Structural Health Monitoring of prestressed concrete beams using different CNNs architectures

Resumo

There are several damage detection techniques that use signal data and can evaluate and ensure the safety
of a structure. Recently, machine-learning algorithms have been used to help classify and detect damages as well as
extract structural features from signal data. Learning algorithms have the advantage of using raw data or data with
minimum pre-processing as input. Convolution neural networks (CNN) is a supervised learning technique that
uses a combination of filters and pooling layer to extract features and classify these types of data simultaneously.
The performance of CNNS can vary drastically due to the architecture and the input shape of the input selected. In
this study it is compared different CNNs proposed by the literature in order to identify damage presented in a set
of eight identical prestressed concrete beams. Tests with sixteen 1D and 2D different CNNS were conducted with
results of accuracy, recall and f1-score varying from 50% to 99%.