Artificial Neural Network-Based Prediction of Total Eccentricities in Existing Buildings Using Ambient Vibration Records

This paper introduces a neural network-based methodology for determining the total floor eccentricity along one axis in existing buildings. The method utilizes ambient vibration-measured natural frequencies and corresponding frequency response amplitudes at two distinct points on the considered floor. The developed feed-forward neural network takes fifteen input parameters: the first three natural frequencies and the twelve corresponding horizontal spectral velocities extracted from modal and steady state analyses of a single floor finite element model. The output of the neural network is the eccentricity in that direction. To cover a wide range of natural frequencies and eccentricities, both the mass and rigidity characteristics of the model are changed. The dataset obtained from these analyses is used to train, validate, and test the neural network. The predicted eccentricities from the neural network closely match the calculated ones. To validate the numerical results, a series of ambient vibration tests are conducted in a laboratory environment using a physical small-scale model of a single floor. The artificial neural network also accurately predicts most of the floor eccentricities in these tests. The proposed methodology has the potential for further extension in future research to determine the total floor eccentricities in both directions, which would be particularly valuable for existing buildings, especially those lacking structural information. Additionally, it can also be applied to assess accidental eccentricities in real buildings with full information and detailed structural drawings available.