Structural damage detection by multi-setup sensing: a direct data-driven approach

Abstract Implementing structural health monitoring systems on civil structures using a multi-setup sensing system presents significant advantages over fixed single-setup systems. The latter, which requires many sensors, can be exhaustive and financially infeasible. However, conventional multi-setup methods for damage detection are limited by their reliance on global mode shape identification from local modes, which lacks sensitivity to damage. This study introduces a novel data-driven multi-setup framework that bypasses traditional modal identification. Instead, it employs a two-stage convolutional neural network (CNN) approach: the first stage learns local damage-sensitive features from individual setups, while the second combines these features across different setups to directly detect structural damage from raw data. The proposed framework has been applied to both an experimental and a real-world structure, and its results have been compared with those of single-setup models. For the experimental structure, the multi-setup approach has improved the accuracy of 1D CNN and 3D CNN models by an average of 42% and 13% respectively, compared to its single-setup counterparts. The multi-setup 3D CNN model achieved an accuracy nearly matching state-of-the-art performance, despite having five times fewer simultaneous measurements on the same dataset. Moreover, for the real structure, the multi-setup approach improved accuracy by 83% compared to the best single-setup model.