Effective Damage Identification (DI) plays a critical role in protecting structures against local or global failures caused by hazards. Real-time DI provides instant damage data and increases the safety and serviceability of civil structures. Real-time DI helps to understand the structure's behavior during extreme events that may be unknown at the design stage. This field needs innovative solutions for training supervised machine learning classifiers in the absence of measured damaged data. This research proposes an unconventional deep learning algorithm for vibration-based DI. The proposed real-time data-driven DI methodology does not require any manual feature extraction and uses Artificial Neural Networks (ANNs) to identify the presence and location of damage in discrete structural systems. The input is the response signals measured through sensors (no model-based input information required). A dropout technique regularizes the network and avoids co-adaptation in hidden layers. The neural network is optimized through 10-fold cross-validation. The proposed method's effectiveness in identifying the presence and location of damages is studied using a 4-story 2D structure subjected to artificial accelerograms. The recorded response signals create the feature space in the dataset. The lateral stiffness of columns is reduced randomly by different percentages resembling different damage severities. Considering the validation dataset results, the accuracy of the damage detection task varies from 84 to 99% for different damage severities, and accuracy for the localization task ranges from 78-98%. The results show the promising performance of ANNs for real-time DI and pave the way for training the classifiers using real-life data from undamaged structures and simulate data from damage scenarios.