ABSTRACT
In this paper a supervised deep learning based approach to extract bridge natural frequencies from acceleration sensors on a passing train is presented. To overcome the problem of obtaining a sufficient amount of acceleration signals with a known corresponding bridge frequency, it is proposed to use transfer learning. Data acquired from simulations is used to learn the features and their weighing is trained on real measurement data. To validate the proposed method synchronous acceleration measurements at an ICE 4 train and a double track steel box girder bridge from a field measurement campaign were used. The model presented in this paper is able to extract the natural frequencies from a single sensor and a single passage even while high speeds. With the experimental data it is possible to extract bridge natural frequencies from acceleration measurements on a passing train under regular operational conditions.
