Data-driven SHM of a timber lamella dome using deep learning

Timber lamella domes are structurally efficient spatial systems whose long-term performance is challenging to monitor due to complex load paths, limited accessibility, and sensitivity to environmental effects. This study presents a data-driven structural health monitoring (SHM) framework that integrates parametric finite-element modelling with machine learning to enable real-time prediction of three-dimensional structural displacements. A Chaos-Integrated Synaptic-Memory Network (CISMN) is trained using benchmark datasets derived from a calibrated finite-element model under a wide range of seismic load scenarios. To support practical deployment, model inputs are defined as sensor-derivable proxy features, including directional load and dominant frequency indicators. Model performance is evaluated beyond point accuracy through generalisation-gap analysis, uncertainty quantification via conformal prediction, feature attribution using SHAP, and threshold-based alarm diagnostics. The results demonstrate high predictive accuracy with stable generalisation under stress-tested conditions. Incorporating predictive uncertainty into alarm logic substantially reduces missed detections at serviceability limits. Benchmark comparisons with standard neural and ensemble models show that CISMN offers improved robustness under identical evaluation conditions. While the framework is evaluated using datasets derived from numerical analysis, the methodology provides a clear pathway toward uncertainty-aware and explainable SHM of complex timber spatial structures using real sensor data.