ABSTRACT
An axial loading test was conducted on a steel column specimen with a box-shaped cross section, for improving structural health monitoring and non-destructive evaluation using machine learning. In the loading test, 8 acceleration sensors were placed on the surface of the column to obtain vibration characteristics at each loading step. A standard autoencoder was trained on intact frequency responses obtained before the loading test. From the results, anomaly scores calculated from the autoencoder were high around the local buckling regions. Furthermore, another autoencoder model trained on data measured from another specimen using applied transfer learning agreed with the results from the standard autoencoder. The effectiveness of coupling an autoencoder-based anomaly detection method and transfer learning when there is a lack of training data was shown.
