As demand for rehabilitation has increased in recent years, there is a rising demand for home-based rehabilitation (HBR). Inducing proper rehabilitation motions in HBR is essential. To analyze more complex and diverse behaviors of patients, a technique, called human activity recognition (HAR), has been studied using artificial neural network methods, such as convolutional neural networks (CNN). CNN-based HAR is used in many studies because of its high accuracy and easy to use. To use CNN-based HAR in real time, data is segmented by time-window. Rehabilitation motions are mainly repetitive motions, so it is necessary to consider the relationship between the period of motions and the size of time-window. A system using a smartwatch was constructed to collect upper limb data. We collected five upper limb rehabilitation motions for various periods and used a 5-fold cross-validation technique to verify the performance of a prediction model over a particular time-window size. The results showed that the size of the time-window maximizing the classification performance is affected by the period of sample data.