Using deep learning models for fault diagnosis of the transmission system helps improve the smoothness and safety of train operation. However, the operating environment of trains is complex and highly variable, and fault signals from the transmission system are easily corrupted or even obscured by noise, which leads to degraded performance of the diagnostic model. Therefore, a multi-level associative memory network is proposed for accurate fault diagnosis of train transmission systems under noisy conditions. First, a dual-domain feature extraction module is proposed to capture and fuse latent information from both the time and frequency domains, thereby enabling the extraction of multi-level features. Second, a feature fragmentation encoder is proposed to partition continuous features into partially overlapping fragments with fixed lengths and strides while embedding positional information to facilitate content addressability. Subsequently, a feature fragment association reconstructor is proposed to perform content-addressable association and prediction across fragments, complete those corrupted by noise, and reconstruct continuous features through windowing and overlap-add. In addition, a gated residual connection unit is incorporated to selectively inject the reconstructed features into the original multi-level features, enhancing detail recovery and noise robustness. Finally, extensive experiments are conducted on both a self-constructed dataset and a public dataset to demonstrate the effectiveness and superiority of the proposed method. Experimental results show that, under various noise interference, the proposed method achieves average diagnostic accuracies of 94.40% and 97.96% on the two datasets, representing improvements of at least 11.15% and 2.41% over seven comparative methods, respectively. The experimental results demonstrate that the proposed method can suppress noise and achieve superior diagnostic performance, indicating promising potential for application in fault diagnosis of train transmission systems under practical operating conditions.