A methodological approach towards evaluating structural damage severity using 1D CNNs
2021-10-26

Monitoring civil infrastructure simplifies and improves reliability of decision-making in asset management. This task is increasingly important in established economies, in which engineering infrastructure has aged thus becoming exposed to various risks affecting structural integrity. In the latest paper, driven by our friends at the University of Leeds, we have explored the performance of 1D CNN in structural damage detection based on numerical simulations.

Convolutional Neural Network (CNN) is a type of machine learning in which feature extraction and identification are performed automatically. Therefore, it could prove to be a powerful tool in structural health monitoring.

Framework overview

A simple cantilever "I" beam, attached at the root by means of an elastic rotational spring, was subjected to various excitation applied at the tip by means of numerical simulations. The excitation signals varied from monoharmonic, through sine chirp to white noise. Virtual accelerometers were positioned on the beam to monitor its behaviour. The damage was introduced by reducing the stiffness of the spring.

Beam with virtual instruments

Monoharmonic excitation applied in resonant or near-resonant conditions was found to produce the best results. For the particular structure tested and instrumentation arrangement, the ease of damage detection was found correlated with the severity of damage. A greater number of sensors was required to obtain reliable results at low damage levels.

Performance of 1D CNN

More information on this exciting line of research, including results pertaining to the applicability of the method to finite element model updating, can be found here:
Almutairi, M., Nikitas, N., Abdeljaber, O., Avci, O., Bocian, M., A methodological approach towards evaluating structural damage severity using 1D CNNs, Structures 34, 2021, 4435-4446.