In this paper we propose a method to estimate working activities from sensor data without user annotation. Estimation of working activities are much more difficult than that of simple activities such as 'walking' and 'still' in terms of their complexity and length. To address this issue, we assume that they are a probabilistic combination of various simple activities and propose a method to discover working activities by multistage estimation, in which firstly classify sensor data into some basic activities and then estimate them by using topic model.We focused on nursing service as one of professional working activities, in whicn visualization by estimation of working activities is critical, conducted the experiment to observe it in the hospital and verified our method.