Induction motors play an important role in a variety of industrial applications but are particularly sensitive to electrical faults, such as rotor-related problems such as broken rotor bars. Eliminating such faults is critical to reducing maintenance costs and preventing serious financial losses. This study presents a method based on detailed feature extraction for identifying broken rotor pull-out faults in induction motors. The process is initiated by generating spectrograms from sensor-based signals. However, instead of using these spectrograms directly, the resulting power spectral density data is converted into an optimized image format suitable for processing by pre-classified deep neural networks. To utilize these networks’ capabilities, the developed features are fed into nearest neighbor (k-NN) and random forest classifiers for fault detection. The programmatic method was tested on a publicly available dataset of a three-phase step-down motor operating under various load conditions. In particular, the DenseNet201 model’s improved features from the mean pooling structure yielded a remarkable accuracy of 99.75% using the random forest classifier. This result demonstrates a powerful and sensitive fault detection tool in induction motors by effectively integrating the conventional circuit techniques with detailed extraction by the proposed method.