Three-phase induction motors are the most widely used electrical machines in industrial applications worldwide due to their robustness, low cost and reliability. Inter-turn short-circuit faults represent one of the most critical incipient failure modes in these machines and can lead to severe performance degradation if they are not detected at an early stage. Although vibration-based condition monitoring techniques have shown promising results, many recent approaches rely on complex time–frequency representations and deep learning models, which increase computational cost and implementation complexity. Therefore, a lightweight and interpretable fault diagnosis framework based on vibration signals is proposed, combining frequency-domain feature extraction and a multilayer perceptron classified. In this work, vibration signals measured by MEMS accelerometers were segmented and processed using the Fast Fourier Transform. From the resulting spectra, a compact set of spectral features, including energy, spectral centroid, bandwidth, kurtosis and skewness, was extracted and used as input to the lightweight multilayer perceptron network. Also, the method was evaluated under healthy operating conditions and multiple inter-turn short-circuit fault scenarios, considering different phases and fault severity levels. Model performance was assessed using accuracy, precision, recall, F1-score and confusion matrix analysis, along with an evaluation of preprocessing, training and inference times. The results demonstrate that the proposed FFT-based MLP framework achieves competitive classification performance while significantly reducing computational complexity when compared to deep learning approaches. These findings indicate that frequency-domain statistical features combined with shallow neural networks provide an effective and efficient solution for vibration-based inter-turn short-circuit fault diagnosis in three-phase induction motors.