In response to the problems of obscured fault characteristics of metro bogie bearings under complex conditions, the susceptibility of original vibration signals to noise interference, and the poor generalization and high complexity of traditional diagnostic methods, an intelligent bearing fault diagnosis method integrating Eulerian Feature Extraction (EFE) and shallow convolutional neural networks is proposed. This method achieves dimensionless normalization and standardization of vibration signals through adaptive interval mapping, converts linear time series signals into a polar-coordinate phase representation using geometric phase encoding, and then constructs two feature expressions, FirstEuler and SecondEuler, through dual Eulerian feature fusion to achieve deep integration of time-domain and phase-domain features, effectively enhancing the discriminability of weak fault features. The features extracted by EFE can be directly used as input to a one-dimensional convolutional neural network or converted into grayscale images as input to a two-dimensional convolutional neural network, enabling the construction of shallow diagnostic models that balance diagnostic accuracy and efficiency. Using the Case Western Reserve University bearing dataset, fault diagnosis experiments were conducted under single-condition, multi-condition, and variable-speed operating conditions, and the proposed method was compared with traditional feature extraction algorithms such as VMD, EMD, and HT. Experimental results show that in a -6 dB strong noise environment, the EFE method improves one-dimensional diagnostic accuracy by more than 7.8% compared with the raw signals. In single-condition two-dimensional diagnosis, the accuracy is approximately 19.5% higher than that obtained using the raw data, and after secondary enhancement of features extracted by VMD, EMD, and other algorithms, the diagnostic accuracy reaches up to 100%. Under multi-condition and variable-speed operating conditions, secondary feature extraction by EFE increases the diagnostic accuracy by more than 15% on average; in particular, VMD combined with FirstEuler achieves 100% diagnostic accuracy for the full variable-speed condition combination. The proposed method effectively suppresses noise interference and highlight weak fault features, demonstrating excellent generalization ability and feature extraction performance under complex conditions such as constant load, variable load, and variable speed. In addition, the shallow CNN model reduces computational complexity, providing a new technical solution for efficient fault diagnosis of metro bogie bearings.