The invention discloses a self-adaptive beam forming method based on covariance reconstruction and guide vector compensation. According to the method, in the situation that desired signals exist in sampling snapshot and signal guide vectors are mismatched, interference can be effectively inhibited, main lobe shape preservation and side lobes of a self-adaptive directional diagram can be reduced, and high output SINR and quick convergence speed can be obtained. According to the method, eigenvalue decomposition is carried out on a sampling covariance matrix, noise sub space is estimated by means of an MDL criterion, the incident angle of a signal source is estimated in a Root-MUSIC method, the incident angle of the desired signals is judged, and then a new interference and noise covariance matrix is reconstructed; mismatching of the guide vectors of the desired signals is compensated by solving a quadratic programming problem with quadratic constraints; ultimately, a self-adaptive weight vector is solved by means of the reconstructed new interference and noise covariance matrix and the corrected guide vectors, a null is formed in the interference direction in a self-adaptive mode, and interference is effectively inhibited.