The real-time dynamic deflection signal separation of the long-span bridge is an effective new method for the long-span bridge fault diagnosis. Because empirical mode decomposition (EMD) can decompose nonlinear and non-stationary signals into a set of linear and stationary intrinsic mode functions, a blind source separation (BSS) method based on EMD is proposed that combines the respective advantages of BSS and EMD for single-channel long-span bridge deflection signal separation, while EMD can decompose the nonlinear and non-stationary signals into a set of linear and stationary intrinsic mode functions. The source number is estimated using singular value decomposition(SVD), and then to composes the multi-channel input signals with the single-channel deflection signal and its intrinsic mode functions according to the source number. Finally, the method separates the source signals using the fast independent component analysis (Fast ICA) algorithm and obtains the separation value of each component of the bridge deflection signal. Simulation research indicates that this is a good solution for estimating the quantity of source number estimation for the independent component analysis (ICA) model and single-channel deflection signal BSS problem.