Abstract: During field operations, combine harvesters operate in complex environments and endure alternating random loads, leading to fatigue failure in critical components such as the front and rear axles and chassis frame. This compromises the overall operational performance and harvesting efficiency of the combine harvester. Therefore, accurately analyzing the time-domain load characteristics of key components during field harvesting operations and deriving load spectra that reflect actual field working conditions are crucial for assessing the fatigue endurance and reliability of agricultural machinery like harvesters. Load spectra form the foundation for accurate fatigue life analysis. Considering the cost of compiling load spectra, it is often feasible only to measure load data over a limited period, followed by load extrapolation to obtain the full-life load spectrum. Currently, load extrapolation primarily falls into two categories: time-domain extrapolation and rainflow extrapolation. Rainflow extrapolation methods can be classified based on fitting approaches into parametric rainflow methods using single or mixed distributions, and nonparametric rainflow methods employing kernel density estimation and extremal theory. However, rain-flow extrapolation methods lose the load time history during the extrapolation process, limiting their application in load extrapolation. The key technology for time-domain extrapolation is using peak over threshold (POT) for extrapolation. Empirical methods suffer from significant variability in extrapolation results due to unreasonable threshold selection, while computational methods face challenges in obtaining accurate thresholds due to overly complex threshold determination. To address the inaccuracy of load extrapolation results when using the image method to select thresholds for POT models in time-domain extrapolation, this study proposes a threshold optimization method based on the mean excess function-least squares method (MEF-LSM) approach, grounded in empirical mode decomposition (EMD). Using measured strain signals from a tracked combine harvester as an example, this study analyzes time-domain load extrapolation methods and validates the proposed approach. This method initially employs EMD to decompose the non-stationary load signals generated during the combine harvester's field operations. By separating these signals into primary load and trend components, it effectively addresses the unsuitability of POT extrapolation for non-stationary time-domain load signals, ensuring the validity of the selected amplitude load for extrapolation. Furthermore, the MEF method is applied to the dominant load obtained from the decomposition to determine candidate threshold intervals. To obtain the optimal threshold, the relationship between the LSM feature parameters of different candidate thresholds and the validation metrics of the corresponding extrapolation results was investigated. It was found that when the feature parameters were minimal (the residual sum of squares was minimal), the overshoot threshold samples exhibited the best fitting performance, yielding the maximum validation metric R² for the extrapolation results. Therefore, this threshold was selected as the optimal threshold. Once again, based on the optimal threshold, the super-threshold load is extracted to obtain the super-threshold sample. The generalized pareto distribution (GPD) is used to fit the distribution of extreme points in the super-threshold sample. The shape parameter and scale parameter are calculated to obtain the GPD distribution function. The GPD distribution function is then used for simulation extrapolation to obtain the extrapolated main load. The main load and trend load are linearly combined to obtain the time-domain extrapolated load for the combine harvester. Finally, comparing the statistical characteristics of the original load and the extrapolated load reveals relative errors of 1.62%, 1.04%, and 3.89% for the standard deviation, root mean square, and kurtosis coefficient, respectively. Rainfall count analysis of both original and extrapolated loads revealed an amplitude correlation coefficient of 0.9953 and a mean correlation coefficient of 0.9936 between the two sets of loads. The results indicate that smaller characteristic parameters correspond to higher accuracy in threshold extrapolation. Specifically, the MEF-LSM method achieved a 1.23% improvement in the R² verification metric for threshold extrapolation compared to the mean excess function image method. This validates the effectiveness of the MEF-LSM method. The findings provide a reference for compiling load spectra for combine harvesters and offer a basis for accurately predicting the fatigue life and reliability analysis of agricultural machinery equipment.