Shape adjustment for uncertain mesh reflectors using machine learning. (15th April 2023)

• Record Type: Journal Article
• Authors: Ren, Zhiwei; Du, Jingli; Bao, Hong; Ge, Dongming; Wang, Feijie
• Is Part Of: International journal of mechanical sciences. Volume 244(2023)
• Publication Date: 2023-04-15
• Subjects: Mesh reflector -- Uncertainty -- Shape adjustment -- Machine learning -- Extreme gradient boosting; Mechanical engineering -- Periodicals; Génie mécanique -- Périodiques; Mechanical engineering; Maschinenbau; Mechanik; Zeitschrift; Periodicals; 621.05
• DOI: 10.1016/j.ijmecsci.2022.108082
• Languages: English
• ISSNs: 0020-7403

Abstract

Highlights: We use machine learning to find the adjustment amount required for the current prototype. The method takes into account structural uncertainty and nonlinearity. The method has higher adjustment efficiency and precision with less workload. We increase the performance of the method by using the prototype position information. This method is not restricted to a particular uncertain factor. Abstract: Mesh reflectors are uncertain structures due to manufacturing and assembly errors. Shape adjustment must be made to meet the surface precision requirement of space missions by changing the lengths of adjustable cables. It is desirable to find the optimum adjustment amount based on the current state of the actual prototype. However, it is hard to build an accuracy simulation model reflecting mechanical behavior of the actual prototype because only part information about the actual prototype can be measured. Therefore, it is necessary to deal with the parameter uncertainty of the actual prototype in order to provide a reliable structure state to guide the shape adjustment. This field has not been well studied because it involves an uncertainty analysis of parameters and the nonlinearity of high coupling of node displacements. To improve the efficiency of shape adjustment of uncertain mesh reflectors, this paper combines finite element model, actual prototype, and machine learning algorithm to develop a prediction model for precision analysis of the mesh reflector under multi-uncertainties using a machine learning algorithm (extreme gradient boosting), with the reflector surface node position as the input and the cable length adjustment amount required to displace it to a high precision surface as the output. First, based on the nominal design model, the sample data is collected by the Latin hypercube design method; the nonlinearity of the structure was taken into account by iterative solution during the data sampling, and the uncertain models containing different random errors are fused from the perspective of data sampling. The Bayesian optimization algorithm is used to obtain the optimal hyperparameters that enables strong generalization of the prediction model. Next, learning from the idea of K-nearest neighbor algorithm, the prediction model is updated based on the actual prototype data, which further improves the prediction ability of the current prototype to be adjusted. The incremental learning algorithm is used to reduce the time-consuming of the prediction model updating. Finally, numerical examples and experiments show that the proposed method has higher adjustment efficiency and precision with less workload. Graphical abstract: Image, graphical abstract