Sparse autoencoders for unsupervised health assessment of a concrete dam

Abstract

This study explores the application of dynamic monitoring combined with Sparse AutoEncoder (SAE) networks applied to acceleration time series for identifying and locating structural anomalies or material damage and deterioration. Compared to previous approaches proposed in the literature, this work introduces a single SAE architecture that involves the response data from all instrumented locations simultaneously. After being trained with data recorded in normal (i.e., undamaged) conditions, the SAE is capable of accurately reproducing the newly collected signals provided that the structure remains unchanged. However, when structural anomaly takes place, the discrepancy between the actual and reconstructed signals, known as reconstruction error, increases so that deviations from the baseline condition are highlighted. In addition, the reconstruction error is expected to increase in areas close to the damage, thereby allowing the detection of the affected regions. The reliability and robustness of the proposed procedure against environmental and operational variations (EOV) is exemplified using two-years of data collected on the Baixo Sabor dam (Portugal), with no structural faults being observed during the considered monitoring period. Subsequently, the same data were used — properly combined with the simulation of material damage provided by a Finite Element model — to detect and localize the deterioration due to material degradation.