A data-driven digital twin model for bridge health monitoring using feature fusion and unsupervised deep learning

The progressive degradation of bridge infrastructure poses significant risks to transportation safety, necessitating advanced monitoring and management strategies. Digital Twins (DTs) have emerged as a transformative technology for real-time structural monitoring, offering a dynamic and data-driven representation of physical assets. When combined with Deep Learning (DL), DTs enhance predictive capabilities by leveraging vast amounts of sensor data to detect structural anomalies, optimize maintenance strategies, and improve decision-making processes. This synergy enables a more accurate, automated, and cost-effective approach to infrastructure health monitoring, ensuring timely interventions and prolonged service life. This paper presents a novel data-driven DT framework for real-time bridge damage detection, incorporating two key innovations. First, a Multi-Domain Fusion Autoencoder (MDFA) that integrates raw acceleration data with feature fusion from time, frequency, and wavelet domains. Second, a synthetic anomaly simulation approach for evaluation. The framework employs a transformer-based Generative Adversarial Network (GAN) to generate realistic synthetic signals, which are modified to simulate various damage scenarios. Evaluations on Australia's Werrington Bridge demonstrate the MDFA's ability to detect subtle anomalies often overlooked by conventional unsupervised DL methods. A comparative analysis confirms its superior detection capabilities. Further validation on Austria's S101 Bridge using real damage data successfully identifies and localizes all damage scenarios without requiring additional optimization. These results highlight the efficiency of feature fusion techniques in improving anomaly detection accuracy and suggest that synthetic anomaly simulation offers a cost-effective alternative to laboratory testing for evaluating unsupervised damage detection approaches.