DOI：10.3969/j.issn.1672-3872.2026.05.031
基金项目：大学生创新创业训练计划项目“智采丰登——基于TOF系统和5GRedCap网络的淮山智能收获机”（S202510566010）；广东海洋大学深蓝智能机电产品创新团队（CXTD2023009）
作 者：袁诗佳 1，2 ，余 江 1，2 ，麦竣深 1，2 ，刘祥源 1，2
（1. 广东海洋大学机械工程学院，广东 湛江 524088；2. 广东海洋大学深蓝智能机电产品创新团队，广东 湛江 524088）
摘 要：【目的】解决现有模型跨工况适应性低、难以部署至边缘设备、对非平稳噪声鲁棒性不足等问题，提升模型预测精度。【方法】文章提出了一种结合时序卷积网络（TCN）和Transformer的轴承剩余寿命预测模型，并设计了Hyperband与Optuna两阶段超参数优化策略：通过Hyperband快速筛选出关键超参数范围，再经Optuna基于贝叶斯进行搜索精细化，实现高效调参，可在轴承退化数据中实现高效特征融合。最后，该模型通过集成FEMTO-ST和XJTU-SY两个公开轴承数据集进行了系统性训练与验证，并与CNN、Transformer、TCN等主流模型进行了对比试验。【结果】该TCN-Transformer模型在多项性能指标上均显著优于传统结构，尤其在复杂退化趋势建模与多工况预测任务中具备更强泛化能力。【结论】该预测模型在农业机械领域展现出卓越的预测性能与良好的跨场景泛化能力，在工业级RUL预测应用中具备较高的部署价值与稳定性。未来可进一步探索其在复杂工业场景下的实时部署能力以及融合自监督学习与迁移学习机制的潜力。
关键词：轴承；剩余寿命预测；TCN-Transformer；时序卷积网络；超参数优化
A Bearing Remaining Life Prediction Model Based on TCN-Transformer and Hybrid Hyperparameter Optimization
Author: Yuan Shijia1,2, Yu Jiang1,2, Mai Junshen1,2, Liu Xiangyuan1,2
(1.School of Mechanical Engineering, Guangdong Ocean University, Zhanjiang 524088, China; 2.Deep Blue Intelligent Electromechanical Product Innovation Team in Guangdong Ocean University, Zhanjiang 524088, China)
Abstract: [Objective] To solve the problems of low adaptability across operating conditions, difficulty in deploying to edge devices, and insufficient robustness to non-stationary noise in existing models, and to improve the prediction accuracy of the models. [Method] The paper proposes a bearing remaining life prediction model combining Time Series Convolutional Network (TCN) and Transformer, and a two-stage hyperparameter optimization strategy using Hyperband and Optuna was designed. The key hyperparameter range was quickly screened through Hyperband, and then refined through Bayesian search using Optuna to achieve efficient parameter tuning and feature fusion in bearing degradation data. Finally, the model was systematically trained and validated by integrating two publicly available bearing datasets, FEMTO-ST and XJTU-SY, and was compared with mainstream models such as CNN, Transformer, and TCN through contrastive experiments. [Result] The TCN-Transformer model significantly outperforms traditional structures in multiple performance indicators, especially in complex degradation trend modeling and multi condition prediction tasks with stronger generalization ability. [Conclusion] This prediction model demonstrates excellent predictive performance and good cross scenario generalization ability in the field of agricultural machinery, and has high deployment value and stability in industrial RUL prediction applications. In the future, we can further explore its real-time deployment capability in complex industrial scenarios and the potential of integrating self supervised learning and transfer learning mechanisms.
Keywords: bearing; remaining life prediction; TCN-Transformer; temporal convolutional network; hyperparameter optimization
引文信息：[1]袁诗佳，余江，麦竣深，等.基于TCN-Transformer与混合超参数优化的轴承剩余寿命预测模型[J].南方农机，2026，57(5)：119-122+129.
查看全文请下载PDF文件↓
