We use an improved wavelet analysis technique to reconstruct the $(U,V,W)$ velocity distribution for $sim 250000$ stars from Gaia DR2, residing in the solar neighborhood of $200$~pc. The 2D wavelet transforms for three bivariate distributions $(U,V)$, $(U,W)$, and $(V,W)$ were investigated. Though most of currently known (low-velocity) stellar moving groups are densely overlapped in these diagrams, our analysis allowed to detect and disentangle about twenty statistically significant 3D groups of stars with high velocities. Most of them appear new. We also discuss the issue of correct noise thresholding in the wavelet transform and highlight the importance of using a global rather than local statistic for that. Using of a local significance measure may lead to an overstated statistical confidence for individual patterns due to the effect of multiple testing.