Article ID: 2026EAP1197
Structural health monitoring increasingly relies on multiple sensing modalities to provide complementary information for reliable state estimation and risk assessment. In bridge monitoring, measurements obtained from different sensors often exhibit substantial heterogeneity in sampling rates, physical dimensions, noise characteristics and spatial coverage, which makes it difficult to integrate multi-source observations directly, while conventional single-source or weighted fusion methods may not adequately account for physical consistency, sensor uncertainty and degradation. To address these issues, this paper proposes a nonlinear convex optimization framework for heterogeneous sensor data fusion and risk assessment. First, multi-source measurements are transformed into a unified spatiotemporal representation, and the fusion problem is formulated as a composite convex optimization problem consisting of sensor-specific data fidelity terms and a physically consistent regularization term, for which strong convexity and coercivity establish the existence and uniqueness of the fused state estimate. Second, structural risk is represented by a nonlinear functional over a convex feasible set of admissible risk states, and the proximal operator and Moreau envelope are introduced to construct a stable and differentiable risk surrogate. Third, a distributed algorithm based on the alternating direction method of multipliers (ADMM) is developed, enabling individual sensor clusters to perform local optimization while exchanging only local estimates, and a Huber loss with residual-based adaptive reweighting is adopted to suppress faulty observations. The framework is evaluated in three representative scenarios: multi-sensor dynamic displacement reconstruction, scour risk assessment under flood loading, and distributed real-time assessment under sensor degradation. The results show that the proposed method achieves higher estimation accuracy and more stable risk assessment than single-source and conventional fusion methods. The root mean square error (RMSE) of the reconstructed displacement is reduced to 0.42 mm, compared with 2.85 mm and 3.10 mm for the GNSS-only and accelerometer-only estimates, and under injected sensor faults the variance of the risk index is reduced from 0.045 to 0.008 without false alarms. These results demonstrate the effectiveness of nonlinear convex optimization for robust multi-source sensor fusion and risk assessment.