Rolling element bearings (REBs) are vital for modern industrial systems, protecting against frictional damage and maintaining shaft positioning in rotating machines. However, the continuous operation under harsh conditions often results in failures, ranging from machine breakdowns to financial losses. Hence, effective diagnosis of the health of REBs is imperative for the safe and normal operations. In recent years, artificial intelligence, particularly deep learning with convolutional neural networks (CNNs), has become pivotal in diagnosing REB faults. However, CNN-based methods struggle with data imbalance, requiring substantial training data and balanced health states, which are challenging to obtain in real-world engineering contexts. To address these, the paper proposed a bearing fault diagnosis method named VI-2D-DCNN, utilizing vibrational imaging (VI) and two-dimensional deep convolutional neural network (2D-DCNN). This method employed the time-domain transform, converting 1D vibration signals into 2D grayscale images without requiring any signal preprocessing methods. The model performance was further enhanced by incorporating batch normalization and dropout operations to improve feature extraction, reduce overfitting, and enhance classification accuracy. The effectiveness was verified using CWRU bearing datasets. Experimental results showcased 100% diagnostic accuracy under 0–3 hp working loads, with 10 unbalanced health states for each condition, and a peak accuracy of 99.90% under 40 unbalanced health states. The method demonstrates excellent diagnostic accuracy and stability in handling data imbalance under varying loading conditions, showing superiority compared to existing CNN methods. Hence, these findings make a substantial contribution to the improvement of bearing fault diagnosis in rotating machines.