Speaker
Description
Environmental variability poses a persistent challenge in the structural health monitoring (SHM) of bridges, often masking or mimicking genuine structural anomalies. This study introduces a novel unsupervised anomaly detection framework that robustly adapts to such variability through a fusion of advanced clustering and generative learning. The proposed methodology integrates three key stages: (1) Gaussian Mixture Modeling (GMM) to isolate temperature-affected data, enhanced through Generative Adversarial Networks (GANs) to address data scarcity; (2) Improved Affinity Propagation (IAP) clustering guided by a Fruit Fly Optimization Algorithm (FOA) for automatic hyperparameter tuning and enhanced clustering stability; and (3) a density-based anomaly detection module leveraging Mahalanobis distance metrics to isolate outliers from clustered data. This GAN-FOA-IAP frame-work is validated using two benchmark bridges, Z24 and KW51, featuring real and simulated modal frequency data. Results show that the method significantly mitigates environmental effects, yielding high detection accuracy, precision, and balanced recall across diverse test scenarios. Comparative analysis against other configurations (e.g., FOA-AP, GAN-FOA-AP) demonstrates the superior adaptability and reliability of the proposed pipeline. The framework does not rely on labeled data, making it highly scalable for real-world SHM systems. By dynamically adapting to environmental influences, it enables timely and accurate anomaly detection, reinforcing its utility for preventive maintenance and safety assurance of bridge structures.
