The time spent in different physical activities per day was found to be highly associated with many health factors, and thus, accurately measuring the time is critical. Currently many supervised learning methods provided high prediction accuracy for activity type, but their usage were limited to several key known activities, such as sitting still and sustained walking. Many less common or not-well-defined activities were ignored due to the difficulty of establishing reliable training data. We proposed an unsupervised learning method to extract a set dominating patterns of signal from the acceleration time series. We further investigated the interpretation of these patterns and established a relationship between them and some well-defined activities. Using this method, we avoided manually defining types or categories of activity and were still able to investigate the association between the time spent in each category and health factors.