Damage Detection in a Steel Arch Bridge Model Using a Convolutional Neural Network Focused on the High Frequency Bands

The detection of bridge damage by learning the acceleration response, frequency response, etc., as images using convolutional neural networks has been shown to be effective. However, how the differences in the damage location and member types affect the dynamic response remains unclear. Clarification of these influences may be useful for improving the accuracy of damage detection and reducing the number of sensors required. This study performed a dynamic analysis under various damage cases for the finite element steel arch bridge model. As damage was observed to affect different frequency bands up to 200 Hz, a convolutional neural network trained on frequency responses in this range was able to identify the location and level of damage as small as a 10% reduction in the thickness using only two sensors. This model was also shown to be effective in cases involving multiple damage locations and levels.