Machine learning surrogate model for finite element analysis of railway vehicles using principal component analysis and multilayer perceptron

• Published in: Kikai Gakkai ronbunshū = Transactions of the Japan Society of Mechanical Engineers Vol. 90; no. 937; p. 24-00133
• Main Authors: FUKUMOTO, Kohei, SAKAKIMA, Hiroki, FURUTANI, Ryo, KAWASAKI, Takeshi, IZUMI, Satoshi
• Format: Journal Article
• Language: English; Japanese
• Published: The Japan Society of Mechanical Engineers 2024
• Subjects: Computational mechanics; Finite element method; Neural network; Railway; Structural analysis; Surrogate model
• ISSN: 2187-9761; 2187-9761
• DOI: 10.1299/transjsme.24-00133

In the initial stages of railway vehicle design, finite element analyses are often repeated while adjusting design variables such as plate thickness, window dimensions, and under-floor equipment installation positions to achieve the desired performance. However, this repetitive process of finite element analysis, which involves the detailed modeling of large and complex vehicle structures, is highly computationally demanding. Therefore, it is necessary to improve the efficiency and speed of analysis. In this paper, we propose a machine-learning-based surrogate model to replace finite element analysis in railway vehicle design. To address the complexity resulting from the vast number of nodal values, this model utilizes dimensionality reduction through principal component analysis (PCA) and a multilayer perceptron architecture. It enables the prediction of critical parameters for railway vehicle designs including maximum deflection, deformation, stress distribution, eigenfrequencies, and eigenmodes, directly from design parameters such as plate thickness, window dimensions, and under-floor equipment loading positions. The model demonstrates high accuracy, with predicted maximum deflection and eigenfrequencies within 0.2% and 1% deviation, respectively, across all input variables. Additionally, nodal displacements, stress distributions, and eigenmodes are also predicted with accuracies of 4.2%, 13%, and 2.5%, respectively. Slightly lower accuracy is observed particularly when inputting loading positions of point loads. This is attributed to the limitation of capturing locally steep changes in shape and stress caused by dimensionality reduction using PCA. The results for any input-output combinations are obtained in approximately 0.005 seconds per case, potentially eliminating the need for setup and conducting finite element analysis, which may take enormous time and effort.