Strain distribution prediction in UHPC beams using deep learning model

Abstract Over the last three decades, ultra‐high‐performance concrete (UHPC) has emerged as a highly innovative cementitious engineering material. This paper utilizes a distributed fiber optic sensor based on optical frequency domain reflectometry (OFDR) combined with the deep learning model to monitor and predict the strain distribution in UHPC T‐beams subjected to two point bending experiment. These sensors are deployed on the side surfaces of the UHPC T‐beams to capture the strain distribution when subjected to vertical loads, facilitating the determination of crack locations based on strain distribution peaks. While time series modeling is widely used in civil engineering to monitor potential damage using sensor data, its application in tracking and predicting known cracks is less explored. To address this gap, a Long Short‐Term Memory (LSTM) neural network model is developed to forecast the increase in the peak value of the strain distribution prior to structural damage, thus predicting crack initiation. The accuracy of the proposed LSTM model in predicting the UHPC strain distribution was thoroughly investigated. The results demonstrate excellent agreement between the predicted strain distributions and those detected by the ODISI 6000 monitoring system. The root‐mean‐square errors of the model‐predicted strains are generally below 10 με, with an average coefficient of determination ( R 2 ) reaching up to 98%, indicating a high degree of accuracy.