Robotic-powered exoskeletons represent a promising avenue for aiding individuals with movement disorders in their daily activities and rehabilitation efforts. However, achieving precise joint torque estimation, particularly during dynamic movements, remains a significant challenge. While machine learning and deep learning techniques have been ex-plored for estimation, their efficacy has been limited, especially in dynamic scenarios. Our target is to improve ankle joint torque estimation during dynamic movements by employing multiple data augmentation techniques. Augmentation methods did not significantly improve cases involving the same subject or session. However, our experiments reveal substantial performance gains when combining spatial and signal augmentation methods, particularly in scenarios involving different subjects. This indicated that when facing an over-fitting problem caused by a lack of subjects, a combined data augmentation method will be a proper solution to improve the predicting performance.