This study presents a fault diagnosis method for automotive generators using deep learning. Vibration signals are collected from sensors and pre-processed by an Arduino, then converted into 224 \(\times \) 224 spectrogram images using Matlab’s STFT algorithm. These images are input into a CNN with the Au-GeNet architecture to classify generator conditions as normal, bent shaft, or faulty commutator. Experimental results demonstrate that Au-GeNet achieves 98.48% accuracy in about 114 s, outperforming AlexNet and SqueezeNet, while maintaining high generalization with only 322 K parameters across 22 layers. Thus, the study not only validates the effectiveness of converting vibration signals into spectrogram images for fault diagnosis but also paves the way for new developments in deep learning-based recognition and classification systems for industrial applications that demand high accuracy and efficiency.