Monitoring and assessing concrete member states using implantable sensing technology and enhanced long short-term memory networks

Abstract

Monitoring the state of concrete structures and assessing their performance are significant tasks for civil engineering. This study proposes a combined technique of novel concrete implantable bar (CIB) transducers and enhanced long short-term memory (LSTM) networks for monitoring and assessing reinforced concrete (RC) beam state over the whole loading process. The CIB can be installed on the inspected structure in an implantable manner. It contains an array of piezoceramic sensing units that can generate and receive ultrasonic waves from a concrete medium. To enhance the LSTM network’s ability to learn very long-series data, a time-shift energy (TSE) strategy was developed. Compared with another existing convolutional neural network (CNN)–LSTM network, the proposed TSE–LSTM network is highlighted to fully consider the ultrasound propagation characteristics in concrete when extracting sample features, instead of using simple convolution operation. A numerical study was conducted to investigate the sensitivity of the TSE feature to different concrete damage levels through mesoscale finite element models. The results provided the best parameter settings of the TSE. Eventually, to validate the feasibility of the proposed technique, an RC beam four-point bending test was carried out, in which two CIBs were implanted into the specimen for emitting and collecting ultrasonic waves in different damage states to establish a dataset. Two schemes including a classification model for predicting RC beam stress stages and another regression model for predicting the carried forces were separately investigated. The experimental results showed that the TSE–LSTM networks can successfully predict the signature of three stages of RC beams and can, in general, predict their carried forces. The comparison to the results obtained by CNN–LSTM networks further highlighted the stability and accuracy of the proposed one in learning long ultrasound series. The combined technique of CIB transducers and TSE–LSTM networks shows a promising application for monitoring and assessing RC structures.