Abstract: A probabilistic fatigue life prediction method based on an improved fatigue cumulative damage model is proposed to achieve an accurate prediction of probabilistic fatigue life under multi-level variable amplitude loading. The generalized polynomial chaos theory for probabilistic fatigue life prediction is introduced. An improved fatigue cumulative damage model with nonlinear characterization terms is constructed based on the intrinsic damage dissipation theory of damage mechanics. Fatigue test data of various commonly used engineering materials under multi-level variable amplitude loading, covering welded joints and smooth specimens, are selected for comparative verification with five classical fatigue cumulative damage models. The S-N curve parameters are treated as random variables, the fatigue life under each stress level is expanded via the generalized polynomial chaos theory, and the improved fatigue cumulative damage model, as well as five classical fatigue cumulative damage models, are introduced to conduct probabilistic fatigue life prediction under multi-level variable amplitude loading, with the goodness of fit quantified. The results show that the improved fatigue cumulative damage model can effectively adapt to fatigue life prediction for different materials, stress loading levels and structural types, and can be applied to actual engineering structures of bogie frames. The prediction range of the probabilistic fatigue life prediction model established based on this model is highly consistent with the actual fatigue test results, providing a reliable solution for probabilistic fatigue life prediction under multi-level variable amplitude loading.