Lack of sufficient flexibility to deal with actual noisy data in data fusion methods is the main concern in this paper. This deficiency comes from two major reasons. Firstly, in fusion methods, all of the collected data are considered useful. Secondly, often some presumed sensor models are used in the fusion process, which do not necessarily match to the true models. As a general solution to address the aforementioned shortcomings, a novel Adaptive Data Fusion Structure (ADFS) is proposed. In ADFS, incorrect data are eliminated; then the remainders are fused, and finally the sensor models are learned by using the final fusion results as internal feedbacks. In particular, as more misleading data are eliminated, more accurate sensor models emerge. By employing ADFS to localize a simulated mobile robot in a highly noisy environment, the results prove its superiority, especially compared to the conventional entropy based methods of localization.