Application of large language models in the gearbox vibration signals prediction of straddle monorail train

The scheme consists of three core steps: (1) the DCBiformerNet model is presented, which enhances the time-series feature extraction capability and improves the vibration signal trend prediction accuracy by integrating GRU, causal convolution, and multi-head attention mechanism; (2) prompt templates for specific and general tasks are designed. Multimodal data is combined and used as input for LLMs, which greatly improves the inference results; (3) DCBiformerNet is combined with the VSP - LLM framework and the prediction output layer is improved to fulfill high - precision vibration signal prediction. The experimental results show that this method surpasses traditional models (e.g., Autoformer, Informer, DLinear, etc.) in prediction accuracy and stability, confirming its high performance.