Fourier neural operator surrogate for automated stress conversion-based structural health monitoring of complex marine infrastructure

Official URL: https://dx.doi.org/10.1016/j.autcon.2026.106836

Real-time structural health monitoring of complex marine infrastructure requires rapid stress estimates as operating conditions change, but finite-element (FE) analyses are too costly for continuous updating. This paper presents a Fourier Neural Operator (FNO) surrogate that maps sectional-load transfer functions on a frequency–heading grid to stress transfer-function fields, enabling automated stress-conversion updating without repeated FE runs. Response-based phase augmentation expands two loading conditions (ballast and full) into field pairs for robust operator learning. On an FPSO case study, the surrogate achieves R2>0.97 for held-out intermediate loading conditions and unseen frequency–heading combinations, while consistently reducing tail errors versus conventional regressors. Integrated into the conversion workflow, FNO-derived matrices yield robust hotspot predictions in frequency and time domains, avoiding accuracy loss when a mismatched design-condition matrix is used. Once trained, inference takes <1 s versus ~4 h per FE run, supporting condition-aware offshore digital twins.