Stator inter-turn faults (SITFs) are electrical abnormalities in the windings of a motor or generator, resulting from short circuits between adjacent coil turns, potentially leading to reduced performance or even catastrophic failures. This paper aims to detect SITFs and classify their level of severity using a combination of prominence-based features and recently developed neural networks that rely on self-attention mechanisms. The approach involves transforming 3-phase currents using the extended Park Vector approach (EVPA), extracting features based on prominence from the frequency spectrum, and studying their geometry to gain important insights about the data. After this feature-engineering and data exploration step, neural-based classifiers have been trained and tested. Through a comparative study with other neural-based approaches, the Transformer Encoder achieves the highest classification accuracy of 97.25% when tested using the experimental data, outperforming other trained networks. The authors also present the importance of self-attention maps for exploring the interpretability of the Transformer Encoder, revealing the significant contribution of the prominence-based features in classification.