Efficient detection of the Physical Random Access Channel (PRACH) is vital for reliable initial access in 5G New Radio (5G-NR), yet it remains challenged by intra- and inter-cell interference. This paper proposes a deep learning-based solution leveraging an Autoencoder (AE) trained on synthetic PRACH data under noisy conditions. The model detects valid preambles by minimizing reconstruction error, effectively distinguishing them from interference. Simulation results demonstrate improved detection accuracy and reduced RMSE, especially with optimized latent dimensions (32–64), offering a practical balance between performance and complexity. These findings support the integration of AI-based detection in future adaptive 5G-NR systems.