Continuous health assessment of civil engineering structures, especially age-old bridges reaching the end of their lifespan, is highly necessary for real-time damage detection and to prevent catastrophic failure. The implementation of artificial intelligence (AI), more preciously deep learning (DL), in the field of structural anomaly detection, is gaining importance over conventional damage detection procedures for several advantages. In this research, a novel deep learning-based real-time damage assessment method is proposed for steel bridges using dynamic acceleration response of the structure generated by moving load. A finite element model of steel truss bridge for undamaged as well as damaged members with varying severity is developed using SAP 2000 software. The vibration response of the simulated bridge is captured from different nodes when a moving train load passes through the bridge. Gaussian random noise is mixed to the numerical data to generate noisy simulated data to resemble practical scenarios. The data is then fed to the proposed deep learning-based algorithms for identifying damage location and severity. A comparative analysis is performed among different DL algorithms called multi-layer perceptron (MLP), long short-term memory (LSTM), and convolutional neural network (CNN) in terms of accuracy, loss, and evaluation matrices to analyse which deep learning algorithm is performing best. The results show that the DL algorithms successfully identified the damage locations and severities of the steel bridge from dynamic acceleration response generated by moving train. The LSTM algorithm performed more accurately and generated highest accuracy for identifying damage in steel bridge compared to other algorithms. The novelty of this research lies in the development of the proposed DL-based damage detection algorithm for identifying damage in steel truss bridge in real-time from raw dynamic response generated by moving train load traversing the bridge.