Unsupervised Phase Extraction using Dual Autoencoder

Phase assignment is usually a part of periodic time-series data processing which needs some manual labeling or a heuristic method for each specific type of signals. Our work uses an unsupervised learning method to make the phase identification process fully-automated and more universal. This method also allows flexibility of input data in the term of position variation and phase progression variation. Four synthetic periodic two-dimensional signals with different shapes are used to explore our method's learning capabilities and some limitations. The proposed method is also tested with an actual periodic movement sequence captured from a Kinect sensor. This method can learn the periodic pattern automatically from a noisy signal and assign correct phases.