<p>Noise interference in Acoustic Emission (AE) signals presents a major challenge to accurate fault detection in railway tracks, especially under real-world field conditions. This study presents the development of an advanced Artificial Intelligence (AI) model specifically designed to address these challenges by directly processing noisy AE data for fault localization. Employing neural networks, the model incorporates a feature-weighting mechanism to prioritize critical AE parameters, enhancing its ability to identify and localize faults with precision. The training dataset was obtained from both laboratory and field experiments, where AE signals were generated through Pencil Lead Break simulations at various positions along the rail section. Each dataset comprises key AE parameters, forming a data matrix of size 1200 × 5, which is sufficiently large and diverse to ensure effective training and generalization of the neural network model. Laboratory experiments under noise-free conditions yielded highly accurate fault localization, while field data showed up to 20% error due to noise. Conventional noise reduction techniques reduced this error to below 2%, but required extensive preprocessing, limiting real-time use. To address this, a time-series neural network was developed to process raw AE signals, achieving fault localization accuracy within 1% regardless of signal-to-noise ratio. This approach eliminates the dependency on time-consuming signal conditioning, enabling real-time application. The proposed AI model marks a significant advancement in railway structural health monitoring by ensuring reliable fault detection in noisy environments. It offers a scalable, efficient, and real-time solution that enhances the safety and operational reliability of railway infrastructure.</p>