High speed or ultra-high speed collision localizations have been found wide applications in many fields, such as the protection of manned spacecraft, robots, vehicles, ships, machine tool collision damage detection and localizations, etc. In this paper, we develop a high speed collision detection and localization algorithm based on discrete wavelet transform (DWT) and kernel extreme learning machine (KELM). The collision localization is finally achieved using the estimated distances of multiple joint sensors. To validate the effectiveness of the proposed method, real vibrations collected from the collisions generated by three kinds of high speed bullets striking on the aluminium alloy plate are employed for experiments.