Image-Driven Intelligent Prediction of Buckling Behavior for Geometrically Imperfect Cylindrical Shells

Abstract

Thin-walled cylindrical shells are key load-carrying components for aerospace structures. Axial compression is the most common load for cylindrical shells; however, critical buckling loads obtained experimentally are significantly lower than the theoretical values, owing to the presence of initial geometric imperfections (GIs). Previous studies often determined only the knockdown factor through the statistics of experimental data; it is an approximate lower limit of the load-carrying capacity and cannot accurately predict the buckling mode in real time. In this paper, an image-driven framework for the intelligent prediction of buckling load and mode based on measured GIs is proposed. First, through measured samples and the random field method, the GI data pool is obtained. Subsequently, two convolutional neural network (CNN) models are trained to predict the buckling load and mode for cylindrical shells, respectively. The active learning strategy is adopted to discern beneficial sample sets and enhance the prediction accuracy and training speed of the CNN models. A numerical example demonstrates that the proposed framework can effectively predict the buckling behavior of cylindrical shells with GIs. Furthermore, using the proposed prediction framework, it was observed that a common feature of unfavorable GIs is a large fluctuation in magnitude along the axial direction.