Signal and Image Processing
Selin Aviyente | aviyente@egr .msu.edu | www.egr.msu.edu/~aviyente
Research in the Singal and Image Processing Laboratory focuses on the development of signal processing solutions to problems in signal representation, detection and classification. Currently, our group focuses on two application areas: study of functional brain networks from neurophysiological data and development of signal processing tools and development of statistical learning algorithms for fault prognosis in electrical drives and machines.
Functional Connectivity of Brain Networks
Our research group’s current focus is on developing a signal processing framework to infer the functional brain networks from neurophysiological measurements, such as the electroencephalogram (EEG) and neuroimaging data such as the functional magnetic resonance imaging (fMRI). Empirical studies from electromagnetic recordings and neuroimaging suggest that human cognition arises from transient synchronization between distant and specific neural populations. The vast majority of previous work relies on the assumption of stationarity of functional connectivity. However, recent studies have shown that the functional connectivity in brain networks may exhibit dynamic changes within short time scales. Recent years have seen a growth in approaches that examine dynamic changes in functional connectivity during the course of an experiment as well as during the resting state. However, most of this work is focused on studying the dynamic connectivity for resting-state fMRI. Recently, we have developed a tensor based framework for tracking the dynamics of functional connectivity networks from EEG data. This two-stage method first identifies the change points where the connectivity networks show significant changes in structure and then summarizes the network state within each time interval through tensor projections. More generally, we are developing online algorithms for robust low dimensional structure learning algorithms for tensor type data.
Another important problem in the area of inferring functional connectivity networks is to quantify the connectivity across brain regions. Phase synchrony has been used to investigate the dynamics of subsystems that make up a complex system such as the human brain. Current measures of phase synchrony are mostly bivariate focusing on the synchrony between pairs of time series. Bivariate measures do not necessarily lead to a complete picture of the global inter-actions within a complex system. Current multivariate synchrony measures are based on either averaging all possible pairwise synchrony values or eigendecomposition of the pairwise bivariate synchrony matrix. These approaches are sensitive to the accuracy of the bivariate synchrony indices, computationally complex and indirect ways of quantifying the multivariate synchrony. Recently, we proposed a general hyper-spherical coordinate system along with a new higher-dimensional manifold representation to eliminate the dependency on the ordering of the signals’ phases. This new framework, referred to as Hyper-Torus Synchrony (HTS), is shown to be equivalent to the root-mean-square of a sufficient set of squared phase-locking values whose phase differences contain information about all oscillators in the network.
Useful Life of Electrical Drive Components
In collaboration with Dr. Strangas and General Motors, we have combined statistical feature extraction algorithms with prediction methods such as Hidden Markov Models (HMM) and Extended Kalman Filtering (EKF) to determine the remaining useful life (RUL) of different components in electrical drives such as bearings. Bearings constitute a large portion of failures in rotational machines. Although many techniques have been successfully applied for bearing fault diagnosis, prognosis of faults, particularly predicting the remaining useful life (RUL) of bearings, is a remaining challenge. The main reasons for this are a lack of accurate physical degradation models and limited labeled training data. We introduce a data-driven methodology, which relies on both time and time-frequency domain features to track the evolution of bearing faults. Once features are extracted, an analytical function that best approximates the evolution of the fault is determined and used to learn the parameters of EKF. In this work, we illustrated that different features may work better under different operating conditions depending on the length of the test data set and gave a detailed description of RUL estimation based on EKF along with a procedure to estimate the confidence intervals.
