Probabilistic autoencoder-based bridge damage assessment using train-induced responses

Abstract

Structural health monitoring (SHM) systems have been increasingly employed to continually assess the current state of bridges. However, the vast amounts of sensor data generated by SHM systems, along with constantly changing environmental and operational conditions, make structural damage assessment a computationally demanding and challenging task. Traditional data-driven approaches primarily utilise machine learning methods for pattern recognition and feature extraction to address this issue. This paper introduces a methodology for assessing bridge conditions using a probabilistic temporal autoencoder (PTAE). The proposed approach effectively extracts features and captures temporal relationships in multi-sensor data collected only during train crossings. By calculating the reconstruction loss and KL divergence-based of damage features, the methodology enables the identification of potential damage of a monitored bridge. An Exponentially Weighted Moving Average (EWMA) filter and a control chart-based threshold mechanism are applied to further refine the damage assessment process, facilitating the distinction between healthy and progressively deteriorating damage cases. The proposed method is adaptable to various monitoring scenarios and sensor configurations, and robust to varying operational and environmental conditions. The effectiveness of the methodology is assessed using numerically generated data and validated with real-world data from the KW51 bridge. The results demonstrate that the proposed method can detect damage with a limited number of sensors, making it a valuable approach to enhance bridge safety.