To improve fault feature representation and diagnostic performance for bearing fault diagnosis, this study proposes a fault diagnosis method that integrates time-frequency analysis, deep feature extraction, feature selection, and ensemble learning. Continuous Wavelet Transform (CWT) converts bearing vibration signals into time-frequency images and preserves fault characteristics in both the time and frequency domains. A pretrained model extracts deep features from the generated images and improves fault feature representation. Feature selection algorithm based on nearest neighbors (ReliefF), maximum relevance minimum redundancy, and least absolute shrinkage and selection operator remove redundant features and retain discriminative features. Random Forest, eXtreme Gradient Boosting, and Light Gradient Boosting Machine (LightGBM) classify bearing faults. This study compares different combinations of feature selection methods and diagnosis models and evaluates their diagnostic performance. Experimental results show that the combination of ReliefF and LightGBM achieves the highest diagnostic accuracy of 95.72%. This method provides reliable technical support for fault diagnosis, condition monitoring, and health management of rotating machinery and promotes the development of intelligent fault diagnosis technologies.