Straddle-type monorail train gearboxes operate in complex environments, and the vibration signals exhibit significant non-stationarity, nonlinearity, and strong noise characteristics. Due to the real operation environment, it is not only difficult to obtain gearbox fault samples, but also the key features are easily masked by noise, leading to overfitting and poor generalization ability of existing deep learning models in time series prediction. Thus, this paper proposes a topology-guided time-series prediction model under small-sample conditions. Firstly, an adaptive period detection mechanism is constructed by using Hilbert-Huang Transform and Teager Energy Operator to overcome the limitations of traditional methods in extracting features from non-stationary signals. Secondly, a topology attention mechanism is introduced. By mapping the signal to a topological point cloud and quantifying multi-scale topological features via persistent homology, generating an attention map with physical interpretability. Finally, a topology attention-guided convolutional neural network model is constructed to achieve high-precision vibration signal time series prediction. Experimental results show that the proposed model has a coefficient of determination R2 that is about 61% higher than that of the TimesNet model, verifying the robustness and efficient performance of the model under small sample conditions.