A time-interval-based incremental learning paradigm for progressive bridge damage detection using convolutional autoencoders

초록

Effective damage detection is critical for reliable bridge maintenance within structural health monitoring (SHM) systems. Conventional data-driven methods are often constrained to binary anomaly detection, limiting their ability to capture progressive damage that evolves incrementally in practice. To address this limitation, this study introduces a time-interval-based incremental learning paradigm employing a convolutional autoencoder (CAE) for unsupervised damage detection. The paradigm periodically updates the baseline distribution and recalibrates decision thresholds at fixed monitoring intervals (Delta T), enabling adaptive tracking of staged damage as monitoring data accumulate. A full-scale field experiment was conducted on a PSC-I bridge in Korea, where progressive tendon damage was simulated by sequentially cutting external tendons. Acceleration and strain data were collected to train and evaluate the CAE under two deployment scenarios that reflected different SHM operational stages. The strain-based CAE achieved detection accuracy exceeding 97.5% with a false negative rate of 0.0%, demonstrating reliable identification of multiple progressive damage stages. These findings highlight the novelty and practicality of the proposed paradigm, which extends beyond conventional binary frameworks. The approach provides a foundation for intelligent, data-driven maintenance strategies that can support resilient and long-term infrastructure management under real-world conditions.

키워드