Remaining useful life prediction of metal oxide semiconductor field effect transistor ( MOSFET ) can prevent gradual degradation or lose efficacy of devices due to long-term conduction. However, the traditional prediction models are difficult to extract the detailed characteristics of nonlinear changes in MOSFETs degradation parameters, resulting in poor prediction accuracy. To address this issue, a remaining useful life prediction method for MOSFETs is proposed, which is based on variational mode decomposition and nonlinear auto-regressive model with exogenous inputs (NARX) neural networks with external inputs. Firstly, the degenerate parameter sequence is decomposed into multiple sets of characteristic components containing nonlinear change information using the VMD method. Secondly, the NARX prediction model is optimized by using Bayesian regularization and Levenberg-Marquardt algorithms, respectively. Finally, integrating multiple sets of feature component prediction values to obtain the remaining life prediction results of MOSFETs. The experimental results show that the root mean square error of the proposed method is less than 0. 003, the mean absolute percentage error is less than 5% , all of which are better than the comparison method. The average error of remaining useful life prediction is less than 5 min, which evaluates the effectiveness of the method.