This study investigates three convolutional neural network (CNN) architectures for bearing fault detection in rotating machinery using spectrogram representations of vibration signals. Vibration data from a public dataset were transformed into 224 × 224 spectrograms via short-time Fourier transform and used to train CNN models of varying complexity. The models were evaluated in terms of classification accuracy and class-wise sensitivity. Results show that even lightweight CNNs, optimized for resource-constrained environments, achieve competitive diagnostic performance, while larger architectures approach near-perfect fault classification. These findings demonstrate the effectiveness of spectrogram-based deep learning models for reliable and efficient bearing fault diagnosis.