The fault diagnosis and state monitoring of the mechanical vibration signal often struggle with large amount of sampled data, large storage capacity, high transmission bandwidth and low signal reconstruction accuracy. In light of this problem, an adaptive compressed sensing of mechanical vibration signals based on sparsity fitting method is proposed. First, the multi-scale wavelet packet transform is carried out on the mechanical vibration signal, and its sparsity is obtained by zeroing the wavelet packet coefficient at a certain threshold value. Then, the iterative method is adopted to obtain the minimum sampling rate that meets the requirements of reconstruction signal accuracy under each sparsity degree, and the sparsity degree and sampling rate are fitted with the least square method to eliminate the signal measurement error for an optimal signal sampling rate. Finally, an over-complete dictionary adapted to each signal block is constructed by K-singular value decomposition algorithm, and the signals are reconstructed by orthogonal matching pursuit algorithm. Experiments show that the signal compression rate and reconstruction accuracy of this algorithm are greatly improved compared with the traditional compression algorithm.