Real-Time Sensor Fusion to Enhance Plume Detection in Urban Zones

Transcription

1 Real-Time Sensor Fusion to Enhance Plume Detection in Urban Zones Tracy Kijewski-Correa 1, Andrew Henderson 2, Luis Montestruque 3, Jonathan Rager 4 1 Associate Professor, Department of Civil Engineering and Geological Sciences, University of Notre Dame, Notre Dame, IN, USA, tkijewsk@nd.edu 2 Development Engineer, EmNet LLC, Granger, IN, USA, ahenderson@heliosware.com 3 Director of Operations, EmNet LLC, Granger, IN, USA, lmontest@heliosware.com 4 Graduate Student, Department of Civil Engineering and Geological Sciences, University of ABSTRACT Notre Dame, Notre Dame, IN, USA, agrager2@nd.edu The intentional or accidental large-scale release of a chemical-biological agent in an urban zone remains a viable threat to Homeland Security. As these silent, invisible and odorless plumes can cause massive casualties that would overwhelm emergency facilities, modeling, tracking and real-time prediction of plume dispersion and lethality is central to developing a user-friendly first responder's tool for evacuating civilians and managing emergency personnel. In response to this need, the authors and their collaborators have developed a wireless sensor network concept that fuses real-time meteorological and contaminant detection data with a simplified computational fluid dynamics model called CT-Analyst to hindcast plume origins and project subsequent dispersion. This paper overviews the development of the sensor network, the interfacing of real-time data to the pre-existing computational model, and the development of a hindcast algorithm to determine likely contaminant sources from measured sensor data. The results of field demonstrations conducted on the campus of the University of Notre Dame are presented to evaluate the effectiveness of embedded sensor networks fused with simplified computational models in tracking airborne contaminants. The paper concludes with a discussion of the scaled-up deployment of a larger network of sensors for demonstration in a major metropolitan area in August INTRODUCTION One of the nation s greatest fears is the intentional or accidental large-scale release of a chemical-biological (chem-bio) agent in a densely populated area. The silent, invisible and odorless plume could cause massive casualties that would overwhelm emergency facilities. As such, modeling, tracking and real-time prediction of plume dispersion and lethality is central to Homeland Security in developing a user-friendly first responder s tool for evacuating civilians and managing emergency personnel. Tracking of plume evolution can be accomplished through one of two mechanisms: 1) Physical detection of the plume by embedded sensor networks. 2) Prediction of plume dispersion by computational models. The former approach provides actual measurements of the plume boundary, though it is limited by sensor sensitivity and cost and lacks the ability to predict plume trajectories. The latter approach, which has received more attention within the Homeland Security community, is completely reliant on the accuracy of building-resolving Computational Fluid Dynamics (CFD) models that have yet to be comprehensively validated in dense urban zones representative of the highest value targets. Computational models such as FAST3D-CT, a Large Eddy Simulation (LES) package developed by the Laboratory for Computational Physics and Fluid Dynamics at

2 the Naval Research Laboratory (NRL) [1], implicitly model unique urban features such as large scale vortex shedding from buildings, convective heating, complex flow patterns in urban canyons and recirculation zones that trap contaminants, dramatically increasing the local concentration of toxic agents as plumes take on rather unexpected paths [2]. Unfortunately, these models require immense computational resources and are thus not extendable to a real-time framework to serve the needs of first responders. A computational compromise is reached by alternative dispersion models, e.g., Gaussian puff models like DTRA s SCIPUFF [3]. Unfortunately, such models average out all turbulent characteristics of the wind field and ensuing plume, losing any evidence of high concentration zones where plume density can be six times the average, yielding an unrealistic picture of the safe lateral boundaries of a chem.-bio plume [4]. Furthermore, discrepancies between Gaussian puff models and high resolution CFD models are more pronounced in denser, taller and highly corrugated urban landscapes [5]. Thus, in order to maintain accuracy while minimizing computational burden, a real-time software called CT-Analyst was recently introduced, which conservatively predicts the hazard area based on multiple pre-computed realizations from FAST3D-CT condensed into dispersion nomographs [6]. Fully-functional versions of this software have already been developed for a number of urban zones, including Chicago, Los Angeles, and Washington DC. Despite the availability of software such as CT-Analyst, the prediction of plume evolution in cities with high profile morphology and a great deal of variation produce hazard areas whose characteristics are quite sensitive to release location and meteorological conditions [5]. As such, models like FAST3D-CT and even CT-Analyst are no more accurate than the wind morphology forecasts feeding into them [6]. In most cases, such meteorological data is reported at nearby airports and deviates considerably from the actual conditions in the downtown area, due to the significant spatial variability of the wind field in urban zones. Thus, despite all the computational advances of LES, their forecasts are still limited by the uncertainty of data input into the model. Initially, CT-Analyst was developed with the expectation that basic meteorological data and evidence of plume location would be input manually as anecdotal data on the event becomes available. This information would then enable not only the source of the attack to be hindcast, but would also enable the forward projection of the plume to be simulated. Unfortunately, this information many not be immediately available, particularly in any quantitative form, thus hindering the utility of this first-responders tool. Reliance on accurate data to drive these models subsequently motivated the current research program: automated interfacing between CT- Analyst and distributed networks of meteorological and chemical sensors capable of providing continuous, quantitative evidence of plume migration and environmental conditions supporting its dispersion [7]. This data fusion is achieved through embedded wireless sensor networks. In this so-called hybrid framework, the CT-Analyst model attempts to offset the limits in practical density of a stand-alone sensor network for plume detection, while the real-time sensor data feeds the predictive model for more accurate delineation of plume boundaries and release points. The proposed network will provide point verifications of plume location as well as accurate real-time meteorological data at various locations in the an urban zone using stationary nodes of low-cost sensors, as well as mobile sensing capable of moving more expensive but sensitive instrumentation along plume boundaries. Such an approach is in direct response to a recent NRC report calling for a more [effective means to] assimilate into models an appropriate range of meteorological data from observing systems as well as real-time data from [chem.-bio] sensors [8].

3 Thus, the objective of this research program is the development of the embedded sensor network, its interfacing with CT-Analyst, including the development of appropriate algorithms for in-network data processing, and a proof-of-concept demonstrated through releases of a benign chemical agent, sulfur hexafluoride (SF6), on the campus of the University of Notre Dame and in a major metropolitan area. In total, this project represents a joint research effort between Civil and Electrical Engineers at the University of Notre Dame, engineers at EmNet LLC, researchers at Naval Surface Warfare Center, Crane Division, and researchers at Naval Research Labs. Thus the larger effort deals with the a variety of fundamental research tasks associated with the development of additional chem-bio sensors, efficient wireless network routing protocols, and mobile sensing platforms, as detailed in [7]. However, this paper will focus only on the development of the sensor network, the interfacing of real-time data to the CT- Analyst computational model, and the development of a hindcast algorithm to determine likely contaminant sources from measured sensor data. The results of field demonstrations conducted on the campus of the University of Notre Dame are presented to evaluate the effectiveness of embedded sensor networks fused with simplified computational models in tracking airborne contaminants. The paper concludes with a discussion of the scaled-up deployment of a larger network of sensors for demonstration in a major metropolitan area in August (a) (b) InfraRan Power Regulator I-node Figure 1: (a) WXT510 weather transmitter interfaced with I-node and (b) prototype of InfaRan chem. sensor with Inode. HARDWARE EmNet, LLC provides the wireless sensor network (WiSN) hardware enabling real-time monitoring of the environment. The WiSN is composed of three distinct types of nodes, which have been extensively field tested as part of a urban deployment in South Bend, IN, to monitor Combined Sewer Overflow (CSO) Events [9]. Instrumentation nodes, or I-nodes, collect data from the environment and transmit it wirelessly to the nearest Gateway node or G-node. The G-node collects the data from the surrounding clusters of I-nodes and uploads it to a secure web portal, where the raw data is processed and displayed, via a secure internet connection. If the distance between the I-node and the G-node is too great, a Repeater node, or R-node can be added, providing a bridge between the two nodes. In this way, the WiSN can monitor large areas of interest, including dense urban zones, in real time. Furthermore, the use of clusters of sensors enables a scalable architecture that permits the WiSN to be applied readily to small demonstration areas, like that at the University of Notre Dame, or large deployments in major

4 cities. The I-node platform is robust enough to accommodate a wide variety of sensors, providing local memory and processing capabilities for basic data acquisition and communication by wireless radio. For the present study, this platform has been interfaced with the InfraRan Specific Vapor Analyzer by Wilks Enterprise, Inc., an infrared-instrument for the detection and measurement of SF6 in ambient air at 20 parts per billion (ppb) levels (Fig. 1a), as well as a inhouse, customized SF6 detector, with a comparable detection thresholds. These instruments will serve as the chemical (chem.) detectors in the WiSN for the purposes of demonstration. It should be noted that the chem. sensors themselves are the limiting factor in this effort, as affordable, high sensitivity sensors, particularly those suited for detecting benign agents represent a very narrow market segment. Environmental conditions regulating plume dispersion are monitored by the Vaisala WXT 510 Meteorological Station with ultrasonic wind velocity measurement, as well relative humidity, barometric pressure and temperature (Fig. 1b). CT-ANALYST PLATFORM Figure 2: CT-Analyst graphical user interface The most powerful capability of CT-Analyst is its ability to provide a rapid forward simulation of a plume s movement from a fixed location with a specific released mass, as a function of time, wind direction, and wind speed using a convenient graphical user interface (GUI) (Fig. 2). The platform allows end users to place sensors (triangles in Fig. 2) at any location modeled area, which will register as hot (red) when the plume (colored contours in Fig. 2) has reached that location, with the projected concentration in (g/m 3 ). Unfortunately, a forward simulation of plume boundaries can only be generated with knowledge of the amount (mass), location and time of the chem.-bio release and current wind conditions (shown in lower left corner of Fig. 2). While the latter is readily available from the WiSN, the requisite information regarding the chem.-bio release is generally not available in an attack. However, even without this information, CT-Analyst possesses the capability, based on the arrangement of hot and cold sensors and current wind field data, to identify a likely source region. End users may then elect to place a source marker (see star in Fig. 2) within that region, specify a released mass amount and then forward project a plume evolution with time, including projected concentrations at any of the sensors upwind of the source. This constitutes a simulation within the CT-Analyst platform. Depending on the size of the likely source region, a near infinite number of source locations and simulations could be investigated by the end user, with no definitive means of

5 EXAMPLE HINDCAST SEQUENCE Cold Sensors Cold Sensors Hot Sensor Actual Source Hot Sensors Hot Sensors Actual Source Actual Source Hindcast Source Hindcast Source Hindcast Stage 1: First sensor goes hot due to plume arrival, first guess hindcast source is generated. Source cannot be reliably identified from a single sensor hit Hindcast Stage 2: Additional sensors go hot as plume disperses, hindcast source location is refined. Hindcast capabilities are enhanced in the presence of multiple sensor hits Hindcast Source Hindcast Stage 3: A sufficient number of sensors go hot as plume disperses, to permit the source location to be identified with sufficient accuracy. Figure 3: Demonstration of source convergence in hindcast algorithm identifying the actual source location. Keep in mind, that not only did the software in its original form provide no means to isolate the source of an attack, but also required the end user to manually input wind conditions and sensor locations and status (hot vs. cold). Thus, the fusion of real-time sensor data and the provision for an automated hindcast represent major advancements in the state-of-the-art for first responders. HINDCASTING SOURCE LOCATIONS With distributed chem. sensors continuously monitoring for concentrations of the target agent (in this case SF6), the computational platform remains dormant until the concentration at a sensor surpasses a minimum threshold (goes hot). At this point, the network is activated and concentration data from all networked sensors and current meteorological conditions are reported to CT-Analyst to initiate the hindcast process. Interfacing between CT-Analyst and the WiSN is achieved by developing a COM server and a C/C+ client program that simultaneously access the EmNet MySQL database server. The hindcasting algorithm that has been developed can be defined as an optimization tool, which correlates simulated and measured concentration data to determine the most likely source of the airborne contaminant. To fully characterize the source of attack, the location, time of release and mass released must be determined. Thus this process is executed in three stages to address each of these unknowns: (1) Search for Release Location

6 (2) Search for Release Time (time elapsed between release of agent to first hot sensor) (3) Search for Release Mass. A cost function is used to evaluate the fitness of the candidate solution for the source location: (1) where the returned values of simulated concentrations from CT-Analyst are combined in a vector,,,, and the actual concentrations being read from the environment via the WiSN are combined into vector,,,,. By analyzing the gradient of the cost function with respect to space and time, it can be shown that the cost is minimized when the actual time and location are specified. Search for Release Location: In order to provide an educated guess of where the hindcast search should begin, a low resolution search of the entire source region is executed, and the cost function is evaluated for each of these points. The location of the absolute minimum of the cost function is then used as the initial location for the high resolution search. The high resolution search first begins by simulating the source of the release at a certain initial location, and evaluates the cost function:,. The algorithm then takes a set number of points around the initial point, within a specified radius, and evaluates these costs functions, saving the minimum cost value as,. If, <,, the hindcasting algorithm will select the coordinates of, as the next location for to chose for a simulated source location. The algorithm repeats this process until,,. In this manner, the algorithm moves the source location closer and closer to the actual release point. Search for Release Time: The gradient of the cost function with respect to time has been observed to show marked differences both before and after the actual release time. With this knowledge, the algorithm begins evaluating the cost function at time t = 0:, 0 and then at the next incremental time step:. If, the time is incremented again, and the process is repeated until. The algorithm then returns to earlier step of calculating the minimum cost function with respect to position. In this way, the algorithm progresses closer and closer to a best fit of the model, ending the search when a stopping condition has been met. This condition is usually given by a negligible improvement on the cost function on two consecutive steps. Search for Release Mass: The final stage of the hindcasting algorithm is to determine the release mass of the airborne agent. Again, a cost function is evaluated at an initial point, starting with mass equal to 0, and. When, the hindcasting algorithm has successfully finished its search for release location, time, and mass. In general, the entire process can be completed in a matter of seconds. This entire process is shown through a series of screen captures in Figure 3 and is repeated with time as additional sensors go hot in the presence of the advancing plume. Thus the process of hindcasting evolves with time and becomes increasingly reliable as more sensors report concentrations, emphasizing the importance of adequate sensor density and the advantages of scalable wireless sensor networks.

7 VALIDATION OF HINDCAST ALGORITHM Figure 4: Sensor placement in benchmark test In order to test this hindcasting algorithm, a benchmark problem was defined that supplied the authors with concentrations at a pre-defined array of sensors in an urban zone (Fig. 4) every 60 seconds for 17 minutes, beginning when the first sensor goes hot. The wind speed and direction were specified as 2 m/s at 210. The final results of the hindcast algorithm are provided in Table 1 and confirm that the hindcasting algorithm was successful in the blind identification of the source location, time of release, and mass (A video demonstrating the hindcast convergence is also available at It is worth noting that the initial position from the low resolution search yielded a starting point of m East, m North (270 meters from actual release). After one iteration in the high resolution search, the hindcast algorithm identified the release location within 49 meters and the release time within 36 seconds. The hindcast algorithm ultimately took a total of 3 iterations after the initial best guess search to identify the source to the levell of accuracy in Table 1. Table 1: Comparison of hindcast source predictions to benchmark problem [min:sec] Benchmark 1:58 Hindcast 2:00 (Difference) (2 sec) FIELD VALIDATIONS Release Time Stamp Release Mass [g] Release Location, Release Location, Easting [m] Northing [m] 1E E (0) (Resultant = 17.5 m) The WiSN described herein was deployed on the campus of the University of Notre Dame in 2008 for a series of field validation studies. The wireless meteorological stationss were deployed on the rooftops of four buildings bounding the demonstration zone called South Quad. The met stations were placed such that each station monitored a different cardinal direction associated with winds approaching the demonstration zone, numbered 1-4 in Figure 5. A fifth met station was placed on the tallest building on campus, well above any surrounding structures (no. 5 in

8 Fig. 5). The met stations were placed on poles and interfaced with I-nodes, which transmitted the data to the network gateway. The original design of this meteorological portion of this network incorporated the use of disposable batteries, and functioned on a 4% duty cycle. Due to the nature of the intended functionality of the WXT510 and power constraints, the authors opted to instead supply uninterrupted power to the Inodes. This allowed for continuous data collecting and more frequent transmission, as well as allow for greater through-put of data between the meteorological and chemical WiSNs. The upgraded station is shown in Figure 5. This emphasizes an important issue with WiSN: many sensors were designed for deployment in traditional sensor networks that are not duty cycled, requiring more frequent operation and renewable power supplies within the WiSN. The integrity of measurements from this array was verified by comparing the data to NOAA measurements recorded at the South Bend Regional Airport. N I-Node and WXT510 Solar Cell Figure 5: Meteorological sensor WiSN topology (left) and photos of deployed sensors on rooftops of buildings at the University of Notre Dame Due to the high value of the chem.-bio sensors, they were deployed at pedestrian levels only when SF6 was released. Therefore ruggedized enclosures with wheels were designed for rapid transport and re-deployment of the units during various tests in South Quad. As commercial-off-the-shelf components were used for the enclosures, the size was somewhat constrained by the available dimensions from the manufacturer. The form can be shrunk considerably, as the limiting factor would then be the size of the chem-bio sensor, given the small size of the Inode. Each enclosure was equipped with a detachable tripod to elevate the units approximately 1 m off the ground (Fig. 6). The chemical sensor Inodes were equipped with two different operating parameters: one designed for manual operation on a mobile sensor and the others for stationary, autonomous operation. Due to cost constraints, only five chem.-bio (InfraRan) sensors could be purchased for the demonstration. Three were mounted in the

9 Source InfraRan + Bag Sampler Bag Sampler Met. Station Figure 6: Topography of integrated demonstration at the University of Notre Dame, black lines delineate path of mobile sensor with numbered sampling locations ruggedized enclosures with tripods and treated as stationary nodes collecting data at 1 Hz, conducting a 30 second average and transmitting that measurement every 30 seconds. The fourth was interfaced with an I-node and mounted on an i2cargo Segway (the fifth had a component failure on the day of testing). As discussed in [7], release protocols for SF6 were first developed at the University of Notre Dame in the summer of With these in hand, a demonstration of all aspects of the sensor network, under the live release of SF6, was held on October 22, 2008 in the South Quad of the University of Notre Dame. The purpose of this test was to incorporate the met station data and chem. sensor data over the wireless sensor network to trigger a response by CT-Analyst in real-time. Two lines of sensors were deployed: the first, placed at a distance of approximately 90 m downwind from the source, consisted of Wilks InfraRan units in parallel with bag samplers. The second line of sensors, bag samplers, was placed at a distance of 145 m downwind of the source. The bag samples were taken at 10 minute intervals, with the first sample taken at the 5 minute mark after initial release. The Wilks units operated continuously over the entire release period and reported their findings every 30 seconds. Figure 6 overviews the demonstration zone with the thirteen numbered measurement points, interconnected by arrows, signifying locations where mobile units carrying one of the chem. sensors recorded concentrations. During the demonstration period, winds were consistent from the East-Southeast with wind speeds averaging approximately 5 m/s and the temperature was 10 C. The release rate of SF6 was 10 L/min (1.1 g/s) for 20 minutes, from a pedestrian height of approximately 1.5 m. The streaming data confirmed sensitivity of the Wilks device to surging concentrations of SF6, as well as the ability of the wireless network to activate, perform a network-level average of the wind speed and direction data, relay this information and the SF6 concentrations to the CT- Analyst server in real-time, and initiate a successful hindcast sequence. Note the responsivity was quite impressive given the concentrations were near the lower detection limit of the InfraRan

10 units. As common with bag samplers, some collected samples were faulty, but the successful samples were analyzed off-line and correlated against the InfraRan outputs streamed over the wireless network. FIELD DEMONSTRATION IN METROPOLITAN AREA The small scale demonstrations in the Fall of 2008 at the University of Notre Dame helped to validate the underlying technologies developed for hybrid plume detection in real-time. The next stage of the project is a demonstration in a major metropolitan area in the late summer of 2009, with aerodynamic features that will present a more faithful validation of the underlying CFD model. The demonstration will entail the distribution of approximately 20 sensors (15 chemical detectors and 5 meteorological stations). The chemical detectors will be deployed on light poles to enable continuous access to power, with the ability to monitor near pedestrian level conditions but with minimal risk of tampering. Meteorological stations will be deployed on the rooftops of four surrounding high-rises at the boundaries of the demonstration zone to measure the direction and speed of upper elevation winds above the urban zone, as well as other relevant meteorological details such as temperature and relative humidity. In addition, pre-existing meteorological installations within and surrounding the downtown area (NOAA) will be incorporated, whenever possible. The network will be deployed and allowed to operate for a period of at least one month to verify its functionality and to observe background levels of SF6 and other agents that may potentially interfere with measurements. During the demonstration period, slated for late August 2009, bag samplers will flank these installations and will analyze, off-line, concentrations recorded at various intervals during the demonstration as a secondary validation. During the demonstration, a line of SF6 canisters will dispense the agent upwind of the network at street level for periods of 20 minutes. The integration of chemical concentration data and meteorological conditions from these sensor networks at a gateway node will allow real-time interfacing to a new Google-Earth-enabled CT- Analyst model for the demo city, updated with the most recent geometries of buildings and developments. To further enhance capabilities, the same high sensitivity chemical detectors will be mounted on mobile platforms, such as Segways, that can move along the forecast regions with greatest uncertainty or highest value to further delineate plume boundaries and offset the limited density of the stationary sensors. This mobility component provides a demonstration of the ability to integrate data from sensors on emergency vehicles. The end result will be a more reliable means to monitor, in real time, the evolution of toxic plumes in complex urban environments and will enable the blind identification of the source location and forward projection of the plume concentrations to permit evacuation and mobilization of emergency services. CONCLUSIONS The integration of wireless sensor networks with computationally efficient software platforms for plume dispersion and hindcasting provides an effective and economical venue for real-time decision making in the event of a chemical or biological attack in an urban zone. While this project represents a major collaboration between researchers at a number of institutions and organizations to achieve this end, this paper focused specifically on the development of the sensor network, the interfacing of real-time data to this computational model, and the development of a hindcast algorithm to determine likely contaminant source locations from measured sensor data. The hindcast model was shown to be effective in identifying contaminant source locations with respect to three key variables in a matter of seconds. The network was field

11 tested successfully at the University of Notre Dame in the fall of 2008 and the extension of this effort to a major metropolitan area in the late summer of 2009 was also overviewed. This hybrid detection scheme has the potential to rapidly provide critical information to aid in the evacuation of areas using more reliable plume predictions derived from real-time, in-situ observations in a distributed wireless sensor network. As the system features a user-friendly GUI, automatic monitoring and alerting of contaminant levels, and rapid interfacing for more reliable plume prediction with little to no user effort, it can be easily used by first responders in the field for the protection of human life. ACKNOWLEDGEMENTS The authors wish to acknowledge the support of Department of Defense Grant N C- 8510, their collaborators at the University of Notre Dame, Naval Surface Warfare Center - Crane Division, and Naval Research Labs. REFERENCES [1] J. Boris, The threat of chemical and biological terrorism: preparing a response, Comput. Sci. Eng. 4 (2002) [2] M.J. Brown, Urban dispersion challenges for fast response modeling, Los Alamos Natl. Lab. Report LA-UR , [3] R.I. Sykes, et al., PC-SCIPUFF version 1.0 technical documentation, Titan Corporation, Princeton, NJ, [4] G. Settles, Fluid mechanics & homeland security, Annu. Rev. Fluid Mech., 38 (2006) [5] J. Pullen, J.P. Boris, T. Young, G. Patnaik, J. Iselin, A comparison of contaminant plume statistics from a Gaussian puff and urban CFD model for two large cities, Atmospheric Environment 39 (2005) [6] J.P. Boris, K. Obenschain, G. Patnaik, T.R. Young, CT-Analyst: fast and accurate CBR emergency assessment, in: Proceedings of the First Joint Conference on Battle Management for Nuclear, Chemical, Biological and Radiological Defense (Williamsburg, VA, 2002). [7] T. Kijewski-Correa, J. Talley, P. Bauer, M. Haenggi, P. Antsaklis, M. Lemmon, N.L. Laneman, L. Montestruque, J. Fulton, G. Patnaik, Real-time plume detection in urban zones using networked sensing data, in: Proceedings of Chem-Bio Defense Physical Science and Technology Conference (New Orleans, 2008). [8] National Research Council (NRC), Tracking and Predicting the Atmospheric Dispersion of Hazardous Material Releases: Implications for Homeland Security, Natl. Academies Press, Washington, DC, [9] T. Ruggaber, J. Talley, L. Montestruque, Using embedded sensor networks to monitor, control and reduce CSO events: a pilot study, Environmental Engineering Science 24 (2007)