Sensor-Outage-Aware Spatio-Temporal Graph Reconstruction of High-Rise Facade Pressure Fields

TL;DR

This paper introduces a sensor-outage-aware spatio-temporal graph reconstruction framework that accurately estimates facade pressure fields of high-rise buildings from sparse, incomplete sensor data, enhancing wind-resistant design.

Contribution

It presents a novel graph-based method that handles sensor outages and reconstructs full pressure fields, including uninstrumented locations, with high accuracy across different facades.

Findings

Reliable outage-tolerant reconstruction at instrumented sensors

Accurate full-field completion at non-instrumented nodes

Preserves dominant temporal and spatial pressure features

Abstract

Time-resolved facade pressure fields are essential for the wind-resistant design and aerodynamic assessment of high-rise buildings. However, dense instrumentation is costly and often impractical, and sensor outages can further reduce data availability. This study proposes a sensor-outage-aware spatio-temporal graph reconstruction framework for completing facade pressure fields from sparse measurements. The method couples temporal feature extraction with graph-based spatial propagation on a unified facade-domain representation and uses an explicit observation-availability indicator to handle temporarily unavailable sensor signals while reconstructing both missing instrumented channels and non-instrumented locations. The framework is evaluated using wind-tunnel pressure coefficient data for a high-rise building across windward, lateral, and leeward facades under multiple wind directions. The results show reliable outage-tolerant reconstruction at instrumented sensors and accurate full-field completion at non-instrumented nodes, with reconstruction generally most accurate on the windward facade and more challenging on the lateral and leeward facades. Time-domain, spectral, and spatial validations further show that the framework preserves the dominant temporal evolution, principal dynamic content, and coherent large-scale pressure-field organization, while the largest residual discrepancies remain localized in higher-frequency or intermittent components. A two-stage predictive extension is also outlined, in which future sensor signals are forecast at available instrumented locations and then mapped to future full-field pressure estimates through the proposed reconstruction model.