Detection of Failures in Civil Structures Using Artificial Neural Networks

This paper presents an approach to failure detection in civil structure using supervised learning of data under normal conditions. For supervised learning to work, we would typically need data of anomalous cases and normal conditions. However, in reality there is abundant of data under normal conditions, and little or none anomalous data. Anomalous data can be generated from simulation using finite element modeling (FEM). However, every structure needs a specific FEM, and simulation may not cover all damage scenarios. Thus, we propose supervised learning of normal strain data using artificial neural networks and make prediction of the strain at future time instances. Large prediction error indicates anomalies in the structure. We also explore learning of both temporal trends and relationship of nearby sensors. Most literature in anomalies detection makes use of either temporal information or relationship between sensors, and we show that it is advantageous to use both.