To reduce the differences between the data scales and volume of multi-source cross-scale point cloud data， this study proposes a multi-scale decomposition method of point cloud data based on wavelet transform. This study examines the multi-scale decomposition of small-scale point cloud data with considerable attention and the application of scale decomposition in cross-scale point cloud data registration. First， the small-scale point cloud is grid modeled， and the global point cloud binary expression function is established. Subsequently， according to the theory of discrete wavelet transformation， three-dimensional wavelet decomposition of the grid point cloud is performed several times， and the low-pass characteristics of the wavelet scale function are used to retain the low-frequency information to obtain the approximate scale data of the original small-scale point cloud. The similarity with the original data is then characterized based on the surface dimension and the difference in body dimension， and the effective wavelet decomposition series is determined. Finally， the point cloud data obtained by decomposition at various levels are accurately registered with the large-scale point cloud data， and the registration relationship is applied to the original point cloud to increase the registration accuracy of the cross-scale point cloud data. The experimental results show that the multi-scale decomposition method proposed in this paper can effectively decompose the data. When applied to the multi-scale measurement of an aero-engine blade， the registration accuracy of the local cooling holes small-scale point cloud data and the overall blade structure light data of micrometry increased by 61.36%. The proposed decomposition method is applied to the multi-scale measurement of blade edge and grid parts， and the registration accuracy is increased by 48.59% and 43.86%， respectively. The proposed multi-scale decomposition method of the point cloud can effectively decompose small-scale point cloud data， and ultimately improve the registration accuracy of cross-scale data.