Digital signal processing and classification algorithms play a crucial role in modern day
biomedical monitoring systems. Fortunately, emerging sensors and stimulators as well as
specialized networking technologies have enabled biomedical devices to advance to new
frontiers. Deep-brain stimulators, for instance, offer unprecedented modalities for delivering
therapy to patients affected by neurological conditions, ranging from Parkinson’s
disease to epilepsy; out-patient monitoring networks raise the possibility of comprehensive
yet cost-scalable healthcare delivery over large populations with increasingly
diverse disease states. The central need, as these systems advance towards intelligent,
closed-loop operation, is the ability to detect specific physiological states of interest
from signals that are available through sensors. A key challenge in closed-loop biomedical
systems is the ability to detect complex physiological states from the patient data
within a constrained power budget. Signal processing and data-driven machine learning
techniques are major enablers for modeling and detection of such states. However, the
computational power scales with the complexity of models required.
This thesis considers the VLSI implementation of basic signal processing techniques
such as fast Fourier transform (FFT), power spectral density (PSD) computation. Reconfigurable
architectures for classification algorithms including support vector machines
(SVM) and Adaboost are also presented. The proposed architectures improve
performance and reduce area/power consumption.
First, we present a novel methodology to design parallel pipelined FFT architectures
using folding transformation and register minimization techniques. Novel parallelpipelined
architectures for the computation of complex valued fast Fourier transform
are derived. The proposed architectures overcome prior bottlenecks and achieve full
hardware utilization. The operating frequency of the proposed architecture can be decreased
which in turn reduces the power consumption. This significantly reduces power
at same speed or increases speed at same power consumption level. The power consumption
can be reduced up to 37% in 2-parallel architectures. Further, we propose a
novel approach to develop pipelined fast Fourier transform (FFT) architectures for realvalued
signals. Novel 2-parallel and 4-parallel architectures are presented for radix-23 and radix-24 algorithms. The proposed radix-23 and radix-24 architectures lead to low
hardware complexity compared to a prior RFFT architecture.
We propose an efficient architecture for memory-based in-place FFT/IFFT computation.
A conflict-free memory addressing scheme is proposed to ensure the continuous
operation of the FFT processor. The proposed architecture requires fewer computation
cycles along with the low hardware cost compared to prior work. We then present a
low-complexity algorithm and architecture to compute power spectral density (PSD)
using the Welch method. The complexity reduction comes at the cost of slight performance
loss in accuracy due to the approximation used for the implementation of the
fractional delay filter. The performance loss is 6-8% using fractional delay filter with 2-3
multipliers. A novel architecture is presented based on the proposed algorithm which
consumes 33% less energy compared to the original method.
We propose a low-energy reconfigurable architecture for support vector machines
(SVMs) based on approximate computing by exploiting the inherent error resilience in
the computation. We present two design optimizations, fixed-width multiply-add and
non-uniform look-up table (LUT) for exponent function to minimize power consumption
and hardware complexity while retaining the classification performance. The proposed
design consumes 31% less energy on average compared to a conventional design. Finally,
we present a novel low-complexity patient-specific algorithm for seizure prediction using
spectral power features. The proposed algorithm achieves a sensitivity of 94.375% for a total of 71 seizure events with a low false alarm rate of 0.13 per hour and 6.5% of
time spent in false alarms using an average of 5 features for the Freiburg database.
The low computational complexity of the proposed algorithm makes it suitable for an
University of Minnesota Ph.D. dissertation. July 2012. Major: Electrical Engineering. Advisor: Keshab Parhi. 1 computer file (PDF); x, 158 pages.
Low-power architectures for signal processing and classification systems..
Retrieved from the University of Minnesota Digital Conservancy,
Content distributed via the University of Minnesota's Digital Conservancy may be subject to additional license and use restrictions applied by the depositor.