Deep learning algorithms have shown significant potential for analyzing time-series data and are increasingly used in vibration-based damage detection. However, most research has focused on damage detection in individual structures, often overlooking the need to monitor and analyze populations of standardized components. This paper focuses on population-based damage detection, a novel approach aimed at identifying damage across populations of similar structures. Specifically, it investigates the use of Bidirectional Long Short-Term Memory networks and Gated Recurrent Units for damage detection in a small population of nominally identical beams. To capture beam-to-beam variability, training data are generated using a random coefficients model, calibrated with experimental data. The model’s hyperparameters are optimized through grid search to achieve the best detection accuracy. Finally, the performance of the optimized models is evaluated using experimental data from the population of beams.