Learning affective state is determinate to online learning. Different affective states are associated with different online learning behaviors. Given the behavioral indicators of different affective states are still to be explored, this study constructed a data-driven online learning affective state detector by analyzing the learning log data of 269 students on an online learning platform. The accuracy and Cohen's consistency validated the reliability of this affective state detector. This detector gives us an estimated probability that the student is in a state of confusion, engagement, frustration, and distraction. This method was used to analyze students' affective states in an online course. It is found that active affective states such as concentration can promote students' learning effectiveness. But negative affective states such as frustration, confusion, and distraction are not conducive to the student's learning effectiveness. Hence, it is necessary to support students' learning for those in negative affective states. Specifically, the learning materials could be refined to provide students in confusion with more detailed tips. Timely support is helpful for those students in frustration. Timely reminders and interventions are useful for distracted students to engage in learning. This study proposes a reliable method for learning affective state analysis and provides teachers with practical suggestions to improve students' online learning.