Abstract: For complicated engineering systems with multiple health indicators, multiple operation and fault modes, a systemlevel failure prediction method is presented based on the synthesized health index (SHI) and the relevance vector machine (RVM). In the offline training phase, the health assessment models for each operation mode are firstly developed using historical data, which then will be utilized to calculate the corresponding SHI sequences for each degradation path. Moreover, the model that has the best fit to the historical SHI sequences is selected with the help of RVM regression. In the online prediction phase, the parameters of the selected model are estimated and updated using the online SHI sequences and the RVM, then timevarying noises are also added to the selected model to represent the uncertainty. Further, the probability density distribution of system remaining useful life is obtained by the model extrapolating in time. The method is successfully applied to the failure prediction of turbine engines.