Abstract: It is of great significance to identify the modal parameters of engineering structures under environmental excitations. However, the stochastic subspace identification method (SSI), as a time-domain method suitable for the identification of modal parameters under environmental excitations, will produce false modes, loss of real modes, computational efficiency, problem in automatic determination of the system order and other problems due to noise and complex excitations. These problems limit the wide application of this method in practical engineering. In this paper, a stochastic subspace identification method based on the Welch method is proposed. The Welch method is used to reduce the effect of noise, environmental excitations and other uncertainties on vibration response in the frequency domain. It aims to highlight the inherent structural modes from the noise and excitation frequencies. Then, a Toeplitz matrix which contains more structural modes is constructed, and the singular value decomposition of the matrix is performed. Finally, the discrete state matrix is calculated and the eigenvalues are analyzed. In order to determine the modal orders automatically and eliminate the false modes, fuzzy C-means clustering analysis (FCM) and modal mean phase deviation (MPD) analysis are performed on the characteristic parameters obtained from the discrete state matrix constructed by different singular value components. The method proposed in this paper is applied to the measured acceleration response analysis of a long-span suspension bridge and the acceleration response analysis of a 70-story high-rise building. The results are compared with those of frequency-domain decomposition method, traditional stochastic subspace identification method and stochastic subspace identification method based on correlation analysis to verify the effectiveness of the proposed method. It is found that compared with the traditional SSI and the SSI based on correlation analysis, SSI based on Welch method has significant improvement in avoiding modal loss and computing efficiency, and has obvious advantages in automatically identifying and eliminating false modes compared with the frequency domain decomposition method.