Consistent Seismic Event Detection Using Multi-Input End-to-End Neural Networks for Structural Health Monitoring

Abstract

Seismic events pose a significant threat to the safety of bridge structures, potentially causing extensive structural damage or collapse. Structural health monitoring (SHM) systems for long-span bridges capture structural response information and generate substantial data but face issues like sensor faults, environmental noise, and data transmission problems that can degrade data quality and hinder accurate seismic response identification. To address the problem, a multi-input end-to-end deep learning method for seismic event detection is proposed. Vibration data of different directions are separately utilized, and the interference of multi-type anomalous data is considered. First, the segmented acceleration time series were transformed into time-domain, frequency-domain, and probability density curve images, respectively, to form three-channel images; then, images from three directions were input to the neural network in parallel. Back-end architectures are constructed based on two fusion strategies, i.e., decision fusion and feature fusion. Consistent detection results across three-dimensional image sets can be obtained by the end-to-end architecture. A global voting process is implemented to further fuse the detection results of different image sets at the same moment. The proposed method is verified using data from two actual seismic events of a cable-stayed bridge. The results show that the proposed method can consistently and accurately detect seismic events even with interference from anomalous data. Among them, the feature fusion method has higher seismic event detection accuracy, while the decision fusion method offers a certain level of interpretability.