Unsupervised machine learning-based approach for damage detection using autoencoder neural networks

Resumo

Damage detection is one of the main functions of structural health monitoring. Large structures such as bridges and overpasses require real-time monitoring to identify anomalies that could compromise their urban function. With this in mind, the present work proposes a Machine Learning approach based on autoencoder neural networks for damage detection. Since this approach is grounded in an unsupervised methodology, it does not require labeled data (which aligns with engineering practice) and is robust enough to overcome challenges such as the cost of modal identification and the need for knowledge about the structure’s excitation source. It only requires signals as training data and can subsequently be applied in practice for real-time monitoring of special engineering structures.