Context mining algorithms from sensor data have been researched and successful results have been shown. However, since these existing works are focused on improving the accuracy of context mining, they are established on the assumption that they can acquire a complete set of necessary data. Therefore, the context mining algorithms do not work sufficiently since the data drops easily in the reality. In this paper, to cope with this problem, we propose a middleware named UDS (Uninterruptible Data Supply System). The system compensates the missing data, creates virtually complete dataset and provides upper layer applications. Applications operating over UDS can work sufficiently with some data actually missing. We have defined two types of characteristic data deficit patterns and created a robust model for both patterns utilizing Bayesian Network. In the evaluation, we show UDS can sustain the quality of context over 80% with 40% data missing.