPI. Finally, the cyclic pulse spectrum of the low-frequency resonance component is established to automatically identify the bearing faults. The results of simulation and applications in the fault diagnosis of EMU axle box bearings show that the proposed method effectively avoids the misjudgment of resonance frequency band caused by strong pulse interferences, and performs well in the synchronous diagnosis of bearing compound faults under complex working conditions, which demonstrates its engineering applicability in the field of bearing fault diagnosis.