The double row tapered roller bearing is widely used in urban rail transit, due to its complex structure, the traditional fault diagnosis is difficult to recognize the early weak fault. In order to solve this problem, a deep learning method for fault diagnosis of double row tapered roller bearing was put forward. In the experiment, the bearing vibration signals were firstly separated into a series of intrinsic mode functions by empirical mode decomposition, then the transient energy was extracted to construct the eigenvectors. In the pattern recognition, deep learning method was used to generate the fault diagnosis classifier by unsupervised study with eigenvectors. There were three states of rolling bearings in experiments, as normal, inner fault and outer fault. The results show that the proposed method is more stable and accurately to identify bearing faults, and the classification accuracy is 98%.