TY - JOUR
T1 - Physics-informed RMS-driven LSTM framework for real-time fatigue damage estimation of reusable spacecraft structures under broadband random vibration
AU - Wan, Wenqing
AU - Li, Jinyuan
AU - Chen, Jinglong
AU - Chen, Zhen
AU - Liu, Yangyang
AU - Wan, Xiaofei
AU - Xu, Xun
N1 - Publisher Copyright:
© 2026 Elsevier B.V.
PY - 2027/1
Y1 - 2027/1
N2 - Real-time fatigue damage assessment for reusable spacecraft structures under broadband random vibration remains challenging due to reliance on full high-frequency strain histories and computationally intensive cycle-counting procedures. This study proposes a physics-informed parameter-modulated LSTM (PM-LSTM) framework that embeds the theoretical power-law relationship D=k⋅RMSm into a recurrent neural network, where the LSTM dynamically modulates the power-law parameters m and log10k based on temporal context and structural location. The framework integrates sliding-window rainflow counting for training data generation, a unified multi-sensor training strategy with learnable sensor embeddings, and real-time prediction from streaming RMS inputs. Experimental validation on a rocket segment mock-up under multi-level broadband random vibration (20–2000 Hz, nine load levels) demonstrates that PM-LSTM achieves superior in-domain prediction accuracy (MAPE = 6.94%) over both theoretical power-law fitting (MAPE = 10.39%) and optimal power-law regression (MAPE = 7.50%), while maintaining good cross-PSD generalization when transferred to an unseen spectral condition without re-training. The physics-informed architecture also provides significantly enhanced noise robustness compared to pure power-law fitting, with MAPE reduced by 45% at 50% noise level. Cumulative damage tracking attains a mean damage ratio of 1.03 (mean relative error of 11.8%) with a ∼3000× computational speedup over conventional rainflow-counting analysis. The framework offers an accurate, physically interpretable, and efficient solution for real-time structural health monitoring in digital-twin applications.
AB - Real-time fatigue damage assessment for reusable spacecraft structures under broadband random vibration remains challenging due to reliance on full high-frequency strain histories and computationally intensive cycle-counting procedures. This study proposes a physics-informed parameter-modulated LSTM (PM-LSTM) framework that embeds the theoretical power-law relationship D=k⋅RMSm into a recurrent neural network, where the LSTM dynamically modulates the power-law parameters m and log10k based on temporal context and structural location. The framework integrates sliding-window rainflow counting for training data generation, a unified multi-sensor training strategy with learnable sensor embeddings, and real-time prediction from streaming RMS inputs. Experimental validation on a rocket segment mock-up under multi-level broadband random vibration (20–2000 Hz, nine load levels) demonstrates that PM-LSTM achieves superior in-domain prediction accuracy (MAPE = 6.94%) over both theoretical power-law fitting (MAPE = 10.39%) and optimal power-law regression (MAPE = 7.50%), while maintaining good cross-PSD generalization when transferred to an unseen spectral condition without re-training. The physics-informed architecture also provides significantly enhanced noise robustness compared to pure power-law fitting, with MAPE reduced by 45% at 50% noise level. Cumulative damage tracking attains a mean damage ratio of 1.03 (mean relative error of 11.8%) with a ∼3000× computational speedup over conventional rainflow-counting analysis. The framework offers an accurate, physically interpretable, and efficient solution for real-time structural health monitoring in digital-twin applications.
KW - Digital twin
KW - Fatigue damage assessment
KW - Physics-informed neural network
KW - Random vibration
KW - Real-time monitoring
KW - Reusable spacecraft
KW - Structural health monitoring
UR - https://www.scopus.com/pages/publications/105052345656
U2 - 10.1016/j.asoc.2026.116551
DO - 10.1016/j.asoc.2026.116551
M3 - 文章
AN - SCOPUS:105052345656
SN - 1568-4946
VL - 204
JO - Applied Soft Computing
JF - Applied Soft Computing
M1 - 116551
ER -