Characterization of the modal response using Deep recurrent neural networks

• Published in: Engineering structures Vol. 256; p. 113915
• Main Authors: González, Wladimir M., Ferrada, Andrés, Boroschek, Rubén L., López Droguett, Enrique
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Ltd 01.04.2022; Elsevier BV
• Subjects: Civil engineering; Damage detection; Deep learning; Environmental effects; Frequency variation; Long short-term memory; Modal Response; Neural networks; Parameter sensitivity; Recurrent neural networks; Sensitivity; SHM; Structural health monitoring; Tracking; Urban areas; Deep learning; Modal Response; SHM; Neural networks; Environmental effects
• ISSN: 0141-0296; 1873-7323
• DOI: 10.1016/j.engstruct.2022.113915

•A building is monitored with a low density/low resolution sensor system.•Deep learning recurrent neural networks are used to predict the modal response.•Temperature and solar radiation based models are tested for the modal frequency.•LSTM and MLP layers are used in the neural network arquitecture. For Civil Engineering System Structural Health Monitoring (SHM), damage identification is typically based on the observation of appropriate response features. A commonly selected feature is the variation of modal frequency due to its high sensitivity to global damage. However, this parameter also has a high sensitivity to variables unrelated to damage, such as the weather and the structure’s usage. This article focuses on the application of Recurrent Neural Networks (RNN) with Long-Short Term Memory (LSTM) blocks to modal tracking in medium-rise buildings, a case study for which there is very little literature despite being one of the most common building types in urban areas. RNN with LSTM blocks are trained to characterize the environmental trend in the modal frequency to identify the most critical variables and to develop models than can be used to detect changes of state or damage. The models are fed with the recent history of the external temperature, sun position and the modal frequency itself. The performance of these models is evaluated in two different ways: for a variable size of the training set of real data and for scenarios with segments in which the modal response is not known at all instants, a typical situation in real structures. A practical application of this approach in a real medium-rise building is presented, showing that these models are capable of capturing with high precision the annual evolution of the modal frequency and performing well even on a daily scale, making it suitable for damage detection. For the cases in which the modal response is regularly identified and tracked, the characterization has a high performance when tracking single modal frequency or several frequencies with a single model. The models are robust for periods where data is not available but quickly deteriorate if this period extends for several days.