ResNet transfer learning for accurate and efficient anomaly detection of bridge vibration data

Dynamic properties extracted from bridge acceleration responses are critical for assessing safety, particularly in the context of long-span cable-supported bridges with main spans exceeding one kilometer. However, the abundance of acceleration sensors in their Structural Health Monitoring (SHM) systems is compromised by frequent failures in harsh operational environments, leading to significant issues of missing or erroneous vibration monitoring data. Recent advancements in deep learning offer promising solutions to diagnose the monitored abnormal bridge vibration data. Existing methods often rely on single-bridge vibration monitoring data, posing challenges in applying models across different bridges. To address these challenges, this study proposes a novel ResNet-based feature extraction method for bridge vibration data anomaly detection, emphasizing time-efficient classification and transfer learning. The timeseries bridge vibration responses are transformed into images to enhance computation efficiency. The proposed methodology leverages a pre-trained ResNet50 network for feature extraction, feeding extracted feature vectors into a k-means clustering algorithm for classification. Transfer learning with labelled training datasets ensures detection performance across different bridges, minimizing the required training data. Validation utilizes long-term vibration monitoring data from the SHM system of Sutong Bridge. The results aim to provide reliable technical support for data-driven condition assessment and maintenance of long-span bridges, addressing challenges in SHM systems and contributing to the safety and sustainability of critical infrastructure.