Towards Building an Open, Robust, Uncertainty-resilient Detection, Identification, and Tracking Sensor-Cyber Network

1. Overview
2. Research Agenda
3. Participants
4. Collaborators
5. Major Publications
6. Demonstration
7. Funding Source
8. Related Links
9. Detailed Description of Research Agenda

Overview

Detection, identification, and tracking (DIT) problems that arise in applications such as protecting people and the environment against radiational, biological, and chemical plumes, can be solved by combining the modalities of sensor and cyber networks. The sensor network provides information about physical space activities, such as the locations and movements of plume sources and targets. The cyber network provides storage and computational resources to analyze and infer, based on realistic dispersion models of plumes with respect to physical phenomena (e.g., terrain and weather effects), where the plume originated, the trajectory of its movement (speed and direction), and the prediction of its future movement. The cyber network also makes decisions about where to task and activate plume sensors regarding their future sensing and communciation activities.

Central to the notion of convergence between the physical and cyber spaces is how intelligence in the networked system of sensors supports tasks such as the spatial-temporal formation of communication groups, well balanced data gathering / dissemination / reporting / fusion, and robust and time-constrained communication of high quality data over a multihop, unreliable, and best effort wireless network. To answer the question, we must address (i) architectural issues including sensor selection, density, and placement, for benefit- and quality-aware network coverage, and (ii) operational issues including sensor tasking, communication, and sensing protocols that are robust against failures, noises in the sensing and communication channels, and other contingencies.

Under the national SensorNet initiative, Oak Ridge National Lab (ORNL), in conjunction with its University collaborators, has carried out the initial deployment of a detection, identification, and tracking sensor-cyber network (DITSCN) in the Washington D.C. and Memphis Port areas. The D.C. network demonstrates the use of RFTrax radiation detectors integrated (through a universal IEEE 1451/USB interface) with open source Linux (with customized software) to report sensor measurements on chemical, biological, radiological, nuclear, and explosive (CBRNE) threats to an operation center. Wide area communication is realized by the system architecture shown in Figure 1, in which sensors gather data and transmit them in an enterprise context for processing by applications. Sensor nodes have basic wireless communication capabilities, namely to send/receive wirelessly over a single hop to a nearby network gateway, which is itself connected to a more distant operation center through the existing Internet.

Figure 1 Washington D.C. deployment architecture

In the Memphis Port project, on the other hand, chemical detectors are placed in the port area to monitor possible chemical pollution of the fresh water. A set of threat scenarios are defined in prior analysis of the port's vulnerabilities. The threats were summarized in nine representative scenarios concerning three different locations and a specific set of materials. The sensor locations were constrained to about 30 places around the port due to their accessibility, availability of power, and other factors. The project concludes that weather is arguably the most important factor in the sensor placement to best address the threats of pollution on the nearby populations. Seasonal changes in the weather should be derived from historical meteorological data, and incorporated into intelligent sensor placement to significantly reduce the time to incidence detection.

While the D.C. and Memphis Port deployments demonstrate initial success in using a DITSCN for national security applications, thereby validating the feasibility and utility of the underlying network system architecture, several research challenges remain to substantially enhance the performance and robustness of these networks. In particular, field experience has shown that the physical space is deemed to behave non-deterministically, either by nature (e.g., variation in the radiation emitted from the source and hence detected by RFTrax) or in the lack of complete and accurate information that characterizes it (e.g., frame loss due to wireless interference/contention and/or environmental factors). The latter is due to the fact that the data communication process is inherently unreliable and measurement errors inevitably produce noise in the data analysis. Acknowledging the existence of uncertainty and enabling data dissemination/collection mechanisms, inference models, and decision making processes to deal with uncertainty will lead to a more robust and effective DITSCN. On the other hand, in spite of the non-deterministic nature of the physical space, the cyber space has to be intimately aware of its deployment, including both its physical and population aspects. This involves adequate modeling of physical phenomena (e.g., plumes with respect to the absorption, propagation, and disperson coefficients), and its incoproation early into the systems and protocol design. Finally, the physical and cyber spaces have to be closely coupled, in the sense that (i) physical characteristics in the physical space should be transported to the cyber space in a timely fashion for judicious analysis and decision making in the cyber space; and (ii) the decisions made in the cyber space should be used to to effect actions, again in a timely fashion. These requirements must be made even within operational constraints imposed by the environment and by available resources.

Research Agenda

The above additional concerns give rise to a fresh set of research challenges which we tackle in this project: 

[C1] Convergence of physical and cyber spaces.

As mentioned above, the physical and cyber spaces have to be closely coupled even within resource constraints and other operational constraints imposed by the environment. To achieve the convergence goal, we address two technical issues. First, we consider how sensors can effectively gather data that characterize the physical space and deliver the most useful and relevant data to the cyber space, considering the given bandwidth, delay, and signal attenuation constraints. Tackling this technical problem boils down to resolving issues such as (i) spatial-temporal formation of ad hoc communication clusters whose data dissemination/fusion/aggregation actions must be coordinated in a distributed manner, and (ii) reliable and temporally guaranteed wireless communication of high-confidence data over unreliable, best-effort multihop wireless networks. Second, we address the issue of how intelligence in the cyber space can take advantage of the sensor data to the maximal possible extent. To tackle this problem, we incorporate computationally efficient models that can accurately characterize the physical characteristics to be detected/tracked. The necessary input to the models is then collected by tasking and activating sensors in a spatial-temporal robust manner.

[C2] Acknowledging and managing uncertainty in physical-cyber space.

Uncertainty in physical-cyber space results from the fact that the physical space is inherently non-deterministic. While eliminating uncertainties in the measurement and inference process is intuitively appealing, we argue that acknowledging the existence of uncertainty and enabling the data dissemination / collection mechanisms, inference models, and decision making processes to deal with uncertainty in a robust fashion is perhaps more appropriate. Hence we build uncertainty early into protocol designs and inference models and allow the use of possibly incomplete / inaccurate sensor data with measurement errors? For example, the density of sensors deployed in a monitored area affects the redundancy of the information gathered and consequently the system's ability to identify and remove statistical outliers. We effectively determine the density and placement pattern of sensors and their sensing / communication schedules so as to reduce the susceptibility of inference results to uncertainty.

[C3] Designing deeply embedded but open system components.

As mentioned above, an effective DITSCN must be intimately aware of its deployment environment, including both its physical and population aspects. In the physical aspect, a radiational plume model depends on parameters such as the propagation and dispersion coefficients. Propagation is determined by the prevailing strengths and directions of winds or water flows. Dispersion depends on factors such as temperature and the contour of plume concentrations. Modeling various physical phenomena has been traditionally the tasks of physicists and ignored by systems and network designers. With the convergence of physical and cyber spaces, they now play a central role in the overall performance of a DITSCN.

The people aspect of DITSCN recognizes the system as ultimately a protection mechanism that can positively impact the lives and well being of human users. For example, a radiational DITSCN protects people from harmful plume effects; its utility can be reasonably measured by the sizes and densities of the covered population areas and ultimately those of the evacuated areas. To this end, we take into account of human population, street layout, and routes for evacuation, early in network and systems design, and explore availability and accessibility of these physical and population data in national data repositories.

While the design of DITSCN requires that sensors be deeply embedded in both the physical and population aspects, the system itself has to be open and accommdating with various sensor modalities. Applications of sensors for different physical phenomena will require that diverse components from different vendors be used for diverse applications. To support such diversity on a large scale, we design system and software components that conform to open and well-defined data and communication interfaces. We make available configuration parameters that define the sensing process, format of captured sensor data, and communication protocol for transporting data between components for use by component manufacturers and software developers.

In the first year of the project, we have carried out three synergistic research tasks to address the challenges C1--C3.

• (T1) Network formation by static sensor selection, placement, and coverage addresses formation of the sensor network to ensure that critical physical space data are available where and when needed for cyber space analysis.
• (T2) Tracking targets with quality by considering the issue of tracking mobile targets with certain level of quality of monitoring (QoM) but with the use of a minimal number of sensors.
• (T3) Systems prototyping of a sensor-cyber network testbed for plume detection, identificaiton, and tracking. Building on the existing DITSCN SensorNet architecture (developed in the D.C. deployment), we engage in the enhancement and systems prototyping of an open SensorNet architecture that allows incorporation of disparate sensor modalities and supports multi-hop wireless communications.

Participants

• Guanghui He (Postdoctoral research fellow 2004-2005)
• I-Hong Hou (Ph.D. student)
• Yong Yang (Ph.D. pre-candidate, supported in part by Vodafone Graduate Fellowship)
• Honghai Zhang (Ph.D. 2006)

Collaborators

• David Yau, Department of Computer Science, Purdue University
• Frank Denap, Nageswara S. Rao, Mallikarjun Shankar, Oak Ridge National Laboratory

Major Publications

1. Honghai Zhang and Jennifer C. Hou, "Is deterministic deployment worse than random deployment for wireless sensor networks,'' Proc. IEEE INFOCOM, March 2006 (acceptance ratio = 18%).
2. Honghai Zhang and Jennifer C. Hou, "An algorithm for maximizing alpha-lifetime for wireless sensor networks,'' International Journal of Sensor Networks (IJSNet), Vol. 1, No. 1/2, pp. 64-71, June 2006 (invited).
3. Ahmed Sobeih, Wei-Peng Chen, Jennifer C. Hou, Lu-Chuan Kung, Ning Li, Hyuk Lim, Hung-Ying Tyan, and Honghai Zhang, "J-Sim: a simulation and emulation environment for wireless sensor networks,'' IEEE Wireless Communications Magazine, Vol. 13, No. 4, pp. 104--119, August 2006.
4. David K. Y. Yau, Jennifer C. Hou, Shankar Mallikarjun, and Nagi Rao, "An experimental network testbed for radiational plume detection and tracking," Proc. Int'l Symp. on Innovations and Real-time Applications of Distributed Sensor Networks, Washington DC, October 2006.
5. David K. Y. Yau, Jennifer C. Hou, Shankar Mallikarjun, and Nagi Rao, "Systems Support for Radiational Plume Detection, Identification, and Tracking Sensor_cyber Networks," Proc. of NSF Cyber-Physical Systems Workshop, Austin, TX, October 2006.
6. Guanghui He and Jennifer C. Hou, "Tracking targets with quality in wireless sensor networks,'' ACM Trans. on Sensor Networks (accepted in April 2007).
7. J. C. Chin, I-Hong Hou, Jennifer C. Hou, Chris Ma, Nagi Rao, Mobit Saxena, Mallikarjun Shankar, Yong Yang, and David K. Y. Yau, "A sensor-cyber network testbed for plume detection, identification, and tracking," IEEE INFOCOM demo, Anchorage, AK, May 2007.
8. J. C. Chin, I-Hong Hou, Jennifer C. Hou, Chris Ma, Nagi Rao, Mobit Saxena, Mallikarjun Shankar, Yong Yang, and David K. Y. Yau, "A sensor-cyber network testbed for plume detection, identification, and tracking," Proc. ACM Information Processing in Sensor Networks (IPSN), pp. 541-542, Cambridge, MA, April 2007.
9. Jennifer C. Hou, Yong Yang, David Yau, I-Hong Hou, Mallikarjun Shankar, and Nagi Rao, "Sensor placement in realistic environments: theory and practice," in Handbook of Wireless Ad Hoc and Sensor Networks, Sudip Misra, Isaac Woungang and Subhas C. Misra (Eds), Springer-Verlag (London), to be published in the 4th quarter of 2007.
10. Yong Yang, I-Hong Hou, Jennifer C. Hou, Mallikarjun Shankar, and Nagi Rao, "Sensor Placement Revisited in a Realistic Environment,'' submitted to IEEE Int'l Symposium on Information Processing in Sensor Networks (IPSN'08).
11. Hyuk Lim, Jennifer C. Hou, Mallikarjun Shankar, and Nagi Rao, "Localization in wireelss sensor networks,'' Wireless Communications and Mobile Computing, special issue of distributed systems of sensors and actuators, scheduled for publication in the first quarter of 2008 (invited).

Demonstration

• Jren-chit Chin, I-Hong Hou, Jennifer C. Hou, Chris Ma, Nageswara S. Rao, Mohit Saxena, Mallikarjun Shankar, Yong Yang, David K. Y. Yau (alphabetical order), "A Sensor-cyber Network Testbed for Plume Detection, Identification, and Tracking", INFOCOM 2007 and IPSN 2007.
  Download: WMV(compressed, about 49MB) or AVI (Non-compressed, about 1.2GB).
  Please use Microsoft Windows Media Player to watch the demo.

Funding Source

• Oak Ridge National Laboratories, 2007-2008

Related Links

• SensorNet: Nationwide detection and assessment of chemical, biological, radiological, nuclear and explosive (CBRNE) threats.

Detailed Description of Research Agenda

I. Sensor placement revisited in realistic environments

While the initial systems prototyping demonstrate success in using a wireless sensor network for national security applications, several research issues remain to substantially enhance the performance of these networks. One of the most important and up-front issues is where to place sensors in order to fulfill certain performance criteria, subject to the number of sensors to be deployed, the distribution of threats, the terrain, land cover and meteorological conditions, and the population distribution. The performance criteria are either to minimize the maximal detection time (e.g., the time interval from the instant when a dirty bomb explodes to the instant the explosion is detected) or to maximize the population evacuation time (e.g., the time interval between the detection time to the time instant the plume reaches a populated area).

We revisit the sensor placement problem in a more realistic setting [10]. We first consider three sensor placement problems and prove their equivalence. We then focus on formulating/solving the third problem as an optimization problem: given the maximum detection time T and the coverage utility requirement C, how to place sensors so as to minimize the number of sensors. We acknowledge non-negligible detection time, allow the sensing area of a sensor (at certain time instant) to be anisotropic and of arbitrary shape, and define the utility function U(·) to model the expected risks of insufficient coverage (or utilities of coverage) in different parts of the monitoring area. Based on the problem formulation that takes into account of the effects of terrain, land cover, meteorological condition, and population distribution, we propose theoretically grounded solution algorithms for both the 1-coverage and k-coverage cases.

As shown in Figure 2, we evaluate the proposed sensor placement algorithms in the real setting of Port of Memphis. We leverage the elevation/terrain data obtained from the GLOBE database, the population distribution obtained from the LandScan 2005, the meteorological data produced by the National Resources Conservation Service (NRCS), and the SCIPUFF model to produce the contours of the dispersion. We also define the utility function to be proportional to the population distribution. Several important observations have been made in the empirical study:

1. The proposed algorithms outperform random and grid placement by almost 50% in terms of the detection time.
2. Partial coverage with a reasonably high coverage requirement can achieve good average detection time, while requiring a much smaller number of sensors than full coverage.
3. Use of the population distribution as the utility function not only has practical implication, but also helps to reduce the number of sensors needed. This is, in part, due to the fact that it is easier to satisfy a skewed utility function.
4. Instead of considering 16 x 7 representations of wind speeds and directions in a wind rose multiplied by their percentage of occurrence, we show that it suffices to consider 2-3 dominating bins of wind speeds and directions if they exist. Once sensors are placed to handle these dominating meteorological conditions, only a few more sensors are needed to accommodate other less frequent meteorological conditions.

(a) Satellite picture	(b) Terrain
(c) Population distribution	(d) Wind rose

Figure 2. The terrain, population, and meteorological conditions in Port of Memphis and its vicinity.

II. Tracking targets with quality in wireless sensor networks

Tracking of moving targets has attracted more and more attention due to its importance in utilizing sensor networks for surveillance. In this part of project [1], we consider the issue of how to track mobile targets with certain level of quality of monitoring (QoM), while minimizing the number of active sensors. We address the target tracking problem by taking into account of both the coverage and the QoM. In particular, QoM ensures that the probability of reporting inaccurate monitoring information (such as false alarm or target miss) should be as small as possible, even in the presence of noises and signal attenuation. We also analytically whether or not the detection/observation made by a single sensor suffices to tracking the target in a reasonably populated sensor network. Our finding gives a confirmative answer and challenges the long-held paradigm that high tracking quality (low tracking error) necessarily requires high power consumption.

To rigorously analyze the impact of target movement on QoM, we derive both lower and upper bounds on the number of sensors (called duty sensors) required to keep track of a moving target. Based on the analysis, we have devised a cooperative, relay-area-based scheme that determines which sensor should become the next duty sensor when the target is moving (Figure 3). As shown in Figure 4, the number of duty sensor required in the proposed scheme is, in the worst case, approximately 1.2 times larger than the lower bound. It also indicates that a trade-off exists among QoM, the number of duty sensors required, and the load balance.

Figure 3. An illustration that show the current duty node determines the relay area. All the parameters needed to be define the relay area are given in the figure.

Figure 4. The number of duty sensors designated by the proposed scheme versus the derived lower and upper bounds.

III. Systems prototyping of a sensor-cyber network testbed for plume detection, identificaiton, and tracking.

Purdue University, UIUC and ORNL have worked closely together to design and implement a sensor-cyber network testbed for plume detection, identification, and tracking. The testbed has been designed as a model platform to address the three grand systems challenges that the team presented in the 2006 NSF workshop on cyber-physical systems: convergence between the cyber and physical spaces, management of uncertainty due to modeling/sensing errors and probabilitistic environmental effects, and interoperability of diverse systems components by open data and control interfaces.

The systems prototype features the notion of convergence of sensing, communication, computing, and control. As shown in Figure 5, the SensorNet node architecture developed by ORNL is augmented by device driver support that integrates RFTrax RS-485 RAD-CZT radiation sensor and WMS-23 anemometer. The RFTrax detects the presence of a radiation source and its intensity, while the WMS sensor detects wind speeds and directions (that will affect the propagation of the plume). Wide area communication is realized by equipping SensorNet nodes with multi-hop wireless communication capabilities, namely to send/receive wirelessly over multiple hops to the cyber center. This is realized by equipping each SensorNet node with a Linksys router running AODV as the routing protocol. The cyber center hosts J-Sim integrated with the SCIPUFF plume model and the GLOBE database of terrain elevation data. This allows the propagation of the plume to be modeled, subject to the distribution of threats, the terrain and meteorological conditions, and the population distribution (Figure 6). Finally, the results derived by J-Sim integrated with SCIPUFF are used for future tasking, i.e., dispatching a group of ER-1 robots along the direction in which the plume propagates. The tasking commands are sent from the cyber center to robots via the Linksys router-based multihop wireless network.

Figure 5. SensorNet Node hardware components. The hardware components are housed in a protective enclosure consisting of a processor unit and hard disk, power supplies, modem, serial interface, and LAN interface.

(a) Wind speed = 1 m/s from North

(b) Wind speed = 1 m/s from South

Figure 6. Sensing areas within detection time T=30, 60 and 90 minutes, under different meteorological conditions. Note that the sensing area are prolonged along the opposite direction of the wind.