The present invention discloses a sparse representation method for dynamic load identification of a mechanical structure for solving ill-posed natures of the dynamic load identification inverse problem and overcoming the disadvantage that the number of primary functions is required to be determined in advance by a current function approximation method. The sparse representation method comprises the following steps of (1) measuring a frequency response function between an action point of dynamic load of the mechanical structure and a response point of the mechanical structure by using a hammering method, and obtaining a transfer matrix through processing; (2) measuring a response signal generated by the dynamic load acting on the mechanical structure; (3) selecting primary functions to construct a sparse representation dictionary according to morphology of the dynamic load; (4) constructing an L1-norm-based sparse representation model of the dynamic load identification; (5) solving the sparse representation model of the dynamic load identification, and obtaining dynamic load sparse representation coefficient vectors; and (6) obtaining the identified dynamic load. Impact and harmonic load acting on the mechanical structure can be effectively identified, and compared with a traditional Tikhonov regularization method based on L2 norm, the sparse representation method has the advantages that identification precision is high and stability is high.