Deep Learning Damage Detection Using Time–Frequency Image Analysis

Abstract

AbstractRecently, vibration-based structural health monitoring (SHM) for bridges has become a trending research topic in civil engineering. However, with the collection of long-term measurement data from different sensors, accurate damage diagnosis has difficulty deciding damage features due to the effect of noise. In this paper, two-dimensional convolutional neural networks (CNN) are applied to the damage detection challenge to solve this problem. CNN is an advanced neural network system that has the ability to extract the abstract features correctly to classify damaged and normal states on the measured data. In this method, vibration signals will be transformed by a short-time fourier transform into the time–frequency representation. To enhance the distinction between bands, the frequency images have been segmented using graph representation. These images are then put to CNN to extract features before going through fully connected layers. Our proposed method is validated with measured beam data, and the results are obtained with a high level of accuracy.KeywordsDamage detectionConvolutional neural networkTime-frequency imageGraph segmentation