Page 87 - 《振动工程学报》2026年第5期
P. 87

第 39 卷第 5 期                       振 动 工 程 学 报                                       Vol. 39 No. 5
               2026 年  5 月                     Journal of Vibration Engineering                       May 2026



                             大   语   言   模   型    在   跨   座   式   单    轨   列   车   齿   轮    箱

                                         振   动    信   号   预   测   中    的   应   用



                                     赵 玲, 吴杭俊, 孟 阳, 巫 刚, 张 娜

                                           (重庆交通大学信息科学与工程学院,重庆 400074)


              摘要:跨座式单轨列车齿轮箱振动信号是监控列车运行状态的关键指标,振动信号的准确预测对于早期故障检测至关重要。
              鉴于大语言模型在模式识别和时间序列数据分析中的优势,本文提出了一种融合大语言模型(LLM)与时序数据处理技术的
              跨座式单轨列车齿轮箱振动信号预测方法,构建了基于 DCBiformerNet 与 VSP-LLM 架构的集成方案。该方案包含三个核心
              步骤: 提出 DCBiformerNet 模型,通过融合 GRU、因果卷积和多头注意力机制,增强对时序特征的提取能力,提高对振动信号
              趋势预测的准确性;设计特定任务和通用任务的提示模板,结合多模态数据作为大语言模型的输入,显著提升推理效果; 将
              DCBiformerNet 与 VSP-LLM 框架结合,改进预测输出层,实现高精度振动信号预测。试验结果表明,该方法在预测精度和稳
              定性方面显著优于传统模型(如 Autoformer、Informer、DLinear 等),验证了其高效性能。
              关键词: 大语言模型;跨座式单轨列车齿轮箱;振动信号预测;Transformer 模型
                             +
              中图分类号:TH165 .3        文献标志码:A        DOI:10.16385/j.cnki.issn.1004-4523.202501019


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

                                    ZHAO Ling,WU Hangjun,MENG Yang,WU Gang,ZHANG Na
                        (School of Information Science and Engineering,Chongqing Jiaotong University,Chongqing 400074,China)

              Abstract: Vibration  signals  from  the  gearboxes  of  straddle  monorail  trains  are  key  indicators  of  monitoring  train  operational  status,  and
              accurately  predicting  these  signals  is  crucial  for  early  fault  detection.  Given  the  advantages  of  large  language  models  (LLM)  in  pattern
              recognition and time series analysis, this paper proposes a method for predicting vibration signals from straddle monorail train gearboxes by
              integrating  LLM  with  time-series  processing  techniques  and  constructs  an  integrated  solution  based  on  the  DCBiformerNet  and  VSP-LLM
              architectures. This solution comprises three core steps: The DCBiformerNet model is introduced, which enhances temporal feature extraction
              capabilities by integrating GRU, causal convolutions and multi-head attention mechanisms, thereby improving the accuracy of vibration signal
              trend prediction. The prompt templates for both task-specific and general-purpose tasks are designed, and multimodal data is incorporated as
              input to the large language model to significantly enhance inference performance. DCBiformerNet is combined with the VSP-LLM framework
              to improve the prediction output layer and achieve high-precision vibration signal prediction. Experimental results demonstrate that this method
              significantly  outperforms  traditional  models  (such  as  Autoformer,  Informer,  and  DLinear)  in  terms  of  prediction  accuracy  and  stability,
              validating its high-performance capabilities.

              Keywords:large language model;gearbox for straddle monorail train;vibration signal prediction;Transformer model

                  跨座式单轨列车(如图          1  所示)作为一种高效且             轮箱发生故障,如齿轮磨损或轴承损坏,会导致振动
              环保的城市轨道交通工具,因占地面积小、爬坡能                            增强、噪声增大,并对车辆的动力传输效率产生负
                                                                      [4]
              力强、转弯半径小等优点,在全球多个城市广泛应                            面影响 。此外,齿轮箱的振动问题还会降低列车运
                [1]
              用 。跨座式单轨列车的平稳运行高度依赖其关键                            行的平稳性,影响乘客舒适度。尤其在高速运行时,
              部件的正常运转,尤其是齿轮箱 ,该部件负责将电                           振动和噪声的增加可能加剧乘客不适感,产生更严
                                           [2]
              动机的扭矩传递至车轮 ,为列车提供动力。如果齿                           重的机械故障 ,威胁列车安全。因此,及时维护和
                                                                             [5]
                                   [3]


                  收稿日期:2025-01-07;修订日期:2025-02-24
                  基金项目:国家自然科学基金面上项目(62073051);重庆市教委重大科学研究项目(KJZD-M202400706);重庆市研究生
                          科研创新项目(2024yjkc0011)
   82   83   84   85   86   87   88   89   90   91   92