The monitoring and fault diagnosis of axle-box bearings in high-speed trains is crucial for ensuring safe train operations. The vibration signals of these bearings exhibit non-stationary and non-linear characteristics. To further enhance the accuracy of identifying rolling bearing faults, a fault diagnosis method is proposed. This method is based on the improved Dung Beetle Optimization (DBO) algorithm for optimizing Variational Mode Decomposition (VMD) combined with Stacked Sparse Autoencoder (SSAE). Firstly, the DBO algorithm is enhanced to improve its optimization precision and global optimization capability. It is then utilized for the adaptive selection of two parameters: the number of decomposition modes and the penalty factor in VMD. These improvements address issues such as mode mixing, signal loss, and excessive decomposition, which arise from poor parameter selection in the traditional VMD method. Subsequently, components of Intrinsic Mode Functions (IMFs) that are highly correlated with the original signal are selected. The time-domain and frequency-domain features of these IMF components are used to construct the dataset. The feature set is then inputted into the deep machine learning model SSAE for training and testing. Through diagnostic experiments on various types and levels of rolling bearing faults, the model demonstrates a higher rate of fault diagnosis recognition.