Due to the importance of induction motors in a wide variety of industrial processes, it is crucial to properly identify abnormal conditions in order to avoid unexpected stops. The inter-turn short circuit (ITSC) is a very common failure produced with electrical stresses and affects induction motors (IMs), leading to catastrophic damage. Therefore, this work proposes the use of the empirical wavelet transform to characterize the time frequency behavior of the IM combined with a self-organizing map (SOM) structure to perform an automatic detection and classification of different severities of ITSC. Since the amount of information obtained from the empirical wavelet transform is big, a genetic algorithm is implemented to select the modes that allow a reduction in the quantization error in the SOM. The proposed methodology is applied to a real IM during the start-up transient considering four different fundamental frequencies. The results prove that this technique is able to detect and classify three different fault severities regardless of the operation frequency.