Induction motors play a vital role in various industrial applications but are prone to electrical malfunctions, particularly rotor issues like broken rotor bars. Detecting these defects is critical to reducing maintenance costs and avoiding significant financial losses. This study introduces a novel methodology that leverages deep feature extraction to identify broken rotor bar faults in induction motors. The process begins with generating spectrograms from sensor-captured signals. Rather than using these spectrograms directly, the derived power spectral density data is transformed into an image format optimized for processing by pre-trained deep neural networks. These networks are employed for feature extraction, and the extracted features are subsequently fed into k-nearest neighbors and random forest classifiers for fault detection. The proposed approach is tested on a public dataset of a three-phase induction motor under various load conditions. Remarkably, features extracted using the DenseNet201 model, particularly from the average pooling layer, achieve an impressive accuracy of 99.75% with the random forest classifier. This result demonstrates that the proposed method effectively integrates deep feature extraction with traditional classification techniques, providing a robust and precise tool for fault identification in induction motors.