This document proposes a method to augment virtual IMU data using a spring-joint model for motion exercise recognition without real data. The method aims to address the limitation of current virtual IMU data which has a limited motion length. It introduces a spring-joint virtual sensor module that can simulate different acceleration distributions and augment the virtual acceleration data spatially and temporally. An experiment tested the method on three aerobic exercises and showed the proposed data augmentation approach improved motion recognition accuracy from 45.5% to 85.3% compared to non-augmented virtual data.