It is difficult to directly measure the dynamic load due to the complex forms of the structure and the excitation. Therefore, how to accurately identify the dynamic load by indirect method is of great significance. The dynamic loads in the time domain are expressed as the superposition of a series of kernel function, and the response of the linear system can be obtained by the convolution integral of the kernel function and dynamic loads. Load identification is transformed into an inverse problem through discretizing the convolution integral into a set of linear equations. To solve the difficulties of load identification from the noisy response, a method of dynamic load identification is proposed based on the combination of the singular entropy denoising and the regularization preconditioned conjugate gradient iteration. In order to improve the accuracy of input data in solving the load identification problem, the singular entropy denoising method is introduced in the noisy response signal. Furthermore, the conjugate gradient iterative algorithm is preconditioned by Tikhonov regularization method to alleviate the ill-posedness from the inverse problem. Since the solution of load identification inverse problem is improved at two respects: the input response data denoising and regularization algorithm, the proposed method can effectively suppress noise and improve the identification accuracy. The numerical example indicates that it can identify dynamic load more accurately and stably compared with the other conventional regularization method under the different noise levels of interference. Finally the validity and the feasibility of the proposed method are demonstrated by the experiment.