Deep Time Series in Structural Health Monitoring of Civil Structures: A Review of Architectures, Applications and Challenges

The last few years have seen an increase in the amount of data collected in Structural Health Monitoring (SHM) systems. This can be attributed to the availability of cheap sensors and means of data transmission. However, this increase in data presents a challenge in meaningful analysis. Recently, practitioners and researchers have turned to deep learning methods for analysis of such data. Deep learning has already proven to be effective at analysing complex, highly dimensional datasets and this suits well with SHM data streams. The common data type in SHM is time series and mostly comes from vibration based SHM. Unlike static data, time series usually depict temporal dependencies and contextual variations. Further to this, time series are usually noisy and contain missing values. These attributes make the analysis of time series more complex. In deep learning, time series analysis is tackled using specialized architectures such as recurrent neural networks or through careful feature engineering. Considering these issues, it is important to understand the deep intricacies of these models for their effective application and development of new robust models.So far, different reviews have been conducted to provide a state of the art status of deep learning in SHM. However, it is clear that most of the reviews are usually application-centric and rarely consider a deep technical discussion of deep learning analysis for time series. Again, issues to do with uncertainty quantification and data augmentation are rarely discussed from a theoretical standpoint. This review seeks to tackle these issues and provide a deep theoretical review of time series, typical applications, and challenges. The goal is basically to provide the necessary background for researchers and practitioners in SHM to develop new models and to effectively apply the existing models to time series problems.