上海交通大学学报

上海交通大学学报

• • 上一篇    

基于双重注意力机制和选择性集成的装备寿命预测方法

  

  1. 1. 上海交通大学 机械系统与振动全国重点实验室 机械与动力工程学院，上海 200240；2. 上海交通大学-弗劳恩霍夫协会智能制造创新中心，上海201306；3. 中国船舶集团有限公司第七二六研究所，上海201108
  • 通讯作者: 夏唐斌，教授，博士生导师，电话(Tel.)：021-34208589；E-mail：xtbxtb@sjtu.edu.cn。
  • 作者简介:范宜静（2000-），硕士生，研究方向为剩余寿命预测。
  • 基金资助:
    国家重点研发计划重点项目(2022YFF0605700)

A remaining useful life prediction method based on dual attention and selective ensemble

  1. 1. School of Mechanical Engineering, State Key Laboratory of Mechanical System and Vibration, Shanghai Jiao Tong University, Shanghai 200240, China; 2. Fraunhofer Project Center for Smart Manufacturing at Shanghai Jiao Tong University, Shanghai 201306, China；3. The 726 Research Institute, China Shipbuilding Industry Corporation, Shanghai 201108, China

PDF (PC)

16

摘要：

剩余使用寿命的精准预测对于大型复杂装备的稳定可靠运行具有重要价值。为了确保预测准确性的同时提高模型的鲁棒性和泛化性，提出了一种基于双重注意力时序卷积网络和选择性粒子群优化集成的预测方法。首先，利用双重注意力时序卷积网络探索监测数据中多类别输入特征和不同时间步长之间的内部关联，从特征与时间两个维度强化退化信息；再利用选择性粒子群优化集成算法修剪不同时间尺度下的基模型，自删除表现不佳模型，自生成最优模型子集并赋予最优权重，加权输出预测结果。所提方法已在航空涡扇发动机退化数据集上开展验证，结果表明预测准确性比其他方法提高了13.9%。

关键词: 剩余使用寿命预测, 集成学习, 双重注意力机制, 选择性集成

Abstract:

Accurately predicting remaining useful life (RUL) is crucial for ensuring the stable and reliable operation of large complex equipment. To enhance prediction accuracy while improving model robustness and generalization, a method based on dual-attention temporal convolutional networks and selective particle swarm optimization ensemble is proposed. Firstly, the dual-attention temporal convolutional networks are employed to explore the internal correlations between multi-category input features and different time steps in monitoring data, enhancing degradation information from both feature and temporal dimensions. Subsequently, the selective particle swarm optimization ensemble algorithm prunes underperforming base models at various time scales, autonomously generating an optimal subset of models and assigning optimal weights for weighted output predictions. The proposed method has been validated on a dataset of aviation turbofan engine degradation, demonstrating a 13.9% improvement in prediction accuracy compared to other methods.

Key words:

remaining useful life prediction, ensemble learning, dual attention mechanism, selective ensemble

中图分类号: