Deep learning-based study of strength variance coefficient for large diameter thin-walled structures

Abstract

Strength variation coefficient is a basic parameter to carry out structural reliability design and assessment, for the large diameter thin-walled structure test to obtain the strength variation coefficient is expensive, this paper is based on the actual measurement of the product information, the application of finite element simulation and analysis methods, comprehensive consideration of the material properties, structural dimensions and geometrical uncertainty factors, put forward a method for the study of strength variation coefficients of thin-walled structures based on multi-head CNN. Taking the multi-wall panel welded stiffened rocket tank cylinder section as the research object, the geometric imperfection of a single panel is used as a sub-sample, and the particle swarm optimisation based inter-wall panel connection coordination method is used to achieve the geometric imperfection random field construction; relying on a small number of experiments, a multi-head convolutional network structure is used to realise the fusion of uncertain features of material properties, structural dimensions and geometrical imperfections at different scales, to efficiently establish an ‘uncertainty-response’ mapping model, and to realise the prediction of strength variation coefficients at low cost. The research results show that the method is able to identify the complex action law of geometrical imperfection on structural bearing, and the accuracy of the prediction of bearing capacity is more than 99.2%; it can realise the accurate quantitative analysis of the coefficient of variation of the strength of thin-walled structure and its influencing factors, and the predicted coefficient of variation of the structural strength of the structure is reasonably encompassed by the upper limit of the test value.