Overview

Sensor networks differ from traditional wireless networks in several respects. Unlike handheld wireless devices which can be recharged at reasonable frequent intervals, sensor nodes must operate autonomously for much longer durations. Energy supply thus remains an open challenge in sensor networks because unfettered deployment rules out traditional wall socket supplies and batteries with acceptable form factor and cost constraints do not yield the lifetimes desired by most applications. One method to improve the battery lifetime of such systems is to supplement the battery supply with environmental energy. Several technologies exist to extract energy from the environment such as solar, thermal, kinetic energy, and vibration energy. However, we lack system level methods to efficiently exploit these resources for optimal performance. Sensor networks are expected to be deployed for several mission critical tasks and operate unattended for extended durations. The autonomous nature of operation makes it imperative that the system learn its own energy environment and adapt its power consumption accordingly. In distributed systems, not only does the energy source vary in time, but also the energy available at different locations, and thus at different nodes of the sensor network differs. In this situation, the performance can be improved by scheduling tasks according to the spatio-temporal characteristics of energy availability. The problem then, is to find scheduling mechanisms that can adapt the performance to the available energy profile.

Approach

Harvesting Theory:

DPS is based on an abstract mathematical model for the energy scavenging process which allows us to determine a bound on the achievable performance for a particular energy source. This mathematical framework, referred to as Harvesting Theory, is based on a simple model for energy sources. This model can capture a wide variety of energy sources, both environmental and artificial (such as robotic energy delivery) in just three parameters. Despite its simplicity, the model is powerful enough to model the fundamental characteristics of the energy source. Thus it can be used to derive performance bounds using a set of theorems that we have proven as part of the basic harvesting theory.

As part of the harvesting theory, we are also determining the optimal routing methods to achieve the maximum lifetime while satisfying the application performance constraints such as data distortion. Since the number and choice of which sensors are used to collect the data not only affects the distortion but also the routing cost of delivering the collected data to the fusion center, we model the energy cost of multi-path routing jointly with the rate-distortion function. Rather than working with asymptotic and random network placements, we include the exact network topology information into our constraints. We then use the joint constraints to set up an optimization problem which yields the best possible choice of routes for a given distortion performance in order to maximize the network lifetime. The optimization problem derived is nonlinear and we also prove heuristics to linearize the constraints in order to utilize linear programming tools for solving the optimization with reasonable computation complexity.

 Adaptive Duty Cycling

In order to achieve performance scaling which has a directly impact on energy consumption, we have developed an adaptive duty cycling algorithm which enables system-wide performance to be adapted to the spatio-temporal characteristics of energy availability based on only local measurements. The algorithm takes advantage of the performance scaling mechanism available on most mote class sensor nodes – sleep mode and has the following three objectives:

• Achieving energy neutral operation, i.e. energy consumption should not be more than the energy provided by the environment.
• Maximizing the system performance based on an application utility model subject to the above constraint,
• Being adaptive to dynamics of energy source in real-time.

A prediction model that enables harvesting sensor nodes to predict future energy opportunities based on historical data was also developed. We also derive a theoretically optimal bound on the maximum performance achievable assuming perfect knowledge about the future.

Systems/Experiments

Significant system development has been carried out for developing the enabling hardware required for energy scavenging, developing the software drivers and protocol implementations for DPS, and carrying out large scale simulations for performance evaluation.

 Hardware:

A complete test-bed, named the Heliomote (shown in the figure below), has been developed. This testbed not only allows harvesting energy from the environment but also enables the sensor network to learn the characteristics of the available energy and hence use task scheduling and load sharing methods which can adapt performance to the available resources. The key hardware component developed as part of the test-bed effort is an embedded circuit that can be added to a low power sensor node (the Berkeley mote) for extracting energy from the environment. This hardware performs the following scavenging functions:

• Charge the battery from solar cells: The circuit performs the battery chemistry specific tasks required for storing solar energy efficiently in the battery.
• Record the solar and battery energy at any instant: The recorded energy is provided over a data interface to the sensor node which can use the information about environmental energy and the battery status to control its performance and schedule tasks such as by using DPS methods mentioned before.

Figure Heliomote – An Solar Energy Harvesting Sensor node.

Figure Solar energy data collected during the period of 71 days (July-Sept)

 Software

Extensive software has been developed for the harvesting test-bed. The software is not only distributed, but, also spans multiple platforms. The mote parts of the software have been written in nesC which runs on TinyOS. The monitoring tool is written in Java and runs on Linux. The software consists of three major components:

• TinyOS drivers which use the information provided by our harvesting board.
• Adaptive Duty Cycling Algorithm under implementation on MICA node.
• DPS based routing and performance control protocols.
• GUI based monitoring tool for performing experiments with the developed system.

A screenshot from the GUI based monitoring tool is shown in the above figure. Each dark green dot represents a node in the network. The small bar and pie chart next to the dots show the tracked battery and solar energy status. The energy efficiency routes back to base station located at the bottom right corner is represented by the green lines.

Accomplishments

The project accomplishments include:

• The Heliomote platform has been selected as one of the winners of the low power design contest at the ACM/IEEE - International Symposium on Low Power Electronics and Design (ISLPED)
• The Heliomote platform was deployed in a residential area in Los Angeles for a total of 72 days from the beginning of June through the middle of August. Valuable data such as energy generated by solar cells and battery level status was collected.
• The data collected from the field test can be used in further study of harvesting aware application.
• A complete new packaging solution has been introduced. This includes a new waterproof, shock-resistance casing and flexible high efficiency solar panels.
• A fully working network protocol which also collects energy data over the network was implemented and installed on a multihop network of Heliomotes.
• Adaptive Duty Cycling algorithm was developed and promising results was shown in simulation. We compared the energy usage efficiency with two other approaches. 1. The theoretical optimal calculated assuming complete knowledge about future energy availability and 2. A simple approach which attempts to achieve energy neutrality using a fixed duty cycle without accounting for battery inefficiency. The figure below shows that the adaptive duty cycling can utilize up to 20% of the environmental energy over a simple approach, and it performs very closely to the optimal.

Future Directions

The adaptive duty cycling algorithm will be tested in the field once the implementation is completed and different system parameters used in the model will be explored. We are also looking into incorporating local forecast service to our model to provide more accurate energy profile prediction. Further, since nodes in network operates on different duty cycles as per their local energy opportunities, a new MAC protocol which supports communication among nodes operating with different duty cycles will need to be implemented, and this will most likely to be based on the UBMAC due to its low overhead for synchronization.

People

Graduate students: Aman Kansal; Jason Hsu; Sadaf Zahedi; Vijay Raghunathan; Jonathan Friedman

Faculty: Mani Srivastava, EE