航空学报 > 2024, Vol. 45 Issue (19): 229969-229969   doi: 10.7527/S1000-6893.2024.29969

固体力学与飞行器总体设计

观测不确定性下变分贝叶斯高效模型修正

陶言和1,2, 郭勤涛1,2(), 周瑾1,2, 马嘉倩1,2, 李效法3   

  1. 1.南京航空航天大学 机电学院,南京 210016
    2.直升机动力学全国重点实验室,南京 210016
    3.中国商飞上海飞机设计研究院,上海 201210
  • 收稿日期:2023-12-12 修回日期:2023-12-26 接受日期:2024-01-17 出版日期:2024-02-06 发布日期:2024-02-02
  • 通讯作者: 郭勤涛 E-mail:guo_qintao@nuaa.edu.cn
  • 基金资助:
    国家自然科学基金(U23B20105)

Efficient variational Bayesian model updating under observation uncertainty

Yanhe TAO1,2, Qintao GUO1,2(), Jin ZHOU1,2, Jiaqian MA1,2, Xiaofa LI3   

  1. 1.College of Mechanical and Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing  210016,China
    2.National Key Laboratory of Helicopter Dynamics,Nanjing  210016,China
    3.COMAC Shanghai Aircraft Design and Research Institute,Shanghai  201210,China
  • Received:2023-12-12 Revised:2023-12-26 Accepted:2024-01-17 Online:2024-02-06 Published:2024-02-02
  • Contact: Qintao GUO E-mail:guo_qintao@nuaa.edu.cn
  • Supported by:
    National Natural Science Foundation of China(U23B20105)

摘要:

考虑试验观测数据不确定性,针对以随机响应为目标的复杂仿真模型不确定性修正问题,提出一种以自回归模型提取信号特征,马氏距离构造近似似然函数,并由变分贝叶斯-蒙特卡洛进行参数后验识别的模型修正方法。对经过平稳性检验的随机信号进行自回归分析,得到模型特征向量实现信息降维;考虑试验观测数据混合不确定性并用概率盒法表征,以数值模拟数据与试验观测数据的特征向量之间的马氏距离构造近似对数似然;基于变分贝叶斯-蒙特卡洛方法求解边际似然,经过很少的迭代次数即可收敛,最终识别出参数的后验分布。在螺栓连接结构仿真案例和某民用飞机机翼模型工程案例中,修正后的模型具有很高精度,且在一定的观测不确定性水平下依然具有良好的修正效果,验证了所提方法对工程结构不确定性模型修正问题的有效性。

关键词: 模型修正, 不确定性量化, 变分贝叶斯, 自回归模型, 马氏距离

Abstract:

A model updating approach is proposed to address uncertainty in complex numerical models with stochastic responses. The approach involves using auto-regressive models for signal feature extraction, utilizing Mahalanobis distance as uncertainty quantification metric, and performing parameter posterior identification through variational Bayesian Monte Carlo. Initially, auto-regressive analysis is conducted on stationary stochastic signals to obtain model feature vectors for dimension reduction. The hybrid uncertainty of experimental observation data is then characterized utilizing the probability-box method. An approximate logarithmic likelihood is constructed based on the Mahalanobis distance between feature vectors of simulated data and experimental observation data. Finally, the variational Bayesian Monte Carlo method is used to solve the marginal likelihood, resulting in the identification of the posterior of parameters after very few iterations. Effectiveness of the proposed method for uncertainty updating in engineering structural models is validated through a numerical case of bolted connection structures and a civil aircraft wing model. The updated model exhibits high accuracy and retains good updating performance under certain levels of observation uncertainty.

Key words: model updating, uncertainty quantification, variational Bayesian, auto-regressive model, Mahalanobis distance

中图分类号: