Energy-Efficient Distributed Compression and Error Control in Wireless Sensor Networks

This is a Wireless@KTH project during 2005-06. The project manager is Prof. Mikael Skoglund, who is with the communication theory lab at KTH/EE/S3.

Introduction

This project will study distributed algorithms for signal compression and error control in wireless sensor networks. This is a new area, to our knowledge not previously addressed within Wireless@KTH, even though it is tightly related to existing activities in ad hoc wireless networks, relay networks and cooperative transmission. Our main motivation is energy-efficiency, to address applications where battery drain in the sensor nodes is a major issue. To introduce the problem we propose to study, consider the setup illustrated in the figure below.

Two sensors measure two different variables, S_1 and S_2, containing information about the same physical entity X. In practice, S_1 and S_2 can be the temperature at two different locations, dependent on an underlying physical entity X, e.g., pressure or radiation. Or, S_1 and S_2 can be spatial sound samples due to the same sound source X, located in the vicinity. Or, S_1 and S_2 can be two different images representing the same object X from two different angles, etc. In many practical applications, there would be a large number M of sensors measuring signals that represent different "versions" of an underlying entity X. For simplicity and purpose of illustration, we show only the case of M=2.

The samples S_1 and S_2 are processed locally and independently in the two sensors, and two different descriptions at rates R_1 and R_2 (bits per second per Hertz) are transmitted to a sink node. In practice this transmission would typically involve relaying through intermediate nodes, but again for simplicity we refer to the simple example in the figure. The sink performs local processing of the received data, and produces an estimate of the entity X, for presentation, further processing or distribution through a core network (e.g. the Internet).

Now, a key observation --- according to a fundamental information-theoretic result, the Slepian--Wolf coding theorem, the same resolution/quality in the reproduction can be achieved in the illustrated system as in the case where the two sensor nodes are perfectly connected to each-other, and can process S_1 and S_2 jointly. That is, the joint processing of S_1 and S_2 can be "moved" from the sensors (that each see either S_1 or S_2) to the sink node (that receives the two descriptions at rates R_1 and R2). Hence, excellent performance can be achieved with a minimum amount of inter-sensor signaling. For this reason, some existing works have already suggested algorithms for distributed compressing, in the sense of the figure, for wireless sensor networks.

Distributed Compression and Error Control

As mentioned, the usefulness of the Slepian--Wolf result in source compression for sensor networks has already been observed. However, previous works have focused solely on the source compression (quantization and/or data compression) problem, and have not taken transmission and error control issues into account. Therefore we propose to study distributed source compression and error protection, where we will handle the compression and error control jointly along the lines of previous work in joint source--channel coding. This will result in a cross-layer approach, involving physical and link layer error control, and application layer source compression and data processing. To our knowledge this is a new area that has not been considered before. Below we outline some more specific areas we intend to study in connection to distributed compression and error control.

Distributed Source--Channel Coding

As mentioned, the work so far on distributed coding for WSNs has focused on the compression problem and has not involved aspects of the robustness and error control problems. Therefore, we intend to focus on distributed joint source--channel coding for WSNs.

Hybrid Digital--Analog Coding

Techniques for hybrid digital--analog coding have recently been demonstrated to show efficient robustness toward design mismatch with respect to an unknown time-varying channel. Indeed, the very fact that these systems are inherently more robust to a mismatch than are "traditional" coded systems, make them particularly suited for low-cost deployment of WSNs. That is, techniques "with an analog touch" are excellently suited for power-efficient communication of analog sources, in time-varying channel environments. The "compatibility problem," i.e. the fact that semi-analog techniques are often incompatible with current technologies, is much less of a problem in sensor applications. This is because many networks will be more or less self-contained and without the need to communicate with the "outside world" (except possibly through the data fusion center); thus more radical communication system solutions can be advocated.

Soft Source Decoding

We will also be interested in another type of cross-layer interaction involving source decoding based on soft reliability information from the application layer. Issues to address include how to convey soft information from the lower layers and cost versus performance comparisons. We will for example consider the tradeoff between encoding and transmission of the soft information itself through the network, e.g. from a receiving node to an end user personal computer, versus decoding and re-quantization.

Multiple-Description Coding

We will investigate techniques for robust source data transmission based on multiple descriptions. These techniques fit the framework of sensor networking well, and there is an interesting connection with routing at the middle layers. For example, diversity against node failure can be achieved by letting nodes convey multiple descriptions at multiple routes through the network.

Energy-Efficiency

One main motivation for the proposed work is efficiency in the sensor nodes' use of the available energy. Such efficiency is of vital importance since these nodes typically operate on batteries, that in many applications cannot be replaced or re-charged. Hence there is often a strict total energy budget. There are four main ways in which the distributed processing we propose will save energy:

Since distributed compression and error control requires no overhead signaling between wireless sensor nodes to benefit from the potential gains of combined inter-sensor data processing, energy spent on communication is held at a minimum.
The distributed approach can interact across layers with energy-efficient protocols suggested for MAC, routing and traffic control.
In short-range networks, energy spent on data processing in the sensor nodes can be of the same order of magnitude as the energy spent on radio transmission. However, as described, the proposed distributed approach will "move" much of the processing to the sink node, which in most practical cases is not energy-limited to the same extent as the sensor nodes. Thus, the proposed approach opens up the potential for savings in both transmission and processing.
For single-user links, several previous works on joint cross-layer compression and error control have demonstrated significant transmit-power savings compared with the traditional separated approach, at a comparable reproduction fidelity.

Besides the focus on energy-efficiency, we emphasize that this work can also be motivated from the perspective of affordability, since our focus on efficient error-robust transmission allows a compensation for deficiencies due to "cheap" hardware solutions. This is tightly related to the motivation behind the research carried out in the AWSI/MAMC project, where we are involved.

The Project Description (pdf) describes the project in more detail, including a budget and references.

Last updated: 2005-09-06 by Mikael Skoglund