With the growth of sensing technologies, the sensors are applied to diversified fields including health care, elderly protection, human activity abnormal detection, and surveillance. The advanced sensors embedded in mobile devices generate a large amount of valuable data. In recent years, to deal with the massive volumes of data, representation learning emerged as an alternative approach to extract the features without manual feature extraction. In this paper, we develop an unsupervised representation learning system for mining features across multiple sensors placed on different parts of the human body for recognizing human daily activities. The unsupervised representation learning approach allows models to learn the feature representation among a large number of unlabeled data samples collected from different parts of the human body. In order to demonstrate the feasibility of our system, extensive experiments on human daily activities recognition are carried out to evaluate the effectiveness of the learned representations.