Although the experiment we have described is encouraging, it is only a preliminary study. There are a large number of possible future directions for research; we have focused only on some that seem the most promising and interesting.
Our planned work for the future falls into three main categories. First, we wish to develop and characterize new time synchronization methods. We also plan to characterize traditional synchronization methods for the purpose of comparison. Second, we would like to implement a working system as proof-of-concept. Specifically, we will develop a testbed that can track the motion of objects through a field of sensors; this application will demonstrate two different time synchronization in different contexts. Finally, we will explore methods that might allow a running system to adapt to new applications and conditions by automatically selecting the time synchronization methods that are most appropriate.
An important goal of our work is to provide a palette of time sync methods to applications that covers a good portion of the parameter space we described in Section 2. Because it is impossible for any single synchronization method to appropriate in all situations, sensors should have multiple methods available. If a node can dynamically trade error for energy, or scope for convergence time, it can avoid ``paying'' for something that it doesn't need. Ideally, the algorithms should also be tunable--allowing finer control over an algorithm than simply turning it on or off.
Towards this goal, we plan to continue the development of sync methods and their characterization along the axes we described earlier. In addition, we plan to more fully characterize the ``traditional'' time sync methods for comparison. Our specific plans for further development and experiments follows.
There are many interesting dimensions of pulse synchronization that we did not explore with our initial experiment. For example, what are the effects on error of scaling up the number of nodes, the duty cycle of the nodes, or changes in the ambient temperature? Many of these questions can be answered with a larger, truly untethered testbed.
An even a more basic question that we would like to answer is: what is the best modality for a synchronization pulse? We should not assume that a radio signal is the best (e.g., most deterministically detected) mode. We plan a simple experiment to test the detection determinism of other modes not yet examined, such as using a strobe light to trigger remote photodiodes. By using a multichannel oscilloscope to compare the dispersion in triggering times of various modes (e.g., the strobe light, radios with various MAC layers, and wired stimuli), we can get an interesting picture of how the detection modalities themselves contribute to the overall system's error budget.
Naturally, each link in this synchronization chain will accumulate error. Our plan is to understand how this error accumulates as the size of the synchronized area increases. This ``maximum error vs. distance'' is analogous to the ``maximum error vs. time'' experiment we described in Section 6.
Although experiments that characterize our methods are important, we also think it is important to implement an actual application using these methods as a proof-of-concept. Therefore, the second major portion of our planned work is the implementation of a testbed that demonstrates our time sync algorithms in the context of a sensing application.
Our proposed testbed application is tracking of cooperative objects through a sensor field. A ``cooperative'' object is one that wants to be tracked; that is, the object has been augmented in a way that makes it easier to localize. The goal is to build sensor network that will report to a user the location and speed of the cooperative object over time. The sensor field will be truly distributed: that is, its geographic span will prevent any single node from broadcasting a radio message globally.
The application we propose will build on Girod's prototype acoustic rangefinder [Gir00]. His rangefinder uses digital signal processing techniques to accurately determine the time of arrival of a specially formed sound. (The sound is a ``chirp'' that is modulated by a sender to make it more easily detected by the receiver.) If the sender and receiver are synchronized in time, Girod's technique allows accurate determination of the time of flight of sound between two points. This, in turn, leads to an accurate range estimate between them.
In cooperation with Girod, we plan to use post-facto time synchronization to facilitate measurement of the time of flight of sound from an audio source to a set of receivers--not just a single receiver--allowing them to trilaterate. At any given instant, this technique will report instantaneous position in a coordinate system defined by the sensors4. As the object moves through the sensor field, many such localization measurements can be taken and integrated to form an estimate of velocity.
This experiment demonstrates two aspects of our time synchronization services:
sec. This is exactly what the
trilateration/localization service needs.
We believe this experiment to be an effective demonstration on several fronts:
There are additional future directions that we wish to pursue, though time may not permit a complete exploration of these details in the context of an actual implementation. We plan to examine some of these issues in the context of simulation or modeling instead. Specifically:
One such ``tuning knob'' that might be possible is in selecting the number of nodes that participate in a federation of synchronized nodes. Many time-standards (e.g., UTC(USNO), the U.S. Naval Observatory's realization of UTC), use a large ensemble of clocks and dynamically choose the best in order to improve precision. Will adding additional beacons make post-facto synchronization better, at the expense of more energy expended by the additional nodes that are participating? If so, this is another interesting energy-error tradeoff.