Vibration-Based Structural Health Monitoring of a Concrete Dam Using Autoencoder Networks

Structural Health Monitoring (SHM) of civil engineering structures is often based on continuous dynamic monitoring and Operational Modal Analysis (OMA), aimed at extracting the modal parameters that are sensitive to mass and stiffness changes. However, increasing attention is currently given to Deep Learning (DL) framework, which excels at processing large volumes of unlabelled data common in SHM applications. Compared to OMA techniques, DL procedures might allow some advantages, such as: (a) reducing the dimensionality of the input data; (b) reducing the computational efforts, as no system identification is required and (c) implicitly modeling the environmental and operational variability inside the input data. The paper focuses on the use of Autoencoder (AE), aimed at detecting and localizing damage; unlike the usual AE applications - where a specific AE network is trained for each data channel - the signals simultaneously collected by all available channels are used to train a unique AE network when the structure is assumed to be in a healthy state. After training, the resulting network should be capable of accurately reconstructing the newly collected data until the structural condition does not change. On the other hand, the increase of the reconstruction error is conceivably expected, corresponding to occurrence of structural anomalies and damage. The effectiveness and accuracy of the developed AE methodology of damage assessment is demonstrated using accelerations data collected in the continuous monitoring of the Baixo Sabor dam (Portugal), with different simulated damage scenarios being superposed to measured data. The optimal network provided excellent results in terms of detection and localization of the simulated damage.