This study examines the impact of time and frequency resolution on the accuracy of bridge damage identification using convolutional neural networks (CNNs) in acceleration spectrogram images. Acceleration responses were measured and transformed into spectrogram images using the short-time Fourier transform, allowing the data to be utilized by CNNs. These images were used to train and evaluate a CNN based on the ResNet18 architecture. The accuracy of damage detection was assessed at a total of 20 different combinations of time and frequency resolutions. All conditions showed an accuracy of more than 94%, resulting in highly accurate identification. Identification accuracy trends to increase with increasing time resolution. Identification accuracy improves with frequency resolution below 0.5 Hz. The highest accuracy was achieved with a frequency resolution of 0.5 Hz and a time resolution of 1/32 second. Although a certain level of identification accuracy can be obtained by learning only the first-order vibration after damping, it is clear that learning the higher-order mode vibration immediately after a blow, which is fast-damped, improves the identification accuracy.
TAKEYAMA et al. (Wed,) studied this question.