Capabilities that prognostics and system health management (PHM) system detect fault, isolate fault and forecast fault directly determine the effectiveness of the maintenance work. With the development of sensor technology and signal processing methods, in order to precisely detect and identify faults, fault diagnosis is a typical multi-sensor fusion problem. New challenges have arisen with regard to obtaining a reliable fault diagnostic result based on multi-source information. From the viewpoint of evidence theory, information obtained from each sensor can be considered as a piece of evidence, and as such, multi-sensor based machines diagnosis can be viewed as a problem of evidence fusion. In this paper we investigate the use of Dempster-Shafer (D-S) evidence theory as a tool for modeling and fusing multi-source information from the machines. We present a preliminary review of evidence theory and explain how the multi-sensor machine diagnosis problem can be framed in the context of this theory. We propose a method for enhancing the effectiveness of basic probability assignment functions in modeling and the method combining pieces of evidence. By introducing importance index, the issues of evidence importance in the practical application of D-S evidence theory are addressed. Finally, we report a case study to demonstrate the efficacy of our method.