The invention provides a vibration signal reconstruction method for mechanical fault diagnosis. In a non negative condition, sparser coding is carried out on original vibration signals, the signal spectrum detection effectiveness is improved, and the method belongs to the technical field of vibration signal processing. The method comprises steps: 1, vibration signals of a mechanical system are sampled, and the original vibration signals are overlapped and segmented to obtain column vectors to form a training sample; 2, a learning dictionary is initialized, sparse coding is carried out on the training sample based on a non negative-base tracking algorithm to obtain an initial sparse coding matrix; 3, the sparse coding matrix is fixed, and the learning dictionary is optimized and updated based on a non negative K singular value decomposition algorithm; 4, the learning dictionary is fixed, and sparse coding is carried out based on the non negative-base tracking algorithm to update a corresponding sparse coding matrix; 5, the former two steps are circulated, alternate iterative updating is carried out to obtain the final learning dictionary and the final sparse coding matrix; 6, an updating sample is calculated, and reconstructed signals are inversely overlapped; and 7, an envelope spectrum for the reconstructed signals is extracted for fault diagnosis. The method of the invention is used for mechanical fault diagnosis.