Convolutional neural network-based real-time event classification on a cable-stayed bridge with spatially distributed IoT sensors

This study proposes an automated causal analysis framework for bridge safety management using deep learning-based Internet of Things (IoT) sensor data analysis. Conventional methods rely on statistically derived thresholds for individual sensors, requiring expert intervention when reading exceeds these limits. However, such approaches hinder real-time safety responses and lack causal analysis capabilities for alarm triggers. To overcome these technical challenges, we developed a convolutional neural network (CNN)-based diagnostic framework for real-time IoT sensor data processing and instantaneous event classification. First, a spatial sensor matrix (SSM) was established from spatially distributed IoT sensors installed on Jindo Bridge, South Korea, followed by comprehensive statistical analysis of 2 years of operational data correlated with historical events. Subsequently, the CNN was trained using time history SSM data augmented with statistically synthetic event-based variations to address rare-event scenarios. The developed framework was validated with untrained real-time sensing data obtained from Jindo Bridge, demonstrating 97.46% accuracy in classifying designed events across 427,326 SSM instances. Key innovations include (1) real-time event classification via data-driven analysis without experts’ intervention, (2) enhanced reliability and scalability using adaptive thresholds for each IoT sensors, (3) improved CNN training efficiency via statistical data augmentation with respect to low-probability historical events. This framework advances bridge health monitoring by enabling proactive maintenance through reliable and real-time diagnostics.