To better understand the self-regulated learning process in online learning environments, this research applied a data mining method, the two-layer hidden Markov model (TL-HMM), to explore the patterns of learning activities. We analyzed 25,818 entries of behavior log data from an intelligent tutoring system. Results indicated that students with different learning outcomes demonstrated distinct learning patterns. Students who failed a problem set exhibited more passive learning behaviors and could hardly learn from practice, while students who mastered a problem set could effectively regulate their learning. Furthermore, we extended the use of TL-HMM to predicting learning outcome from behavior sequences and checked through cross-validation. TL-HMM is demonstrated helpful to gain insight into learners' interactions with online learning environments. In practice, TL-HMM could be embedded in intelligent tutoring systems to monitor learning behaviors and learner status, so as to detect the difficulties of learners and facilitate learning.