With a focus on the problem of the typical evidence weight calculating method that it cannot define the weight and complete the decision fusion in the high conflict evidence D-S fusion problem where only a few sensors make the judgments correctly while the most incorrectly, the paper proposes an evidence weight calculating method based on fault sensitivity. Firstly, kernel principal component analysis (KPCA) is used to obtain the nonlinear sensitive feature; then, calculating the sensitivity based on such feature and obtaining sensor decision weight based the fault detection sensitivity; applying the weight obtained above and that obtained from equal weight method and decision-making distance based method to the fusion diagnosis of the simulated fault in the rotor where three sensor are installed. The results show that: the weight obtained through the method proposed in the paper can reflect the sensitivity of different sensors when detecting the faults. High weight is given to the sensors which contain much fault information and sensitive to fault and low weight will be given to the sensors of the opposite kind. In the method proposed, evidence weights play the role as a “regulator”, which makes it possible to obtain better decision fusion result whether in the cases where only a few sensors manage to find the fault and give the correct diagnosis or in the cases where few or no conflict exists.