Predicting low-cycle fatigue-induced fracture in reinforcing bars: A CNN-based approach

In low-damage structures such as rocking columns in bridges where the column ends are protected from spalling through confinement, the longitudinal reinforcing bars undergo reversing strain cycles, leading to fracture due to low-cycle fatigue. Several challenges arise due to the unsuitability of strain gauges to measure high strain ranges, in addition to the inability to visually inspect reinforcing bars after seismic events due to the use of confining details at the ends. Building on previous research that identified the fracture of reinforcing bars using column ends' rotation to estimate reinforcing bars' strains in conjunction with low-cycle fatigue models, this paper presents substantial development to predict the low-cycle fatigue-induced fracture of reinforcing bars using convolutional neural networks (CNNs) solely from strain time series data. The fracture was identified through strain measurements measured from a shake table testing of a quarter-scale, two-span resilient bridge specimen subjected to seismic excitations at the Earthquake Laboratory at the University of Nevada, Reno. The developed CNN model utilizes the Markov Transition Field technique to encode the strain time series into images and then uses the encoded images as input for channels in the input layer. These images are stacked in sequence to create a 3D array with 11 channels, one for each of the 11 different earthquake excitations that caused the damage. A three-layered CNN architecture with Adam optimizer was employed in training the model, achieving an accuracy of 100 % during training and more than 96 % during testing. To evaluate the model's performance, three distinct training/testing scenarios are proposed. The results demonstrate the efficacy of using CNNs to detect and characterize damage in structural elements using strain data. This approach has the potential to revolutionize the estimation of material fracture due to low-cycle fatigue in scientific fields using only the recoded strains.