This study focused on the effectiveness of TA-MSF (Targeted Abnormality Adaptive Multi-Stage Framework) as a bridge anomaly identification method using a neural network. In this paper, to improve identification accuracy, TA-MSF method, which has two types of CNN adaptive to overall and targeted abnormality, is applied as identification method. To evaluate the effectiveness of TA-MSF method, acceleration data of vibration test measured from experimental bridge was used. The measured accelerations were converted into spectrograms and used as input images for the CNNs. In TA-MSF method, two CNNs (H-H&B-B, I-R-T) were trained to improve the identification of combinations of easily confused abnormal conditions (H and H&B, I and R, I and T). TA-MSF resulted in an improvement in probabilistic identification error of approximately 28% for H, H&B, and B, 74% for I, R, and T, and 39% overall.
SAGANO et al. (Wed,) studied this question.