失谐
人工神经网络
振动
联轴节(管道)
有限元法
过程(计算)
声学
控制理论(社会学)
计算机科学
工程类
结构工程
物理
人工智能
机械工程
操作系统
控制(管理)
作者
Daosen Liang,Jianyao Yao,Zichu Jia,Zhifu Cao,Xuyang Liu,Xuzhen Jing
出处
期刊:AIAA Journal
[American Institute of Aeronautics and Astronautics]
日期:2022-09-26
卷期号:61 (1): 391-405
被引量:7
摘要
Inevitable mistuning in cyclic bladed disk structures would cause vibration amplification phenomena that seriously reduce the reliability of the bladed disk. The ability to accurately and quickly predict the dynamic responses is critical to investigating the dynamic behavior of the mistuned system. However, it is still challenging because the mistuned responses are extremely sensitive to the random mistuning parameters. In this work, a novel mistuned system deep neural network model (MS-DNN) is presented to predict the dynamic responses of mistuned bladed disks through the mistuning parameters for both the lumped parameter model and the large-scale finite element (FE) model, which decouples the vibration equations of the mistuned system and uses a neural network to replace the coupling process. MS-DNN is divided into two levels, namely, the blade and the disk. The blade-level neural networks are used for forward and backward propagation of mistuning parameters in the different blades, and the disk-level neural network is used to replace the physical coupling process in the disk of multiple mistuning parameters from individual blades, with data transmission between the neural networks via blade–disk boundary nodes. The expected physical response of the blade tip is predicted through MS-DNN. All neural networks in MS-DNN show high prediction accuracy on both training sets and unknown test sets. For the FE model of the industrial bladed disk, the effect of the number of boundary nodes selected as the data interface between neural networks on the prediction accuracy is also investigated. The results show that, for unknown test data, the predicted response has an [Formula: see text] value of 0.998 versus the actual response with an amplification factor error of less than 0.388%.
科研通智能强力驱动
Strongly Powered by AbleSci AI