Data-efficient memory-augmented autoencoder for real-time daily anomaly detection in bridge health monitoring

Daily anomaly detection in bridge structural health monitoring (SHM) demands models that can operate under limited labeled data, severe class imbalance, and strict real‑time constraints. This paper presents a data‑efficient memory‑augmented autoencoder (AE) framework designed for real‑time daily anomaly detection in operational SHM environments. The method employs an AE to extract compact latent features from raw sensor data, applies synthetic minority oversampling in latent space to address imbalance, and integrates a simplified memory‑augmented classifier capable of rapid adaptation from few samples. An adaptive daily update loop incorporates relabeled misclassified instances for retraining, enabling same‑day deployment. Evaluations on the IPC−2020 Sutong Bridge dataset and a newly labeled Yonghe Bridge dataset demonstrate that the proposed approach matches or exceeds the accuracy of state‑of‑the‑art baselines while reducing daily update time to under 40 s. By combining smart sensing, efficient learning, and scalable deployment, the framework supports resilient SHM systems capable of timely fault detection and continuous infrastructure safety assessment.