Abstract
Structural health monitoring (SHM) is essential for achieving Sustainable Development Goals (SDGs) and establishing Business Continuity Plans (BCPs). SHM fundamentally involves a comparison of measured physical properties with damage indices. However, setting these indices can be a challenge due to inherent uncertainties in structures. Deep learning techniques have recently been incorporated in SHM. Previous methods primarily employed response data, although the use of external force data has also shown promise. Notwithstanding, a critical limitation of deep learning is its high dependency on large data sets. This requirement may pose a hurdle given the possible constraints in obtaining extensive structural observation data. In this paper, authors introduce a novel SHM method utilizing deep learning. Our proposed method leverages the structure's response acceleration and the acceleration of the input seismic ground motion. The wavelet transform outcomes of each acceleration are integrated. The ensuing combined data, containing three-dimensional information, necessitates an ensemble of Convolutional Neural Networks (CNNs). The classification outcomes from each CNN are subsequently connected using neuro-fuzzy to deliver the network's output. The efficiency of the suggested method was confirmed through an analysis of experimental data, achieving an evaluation accuracy exceeding 90% during validation. Furthermore, we found that each CNN's output could be utilized to predict the occurrence timing of anomalies, demonstrating the applicability of our method.
