Performance assessment of deep learning models for reconstructing missing pressure data on the high-speed train from wind tunnel tests

Abstract

Missing pressure data frequently occurs in surface pressure distribution tests. Traditional methods, including mean, median, or regression imputation, rely on statistical measures or relationships to fill missing values. Comparatively, deep learning offers innovative approaches through autoencoders or generative adversarial networks, which can learn dataset patterns and generate plausible data, thereby enhancing completeness. Relying on pressure measurements from wind tunnel experiments, with inflow speeds varying between 2 and 18 m/s, this paper analyzes the train surface pressure distribution and explores deep learning-based reconstruction for missing data. Results show that as wind speeds increase, absolute pressure values rise proportionally, while distribution patterns remain consistent with prior findings. Moreover, deep learning models exhibit notable reconstruction performance, albeit with differing accuracies. Specifically, the Kolmogorov-Arnold Network achieves the highest precision, recording a 3.26 % average error, followed by the Multilayer Perceptron (7.08 %) and the Long Short-Term Memory network (8.90 %). Such discrepancies underscore the unique capabilities and limitations of each model. These findings demonstrate the efficacy of deep learning techniques in recovering missing pressure data and highlight their potential to augment wind tunnel datasets, ultimately offering a valuable reference for future investigations seeking to address data gaps in aerodynamic research and strengthen data-driven analyses and engineering practice.