This paper explores the Deep Belief Networks (DBNs) in the application of high-speed train vibration signals processing.Firstly, a new method based on DBNs is proposed.The vibration signals are preprocessed by Fast Fourier Transform (FFT) and then the FFT coefficient-vectors are used to set the states of the visible units of DBNs.DBNs can be trained greedily, layer by layer to learn highlevel features, using a model referred to as Restricted Boltzmann Machine (RBM).For more efficient learning, a model called K-DBNs is proposed by combining the respective advantages of K-Nearest Neighbor (KNN) and DBNs.Finally, a method of optimizing each layer in deep network is introduced by integrating unsupervised learning with supervised learning.Experiments on real datasets and simulation datasets confirm that the proposed methods may learn useful high-level features and diagnose different faults of high-speed train.