While a variety of existing signal processing approaches commonly used for rolling bearing fault diagnosis lack of adaptability, the overcomplete atomic decomposition has a flexible adaptive ability. With the application of CP(Cosine packet) and WP(Wavelet packet) fast algorithms in matching pursuit, the algorithm of CPWP merged atomic decomposition was proposed. The simulated signal of rolling bearing fault is analyzed by using CP, WP and CPWP atomic decomposition, respectively. The results illustrate that CPWP merged dictionary decomposition has higher revolution factor and provides clearer and more comprehensive characters of amplitude modulation signal of impact than CP or WP atomic decomposition. Moreover, the real signal of outer race fault of rolling bearing is analyzed by using the aforementioned techniques. It is validated that the combination of different atomic dictionaries, which is sensitive to different features of signal, will enhance the ability of adaptive signal recognition. The CPWP merged dictionary analysis can obtain more information of impact and modulation than either of CP and WP dictionary decompositions, and has better performance in extracting rolling bearing fault features.