Aarhus University Seal / Aarhus Universitets segl

Nonstationary Signal Processing (NSiP)

Signal Processing Lies at the Heart of New Data Revolution

Signal processing, that deals with the acquisition and manipulation of data to extract useful information, remains a key enabling technology that is driving the recent digital technological revolution. Its impact on our daily lives, healthcare and economy are far-reaching. For instance, signal processing is used for the analysis and deciphering of brain signals (EEG) and medical imaging (MRI, CT Scan, X-ray) data; for introduction of smart operations (e.g., movement tracking, speech recognition) within wearable sensors and digital assistants; and for delivering smart solutions in renewable energy sector (e.g., wind power forecasting).

Despite those achievements, with the explosion of digital data at unprecedented levels along with the increasing interest in smart operations of digital devices at human-level intelligence, the demand for more sophisticated, reliable and accurate signal processing methods is ever-growing. For example, novel signal processing algorithms are sought for i) on-the-fly detection of serious medical conditions through multimodal data from heterogeneous sensors on wearable devices, and ii) real life brain monitoring via ear-EEG technology, reliably.

Multiscale Signal Processing: Making Sense of Nonstationary Data

Among the key signal processing approaches that would aid in accomplishing the above-mentioned feats are those that rely on retrieving principal data components via multiscale algorithms and transformation of complex nonstationary data into more descriptive domains, such as joint time-frequency plane. The necessity for those algorithms stems from the very nature of natural systems and their associated observations - they are inherently nonstationary and multiscale. For instance, sleep spindles in brain signals (EEG) correspond to multiple brief oscillations, carrying information about different aspects of sleep quality and voiced speech signal consists of fundamental frequency component (pitch) and formant which is a combination of multiple amplitude-frequency modulated signals. Indeed, gaining a clear insight into the multiscale characteristics of such natural data and its subsequent utility in the decision-making process is a vital prerequisite to obtaining reliable next generation systems.

About our Research: Towards Accurate, Robust and Reliable Representations for Nonstationary Data

The main research interests of the group lie in pioneering fully data driven multiscale signal decomposition and time-frequency algorithms that are capable of providing accurate and faithful representation of nonstationary data, both in time and joint time-frequency domains. The key goal is to address fundamental problems inherent to standard approaches which ‘temper’ or ‘color’ true signal representation owing to their adoption of fixed template functions. Further, we strive to develop signal decomposition and time-frequency methods that are inherently robust to noise and are provably convergent (reliable). To accomplish that, we employ modern tools and methods from statistics, optimization theory, multidimensional geometry and linear algebra. In addition to development of novel approaches, several applications of those methods have also been put forward by our group e.g., signal and image denoising, data fusion, biomedical engineering and wind power forecasting.