lazyzhaoyang / masterthesisproject

Code repository of ZhaoyangLi's master thesis.

Optimization and Recognition of Deformation Patterns of Thin-walled Structures based on Point Cloud Sequence and Deep Clustering

Code Implementation of ZhaoyangLi's Master's Thesis

Abstract

$\qquad$ The collision plastic deformation process of thin-walled energy-absorbing structures is very complex and variable, and a variety of undesirable collision deformation modes such as buckling and instability often occur in the process of practical engineering applications and structural design. However, the traditional structural crashworthiness optimization methods cannot represent the structural collision deformation process, and it is difficult to construct structural collision deformation constraints in the optimization process. So it cannot ensure that the final optimization results conform to the expected collision deformation modes. It is of great practical importance to study the structural crashworthiness optimization under the collision deformation mode constraint. In this thesis, we focus on the problem of structural crashworthiness optimization under collision deformation mode constraint, take thin-walled circular tube as the object of crashworthiness optimization, and carry out the research on the deep clustering optimization method for collision deformation mode of thin-walled structure. The main research work of this paper is as follows:

$\qquad$ By analyzing the reasons why it is difficult to effectively quantify the collision deformation process in traditional structural crashworthiness optimization, we analyze the technical difficulties of structural crashworthiness optimization under the constraints of collision deformation mode, and give the solution ideas of each technical difficulty. The basic architecture and process of the collision-deformation mode deep clustering optimization method for thin-walled structures are proposed. The main idea of this method is to predict and identify the structural collision deformation pattern through the agent model and the structural collision deformation pattern deep clustering model, and then construct the collision deformation pattern constraint to control the generation of population individuals in the process of structural collision resistance optimization to ensure that the final optimization results conform to the expected structural collision deformation pattern.

$\qquad$ A structural collision deformation representation method based on finite element grid node coordinate point cloud data is proposed, which takes the structural collision process finite element grid node point cloud sequence as the representation form of structural collision deformation process. And based on this method, the structural collision analysis dataset of a thin-walled metal circular tube is constructed using the finite element simulation results verified by collision tests.

$\qquad$ Combined with the theory related to multimodal fusion in the field of machine learning, a multimodal fusion prediction method for the collision process of thin-walled structures is proposed. It is proved that the prediction results of the proxy model constructed by the method for the structural crashworthiness response and crash deformation representation are within 5% of the finite element simulation results, but the computational speed is 8169 times faster than that of the finite element simulation, which greatly improves the efficiency of structural crashworthiness optimization under the constraints of crash deformation mode.

$\qquad$ A pseudo-label-based deep clustering method for structural collision deformation modes is proposed. And the difference between its clustering results and manual classification results for collision deformation patterns of thin-walled circular tubes is investigated by experimental design. The experimental results show that the accuracy of the method reaches up to 92.17% when using the Simsiam pre-trained model as the feature extractor, which is not only close to the accuracy of 98.50% under supervised learning, but also still has 16.84% improvement compared with traditional deep clustering methods. In addition, the effect of different number of clusters on its clustering effect is studied, and the results show that the method can effectively identify samples with similar deformation processes and has the ability to find new deformation patterns when the number of clusters is larger than the number of manual labels.

$\qquad$ Based on the above research results, a multi-objective optimization of the impact resistance of thin-walled circular tubes under the constraints of axial collision deformation patterns is carried out. A multi-objective optimization method based on NSGA II genetic algorithm is proposed to optimize the impact resistance of thin-walled circular tubes with and without collision deformation mode constraints by constructing the collision deformation mode constraints of thin-walled circular tubes, and the optimization results are verified by finite element simulation. The experimental results show that the method can effectively control the collision deformation mode of the structure under the collision deformation mode constraint compared with that without the collision deformation mode constraint, and make it conform to the desired collision deformation mode.

Keywords

Thin-walled energy absorbing structure; Crash resistance optimization; Structural collision deformation pattern recognition; Deep clustering