Online collaborative learning (OCL) has become a common instructional strategy in higher education for developing students' skills in collaboration, problem-solving, and critical thinking. Cognitive engagement in OCL evolves dynamically, but we do not yet fully understand which patterns of cognitive engagement are conducive to OCL and when to promote them. This study used entropy analysis, sequential pattern mining, and temporal network analysis to examine the online discourse of 44 college students who participated in three OCL tasks. Results showed that, compared with the low-performance groups, the high-performance groups exhibited patterns of continuous "perspective elaboration" and "low-level regulation," as well as frequent shifts from "perspective expression" to "perspective elaboration." In addition, there were differences in the longitudinal evolution patterns of cognitive engagement between the high- and low- performance groups. These findings have important implications for learning tool design and improving collaborative learning design.