Recently, interests in learning action models have been increasing. Although non-deterministic planning has been developed for several decades, most previous studies in the field of action model learning still focus on classical and deterministic action models. This paper presents an algorithm for identifying non-deterministic actions, including effects and preconditions, in partially observable domains. It can be applied when people know nothing about a transferring system and only the action-observation sequences are given. Such scenarios are common in real-world applications. This work focuses on problems in which actions are composed of simple logical structures and features are observed under some frequency. The learning process is divided into three steps: First, compute the probability of each proposition which holds in a state. Second, extract effect schema from propositions and then extract preconditions. Third, cluster effect schema to remove redundancy. Experimental results on benchmark domains show that action model learning is still useful in non-deterministic and partial observable environments.