<p>The train axle has complex structures and works under various non-stationary operating conditions. The acoustic emission (AE) signals of a train axle are complicated and usually polluted by noise and interference. It is difficult to extract effective features of fatigue cracks. In addition, there are some unintelligent fatigue crack identifications for traditional AE-based methods. Aiming at these problems, an intelligent method based on acoustic emission-stacked denoising autoencoder (AE-SDAE) is proposed to identify fatigue cracks. The proposed method leverages deep learning to autonomously extract discriminative features from raw AE data, overcoming the subjectivity and inefficiency of manual feature selection commonly criticized in conventional non-destructive evaluation techniques. The proposed method eliminates the need for manual feature extraction by directly processing raw AE signals through a deep learning network, enabling automated and intelligent crack classification. Experimental validation was conducted using an acoustic emission test bench, where AE signals were collected from train axles under simulated loading conditions. The SDAE network was trained on preprocessed data, and its performance was compared with other models. Results demonstrate that the proposed method achieves a crack identification accuracy of over 98%, significantly outperforming traditional approaches. Using kurtosis-guided segmentation, the framework identifies four crack stages via AE kurtosis jumps, achieving 99.67% accuracy. These experimental results validate the effectiveness of the AE-SDAE method for fatigue crack detection and stage identification in railway axles.</p>