Anomaly detection in concrete structures using autoencoder models customized by transfer learning

Abstract

Abnormality detection in aged reinforced concrete structures using local vibration measurements is investigated. In addition to the conventional local vibration method focused on the localized resonant frequency, three machine-learning-based models are evaluated: (a) a dedicated model, (b) a generalist model, and (c) a novel approach that enhances the capabilities of the generalist model using transfer learning. Experimental results on a damaged pre-stressed slab indicate that machine-learning methods may effectively evaluate structures under various damage patterns and complex boundary conditions. The findings affirm the potential of the transfer learning approach for effective abnormality detection in concrete structures, particularly where detailed initial data is unavailable.