计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (16): 274-283.DOI: 10.3778/j.issn.1002-8331.2111-0448

• 工程与应用 • 上一篇    下一篇

基于改进CNN-LSTM的飞控系统剩余寿命预测

李梦蝶,赵光,罗灵鲲,胡士强   

  1. 1.上海交通大学 航空航天学院,上海 200240
    2.中国商用飞机有限责任公司,上海 200126
  • 出版日期:2022-08-15 发布日期:2022-08-15

Remaining Useful Life Prediction of Flight Control System Based on Improved CNN-LSTM

LI Mengdie, ZHAO Guang, LUO Lingkun, HU Shiqiang   

  1. 1.School of Aeronautics and Astronautics, Shanghai Jiao Tong University, Shanghai 200240, China
    2.Commercial Aircraft Corporation of China, Ltd., Shanghai 200126, China
  • Online:2022-08-15 Published:2022-08-15

摘要: 数据驱动的剩余寿命(remaining useful life,RUL)预测是复杂系统健康管理的重点研究内容,然而数据集的缺乏制约了不同系统上RUL预测的研究。针对这一问题,以飞控系统为例,提出一种仿真模型和数据混合驱动的RUL预测方法。该方法通过模型仿真提供充足的故障数据,并结合改进CNN-LSTM网络实现高质量的故障信息提取。首先对系统及其故障模式建立仿真模型,利用蒙特卡罗方法生成随机故障时间序列并依次注入故障,根据仿真响应和失效阈值确定序列的寿命标签,即可生成包含多组随机序列的系统失效数据集;其次利用长短时记忆网络(long short-term memory,LSTM)提取系统状态参数时间序列的故障信息,结合一维卷积神经网络(1D-CNN)提取不同状态参数之间的关联特征,从而形成时序-空间相结合的剩余寿命预测网络。充分的实验结果证明了所提方法对不同系统均能帮助达到动态和准确的剩余寿命预测。

关键词: 剩余寿命预测, 建模仿真, 长短时记忆网络, 一维卷积神经网络

Abstract: Recently, data-driven remaining useful life(RUL) prediction techniques play central role in health management of complex systems, while lacking of the dataset restricts the wide usability of the RUL across different systems. To remedy the addressed issue, this paper takes the flight control system as an example, and a simulation model and data hybrid-driven RUL prediction method is proposed. The simulation model can provide sufficient fault data, while the improved CNN-LSTM network enables the highly qualified fault information exploration. Initially, a system simulation model with its failure modes is established, and random failure time sequences generated by Monte Carlo method are supposed to be injected into the model iteratively. After the life labels of the sequences are distinguished according to the simulation response and the failure threshold, the failure dataset with multiple random failure sequences is well prepared. Then, this paper uses long short-term memory network(LSTM) to extract the fault information of the state sequence from time dimension, and combines 1-dimensional convolution neural network(1D-CNN) to get the correlation features among state parameters, thereby the temporal-spatial RUL prediction network is formulated subsequently. This paper proposes sufficient experimental results to demonstrate that the proposed method is helpful in achieving dynamic and accurate RUL prediction generally across different systems.

Key words: remaining useful life prediction, model simulation, long short-term memory network, 1-dimensional convolution neural network