Improvement of recognition rate is ultimate aim for fault diagnosis researchers using pattern recognition techniques. However, unique recognition method can only supports a limited classification capability, which is insufficient for real-world application. An ongoing strategy is the decision fusion techniques. In order to avoid the shortage of single information source coupled with unique decision method, the new approach is required to generate better results. This paper proposes a decision fusion system for fault diagnosis, which integrates data sources of different types of sensors and decisions of multiple classifiers. First, non-commensurate sensors data sets are combined using an improved sensor fusion method at decision-level by using relativity theory. The generated labels vectors are then selected based on correlation measure of classifiers in order to find an optimal sequence of classifiers fusion, which can lead to the best fusion performance. Finally, multi-agent classifiers fusion algorithm is employed as the core of the whole fault diagnosis system. Also different fusion methods at decision-level are compared. The efficiency of the proposed system was demonstrated through two experiments. The first one is fault diagnosis of induction motors using test rig designed by our intelligent mechanics lab. In the second experiment, faults data of elevator motor received from Korea Elevator Safety Center (KESC) were employed and regarded as a particular diagnosis example. The results of the two experiments show that the proposed system can take super performance when compared with the best individual classifier with single source data.