Review on deep learning in structural health monitoring

Road bridges are fundamental and most critical elements of land transportation routes which allow to overpass many physical obstacles. Therefore, these elements have to be preserved in order to maintain structural performances over time degradation actions. At the beginning, the first monitoring techniques which were developed are related to dynamic identification which was proven to provide reliable indicators of the current health state of a structural system. Analysing the variations of modal properties (natural frequency, damping ratio and mode shapes) over a certain period of time it is possible to identify if some events or damages occurred in the structural system which determine some changes in structural properties, safety levels and structural performances. Most adopted methodologies are based on frequency domain as frequency domain decomposition and even time-domain approaches are usual such as autoregressive models, moving average models, their combination with or without exogenous term and stochastic subspace identifications. Nowadays, structural health monitoring (SHM) techniques are classified into different levels based on the level of depth of information which is provided from the only damage detection until the accurate structural diagnosis with damage identification and localization and structural prognosis. In recent years, machine learning tools have provided innovative and vibrant developments in this field especially through the deep learning (DL) approaches. These approaches provide a change of paradigm of the feature engineering approach because the feature extraction is automatically conducted in the learning phase of the network. In the present work, the most recent deep learning architectures such as convolutional neural networks, capsule neural networks, recurrent neural networks and neural transformers adopted in the SHM field are analysed and described in order to focus on the most advantages of the state-of-art approaches and to address future direction of further developments of these outstanding new technologies.