Rolling element bearings are crucial for supporting rotating shafts in industrial machines. However, their health status significantly impacts machine reliability and safety, highlighting the need for effective fault diagnosis. Recently, deep learning, particularly convolutional neural networks (CNNs), has become pivotal in diagnosing bearing faults. However, current CNN-based approaches face challenges such as excessive trainable parameters, complex hyperparameter tuning, and long training times. In addition, models require a large amount of data with balanced health states, posing challenges in real-world scenarios. To address these issues, this paper proposes a new approach using pretrained deep transfer learning with a Visual Geometry Group network of 19 layers and global average pooling (VGGNet19-GAP). The framework presents four key improvements: (1) it transforms raw vibration signals into RGB images using a simple continuous overlapped window stride segmentation; (2) integrates global average pooling in VGGNet-19 to reduce parameters and prevent overfitting by removing fully connected layers; (3) applies batch normalizations to top networks to enhance stability and training speed; and (4) adopts a training strategy where bottom network parameters are frozen, except the last two layers, while fine-tuning top layers for improved training efficiency. The effectiveness was evaluated on the CWRU motor bearing datasets, achieving diagnostic accuracy rates of 99.70%, 99.66%, 99.94%, and 100% under four varying loading conditions (0–3hp), respectively. The method demonstrated excellent diagnostic accuracy and stability in handling data imbalance across varying loading conditions, outperforming existing CNN variants. These findings significantly contribute to the advancement of bearing fault diagnosis in rotating machines.