航空学报 > 2023, Vol. 44 Issue (7): 426991-426991   doi: 10.7527/S1000-6893.2022.26991

基于PSO⁃BiLSTM神经网络的机身筒段应力预测

杨超1,2(), 张开富1   

  1. 1.西北工业大学 机电学院,西安  710072
    2.中国航空工业集团公司北京长城航空测控技术研究所,北京  101111
  • 收稿日期:2022-01-24 修回日期:2022-02-18 接受日期:2022-04-01 出版日期:2023-04-15 发布日期:2022-07-08
  • 通讯作者: 杨超 E-mail:yangc@avic-bmc.com
  • 基金资助:
    国家级项目

Stress prediction of fuselage tube section based on PSO⁃BiLSTM neural network

Chao YANG1,2(), Kaifu ZHANG1   

  1. 1.School of Mechanical Engineering,Northwestern Polytechnical University,Xi’an  710072,China
    2.AVIC Beijing Chang Cheng Aeronautic Measurement and Control Technology Research Institute,Beijing  101111,China
  • Received:2022-01-24 Revised:2022-02-18 Accepted:2022-04-01 Online:2023-04-15 Published:2022-07-08
  • Contact: Chao YANG E-mail:yangc@avic-bmc.com
  • Supported by:
    National Level Project

摘要:

飞机机身部件在装配对接前,通常需要使用形状控制工装调节机身筒段形状以达到对接标准,为避免出现控形过程中筒段局部应力过大损坏机身的情况,需实时监测机身筒段应力变化,为此提出基于粒子群优化算法和双向长短期记忆 (PSO-BiLSTM) 神经网络的应力预测方法。通过机身筒段控形历史数据对应力预测模型进行训练和测试,采用粒子群优化算法迭代优化神经网络超参数,实验结果表明,PSO-BiLSTM神经网络凭借长记忆细胞和高模型容量优势在处理序列式数据方面具有明显优势,应力预测损失在0.3%的均方根误差范围内收敛。与竞争模型循环神经网络(RNN)、标准长短期记忆 (LSTM) 神经网络和双向长短期记忆 (BiLSTM) 神经网络相比,PSO-BiLSTM神经网络模型不仅预测结果更准确,训练效率也明显提高。

关键词: 装配, 机身筒段, 形状控制, 实时监测, 应力预测, 粒子群算法

Abstract:

Before aircraft fuselage components are assembled and docked, shape control tools are usually used to adjust the shape of the fuselage tube section to meet the docking standard. In order to avoid damage to the fuselage due to excessive local stress during the shape control process, it is necessary to monitor changes in stress. This research proposes a stress prediction method, which combines Particle Swarm Optimization and Bidirectional Long Short-Term Memory (PSO-BiLSTM) neural network. In this paper, the model input data is converted into sliding window sequential data with extremely high data correlation, and the stress data set is trained and tested. The experimental results show that the PSO-BiLSTM neural network has obvious advantages in processing sequential data. This is because the PSO-BiLSTM network has long memory cells and high model capacity. Stress prediction loss converges within 0.3% Root Mean Square Error (RMSE) error range. Compared with the competitive model RNN network, the standard Long Short-Term Memory (LSTM) network and the Bidirectional Long Short-Term Memory (BiLSTM) neural network, the PSO-BiLSTM model not only predicts more accurately, but also significantly improves training efficiency.

Key words: assembly, fuselage tube section, shape control, real-time monitoring, stress prediction, particle swarm optimization

中图分类号: