To address the rollover risk of crawler-type sugarcane harvesters operating on hilly terrain-caused by a high center of gravity and narrow track width—this study establishes a real-time vibration measurement and prediction framework. Vibration acceleration signals of the harvester frame are collected during tilt tests on a dedicated experimental platform. The signals are processed in the frequency domain to extract key vibration features that characterize different inclination states. A hybrid BiLSTM-Transformer model is proposed to predict potential rollover conditions.In the proposed method, vibration acceleration data are first preprocessed and decomposed using empirical mode decomposition (EMD) to obtain denoised and reconstructed time-frequency components. The BiLSTM network effectively captures long-term temporal dependencies in the vibration sequences, while the Transformer module focuses on extracting local temporal and attention-based contextual features. The complementary strengths of these two networks enhance both learning efficiency and predictive stability.Experimental results demonstrate that the proposed hybrid model achieves a prediction accuracy of 95.39% with an average response time of 11.87ms, meeting real-time monitoring requirements. To further validate model effectiveness, t-SNE dimensionality reduction visualization and confusion matrix analysis are performed, confirming the model’s discriminative capability across different tilt states.This research provides a reliable theoretical and technical foundation for the development of real-time rollover warning and automatic leveling control systems for crawler-type sugarcane harvesters in complex hilly environments.