risk-aware sensor networks for critical infrastructure ...

th

11 International Conference on Shock & Impact Loads on Structures 14-15 May 2015, Ottawa, Canada

RISK-AWARE SENSOR NETWORKS FOR CRITICAL INFRASTRUCTURE MONITORING Jamieson McCausland†, Rami Abielmona†, Rafael Falcon† and Emil Petriu‡ Research & Engineering, Larus Technologies Suite 310 – 170 Laurier Ave. W., Ottawa, Canada, K1P 5V5 e-mail: webpage: http://www.larus.com

Keywords: risk management; risk assessment; critical infrastructure protection; robotic sensor networks; situational awareness; high-level information fusion

Abstract. The use of robotic sensor networks (RSNs) for Critical Infrastructure Protection (CIP) has been quite popular in recent years. In this paper, an auction-based node selection technique is considered for a risk-aware RSN applied to CIP; its main goal is to maintain a secure perimeter around the CIP, which is done by detecting high-risk network events and mitigating them through an optimized response. Such a response will only involve the most suitable robotic nodes and must successfully counter any detected vulnerability in the system. This paper is a study of the role played by multi-objective optimization (MOO) in the elicitation of responses from a risk-aware RSN that is deployed around a critical infrastructure. The MOO will produce a set of optimized responses for each network segment for a security operator to pick the most suitable response. 1

INTRODUCTION

A sensor network is rarely referred to as a “sensor network”, but called by one of many specific sensor network types: Wireless Sensor Networks (WSNs), Mobile Sensor Networks (MSNs), Robotic Sensor Networks (RSNs), Sensor and Actuator Networks (SANs), Wireless 1 Sensor and Actor Networks (WSANs) and more . Each type of sensor network represents a certain level of functionality and sophistication, which must be considered when choosing one for a specific use case. For example, a recent SAN use case is that of a mobile robot team assisting 2 a network of static sensor nodes in a variety of ways . Typically, the most functionality in a sensor network will be found within an RSN. A network of robotic nodes will represent many onboard sensing instruments, actuators, and processing capabilities. In order to manage high-volume sensor data, coordinate actions performed by actuators, and execute complex algorithms to enable autonomy, powerful embedded software are required to be integrated within these sensor nodes. A robotic sensor node can be decomposed into four primary components: sensing, computation, communication, and actuation. Compared to a conventional sensor network, it is expected that an RSN will be more demanding in terms of computational power, memory, and power consumption. Despite these drawbacks, an RSN can be programmed to process the data from its sensors that come in the form of raw data features. Such data features could be used to extract risk features and perceive whether a risky situation is occurring. More interestingly is that an RSN can be programmed with † ‡

Research & Engineering, Larus Technologies School of Engineering and Computer Science, University of Ottawa

295

Jamieson McCausland, Rami Abielmona, Rafael Falcon and Emil Petriu

the ability to react in the environment when exposed to a risky situation. An RSN could respond using the actuators available to the robotic sensor nodes, such as differential drive platforms that enable mobility. This would allow the network topology to change in response to events either in the sensor network itself or in the deployment environment. This ability in a sensor network has proven to be very useful in many applications, including those related to Critical Infrastructure 3 Protection (CIP) . 4 The Canadian National Strategy for Critical Infrastructure categorizes critical infrastructures as: energy and utilities, finance, food, transportation, government, information and communication technology, health, water, safety, and manufacturing. These items are essential for the operations of the country, thus motivating numerous investments to safeguard them. The role of sensor networks in CIP is well documented in literature; there are numerous 5,6,7,8,9,10 examples where researchers were able to produce promising results. Defending critical infrastructures can be accomplished in many different ways such as deploying robotic sensor nodes to monitor the surrounding area perimeter for trespassers and suspicious vehicles. The real world imposes countless risks and the world is inherently full of threats producing risk. It is possible for robotic nodes to be able to sense the risk in their environment. An example of this may be a robot that is unable to function properly or lacks the necessary battery power to remain operational. In the domain of CIP, this could mean loss of essential coverage of a highinterest region, which could potentially be exploited by trespassers. Robotic nodes are not immune to hardware glitches, manufacturing mistakes, harsh environmental conditions, or continuous wear and tear. A collection of robotic nodes working in collaboration to self-organize, can provide a means to mitigate risky situations. With a sophisticated RSN and perhaps a large network (that is difficult to manage) there is a need for the capability to identify high-risk sources in the network. Only a perfect system in an ideal environment will be risk-free, so it can be safely assumed that in actual deployment all nodes in an RSN will be subject to some degree of risk. The ability to assess risk through risk features can identify potentially disruptive events. The advantages of knowing these in a defence system provide substantial motivation to pursue this research endeavour. The deployment and management of a sensor network can become less dependent on a subject-matter expert (SME). In the event of a sensor failure or an increase in frequency of intruder activity, an SME must consider communication ranges, sensor ranges, and expected lifespan of the sensor nodes to deploy a sensor network in the most efficient manner. 11,12 to develop a Risk Management Framework (RMF) This paper extends two existing works for an RSN, which in turn, creates a risk-aware RSN capable of assessing and mitigating risk. Amongst the wide range of applications for a risk-aware RSN, CIP was chosen as a case study. The RMF’s modular components will be extended to include: (1) a risk awareness modules for RSNs in the domain of CIP, (2) a Fuzzy-Auction Multi-Robot Task Allocation (MRTA) technique for an RSN in the domain of CIP, and (3) a response selection technique using Evolutionary Multi-Objective (EMO) optimization. The contributions of this paper are explicitly defined as: 1. A risk feature extraction module is designed and implemented. Its role is to receive raw data features and extract a set of risk features. A specialized risk model for an RSN for CIP has been implemented. 2. An innovative Fuzzy-Auction MRTA technique had been implemented for selecting nodes to participate in a risk mitigation response. With this scheme, each robotic node will 13 employ a fuzzy-based bid metric calculation; 3. A response generation process using the popular Non-Dominated Sorting Genetic 14 Algorithm II (NSGA-II) with a customized genetic encoding structure to generate optimized robotic node configurations is developed. The combination of these contributions give rise to a system that is capable of assessing risk over a large spatial region and responding to concurrent risky events without the requirement of centralized robot management. The rest of the paper is structured as follows. Section 2 reviews some relevant works. Section 3 elaborates on the proposed optimization and bidding methodology. Experimental results are discussed in Section 4 while conclusions are outlined in Section 5. 2

RELATED WORK

In this section, we will review the topics of risk modelling and assessment, MRTA and selforganizing sensor networks. These will form the foundations of the presented work.

296


2.1 Risk-Aware Robotic Systems for CIP Interesting dynamic behaviours and consequences arise when attempting to meet two very conflicting objectives: maximum operation time on a single battery charge and maintenance of the wireless communications channel. Developing a risk model that is capable of successfully transforming observable states of a distributed system into a structured unified format, is a challenging task. Such a format can be labelled as “risk features” and can allow for a comparison against some criteria to identify the risk. Techniques using Hidden Markov Models (HMMs) 15,16 approach this challenge , but require a large amount of a priori information. 15 Tan et al. propose a real-time risk management framework, including a risk assessment module as shown in Figure 1.

15

Figure 1: Real-Time Risk Management Framework. Extracted from the reference . 8

An RSN for CIP is proposed in the following article . The novelty with this approach is that the 10 RSN implements the risk management framework put forth in the following article . From the information submitted by each RSN node, three risk features are extracted: degree of distress, intruder proximity factor and terrain manoeuvrability. The overall risk posed by any RSN node is assessed on the basis of these local risk features and those nodes exceeding a permissible risk threshold are flagged as “nodes in distress” (NIDs), i.e., robotic entities likely to originate a security breach in the perimeter. The response from the RSN is to self-organize in order to maintain as much coverage as possible around the critical infrastructure while minimizing the cost of doing so. For each NID, a restricted group of candidate topologies are evolved with NSGA-II as the evolutionary Multi-Objective Optimization (MOO) algorithm of choice and then ranked following some network operator preferences. The network manager then decides on the most suitable response topology, which is enacted upon the environment. The above framework could only handle a single response to a single event. That is, a network response must be effectuated before another response (such as one to another event) 11 could be orchestrated. With this limitation in mind, the authors later augmented this approach to handle multiple concurrent events, hence creating several independent network regions which can autonomously evolve their own response sets and thus mitigate their local risks. Each response coordinator (i.e., the perimeter node that best perceives the event) initiates an auction to all RSN nodes by advertising the event location. The nodes calculate their availability as 14 responders by using a Sugeno fuzzy inference system (FIS) and decide whether to bid or not. The response coordinator selects the winners according to the rules of the first-price sealed bid auction and then evolves their final positions with NSGA-II. Finally, the winner nodes relocate to these positions along the perimeter. 2.2 Multi-Robot Task Allocation Techniques Many of the above risk management framework processes do not stop after risk assessment, but proceed to produce solutions to the identified risky situations. In the case of an RSN, a risky sensor node may require replacing, reconfiguration, or maybe assistance from other nodes. This section reviews existing literature on Multi-Robot Task Allocation (MRTA) problems. 17 Gerkey and Matarić perform a formal analysis and taxonomy of MRTA. These authors explore the various types of MRTA and the research applied in each group. MRTA problems are typically described by the following criteria:

297


•

Robots o Single-Task (ST): robots capable of a single-task at a time; o Multi-Task (MT): robots capable of multiple concurrent tasks. • Tasks o Single-Robot (SR): A task requiring a single robot to accomplish; o Multi-Robot (MR): A task requiring multiple robots to accomplish. • Task Assignment o Instantaneous Assignment (IA): Available information, the task assignment conducted at a particular time instance; o Time-Extended Assignment (TA): Information regarding robots and tasks can be received at a later point in time, where they need assignment as they are received. Given the above criteria, the authors discuss eight types of MRTA problems: ST-SR-IA, STSR-TA, ST-MR-IA, ST-MR-TA, MT-SR-IA, MT-SR-TA, MT-MR-IA, and MT-MR-TA. This research study will consider the case of ST-MR-IA, which is single-task robots with tasks requiring multiple robots assigned instantaneously upon a high-risk event in the network. The robotic nodes can only participate in a single optimization run at any given time, hence the ST; optimizing the network’s topology will typically require the involvement of more than one robot (hence the MR); and robot nodes are assigned tasks shortly after the detection of the high-risk event. According to 17 the analysis , ST-MR-IA MRTA problems typically become a task of coalition formation. Given a collection of robots in a family, F, a partitioned set E is acquired through a combinatorial optimization algorithm, to form a task-specific coalition. 18 A multi-robot cooperation system is introduced in another paper by the same authors called MURDOCH. The MURDOCH multi-robot cooperation system best fits the MT-MR-TA classification of MRTA problems. MURDOCH pursues intentional and emergent cooperation, such that each robot does not work explicitly with one another, but cooperate for the purpose of task allocation. MURDOCH is developed as a general task allocation system based on a publish/subscribe communication model and is considered a variant of the Contract Net Protocol 19 (CNP) where simple auctions are used to allocate tasks among the family of robots. Main discussion points of MURDOCH are the publish/subscribe messaging and auction protocol. The developers of MURDOCH use anonymous messages, which are addressed by content rather than message destination. The sender publishes a message with content and the type of content in the subject of the message. Subject namespaces are also used in MURDOCH, where subjects of messages can pertain to specific namespaces. Messages can be then directed to a particular group of robots listening for certain namespaces only. The primary components of the MURDOCH task allocation system is based on the underlying distributed negotiation protocol. The tasks are allocated after a first-price and singleround auction session. When a task becomes available, MURDOCH will communicate this with the family of robots using a task announcement message. This is conducted by the auctioneer who then will await bids from the robot population. Before a robot member of the population can submit a bid, they must execute a metric evaluation, which is a utility metric quantifying their fitness of the announced task. Post metric evaluation, a robot member can execute bid submission – a response (bid) message is sent to the auctioneer. The auctioneer will wait for a sufficient amount of time to process each received bid. Once a winner is determined by the auctioneer, it will broadcast a “close of auction” message to the robot population. In MURDOCH only a single winner is chosen among the received bids. The selected winner has now acquired a contract to complete the task initially announced. The auctioneer will monitor the progress of the winning individual and renew the contract as appropriate. It is through this procedure that if the robot carrying out the allocated task is doing so insufficiently, it can be reassigned through an independent auction session. 2.3 Self-Organizing of Robotic Sensor Nodes The development of self-organization techniques for the various types of sensor networks has been pivotal. The concept of a sensor network self-deploying and correcting deployments automatically is a mutual goal among many researchers. State-of-the-art techniques in this research domain will be discussed. 20 The authors propose a new dynamic model for managing the mobility of a mobile sensor network (MSN). A Parallel and Distributed Network Dynamics (PDND) algorithm is proposed, which can execute on each robot individually. The PDND uses a mathematical model to define the laws of motion for each robot formulated using the steepest descent method in optimization.

298


With a MSN in a non-optimal configuration, the system is said to have potential energy. The PDND algorithm executes to reduce this potential energy, allowing the network to “settle” into an optimized state. The authors propose that an MSN could be attracted to a region of interest or even a moving target; through the execution of the PDND algorithm, the robots will establish a new topology based on the defined network dynamics. Opposite (i.e., repulsive) forces can also exist, repelling a member of the MSN away from a region or specific object. These attraction and repellent forces are the result of specific potential functions, the latter providing a measure for the amount of resistance a mobile unit applies against changing positions. 21 Lizhong et al. use a Multi-Objective Differential Evolution Algorithm (MODEA) to satisfy a set of objective functions. These objective functions are defined as: • Area coverage rate; • Network redundant coverage area; and • Energy consumption of sensing The average coverage rate, is described by Equation (1), which provides a ratio of the grid points covered by a sensor node C and the total area of grid points representing a particular region of interest, m x n. ∑ (1) ∑ , , = The sensor network redundant coverage area, , is defined in Equation (2). This function evaluates how many times a specific grid points at (x,y) is being surveyed by other nodes, measuring the amount of redundant coverage.

&

= ! , , "# − 1'

(2)

#

where ! is defined in Equation (3). This equation returns the value one if robot node "# and any other node both cover the location at x and y, otherwise a zero is returned. 1 )*, ∈ "# , ,- , ∈ ". , ), / ∈ 01, 12, ) ≠ /,: (3)

! , , "# = ( 0 56ℎ89 The third objective function these authors consider is the amount of energy being consumed to operate the sensors. The energy demands of the sensor are proportional to the desired range of the sensor. Equation (4) captures the energy demands of the sensor network. &

;

(4)

#

where u represents the sensor parameters and 9# is the sensor radius of node i. The authors use an overall fitness function based on a weighted-sum approach as described in Equation (5). * ?. = @ * + @> > *> (5)

Let @ and @> be the weight coefficients and and > are used as normalizing coefficients for 19 the functions * and *> , respectively. The work employs a genetic encoding of robot coordinate values (i.e., real numbers), a coordinate for each sensor node in the network. Sensor nodes can be disabled by assigning the coordinate (0, 0) for a particular asset in the chromosome. Thus, the optimization algorithm will search for a solution using a subset of sensor nodes, if possible. However, the optimization algorithm always considers all sensor nodes which can become very cumbersome for very large sensor networks. Our proposed methodology optimizes the number of nodes that are involved in a particular response to an event reducing the computational overhead as well as the response latency. 3

PROPOSED METHODOLOGY

In this section, we present the proposed methodology involving MOO and market-based robot coordination for a risk-aware RSN in a CIP scenario. 9,11

As envisioned , the robotic nodes surrounding the critical infrastructure play multiple roles in the system. First, they can detect high-risk events that could jeopardize the integrity of the surveillance conducted by the RSN upon the region of interest. Second, they can become auctioneers for a particular event, which means they are responsible for advertising some event-related information to the rest of the RSN nodes, gathering the received bids and clear off the auction, i.e., announce the winner(s). Finally, they could participate as bidders in any auction circulating around the network.

299


3.1 Risk-Driven Event Detection The RSN nodes can detect two types of events that are quite relevant to the CIP context: (1) another node’s failure (e.g., because of battery depletion) or (2) an attempted perimeter intrusion. Both events demand a corporate RSN response in the form of a topological self-organization, whether to fill the coverage gap brought forth by the failed node or to increase the sampling capabilities around a portion of the perimeter where the intrusion attempt might have taken place. 10

The Risk Management Framework (RMF) is the backbone of the event detection phase. From the raw data stream periodically submitted by each RSN node to the operations centre, a parallel risk stream is dynamically extracted. This risk stream consists of a collection of userdefined risk features which are modelled after different constructs. In our example, each RSN node reports: (1) its current battery level, in percentage; (2) its distance, in meters, from a potential intruder and (3) the terrain maneuverability index associated with its geographical location. The distance from a potential intruder can be gauged with the help of a laser range finder in order to provide the RSN node with depth perception. The terrain maneuverability index corresponding to the node’s location can be queried from the Knowledge Base of the RSN’s deployment environment. In this paper, a random number in [0; 1] has been used for each location. Table 1 depicts the raw-feature-to-risk-feature mapping in our CIP use case.

Raw Feature

Risk Feature

Battery Level (%)

Degree of Distress ([0;1])

Distance to Potential Intruder (m)

Intrusion Risk ([0;1])

Geographical position (x,y)

Terrain Risk ([0;1])

Modeling Construct Fuzzy set with a triangular membership function Fuzzy set with a trapezoidal membership function Linear mapping

Parameters / Expression A=0 B=0 C = 100 A=0 B=0 C=1 D=5 1 – terrain maneuverability index of the node’s position

Table 1. Risk feature extraction The raw data input vector submitted by each RSN node is mapped onto an output risk vector, defined as a collection of local risk values, one per risk feature. The local risk value of any risk feature is determined by applying its modeling construct and parameters/expression to the corresponding input (raw feature data) in Table 1. The overall risk assessment for an RSN unit is obtained after the application of an FIS to the collection of local risk values. For simplicity, our Mamdani-type FIS consists of a single fuzzy rule that reads as follows: IF Degree of Distress is DD-HIGH OR Intrusion Risk is IR-HIGH OR Terrain Risk in TR-HIGH THEN Overall Risk is OR-HIGH The linguistic terms DD-HIGH, IR-HIGH, TR-HIGH and OR-HIGH are modeled as fuzzy sets by taking into account the network manager’s knowledge about these local risks. If the overall risk of any RSN node exceeds a user-set threshold, then it is flagged as a “node in distress” and the pursuit of a corporate risk mitigation strategy to assist that node is initiated. 3.2 Risk-Mitigating Task Announcement The node discovering the high-risk event assumes the role of auctioneer (switching from the available role). Let all possible roles be: available, auctioneer, and participant. As an auctioneer, the primary concern is in forming a network segment and then orchestrating the robotic nodes within. The first step is announcing the auction session to all robotic nodes in the network. This is accomplished by generating a task announcement message to be broadcast to all available nodes. Assuming that the robotic sensor nodes communicate wirelessly over ad hoc networks,

300


neighboring sensor nodes will be able to exchange network messages. This announcement message incorporates the necessary risk details so that each receiving node can appropriately bid on the task. All robotic nodes in role available will receive and accept a task announcement message. Other nodes will forward the message to all neighbors neighbors and then discard the task announcement. As an auctioneer, auctioneer a node can be in one of three states: AWAIT-BIDS, AWAIT CLOSEAUCTION, and OPTIMIZE. The initial state of the auctioneer is AWAIT-BIDS AWAIT where the announcement message has been sent and we need the bids to continue forward. 3.3 Fuzzy-Based Based Bid Evaluation Upon the arrival of the task announcement message, a robotic node becomes aware of the high-risk risk event. Each robotic node then evaluates their availability metric (bid) to the announced task. This availability vailability metric is based on three primary data sources: battery level, level distance to the event, and coverage redundancy. redundancy . The availability metric is evaluated through the use of fuzzy 22 logic. Each robotic node will utilize a Sugeno fuzzy system to produce a bid response to the auctioneer indicated in the network header of the task announcement message. Figure 2 depicts the fuzzy system available on each sensor node.

Figure 2: Sugeno fuzzy model The battery metric is an incentive to bid based on the node’s remaining energy. The distance metric is the robotic node’s incentive to bid based on the distance of the node from the risk event. The closer the robotic node is to the event, the more incentive to bid. Finally, Finall the redundancy metric is the robotic node’s incentive to bid based on the amount of redundant sensor coverage of the user defined security perimeter. Redundant sensor coverage is the amount of sensor fieldfield of-view view overlap resulting in security perimeter points being surveyed by multiple robotic nodes and promotes the incentive to bid. A robotic sensor node must submit a bid to the auctioneer unless its availability is 0.0 (reject). This is accomplished through the sending of a bid message; the auctioneer is still in state AWAIT-BIDS BIDS and must receive a minimum of N bids before entering state CLOSE-AUCTION. CLOSE 3.4 Auction Closing and Ensuing Optimization The final stage of the auction-segmentation auction segmentation process is closing the auction. This is performed by the auctioneer neer by transmitting an auction closed message to all robotic nodes. This message contains a list of node ids indicating the winner bidders in the auction and will now be locked in an optimization session. As a result, the role of the robotic nodes changes from available to participant and cannot be an asset in any other optimization session or bid in any auction. This is the case until the role of the node returns to available. The auctioneer – also the response leader of the segment – begins an optimization optimizati session using the NSGA-II II to search for new segment topologies (solutions) to mitigate the risk. The auctioneer then proceeds to optimize the number of the winner nodes that are actually required to respond to the event and their target locations via NSGA-II. NS The latter initializes a population of individuals encoded with a genetic schema to represent a single possible node topology. Chromosomes are coded with integer values representing coordinate location indices while the integer-based genes within the chromosomes can also be enabled or disabled using a single bit value. Each population individual (risk mitigation strategy) is evaluated according to two mutually conflictive objectives: the perimeter coverage it provides and the energy cost involved in i enacting its response, i.e., the energy spent in relocating all participant nodes to their target positions. The optimization algorithm does not explore an infinite search space, but is limited to a set of potential responses. A circular response region, centralized about the robot node’s center and radius is a function of maximum mobility range. Random coordinates are generated within these regions, representing all possible locations an asset can traverse to.

301


4

EXPERIMENTAL VALIDATION 23

In our experimentation, Microsoft Robotics Developer Studio (MRDS) was used to simulate an RSN for CIP. The setting is a fenced compound with two access points at paved roads for vehicles. A sensor network of 115 nodes is deployed along the perimeter of the compound’s fence.

Figure 3: A CIP scenario in MRDS showing a fenced facility encapsulated by a RSN. The RSN will establish a virtual fence, encapsulating the physical fence around the compound. The main objective of the RSN is to be aware of any movements along the security perimeter, while detecting any risks within the network. Each sensor node uses a simulated Global Positioning System (GPS) module as a primary source of localization and a virtual Laser Range Finder (LRF) as an instrument to detect intruders. The sensing field of view is modeled as a circular region centered on the node’s location with a radius of 3.5 m. Figure 4 show a 2-D graphical representation of the simulation.

Figure 4: A 2-D representation of the simulated CIP scenario. A security perimeter (shown in Figure 4 as a finely dotted contour around the CI) is defined manually by the user. Sensor nodes will evaluate their coverage metrics based on surveillance of this contour. The contour is discretized into 750 perimeter points. The sensor nodes are initialized in a distributed manner along the security perimeter with some overlapping sensor coverage. The batteries are initialized using a uniform random distribution between 20% and 100%. In one experimental run of the simulation, two nodes identify themselves as high-risk due to low battery levels. Node 1 is in distress with battery level 19% and node 31 is has battery level 22% and is also in distress. The potential coverage loss of node 1 is 4.61 m and 1.54m from node 31. These nodes require recharging, but until they can be replaced the network must create independent autonomous groups and fill the combined coverage gap of 6.15 m.

302


To form the autonomous groups, node 1 and node 31 assume roles as auctioneers and announce auction sessions. Both of these nodes seek the most suitable robotic nodes to fill in the coverage gap and so they enter the AWAIT_BIDS state. They remain in this state until 10 bids are received. The NSGA-II has been successful in exploring the solution space for highcoverage, but low energy cost solutions. The network response for node 1 is shown at the top of Figure 5. Here a coverage metric of 99.3% was obtained with a cost of 1.19. The optimization has selected new target locations for the node in the group at the top of Figure 5 marked in red, which has allows them to shift over to fill in the coverage gap. Six of the robotic nodes were excluded in this solution and remained at their initial position. In the group managed by node 31 on the right side of Figure 5, the optimization algorithm has generated a feasible solution with a coverage metric of 98.9% and energy cost of 0.496. It can be observed that one the node close to the auctioneer as shifted over to fill in the coverage gap. Most of the nodes were excluded from the solution NSGA-II converged on a solution that provided optimal coverage with the least energy exerted.

Figure 5: A 2-D representation of the optimized solutions. 5

CONCLUSIONS

Having an RSN deployed for CIP is a challenging task, especially to maintaining a secure perimeter in a potentially harsh environment. The importance of being able to detect and mitigate risks in the network in an intelligent manner is paramount. In this paper, we have been successful in simulating an RSN for the protection of a remote facility. The network was able to identify multiple sources of risk, self-organize into small autonomous groups, and optimize the topology of each group. Due to the nature of the auction protocol implemented, this process occurred in a decentralized manner, without the need of a base station and central management system. Future work would include more involved modeling of a realistic communication protocol and expanding the types of network events and in turn considering other types of network responses. REFERENCES [1] Langendoerfer, P. Wireless sensor and actuator networks for critical infrastructure protection. 2011. URL: http://www.wsan4cip.eu. [2] Xu Li, Rafael Falcon, Amiya Nayak and Ivan Stojmenovic, “Servicing Wireless Sensor Networks by Mobile Robots”, IEEE Communications Magazine, Vol 50 No. 7, pp. 147-154, 2012. [3] Public Safety Canada, "Action Plan for Critical Infrastructure," 2009. [4] J. Aubert, T. Schaberreiter, C. Incoul, D. Khadraoui and B. Gateau, "Risk-Based Methodology for Real-Time Security Monitoring of Interdependent Services in Critical Infrastructures," Availability, Reliability, and Security, ARES '10 International Conference on, pp. 262-267, 2010.

303


[5] L. Coppolino, S. D'Antonio, L. Romano and G. Spagnuolo, "An Intrusion Detection System for Critical Information Infrastructures using Wireless Sensor Network," Critical Infrastructure (CRIS), 2010 5th International Conference on, pp. 1-8, 2010. [6] L. Gomez and C. Ulmer, "Secure Sensor Networks for Critical Infrastructure Protection," Sensor Technologies and Applications (SENSORCOMM), 2010 Fourth International Conference on, pp. 144-150, 2010. [7] A. Attwood, M. Merabti, P. Fergus and O. Abuelmaatti, "SCCIR: Smart Cities Critical Infrastructure Response Framework," Developments in E-systems Engineering (DeSE), pp. 460464, 2011. [8] J. McCausland, G. Di Nardo, R. Falcon, R. Abielmona, V. Groza and E. Petriu, "A proactive riskaware robotic sensor network for Critical Infrastructure Protection," Computational Intelligence and Virtual Environments for Measurement Systems and Applications (CIVEMSA), 2013 IEEE International Conference on, pp. 132-137, 2013. [9] L. M. Feeney and M. Nilsson, "Investigating the energy consumption of a wireless network interface in an ad hoc networking environment," INFOCOM 2001. Twentieth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, vol. 3, pp. 1548-1557, 2001. [10] R. Falcon, A. Nayak and R. Abielmona, "An evolving risk management framework for wireless sensor networks," Computational Intelligence for Measurement Systems and Applications (CIMSA), 2011 IEEE International Conference on, pp. 1-6, 2011. [11] R. Falcon and R. Abielmona, "A response-aware risk management framework for search-andrescue operations," Evolutionary Computation (CEC), 2012 IEEE Congress on, pp. 1-8, 2012. [12] J. McCausland, R. Abielmona, R. Falcon, A.-M. Cretu and E. Petriu, "Auction-Based Node Selection of Optimal and Concurrent Responses for a Risk-Aware Robotic Sensor Network," IEEE Int. Symp. Robotic and Sensors Environments (ROSE), pp. 136-141, 2013. [13] Negnevitsky, M. 2005. Artificial Intelligence: a guide to intelligent systems. Pearson Education Ltd. Essex, England. [14] K. Deb, A. Pratap, S. Agarwal and T. Meyarivan, "A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II," Evolutionary Computation, IEEE Transactions on, vol. 6, no. 2, pp. 182-197, 2002. [15] X. Tan, Y. Zhang, X. Cui and H. Xi, "Using hidden markov models to evaluate the real-time risks of network," Knowledge Acquisition and Modeling Workshop, 2008. KAM Workshop 2008. IEEE International Symposium on, pp. 490-493, 2008. [16] X. Cheng and Y. Ni, "The Research on Dynamic Risk Assessment Based on Hidden Markov Models," Computer Science & Service System (CSSS), 2012 International Conference on, pp. 1106-1109, 2012. [17] B. P. Gerkey and M. J. Mataric, "A formal analysis and taxonomy of task allocation in multi-robot systems," The International Journal of Robotics Research, vol. 23, no. 9, pp. 939-954, 2004. [18] B. P. Gerkey and M. J. Mataric, "Sold!: Auction methods for multirobot coordination," Robotics and Automation, IEEE Transactions, pp. 758-768, 2002. [19] R. Smith, "The contract net protocol: High-level communication and control in a distributed problem solver," Computers, IEEE Transactions on, vol. 100, no. 12, pp. 1104-113, 1980. [20] K. Ma, Y. Zhang and W. Trappe, "Managing the Mobility of a Mobile Sensor Network Using Network Dynamics," Parallel and Distributed Systems, IEEE Transactions on, vol. 19, no. 1, pp. 106-120, 2008. [21] J. Lizhong, J. Jia and D. Sun, "Node Distribution Optimization in Mobile Sensor Network Based on Multi-Objective Differential Evolution Algorithm," Genetic and Evolutionary Computing (ICGEC), 2010 Fourth International Conference on, pp. 51-54, 2010. [22] Sugeno, Michio, and G. T. Kang. "Structure identification of fuzzy model." Fuzzy sets and systems 28.1 (1988): 15-33. [23] Microsoft: “Microsoft Robotics Developer Studio (Version 4)”. Available from http://www.microsoft.com/robotics.

304