A spatiotemporal hybrid framework integrating Hypergraph-enhanced LSTM for measured response reconstruction in structural damage identification

The effectiveness of data-driven Structural Damage Identification (SDI) heavily relies on high-quality response data. However, data loss caused by factors such as measuring sensor failures severely limits its practical application. Although existing deep learning-based response reconstruction offers effective solutions to address data loss, most studies primarily focus on the long-term dependencies of temporal series, with insufficient exploration of the spatiotemporal characteristics of sensors. To address this challenge, this work proposes a novel spatiotemporal hybrid framework for response reconstruction based on a Hypergraph-enhanced Long Short-Term Memory (HeLSTM) network, which achieves high precision through “temporal-hypergraph” dual-stream framework. Specifically, a double-layer LSTM is employed to extract dynamic features from temporal sensor data, while hypergraph is innovatively used to model the implicit higher-order spatial relationships among sensors. Additionally, double-layer Hypergraph Convolutional Network (HGCN) is designed to capture the feature aggregation and dependency patterns of measurement points in non-Euclidean space, enabling deep integration of multi-sensor spatiotemporal features. The experimental results with two steel structures demonstrate that HeLSTM effectively leverages the advantages of hypergraph to integrate higher-order spatiotemporal feature correlations among measurement points, achieving exceptional accuracy in response reconstruction. Furthermore, the contribution assessments to SDI show that supplementing training data with reconstructed signals improve damage identification accuracy by 7.86% and 7.82% in the two cases, respectively. This indicates that the reconstructed signals, derived from multi-sensor spatiotemporal knowledge learning, not only align with the response patterns and trends of the target measurement points but also provide important feature supplements for data-driven SDI, thereby enhancing overall identification performance.