Deep learning of electromechanical admittance data augmented by generative adversarial networks for flexural performance evaluation of RC beam structure

Abstract

Deep learning networks facilitate automated damage identification and performance evaluation for concrete structures using electromechanical impedance/admittance (EMI/EMA) technique, while data quantity and quality limit the performance of such a data-driven network. For the first time, this paper proposed a data-augmentation approach using deep-convolutional admittance generative adversarial networks (AdmiGAN) to solve data deficiency and measurement inefficiency for deep learning-based flexural performance evaluation of reinforced concrete (RC) structures. In the approach, a new data normalization procedure was developed to collaboratively foster AdmiGAN-based EMA data synthesis, and synthetic datasets were fed into an adaptive convolutional neural network (CNN) for deep learning. Proof-of-concept experiment was conducted on a four-point bending RC beam structure, which was continuously monitored from initial loading to final failure by three surface-bonded piezoelectric ceramic lead zirconate titanate (PZT) patches. Qualitative detection of stress and damage was performed by traditional feature analysis of EMA signatures, automated performance evaluation was attempted by using CNN approach. Results demonstrated that the AdmiGAN required merely 5 groups of EMA signatures to generate high-accuracy dataset with 174 times of speed faster than conventional measurement method, and the AdmiGAN cooperated with CNN provided a new paradigm of data-driven structural performance evaluation with high accuracy, efficiency, and intelligence.