Automated classification of linear bifurcation buckling eigenmodes in thin-walled cylindrical shell structures. (November 2022)

• Record Type: Journal Article
• Authors: Sadowski, Adam J.
• Is Part Of: Advances in engineering software. Volume 173(2022)
• Publication Date: 2022-11
• Subjects: Shell structures -- Buckling -- Automated parameter sweeps -- Convolutional neural network -- Image classification -- Chromatic signatures; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2022.103257
• Languages: English
• ISSNs: 0965-9978

Abstract

Abstract: Computational problems in structural engineering are growing ever larger and solutions must increasingly be based on correspondingly large datasets obtained from detailed parametric sweeps. However, the acquisition of computational datasets of useful size is also becoming increasingly unfeasible without extensive use of automation. In computational shell buckling studies, particularly those of thin-walled shells under complex loading conditions, an important qualitative piece of information is the class of buckling mode which reveals the dominant destabilising membrane stress components. Unfortunately, the diversity of geometries that can be encountered in computational shell buckling studies is truly vast, and there is currently no way to rapidly assess the buckling mode without laborious direct human observation of the model output. This paper presents an automated classification tool for linear bifurcation buckling eigenmodes in cylindrical shells such as those found as wind turbine support towers, chimneys, silos, tanks, piles and pipelines. It is based on a convolutional neural network implemented using the PyTorch machine learning framework. The adopted network architecture is based on those widely adopted for image classification and recognition tasks, chosen based on a stratified five-fold cross-validation exercise. The network is trained on a purposefully generated basic dataset of 13, 392 linear bifurcation buckling eigenmodes modes encoded as chromatic signatures in .jpg images (enhanced to 25, 726 by transformations). An example parametric sweep of a cylindrical shell under unsymmetrical wind loading illustrates the performance of the classifier. A GitHub repository offers Python scripts and instructions on how to download the dataset and trained network. Highlights: Buckle classifier developed based on a convolutional neural network architecture. Eigenmodes classified into eleven classes based on dominant destabilising stresses. Automated generation of prototypical dataset for cross-validation and training. Enhancement of dataset for release training via probabilistic transformations. Classifier used in ongoing development of the RRD method for EN 1993-1-6