Development of a Deep Learning-Based Anomaly Detection System for Structures

Anomaly detection in structures is a vital area in structural health monitoring (SHM) especially after some extraordinary events such as earthquakes, severe accidents and so on. Recent advances in sensor technology, data science and deep learning have promoted the use of aforementioned fields in SHM. In this study, a deep learning-based anomaly detection system using actual structural ambient vibration data is proposed. Acceleration data of an ample duration in the healthy state of the structure was collected as a reference data. And another set of acceleration data of same duration was collected immediately after an extraordinary event. The time domain acceleration data was converted into spectrograms using Fourier synchrosqueezed transform. The compressed images dataset was then fed to a convolutional neural network model for training and validation. The well-trained network was able to detect the presence of any anomaly by detecting the change in time-frequency spectrograms. The proposed method exhibited excellent performance with high accuracy to predict the anomalies caused by rebars yielding.