To analyze fuzzy data in uncertain evidential reasoning, some researchers have recently extended the Dempster-Shafer (D-S) Evidence Theory to fuzzy sets. But there are some insufficiencies in the definition of the fuzzy belief function and the combination rule on fuzzy sets of the D-S evidence theory. This paper describes a new definition of the similarity degree between two fuzzy sets and presents an improved extension combination rule of the evidence theory on fuzzy sets. It also presents the corresponding mathematical proof to validate the improved combination rule. To research the application of this method to coal mine gas outburst prediction, we add olfaction information as a real-time and dynamic data source with a self- navigating robot. There is variance or conflict between information from different kinds of sensors which serves as the evidence of the D-S fusion. So firstly, before the decision-making with the D-S Evidence Theory, we use a neural network to pretreat and extract the nonlinear information. Secondly, we adopt the AND-algorithm to combine the coherence evidence, and reallocate the conflicts to various focal elements according to the credibility of the coherence evidence. This improved combination rules for the D-S Evidence Theory for resolving the problem of evidence conflicts was proved to be effective. Compared with other generalizing combination rules, the results of the numerical and practical experiments show that the new combination rule in this paper can acquire more changing information about the change of fuzzy focal elements more effectively, and it overcomes the insufficiencies of other existing combination rules and effectively enhances the robustness of fusion decision systems.