Novelty detection and localization using vibration monitoring and sparse Autoencoders

Dynamic monitoring of structures and infrastructures, complemented by the automation of Operational Modal Analysis (OMA), represents a classic strategy of Structural Health Monitoring (SHM), mainly based on the correlation between modal and structural parameters. More recently, an increasing interest is growing in applying Deep Learning (DL) frameworks to SHM. This interest is driven by the advanced processing power of DL networks, which can model essential data features using layers of artificial neurons. Among DL algorithms, the sparse Autoencoder (SAE) has shown promise for SHM because SAE networks: (a) can discover hidden features by reducing the dimensionality of input data and (b) are believed to inherently handle the environmental and operational variability (EOV), so that additional steps of cleansing features are not needed. Unlike traditional methods, which train separate networks for each channel of data, the procedure herein discussed trains a single network using simultaneously all available channels from the structure. Once trained, the SAE can be fed with new data and reconstruct the original signals: if the structure’s condition and EOV did not change after training, the SAE should accurately reconstruct the input data. Hence, the structural health is evaluated by calculating the Mean Absolute Error (MAE) between the input and reconstructed data, with deviations indicating potential damage. In addition, higher MAE values are expected in damaged areas, assisting in identifying the presence of critical regions. The application of the proposed SAE-based methodology is demonstrated using data collected on one benchmark bridge.