The invention relates to a multi-fault identification method and device of a sparse tight frame dictionary learning model. According to the method, tight frame constraint is added in a traditional sparse dictionary learning process, different structured fault feature information is decomposed into a series of subspaces in an adaptive mode, most white Gaussian noise is eliminated by means of the analysis sparse tracking algorithm, then a fault sensitivity index set is researched and designed by means of physical fault dynamic experiments, composite fault signals are decoupled by means of the optimal subspace matching technology, and finally composite fault modes are identified through the spectral analysis method. By means of the method and device, decoupling of multi-source feature subspaces and identification of multiple fault modes are effectively achieved, and guiding suggestions are provided for core assembly maintenance of a mechanical system and arrangement adjustment of overhaul plans.