Aiming at the problem that it is difficult to extract the fault features of railway locomotive bearings in a real complex environment, which leads to the difficulty of fault diagnosis, an improved multiscale symbolic dynamic entropy（IMSDE） fault diagnosis method is proposed. Firstly, the MSDE is improved by utilizing neighborhood slip averaging, which overcomes the defects of entropy deviation caused by traditional coarse -graining. Then, IMSDE is used to fully extract the key fault features of vibration signals at different scales. Finally, the identification of different fault types and degrees of railway bearings is achieved by combining with an extreme learning machine（ELM）. On this basis, three separate sets of tests were analyzed. The results show that the method has an accurate fault identification effect for both artificially constructed bearing faults and bearing faults generated by engineering reality, and the fault identification rate is higher compared with the other four methods, which verifies that the method has a certain value of practical application in engineering.