Aiming at the problem that the vibration signals in train operation are complex and nonlinear, and the information of single channel signal is incomplete, a fault diagnosis method of yaw damper based on multi-channel signal fusion on car body and bogie was proposed. Firstly, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) was performed on the signals of multiple train channels, and the intrinsic mode function (IMF) was extracted to compose the feature set of refined composite multiscale dispersion entropy. Secondly, kernel principal component analysis (KPCA) was used to reduce the dimensionality of the extracted feature set. Finally, the optimal feature subset was inputted into the snake optimized kernel extreme learning machine(SO-KELM) to diagnose the yaw damper fault types. The experimental results show that the multi-channel fusion feature set optimized by kernel principal component analysis can accurately reflect the signal characteristics of different fault types of yaw damper, and realize the fault diagnosis of yaw damper. The superiority of this method is verified by comparing with other models.