According to neuro-rehabilitation practice, active training is effective for mild stroke patients, which means these patients are able to recovery effective when they perform the training to overcome certain resistance by themselves. Therefore, for rehabilitation devices without backdrivability, implementation of human-machine synchronization is important and a precondition to perform active training. In this paper, a method to implement this precondition is proposed and applied in a user's performance of elbow flexions and extensions when he wore an upper limb exoskeleton rehabilitation device (ULERD), which is portable, wearable and non-backdrivable. In this method, an inertia sensor is adapted to detect the motion of the user's forearm. In order to get a smooth value of the velocity of the user's forearm, an adaptive weighted average filtering is applied. On the other hand, to obtain accurate tracking performance, a double close-loop control is proposed to realize real-time and stable tracking. Experiments have been conducted to prove that these methods are effective and feasible for active rehabilitation.