For complex engineering systems, such as trains, planes, and offshore oil platforms, load spectra are cornerstone of their safety designs and fault diagnoses. We demonstrate in this study that well-orchestrated machine learning modeling, in combination with limited experimental data, can effectively reproduce the high-fidelity, history-dependent load spectra in critical sites of complex engineering systems, such as high-speed trains. To meet the need for in-service monitoring, we propose a segmentation and randomization strategy for long-duration historical data processing to improve the accuracy of our data-driven model for long-term load-time history prediction. Results showed the existence of an optimal length of subsequence, which is associated with the characteristic dissipation time of the dynamic system. Moreover, the data-driven model exhibits an excellent generalization capability to accurately predict the load spectra for different levels of passenger-dedicated lines. In brief, we pave the way, from data preprocessing, hyperparameter selection, to learning strategy, on how to capture the nonlinear responses of such a dynamic system, which may then provide a unifying framework that could enable the synergy of computation and in-field experiments to save orders of magnitude of expenses for the load spectrum monitoring of complex engineering structures in service and prevent catastrophic fatigue and fracture in those solids.