Abstract: The automatic modal parameters identification has significant implications for real-time structural health monitoring systems. In response to the drawback of a large number of false modes in the stable diagram obtained by the stochastic subspace identification method, this study proposes a spectral clustering algorithm based on weighted distance function, and applies it to the automatic identification of stable diagrams, aiming to improve the accuracy of modal parameter identification. The principles of the spectral clustering algorithm and the stochastic subspace method are introduced, and the weighted distance function is used to calculate the similarity matrix of the nodes to be clustered, taking into account the contribution of modal data comprehensively. To improve clustering efficiency, the modal parameters of the stable diagram are pre-processed by modal verification criteria to eliminate obvious false modes. The method proposed is applied to the calculation of modal parameters of a plane truss model and of a pedestrian bridge, and the computational results show that the method proposed can effectively eliminate false modes in the stable diagram and has good robustness.