An on-line algorithm for T-S model identification is proposed, which automates the structure and parameter identification simultaneously based on input-target samples. Firstly, several subspaces are produced in the input space, and  their shape and distribution are optimized on-line. Then the recursive least square algorithm is employed to update the parameters of each local model, which makes each local model approximate the real system. If a new subspace is produced or the shape of the subspace is changed, an adjustment method is used for each local parameters and data matrix. Finally, a simulation experiment on a dynamic nonlinear system and the gas furnace data is given to verify the effectiveness of the proposed method.