A Privacy-Preserving Framework Using Federated Learning for Structural Health Monitoring with Miter Gates Application

Abstract

Structural Health Monitoring (SHM) systems play a critical role in maintaining the safety and operational efficiency of infrastructure such as miter gates managed by the United States Army Corps of Engineers (USACE). These gates, integral to navigation and flood control systems, demand robust and efficient monitoring techniques to predict failures and thus optimize maintenance schedules. Any unexpected shutdown costs nearly three million dollars per day to the US economy. The existing traditional data centralized machine learning approaches for SHM often face challenges, including data privacy concerns, high communication costs, and computational limitations. This study mainly explores the application of Federated Learning (FL) techniques to SHM systems using three finite element miter gate models for the loss-of-contact damage detection problem. This approach addresses these challenges by enabling decentralized training of machine learning models across multiple assets. FL ensures data privacy by keeping sensitive information local while leveraging shared global models through aggregation methods. Our results demonstrate that FL can achieve comparable prediction accuracy to centralized methods while maintaining data privacy and reducing communication overhead. This framework improves model accuracy by incorporating diverse data distributions from different gates and their operational conditions. The study also highlights the scalability of FL in handling large-scale SHM systems, making it a viable solution for USACE to extend the lifespan of their critical assets.