Data-driven methods based on machine learning have been employed to predict fatigue crack propagation. However, existing studies have largely overlooked the multi-scale nature of this process. Relying solely on macroscopic data for long-crack prediction often fails to capture the complete crack growth process, potentially resulting in non-conservative predictions. Moreover, purely data-driven models often lack interpretability, exhibit limited generalization capabilities, and struggle to adhere to physical laws. The challenge of integrating modeling with reasonable interpretations, driven by both data and mechanisms, remains a significant issue for researchers. In this study, we selected 304 austenitic stainless steel as the research object. To identify the algorithm with the best predictive performance, firstly, the fatigue crack propagation prediction capabilities of three algorithms were compared: K-nearest neighbor regression (KNN), support vector machine regression (SVR), and random forest regression (RF). The most effective algorithm and implemented data augmentation strategies based on three fatigue crack propagation models were selected, namely, long cracks, short cracks, and multi-scale, to enhance prediction accuracy. Finally, using a dual-drive model framework that incorporated the data augmentation strategy, predictions of multi-scale fatigue crack propagation under different loads were conducted. Compared to the SVR and KNN algorithms, the results indicated that the RF algorithm had a lower RMSE value and higher coefficient of determination (R²), making it more suitable for predicting fatigue crack propagation. Under a load of 370 MPa, the prediction accuracy for the training set ranked in the order of RF > KNN > SVR. By contrast, the accuracy for the test set was in the order of RF > SVR > KNN. The multi-scale fatigue crack propagation model effectively captured the entire process of crack growth, whereas the long- and short-crack models accurately represented only parts of it. Data enhancement based on the multi-scale model demonstrated significantly better results than the other two models, with increases in accuracy of 25.76% and 71.74% for the training and test sets, respectively. The dual-drive model based on the RF algorithm and data enhancement strategy exhibited strong generalization capabilities. Under a load of 350 MPa, the RMSE values for the training and test sets were 0.046 and 0.111, respectively, and R² reached 0.995 and 0.961. Under a load of 330 MPa, the RMSE values for the training and test sets were 0.171 and 0.081, respectively, and the R² reached 0.911 and 0.721. Finally, compared with the pure data-driven model based on the RF algorithm, the predictions from the dual-drive model were found to be significantly more accurate.