Design of Energy-efficient Sensing Systems with Direct Computations on Compressively-sensed Data

Abstract

The aim of this thesis is to explore the energy limits that

can be achieved by signal-processing systems when they explicitly

utilize signal representations that encode information

efficiently. Compressive sensing is one method that enables us to

efficiently represent data. The challenge, however, is that in

compressive sensing, signals get substantially altered due to the

random projections involved, posing a challenge for signal

analysis. Moreover, due to the high energy costs, reconstructing

signals before analysis is also often infeasible. In this thesis, we

develop methodologies that enable us to directly perform analysis on

embedded signals that are compressively sensed. Thus, our approach

helps potentially reduce the energy and/or resources required for

computation, communication, and storage in sensor networks.

We specifically focus on transforming linear signal-processing

computations so that they can be applied directly to

compressively-sensed signals. We show that this can be achieved by

solving a system of linear equations, where we solve for a projection

of the processed signals as opposed to the processed signals

themselves. This opens up two approaches: (1) when the projection

matrix is the random projection matrix used in compressive sensing,

where we show that the linear equations can be solved with a

least-squares approximation, and (2) when the projection matrix is an

auxiliary matrix, where we show that the equations become

underdetermined, allowing us to obtain either high-accuracy or

low-energy solutions based on two designer-controllable knobs. We

study our methodologies through information metrics, validating their

generality, and through application to biomedical detectors, utilizing

clinical patient data. Through a prototype IC implementation, we also

demonstrate a hardware architecture that exploits the two knobs for

power management. Further, we also explore options for hardware

specialization through architectures based on custom-instruction and

coprocessor computations. We identify the limitations in the former

and propose a co-processor based platform, which exploits parallelism

in computation as well as voltage scaling to operate at a subthreshold

minimum-energy point. We show that the optimized coprocessor reduces

the computational energy of an embedded signal-analysis platform by

over three orders of magnitude compared to that of a low-power

processor with custom instructions alone.