TY - JOUR
T1 - Vibration2Image
T2 - Transfer learning for amplitude-dependent FRF prediction in progressive random vibration tests
AU - Wan, Wenqing
AU - Chen, Jinglong
AU - Huang, Jinyi
AU - Wan, Xiaofei
N1 - Publisher Copyright:
© 2026 Elsevier Masson SAS.
PY - 2027/1
Y1 - 2027/1
N2 - Progressive multi-level random vibration tests are essential for spacecraft qualification, yet engineers must push toward dynamic strength limits without triggering vibration-induced damage and costly test losses at high excitation levels. Therefore, practitioners seek to anticipate high-level responses from early low-level data, yet nonlinear, amplitude-dependent evolution emerges as excitation levels increase, making such extrapolation inherently difficult.To address this challenge, this paper presents FRF-UNet, which reframes cross-level frequency response function (FRF) prediction as an image-to-image task by encoding tri-axial FRFs into RGB images. Unlike standard U-Net, the proposed architecture employs asymmetric convolutions and frequency-only downsampling to preserve amplitude-level correlations. A two-stage strategy-masked pretraining on similar structures followed by fine-tuning-learns transferable cross-level patterns. Experimental validation demonstrates 8.5% RMSE improvement over standard U-Net and 14.6% over CNN. Ablation studies confirm contributions from masked pretraining (18.7% RMSE reduction) and positional encoding (18.3%). With only three low-level tests, the framework maintains R2 > 0.81, enabling reliable high-level prediction before potentially destructive qualification tests.
AB - Progressive multi-level random vibration tests are essential for spacecraft qualification, yet engineers must push toward dynamic strength limits without triggering vibration-induced damage and costly test losses at high excitation levels. Therefore, practitioners seek to anticipate high-level responses from early low-level data, yet nonlinear, amplitude-dependent evolution emerges as excitation levels increase, making such extrapolation inherently difficult.To address this challenge, this paper presents FRF-UNet, which reframes cross-level frequency response function (FRF) prediction as an image-to-image task by encoding tri-axial FRFs into RGB images. Unlike standard U-Net, the proposed architecture employs asymmetric convolutions and frequency-only downsampling to preserve amplitude-level correlations. A two-stage strategy-masked pretraining on similar structures followed by fine-tuning-learns transferable cross-level patterns. Experimental validation demonstrates 8.5% RMSE improvement over standard U-Net and 14.6% over CNN. Ablation studies confirm contributions from masked pretraining (18.7% RMSE reduction) and positional encoding (18.3%). With only three low-level tests, the framework maintains R2 > 0.81, enabling reliable high-level prediction before potentially destructive qualification tests.
KW - Deep learning
KW - Digital twin
KW - Frequency response function (FRF)
KW - Nonlinear structural dynamics
KW - Random vibration testing
KW - Spacecraft structural qualification
KW - Transfer learning
UR - https://www.scopus.com/pages/publications/105049067591
U2 - 10.1016/j.ast.2026.113626
DO - 10.1016/j.ast.2026.113626
M3 - 文章
AN - SCOPUS:105049067591
SN - 1270-9638
VL - 180
JO - Aerospace Science and Technology
JF - Aerospace Science and Technology
M1 - 113626
ER -