Automated identification of compressive stress and damage in concrete specimen using convolutional neural network learned electromechanical admittance

Abstract

For the first time, this paper proposed a simple two-dimensional convolutional neural network (2-D CNN) integrated with electromechanical admittance (EMA, inverse of impedance) to detect compressive stress and load-induced damages of concrete cubic structure subjected to applied loading, which was attempted to solve the difficulty for stress/damage quantification merely based on traditional EMA technique. It automatically resolved the raw EMA signatures with no need of complex transformation or manual feature extraction. In the experimentation, the whole process of the structural states from initial loading to fatal failure was monitored via using surface-bonded piezoelectric ceramic lead zirconate titanate (PZT) transducer. Independent EMA databases were established at each stress/damage state of the structure. Qualitative detection of stress development and cracking damage progression as three stages of failure mode was analysed via using characteristics of EMA spectra encompassing resonant frequency and amplitude values. Quantifiable assessment of stress level and damage severity was attempted using the raw EMA signatures learned by the proposed CNN model without signal pre-processing. Experimental results demonstrated that EMA signature was sensitive to stress variations in concrete and its cracking, expansion and propagation. The proposed approach possessed excellent accuracy and efficiency superior to traditional root mean square deviation (RMSD)-based one for quantification on compressive stress and damage state changes in the specimen, which potentially provided an alternative paradigm of data-driven stress/damage identification in concrete structures.