Integrating Bi-Bi-LSTM-Based Virtual Sensing and VAE for Sensor-Efficient Unsupervised Structural Damage Detection

A key component to determining bridge conditions and functionality is implementing structural health monitoring (SHM) systems that analyze structural behavior under variable loading conditions. Effective data-driven SHM requires distributed data acquisition across a structure to ensure comprehensive assessment and early damage detection. Strain-based SHM systems allow for localized monitoring because they are sensitive to minor defects. However, deploying dense sensor networks to achieve high localized damage detection sensitivity on large structures is often impractical due to installation and maintenance logistical challenges and associated costs. Virtual sensing techniques offer a solution by estimating response at key locations that are not instrumented by using data-driven techniques, which leverage statistical and machine and deep learning algorithms to map patterns and dependencies. While virtual sensing methods have been developed to estimate structural responses at noninstrumented locations, and unsupervised deep learning models have been employed for damage detection, no previous study has integrated these two approaches into a unified, sensor-efficient framework capable of detecting damage without prior knowledge of damaged states. This study proposes a novel, data-driven integrated framework that employs direct and virtually sensed strains into an SHM system. A bidirectional, bidirectional (Bi-Bi) architecture was developed and integrated into long short-term memory (LSTM) networks to predict virtual strains at locations where physical sensor deployment is impractical. By leveraging spatiotemporal relationships, Bi-Bi-LSTM significantly reduces the number of physical sensors without compromising accuracy. Virtual strains were combined with limited real-time measurements and used as input for a variational autoencoder (VAE)-based damage detection model that identifies potential damage by determining deviations between healthy state and new, unknown, state responses. Methodology performance was assessed utilizing data acquired from a series of tests of a single-span, simply supported, prestressed concrete bridge in central Nebraska, with strain time histories under variable loading conditions predicted at 11 “unmeasured” locations using data recorded from 25 measured locations. Two types of damage were introduced to the actual bridge, and results indicated that the proposed framework could accurately identify induced damage using a reduced number of sensors. By integrating virtual sensing and unsupervised damage detection into a single SHM framework, this work offers a practical, scalable, and sensor-efficient approach for real-time bridge health monitoring.