Non-Stationary Time-Series Analysis based on Geometric Signal Modeling

High dimensional signals may arise in many fields of science. For example, biomedical signals such as EEG and EMG can be modeled by few latent processes measured by a large set of noisy sensors. In such applications the goal is to identify the latent intrinsic variables, which describe the true, intrinsic state of the system, e.g. disease state, sleep stage, performed action. Such real signals are commonly hard to analyze due to nonlinearities and rapid state changes in time (non-stationarity). In this project we plan to address these difficulties and construct a framework which is based on manifold learning techniques. These techniques provide a compact and reliable representation of the system by exploiting the structure and patterns in the data. This compact representation describes intrinsic properties of the analyzed system. We plan to develop new mathematical tools which will give rise to a new analysis framework for problems involving rapidly changing signals that existing tools fail to handle.

Faculty Supervisor:

Hau-tieng Wu

Student:

Partner:

Technion – Israel Institute of Technology

Discipline:

Engineering

Sector:

Education

University:

University of Toronto

Program: