Time‐frequency enhanced CNN models for nonlinear strain prediction in RC structures

This study proposes convolutional neural network (CNN)‐based models for predicting the nonlinear strain responses of reinforced concrete (RC) structures subjected to seismic loads. Using convolutional neural networks (CNNs), the study establishes the relationships between the global structural responses—such as acceleration and displacement responses—and local strain responses in RC structural members. These global responses are transformed into the time‐frequency domain using continuous wavelet transform (CWT) to capture nonlinear characteristics more effectively. Eight prediction models were developed with varying input configurations, including response amplification near the natural frequency of target structures and combinations of acceleration and displacement data. The models were trained and tested using both nonlinear time‐history analyses and shaking table experiments. Comparative evaluations revealed that models incorporating both time and frequency domain responses—particularly the model combining CWT data with acceleration and displacement inputs—achieved the most accurate strain predictions. The best‐performing model reduced prediction error by up to 40.20% under strong seismic input compared to the baseline model using only time‐domain data. These findings highlight the potential of time‐frequency enhanced deep learning for seismic strain prediction in RC structures.