Investigating Low-complexity Architectural Issues under UBSS

Reddy, P S (2014) Investigating Low-complexity Architectural Issues under UBSS. Masters thesis, Indian Institute of Technology, Hyderabad.

Preview

Text
EE12P1007.pdf - Submitted Version
Download (1MB)

Abstract

Our Project aim is to develop a real time chip to process the sensor signals and separating the source signals, which is used in Health care like Autism. Autism is a disease which aects the child mental behavior. So If we analyze the signals form the brain so we can observe the how eectively the disease is cured. So to analyze the Autism we need EEG signals from almost 128 Leads from the scalp of child, which is dicult to do so. Thus we have to reduce the number of Leads used and at the same time we should get the all information as in the case of 128-Leads. Thus solving our problem is to solve Underdetermined Blind Source Separation (UBSS). And in some other cases we may have only one mixture signal (M=1), which is extreme case of UBSS, from which we have to extract the unknown sources, which is called Single channel Independent Component Analysis also called SCICA. In SCICA if we have N source signals then it is called ND-SCICA. In real time UBSS or SCICA problem we require a Digital chip which will separate the sources in real time case. So we require a chip which is High speed so that it will be suitable for real time applications and also it should be Recongurable so that it can work for dierent type of applications where the frame length of signals vary. So rst we investigated the architectural issues of Recongurable Discrete Hilbert Transform for UBSS where M is greater than one. Thus we proposed a high-speed and recongurable Discrete Hilbert Transform architecture design methodology targeting the real-time applications including Cyber-Physical systems, Internet of Things or Remote Health-Monitoring where the same chip-set needs to be used for various pur- poses under real-time scenario. By using this architecture we are able to get Discrete Hilbert Transform for any given M-point by re-using N-point Discrete Hilbert Trans- form as a kernel. Here N and M are multiple of 4 and N respectively. Subsequently we provide the architecture design details and compare the proposed architecture with the conventional state-of-the-art architecture. Thorough theoretical analysis and ex- vi perimental comparison results show that the proposed design is twice as fast and recongurability is also achieved simultaneously. After DHT, we proposed a new algorithm for ND-FastICA which is used for ex- treme case of UBSS where the number of mixture/sensor signals are only one. In this algorithm we used CORDIC based ND-FastICA which is recongurable so that the same chip can be used for dierent dimensioned FastICA.

[error in script]

IITH Creators: