This study presents a hybrid deep learning framework for vibration-based damage detection and localization in reinforced concrete (RC) bridge girders using finite element (FE) simulations and discrete crack modelling in Abaqus software. A wide range of crack scenarios including undamaged (UD), shear (SH) and flexural (FL) cases were generated from a scaled-down girder with varying crack parameters like location, depth, width and inclination. The acceleration time history response generated for each of these cases are then subjected to multiple forms of artificial noise of up to 10% variations to emulate real field conditions. A dual-input hybrid neural architecture was implemented that integrates handcrafted vibration features with automatically learned temporal features from a 1D-CNN. The model was trained using class balanced weighting and regularization techniques, achieving a classification accuracy of more than 97%. Dimensionality reduced visualizations like PCA and t-SNE revealed clear representation of separability of damage and undamaged states. The results highlight the ability of the proposed framework to discriminate between UD, SH, and FL cases, while exhibiting strong robustness under noise-perturbed conditions. This work demonstrates that combining FE-generated synthetic datasets with hybrid deep learning provides a scalable, automated, and highly accurate approach for SHM of RC bridge girders, with strong potential for real-world deployment.