Adaptive sparse learning-based data assimilation framework for predicting deformations induced by deep excavations adjacent to existing tunnels

Accurate deformation prediction is critical for proactive risk management of deep excavations adjacent to existing tunnels. Conventional prediction frameworks that rely on the back-analysis of soil parameters are plagued by three limitations: prohibitive computational cost, inherent ill-posedness of inverse problems, and inadequate multi-source data fusion. To avoid these issues, this paper proposes an adaptive sparse learning (ASL)-based data assimilation framework for rapid and high-fidelity deformation state estimation of the excavation-tunnel system using multi-source field monitoring data. A physics-guided dictionary is first constructed from offline numerical simulation results. Orthogonal matching pursuit (OMP) algorithm is enhanced via a proposed subspace matching metric (SMM) to enable adaptive optimal atom selection for multi-source data fusion, followed by a relevance vector machine (RVM) model for multi-task probabilistic prediction of retaining wall lateral deflection, ground settlement, and tunnel deformation. The framework is validated via a deep excavation case in Fuzhou, China. Results show that the proposed ASL framework achieves superior prediction accuracy with limited monitoring data, yields physically consistent deformation patterns, and outperforms conventional back-analysis methods in balancing multi-task prediction performance and computational efficiency. The SMM criterion enables comparatively superior adaptive atom selection, rendering it particularly suitable for multi-source data fusion scenarios.