A Deep Learning-Based Adaptive Fusion Method for In-Situ Seismic Event Detection

Seismic events pose a critical threat to the safety and stability of bridge structures, with the potential to cause severe structural damage or even collapse. Structural health monitoring (SHM) systems for large-span bridges are designed to capture structural response data; however, challenges such as sensor malfunctions, environmental interference, and data transmission issues can compromise data quality and hinder accurate seismic response identification. This study introduces an adaptive weighted decision fusion deep learning method for seismic event detection, which exploits inter-sensor correlations in all directions, especially under sensor failure conditions, to improve detection accuracy. The proposed method constructs an acceleration database using time-domain, frequency-domain, and probability density curve images as model training inputs aimed at classifying sensor anomalies. Additionally, the model integrates the seismic event recognition probabilities from sensors in multiple directions, significantly enhancing the seismic event recognition capability. A global voting strategy is implemented to refine decision-making during the prediction phase. Experimental results based on real-world seismic response data from a cable-stayed bridge demonstrate the method's efficacy in accurately detecting sensor anomalies and identifying seismic events.