Hybrid Physics-Data Framework for Next-Generation Bridge Structural Health Monitoring
Francesco Basone, M. Longo, D. La mazza, Paola Daró, Giuseppe Mancini
Structural Health Monitoring (SHM) is becoming essential in civil engineering due to its ability to continuously assess the condition of infrastructures and detect potential damage. SHM techniques are generally categorized into data-driven (DD) and model-driven (MD) approaches. DD methods exploit long-term monitoring data to identify the standard structural behaviour and define relevant thresholds for anomaly detection. However, these methods typically lack labelled damage examples to train the algorithms. On the other hand, MD methods rely on updated finite element (FE) structural models to define the expected performance, paying the price of being computationally demanding, requiring substantial technical and economic resources. Recent research has therefore focused on transfer learning (TL) strategies that integrate the strengths of both paradigms. TL provides a powerful knowledge-generalization framework that allows information learned from an abundant, labelled source domain to be transferred to a related target domain where labelled data are scarce. Recent research has therefore focused on hybrid approaches (HAs) that integrate the strengths of both DD and MD paradigms. These strategies are designed to overcome the inherent limitations of each method, compensating for the lack of physical interpretability while mitigating the high computational demand. By merging physical consistency with the flexibility of statistical learning, HAs provide a more robust and reliable tool for SHM. In this work, a HA is applied to a span of a viaduct monitored with multi-year clinometer measurements. First, a FE model is developed and used to generate a large labelled dataset covering multiple damage scenarios. Subsequently, instrumental noise is stochastically injected into the synthetic data. This augmented dataset is then used to train and validate Physics-Aided Surrogate Models (SMs) based on neural networks, to ensure physical consistency while achieving high computational efficiency. Once trained, SMs are applied to real clinometer rotations to infer structural damage indicators. To further demonstrate the reliability of the system, the process is validated by integrating simulated damage scenarios with experimental monitoring data, thereby testing the framework’s robustness against complex, real-world conditions. Results, tested on a multi-set of real bridges, highlight the potential of this hybrid framework as a scalable and effective tool for next-generation SHM systems in bridges and large civil infrastructures.