Ensemble predictive structural health monitoring of space frame structures using deep learning models

This study develops a simulation-based surrogate modeling framework for predicting the structural response of space frame structures under predefined geometric, loading, and damage scenarios. The contribution lies in integrating a space-frame-specific parametric modeling workflow with Grasshopper–Karamba3D finite element analysis to generate 28,297 labelled configurations and evaluate machine-learning models for multi-output response prediction. The input variables include member geometry, loading conditions, and prescribed damage-location/radius parameters, while the outputs are total structural mass and maximum displacement. XGBoost, LightGBM, CNN, LSTM, MLP + LSTM, and a fixed equal-weight CNN–LSTM ensemble were compared using the same held-out test set and consistent metrics. The CNN–LSTM ensemble achieved the strongest overall R² and RMSE performance, predicting mass with R² = 0.9985 and RMSE = 0.9278 tons, and displacement with R² = 0.7542 and RMSE = 1.9997 cm. However, the ensemble did not dominate every metric, and the results show that gradient-boosted tree models are strong baselines for structured tabular finite element data. The high mass accuracy mainly reflects the geometry- and material-governed nature of the mass target, whereas displacement prediction is the more demanding and SHM-relevant task. The framework is positioned as a computational tool for rapid simulation-based scenario assessment, not as an inverse damage-localization or field-deployment-ready SHM system. Operational monitoring and inverse damage identification from measured response data remain subjects for future work.