Deep learning for structural health monitoring under environmental and operational variations

Structural Health Monitoring (SHM) techniques can be classified into global and local monitoring strategies. Regarding the global strategies, the aim is to monitor the structure for any change that can be related to damage globally, whereas, the local damage detection schemes usually aim at detecting damage at a very confined area on the structure. The global techniques, which are also sometimes termed as vibration methods, are very sensitive to Environmental and Operational Variations (EOV). These variations can affect the structural response and, subsequently, mask any changes in vibration signals that can be referred to as damage. It is known that temperature has the most effect on the natural frequencies of any structure. In this paper, an inverse strategy is proposed that aims to predict the temperature variation using frequency time series of the structure. To this end, a Recurrent Neural Networks (RNN) is exploited in which the natural frequency time series obtained from the intact structure along with recorded temperatures are respectively used as training features and label. The trined RNN is then tested on data obtained from the damaged structure. It is shown that the error in predictions increases as the damage occurs. A numerical example is presented to demonstrate the applicability of the proposed method.