A Physics-Guided Feature Fusion Network for Vibration-Based Bridge Scour Monitoring

Abstract

Bridge scour poses a significant risk to infrastructure safety, yet traditional underwater monitoring methods are often unreliable and impractical during critical flood events. To address these limitations, this study proposes a physics-guided feature fusion network (PG-FFN) for estimating scour depth from bridge pier acceleration data. The PG-FFN features a novel dual-branch architecture design. A temporal feature branch uses a bidirectional long short-term memory (BiLSTM) network to extract latent patterns from raw vibration signals. Concurrently, a physical feature branch processes a set of engineered, physics-based descriptors using a multilayer perceptron. These distinct feature sets are integrated through a fusion module to produce a comprehensive representation of the structural response. To ensure the model’s predictions adhere to fundamental physical principles, a composite loss function is introduced to strictly enforce monotonicity constraints, while the non-negativity of the scour depth is explicitly guaranteed by the network architecture. For proof of the concept, the proposed model was trained and evaluated on a comprehensive synthetic data set generated from a calibrated numerical model. Results show the PG-FFN achieved a normalized root mean square error (NRMSE) of 7.60% on the unseen test set, representing a 33% improvement over a standard baseline BiLSTM model, which yielded an 11.31% NRMSE. An ablation study confirmed that the feature fusion mechanism was the primary contributor to this enhanced performance. Moreover, field validation on a real flood event demonstrated the model’s robustness in capturing scour trends under practical conditions. The findings demonstrate that integrating physical constraints and domain knowledge with deep learning provides a more accurate and physically consistent framework for vibration-based bridge scour monitoring.