In order to perform structural fatigue analyses of vehicles, it is necessary to extrapolate the time-domain signals of dynamic stresses measured over a short period of time into a longer history using extrapolation techniques. Taking some measurement points of a certain locomotive body as an example, a method to determine the optimal threshold based on the entropy weight method and the Topsis method is proposed. Firstly, after the initial time-domain signal is pre-processed by de-zero drifting and other pre-processing, the entropy weighting method is used to assign weights to the threshold evaluation indexes of the peak-over-threshold (POT) model; secondly, the Topsis method of multiple indexes is used to conduct a comprehensive evaluation, and the optimal threshold is obtained by comparing the relative proximity. The parameters of generalized Pareto distribution were fitted to the excess of peak and valley values, and the extreme value was reconstructed and replaced the original samples to finally obtain the extrapolated samples. From the P-P and Q-Q plots, it can be seen that the peak and valley samples are fitted better; from the cumulative distribution plot, it can be seen that the extreme value of the sample after extrapolation by a factor of 5 and 10 becomes larger, and the curve is close to the actual sample curve. After the damage calculation of the time-domain signal after extrapolation, it is found that the fatigue life assessment method is more biased towards safety.