Predicting seismic responses of non-specific steel structures via learned representation of structural dynamic behaviors

Recent advances in structural engineering highlight deep learning (DL) as a promising approach for developing surrogate models capable of rapidly predicting structural responses to dynamic excitations. Prior studies have demonstrated that DL models can effectively approximate the mapping between exc...

Ausführliche Beschreibung

Gespeichert in:

Bibliographische Detailangaben
Veröffentlicht in:	Engineering structures Jg. 344; S. 121273
Hauptverfasser:	Pan, Zeyu, Shi, Jianyong, Jiang, Liu
Format:	Journal Article
Sprache:	Englisch
Veröffentlicht:	Elsevier Ltd 01.12.2025
Schlagworte:	Deep representation Encoder-decoder architecture Hysteresis curve Structural dynamic behavior representation Structural seismic responses prediction MRF EOM PIDR LSTM PRD Encoder-decoder architecture NSDOF DL Structural dynamic behavior representation Hysteresis curve Structural seismic responses prediction PFA GRU PGA PGD Deep representation CAV HB DOF HC PGV ML
ISSN:	0141-0296
Online-Zugang:	Volltext
Tags:	Tag hinzufügen Keine Tags, Fügen Sie den ersten Tag hinzu!

Beschreibung
Zusammenfassung:	Recent advances in structural engineering highlight deep learning (DL) as a promising approach for developing surrogate models capable of rapidly predicting structural responses to dynamic excitations. Prior studies have demonstrated that DL models can effectively approximate the mapping between excitation inputs and structural responses, capturing inherent nonlinearities in structural behavior and non-stationarities in input excitations. However, broader adoption of DL surrogates requires addressing critical challenges, particularly their adaptability to diverse structural configurations. Drawing inspiration from representation learning techniques in domains such as natural language processing (NLP) and computer vision, this study investigates their application to enhance the generalizability of DL surrogates by leveraging learned representations of structural dynamic behaviors. This paper proposes a denoising autoencoder (DAE)-based approach to extract deep representations of structural dynamics and subsequently apply these representations to seismic response prediction tasks. To validate the approach, we conduct numerical experiments on two distinct structural systems: a nonlinear single-degree-of-freedom (NSDOF) system and a multi-story steel frame structure. The DAE encodes hysteretic behaviors into real-valued vectors of fixed dimensionality, analogous to embeddings in NLP. Analysis of the embedding space reveals strong correlations between vector similarities and corresponding hysteretic behaviors. Results confirm the efficacy of the proposed surrogate model in seismic response prediction through learned structural representations. Despite its simplicity, the framework exhibits potential for generalization across non-specific structural configurations, offering a practical solution to improve the versatility of DL-based surrogate models in structural engineering applications. •Deep learning based seismic response prediction for non-specific structures is proposed.•A general input paradigm of hysteretic behavior is presented for different structures.•Learned representations of hysteretic behavior can be clearly explained in this model.•Strategies alleviate the data magnitude issue regarding different structural responses.•Arbitrary ground motion records with different sampling rates are allowed as inputs.
ISSN:	0141-0296
DOI:	10.1016/j.engstruct.2025.121273