Struct-WNO: Wavelet neural operator for seismic response prediction of nonlinear structural system

TY - JOUR

T1 - Struct-WNO

T2 - Wavelet neural operator for seismic response prediction of nonlinear structural system

AU - Hu, Yao

AU - Guo, Wei

AU - Tsang, Hing Ho

AU - Li, Sheng

N1 - Publisher Copyright: © 2025 Elsevier Ltd

PY - 2025/10/1

Y1 - 2025/10/1

N2 - Recent advancements in neural operators have demonstrated great promise for addressing the challenges of nonlinear pattern recognition and mapping between function spaces. The emergence of wavelet neural operator (WNO), characterized by both spatial and frequency localization, proves particularly useful for extracting features from non-stationary signals (e.g., seismic excitations). This paper introduces a novel framework of the Struct-WNO model that integrates WNO with structural knowledge to efficiently predict seismic responses, offering an alternative to traditional methods like finite element analysis by avoiding iterative computations and complex nonlinear modelling. The framework comprises three core components: (i) a multi-resolution WNO module, (ii) equivalent linearization-based mode decomposition, and (iii) pole-residue method applied in the Laplace domain. The WNO module is initially used to capture the nonlinear relationships between input features (i.e., seismic excitations) and output responses (i.e., nonlinear displacements). The nonlinear system is then linearized by treating the nonlinear displacement responses as pseudo-external loading. Mode decomposition theory is applied to reduce the multiple degree-of-freedom (MDOF) system to a set of single degree-of-freedom (SDOF) systems. The dynamic behaviour of the SDOF system can be solved using the pole-residue method in the Laplace domain. Finally, full structural responses are reconstructed by superimposing the response of individual modes of the system. Numerical simulations are conducted to assess the generality, reliability, robustness, and inference capabilities of the proposed Struct-WNO model. Additionally, experimental tests on a frame building and capacitor voltage transformer are utilized to further validate the effectiveness of the proposed framework.

AB - Recent advancements in neural operators have demonstrated great promise for addressing the challenges of nonlinear pattern recognition and mapping between function spaces. The emergence of wavelet neural operator (WNO), characterized by both spatial and frequency localization, proves particularly useful for extracting features from non-stationary signals (e.g., seismic excitations). This paper introduces a novel framework of the Struct-WNO model that integrates WNO with structural knowledge to efficiently predict seismic responses, offering an alternative to traditional methods like finite element analysis by avoiding iterative computations and complex nonlinear modelling. The framework comprises three core components: (i) a multi-resolution WNO module, (ii) equivalent linearization-based mode decomposition, and (iii) pole-residue method applied in the Laplace domain. The WNO module is initially used to capture the nonlinear relationships between input features (i.e., seismic excitations) and output responses (i.e., nonlinear displacements). The nonlinear system is then linearized by treating the nonlinear displacement responses as pseudo-external loading. Mode decomposition theory is applied to reduce the multiple degree-of-freedom (MDOF) system to a set of single degree-of-freedom (SDOF) systems. The dynamic behaviour of the SDOF system can be solved using the pole-residue method in the Laplace domain. Finally, full structural responses are reconstructed by superimposing the response of individual modes of the system. Numerical simulations are conducted to assess the generality, reliability, robustness, and inference capabilities of the proposed Struct-WNO model. Additionally, experimental tests on a frame building and capacitor voltage transformer are utilized to further validate the effectiveness of the proposed framework.

KW - Laplace transform

KW - Mode decomposition

KW - Pole-residue method

KW - Structural dynamic analysis

KW - Wavelet neural operator

UR - https://www.scopus.com/pages/publications/105010120042

U2 - 10.1016/j.jobe.2025.113381

DO - 10.1016/j.jobe.2025.113381

M3 - Article

AN - SCOPUS:105010120042

SN - 2352-7102

VL - 111

JO - Journal of Building Engineering

JF - Journal of Building Engineering

M1 - 113381

ER -