Steel strand is one of the most important and indispensable stress components in longspan bridges, however, there is still lack of effective method to detect and monitor the stress of the steel strand in existing bridges due to the influence of necessary anticorrosion measures. The ultrasonic guided wave carries obvious stress characteristics in the propagation along the steel strand. Through the wavelet packet decomposition of the guided wave signal in the timefrequency domain, the coefficient matrix of wavelet packet decomposition is extracted under different stress states. Then, the singular value vector of the coefficient matrix is used as the characteristic parameter, the support vector regression model with learning ability is established to detect the stress value of the steel strand. The results show that the singular value vector of the guided wave is an effective characteristic parameter in presenting stress state, and the relationship between the singular value vector distance and steel strand stress value shows a monotonous linear law in stepwise loading process. While adopting the support vector regression model established with the singular value vector to predict the steel strand stress, the determination coefficient of the results reaches to 0973 9, the prediction result obtained using the support vector regression model is more stable than that using neural network method.