<p>Due to the complex structure of helicopter transmission systems and the variability of operational environments, traditional numerical methods often struggle to achieve high predictive performance with multi-point, multi-modal vibration data. The vibration data from these systems contain essential temporal and spatial features that are critical to effective prediction. However, current methods for predicting physical structure responses rarely consider sensor placement as a graph structure or utilize the distances between sensors to establish node weight coefficients. To address this limitation, we propose a graph attention-former (GA-former) network to capture spatio-temporal graph data that aligns with actual mechanical systems. The spatial model features an interpretable node weight coefficient and utilizes a graph convolutional network (GCN) that adaptively extracts frequency information. Meanwhile, the temporal model employs an encoder block for time series prediction, which enhances computational efficiency. Compared to existing methods, our proposed approach demonstrates significantly improved performance in predicting the vibration responses of helicopter systems. Moreover, the effectiveness of each block in the proposed model is verified through ablation experiments. Hyperparameter analysis is performed to find the optimal parameter set. Vibration response prediction experiments with different signal-to-noise ratios are undertaken to analyze the noise tolerance of the proposed model. This study provides a new perspective for the intelligent prediction of the vibration response of helicopter transmission systems.</p>