<p>Blind source separation is a popular and widely used technique for output-only modal identification. However, blind source separation cannot be used to separate modal information in an under-determined problem. This study proposes a weighted joint spectral density method to more accurately and automatically separate modal information in under-determined systems. Firstly, an adaptive modal window selection based on modal assurance criteria is designed to enable the window’s bandwidth to be self-located and self-adjusted. Secondly, a tensor enhanced weighted joint spectral density is used to determine the optimal weighting matrix for obtaining high-precision modal information. Finally, numerical simulations of an under-determined spring-mass-damper system were used to demonstrate the advantages of the proposed weighted joint spectral density method over traditional weighted spectral density method in terms of accuracy and robustness to noise. The proposed weighted joint spectral density method was then applied to the measured data of Gaojia Garden Bridge. Results showed that the proposed weighted joint spectral density method reduced modal frequency identification error by 1.3% and improved the modal assurance criterion of the modal mode by 0.144 compared with the traditional weighted spectral density method.</p>