Deep learning-based approaches for damage detection and localisation in large-scale civil infrastructure using vibration-based monitoring methods: a review

Structural damage detection in large-scale civil infrastructure such as bridges and buildings in a timely and accurate manner is of great importance, as failure to do so can lead to catastrophic consequences. Structural health monitoring is a broad multi-disciplinary field interested in avoiding these catastrophic events and one of its main purposes is damage detection. This paper focuses on a quantitative type of structural health monitoring known as vibration-based monitoring. The emphasis is on the application of deep learning algorithms to detect and localise damage in large-scale civil infrastructure using vibration-based methods. Four types of deep learning architectures were considered in the review. These are one-dimensional convolutional neural networks, two-dimensional convolutional neural networks, deep auto-encoders and long short-term memory neural networks. From these four types of networks, convolutional neural networks were found to be most used in vibration-based methods, particularly one-dimensional convolutional neural networks. This is due to the one-dimensional nature of the time-series data acquired from vibration-based data acquisition methods, therefore allowing the data to be used as inputs to the deep learning network even in its raw form. The use of deep learning algorithms to detect and localise damage when using vibration-based monitoring methods presents many merits, such as real-time damage detection and localisation, damage detection in the presence of noise etc. However, there are also some shortcomings such as the lack of availability of damaged data for real-world structures, and the need for large datasets. The use of finite element methods to generate data and the use of transfer learning can help overcome the issues of lack of damage data and the need for large databases, respectively. To fully exploit the capabilities and power of deep learning, the algorithms must be applied in a manner that allows it to go beyond just a damage classification tool. Since vibration-based monitoring methods only focus on detecting and localising damage at a global level, the authors of this paper suggest and are currently developing a hybrid vibration-based monitoring and computer vision method. This hybrid method will be capable of autonomous damage detection, localisation and damage quantification of an entire structure at global and local levels.