Stochastic subspace algorithm is widely used in bridge structure modal parameter identification. However, there remains problems like: difficulty to determine the order of a system, limitedapplication on time-invariant structures, and the subjectivity in real modal selection. In light of these problems, an improved algorithm is proposed. First, the order of the system is decided intelligently by singular entropy increment first order derivative method. Then, the parameters of the time-varying structure are identified by processing the input signal with sliding window technique. Furthermore, the real modes are identified from a similarity matrix generated based on the general rules. Finally, the signal acquisition, signal preprocessing and the improved stochastic subspace algorithm are combined, and the comprehensive method is applied to the shaking table test of a large cable-stayed bridge to verify the reliability of the proposed algorithm. The results show that the proposed algorithm can be used in bridge health monitoring, and the recognition results are credible.