A technique of closed-loop identification for multi-variable processes is proposed to solve the problems of identification for model predictive control. On the basis of multi-variable process input and output data during the normal closed-loop operation, the process frequency-response matrix is estimated with signal decomposition and frequency spectrum analysis, and the second order plus time delay model matrix is identified by using the linear least-squares method. Simulation examples show the robustness and accuracy of the identification method for a wide range of multi-variable processes.