ABSTRACT
Convolutional Neural Network (CNN) has demonstrated its excellent prospects in vibration-based structural damage detection due to its strong ability of feature learning and processing big data. However, most existing CNN-based damage detection methods can only allow damage pattern recognition, which may lead to the inability to identify and quantify unknown damage patterns. In this study, a new method called modified convolutional neural network based on Short-Time Fourier Transform (STFT) is developed for structural damage detection. The IASC-ASCE benchmark is used to evaluate the proposed method. Firstly, the vibration signals of different damage patterns of the benchmark structure are simulated based on its finite element model; secondly, STFT is performed to obtain their corresponding spectrograms; thirdly the spectrograms related to different damage patterns are fed into the CNN training; finally, a condition-based damage function is proposed to estimate the severity of any damage mode.
