Real-time structural response reconstruction via window cached attention and structured state-space layer

Deep learning methods have markedly advanced structural response reconstruction. However, most existing approaches rely on the sliding window strategy to process streaming measurements, introducing computational redundancy and latency that are incompatible with real-time requirements. This study proposes WincaS4, a neural network that integrates a novel Window Cached Attention mechanism with the structured state-space (S4) model for real-time response reconstruction. The Window Cached Attention maintains an incrementally updated key–value cache, enabling stepwise processing and reducing redundant computation over historical measurements. Evaluation on simulated and experimental datasets shows that WincaS4 achieves superior reconstruction accuracy while reducing the average inference latency to 0.003 s, thereby meeting real-time constraints under sampling rates up to 90 Hz. These results demonstrate the potential of WincaS4 as an effective solution for real-time structural health monitoring and other latency-sensitive applications.