Damage diagnosis in beam-like structures by artificial neural networks

Creative Commons License

Aydin K., Kisi O.

JOURNAL OF CIVIL ENGINEERING AND MANAGEMENT, cilt.21, sa.5, ss.591-604, 2015 (SCI-Expanded) identifier

Özet

Applicability of artificial neural networks is examined in determining the natural frequencies of intact beams and crack parameters of damaged beams. Multi-layer perceptron (MLP) and radial basis neural networks (RBNN) are utilized for training and validation of input data. In the first part of the study, the first four frequencies of free vibration are predicted based on beam properties by the networks. Showing the effectiveness of the neural networks in predicting the vibrational frequencies, the second part of the study is carried out. At this stage of the inverse problem, the frequencies and mode shape rotation deviations in addition to beam properties are used as input to the networks to determine the crack parameters. Different hidden nodes, epochs and spread values are tried to find the optimal neural networks that give the lowest error estimates. In both parts of the study, the RBNN model performs better. The robustness of the network models in the presence of noise is also shown. It is shown that the optimal MLP network predicts the crack parameters slightly better in the presence of noise. As a conclusion, the trained RBNN model can be used in health monitoring of beam-like structures as a crack identification algorithm.

Applicability of artificial neural networks is examined in determining the natural frequencies of intact beams and crack parameters of damaged beams. Multi-layer perceptron (MLP) and radial basis neural networks (RBNN) are utilized for training and validation of input data. In the first part of the study, the first four frequencies of free vibration are predicted based on beam properties by the networks. Showing the effectiveness of the neural networks in predicting the vibrational frequencies, the second part of the study is carried out. At this stage of the inverse problem, the frequencies and mode shape rotation deviations in addition to beam properties are used as input to the networks to determine the crack parameters. Different hidden nodes, epochs and spread values are tried to find the optimal neural networks that give the lowest error estimates. In both parts of the study, the RBNN model performs better. The robustness of the network models in the presence of noise is also shown. It is shown that the optimal MLP network predicts the crack parameters slightly better in the presence of noise. As a conclusion, the trained RBNN model can be used in health monitoring of beam-like structures as a crack identification algorithm.