A Comparison of Deep Learning Algorithms for Anomaly Detection in Discrete Mechanical Systems

The application of intelligent systems for structural health monitoring is investigated. A change in the nominal configuration can be related to a structural defect that has to be monitored before it reaches a critical condition. Evidently, the ability to automatically detect changes in a structure is a very attractive feature. When there is no prior knowledge on the system, deep learning models could effectively detect a change and enhance the capability of determining the damage location. However, the acquisition of data related to damaged structures is not always practical. In this paper, two deep learning approaches, a physics-informed autoencoder and a simple data-driven autoencoder, are applied to a test rig consisting of a small four-storey building model. Modifications to the system are simulated by changing the stiffness of the springs. Both the machine learning algorithms outperform the traditional approach based on an experimental modal analysis. Moreover, the increased potential of the physics-informed neural networks to detect and locate damage is confirmed.