The inverse problem of force identification is always ill-posed due to the inverse process of structural matrix and measurement noise in practical engineering and the ordinary least square method always obtains the meaningless solution. An improved regularization method is proposed to overcome these problems. The loads can be expressed as a series of kernels of impulse function in time domain and the total response of the system can be obtained using the product of the convolution integral of the kernel response and the loads. The essence of ill-posedness of force identification problem is further disclosed in detail through the singular valued decomposition technique, and the corresponding regularization technique is presented to solve this inverse problem in order to alleviate the ill-posedness. A new method for choosing regularization parameter based on the Morozov compatibility principle is also used. The loads are identified in single input and single output system and two inputs and two outputs system respectively. Simulation results show that the identification method proposed is effective and it can meet the engineering requirements of stable approximate solution.