In order to improve the fault diagnosis accuracy of check valve, a new method combining the complete ensemble empirical mode decomposition (CEEMD), compound screening method and improved second-order particle swarm optimization (SecPSO) is proposed. In order to select useful IMF from the intrinsic modal function (IMF) obtained by CEEMD, this paper uses the compound screening method to remove false components. In order to improve the search ability of the optimization algorithm, this paper uses the asynchronous learning factor to improve the SecPSO. Firstly, the CEEMD is used to decompose the vibration signal. Secondly, the compound screening method is used to select useful IMF and the energy features and Lempel_Ziv complexity features of these IMF are extracted. Then 10 time-domain features and 5 frequency-domain features of the IMF reconstruction signal are extracted, and all the features are combined into a mixed domain feature set. Finally, the improved SecPSO is used to optimize the parameters of least squares support vector machine (LSSVM). The proposed method is compared with various fault diagnosis methods, and the method can obtain high accuracy. The results show that the method not only removes the false components, but also improves the search ability and fault diagnosis accuracy.