To address the challenges of sparse labels and insufficient data in mechanical fault diagnosis while enhancing diagnostic performance and model interpretability, this paper introduces a time-frequency convolutional neural network (TF-CNN) model embedded with time-frequency analysis. By integrating advanced data augmentation techniques with a convolutional architecture leveraging time-frequency transformation, the model effectively extracts multi-scale, key time-frequency features from vibration signals. The time-frequency convolutional layer combines the physical interpretability of time-frequency analysis with the autonomous feature extraction capabilities of convolutional neural networks, enabling precise identification of critical signal characteristics and fault diagnosis through classification mechanisms. Experimental validation on the CWRU dataset demonstrates that the TF-CNN model achieves a diagnostic accuracy of 99.8%, significantly outperforming baseline methods. Additionally, frequency response analysis confirms the model's ability to emphasize key signal frequency bands, further strengthening its physical interpretability. By seamlessly integrating the strengths of time-frequency analysis and deep learning, the TF-CNN model offers an innovative, efficient, and interpretable approach to industrial fault diagnosis. This work provides valuable insights and practical guidance for advancing fault diagnosis techniques, paving the way for robust applications in complex industrial scenarios.