Accurate fault diagnosis of electric drive rolling bearings is crucial for ensuring the reliability and safety of new energy vehicles. However, obtaining sufficient labeled fault data for training robust diagnostic models remains a significant challenge, particularly for newly deployed systems or those operating under diverse conditions. This paper proposes a novel fault diagnosis method for electric drive rolling bearings using small sample data and transfer learning. The method leverages the SqueezeNet model as a feature extractor, combined with a transfer learning strategy, to enhance both the accuracy and computational efficiency of rolling bearing fault diagnosis. First, acceleration vibration signals from the rolling bearing are acquired in real-time under its current operating conditions, followed by signal preprocessing. The Continuous Wavelet Transform (CWT) is then applied to convert the processed signals into time–frequency images, which serve as inputs for the SqueezeNet model. The model, pre-trained on a large-scale dataset, has its convolutional layers frozen while the final fully connected layer is fine-tuned to adapt to the task of rolling bearing fault diagnosis. Finally, the model is trained and validated using the rolling bearing fault dataset, demonstrating the diagnostic performance under various working conditions and loads. This approach provides a promising solution for practical applications where collecting large amounts of labeled fault data is challenging or cost-prohibitive.