Sheet Music

14 views

Opportunistic Information Fusion: A New Paradigm for Next Generation Networked Sensing Systems

Opportunistic Information Fusion: A New Paradigm for Next Generation Networked Sensing Systems
of 8
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Share
Tags
Transcript
  2005 7th International Conference on Information Fusion (FUSION) OpportunisticInformation Fusion: A New Paradigm for Next Generation Networked Sensing Systems Subhash Chalia, Tauseef Gulrez, Zenon Chaczko, T.N. Paranesha Information andCommunications Group Faculty Of Engineering University of Technology Sydney, Australia. {schalla, tgulrez, zenon} eng.uts.edu.au Abstract - Traditionally, Information Fusion systems assume that the information is gathered from known sensors over proprietary communication networks and fuse usingfixed rules of informationfusion and designated computing and communication resources. Emerging technologies like wireless sensor networks, TEDS enabled legacy sensors, ubiquitous computing devices and all IP next generation networks are challenging therationale of conventional information fusion systems. Tae technology has matured to a point where it is reasonable to discoversensors based on the context, establish relevance, queryfor appropriate data, andfuse it using the most appropriatefusion rule, using ubiquitous computing and communicationenvironment in an opportunistic manner. We define such fusionsystems as opportunistic information fusion systems. In this paper we introduce this new paradigm for information fusion and identify plausible approaches and challenges to design, develop and deploy the proposed next generation opportunistic informationfusion systems. 1 Introduction Multi-sensordata/information fusion is a rapidly maturing technology, concerning the problem of combining data/information from multiple sensors to servedifferent applications. It is a core component of all networked sensing systems, which is used either to: - Join/combine complementary information produced by sensors to obtain a more complete picture or - Reduce/manage uncertainty by using sensor information from multiple sources. Several end applications including environment monitoring,automatic target detection and tracking, battlefield surveillance, remote sensing, etc. depend critically upon data/information fusion [37]. Traditional data/information fusion solutions are designed to work with dedicated sensor and information sources. Emerging technologies like wireless sensor networks [1], standards enabledlegacy sensors [20,21], ubiquitous computing and communication systems [2,3] and all internetprotocol (IP) next generation networks [4] are leading to a new paradigm where sensors can be treated as a shared infrastructure and a common resource with an ability to servemultiple simultaneous data/information fusionapplications concurrently. This is in stark contrastto the dedicated sensor, computing andcommunicating infrastructure serving a specific data/informationfusion application. There has been significant progress in dealing withshared computing andcommunication resources [19]. However, little has been done in regards to treatingsensor resources also as a shared resource. Thus, thetraditional data/informationfusion models,methods, middleware and techniques need to be enhanced/modified to work with sensors that are part of a shared infrastructure. In this paper, we introducea new paradigm of data/information fusion - the opportunistic information fusion (OIF) paradigm - to handlesuch multi-purpose sensor networks and services. In addition, we identify several challenges and propose innovations to realize such next generationinformation fusion systems. The paper is organized as follows after introduction in section 1, section 2 proposes a new OIF model derived from the well established JDL Model. Section 3, reviews the recently proposed IEEE P1451/TEDS sensor standards and proposes enhancements to support the OIF services. Section 4 proposes new methods of fusing information modeled usingheterogeneous uncertainty measures and section 5 proposes opportunistic middleware solutions that can enable the development and deployment of secure,context aware, scalable, flexible and heterogeneous OIF applications.Section 6 illustrates the potential of OIF paradigm using a couple of practical applications. Finally,section 7 concludes the paperand its contributions. 0-7803-9286-8105/ 20.00 ©2005 IEEE720  opportunistic services to multiple applications at the same time and deliver new, non-zero sum benefits of information fusion. 2 The New OIF Model In the traditional data/information fusion Joint Director of Laboratories (JDL) model is the most popular model in use - especially in the defense applications [5,6]. The model breaks down thedata fusion processes intofivelevels ranging from (0-4)consisting of sensor specific signal processing, objectrefinement and tracking, situation,threat and impactassessment augmented with appropriatedatabases and fusion rules. One of the fundamental, implicit, assumptions of the JDL model is that the sensors   the corresponding sensor models are known a priori and are available to serve a particular application in a dedicated manner. Due to this, thedata fusion systems, once developed, cannot be re-deployed to serveincontexts other than what it is built for. Thus, when sensors, computingand communications become shared infrastructure and there is no dedicateddatafusion application in place, this model is no longer adequate. Conventional OpportunisticInformationFusion InformationFusion(OIF) Model Sensors are assumed Sensor need to be discovered Apriori info on sensor Sensor model not known model apriori Proprietary networks Next Gen IP Networks Pre-defined Fusion rules Determine Fusion Rule Models of Fusion JDL, New/Modified Fusion OODA Models Pre-defined Compt. load Adhoc Computation Load Fixed middleware Flexible middleware Table -1: Old Paradigm v/s New ParadigmHence a new datafusion model - defined as the Opportunistic Information Fusion Model (OIFM) - to deal with the challenges of realizing thefusion systems in this new paradigm is proposed in figure 1. The key additions to JDL model that leads to an OIFModel are a new process forestablishing relevance and enabling sensor discovery at the lowest level and context enabled re-configuration at all levels of thedatafusion process. Even therecently proposed enhancements to JDL model [5] does not support theseprocesses. This new OIF model will enable sensors in a shared infrastructure mode to provide Sensor Resource discovery can be achieved by methods like [7] and its relevance can be establishment forthe application using the sensors selfidentification capabilities. Until recently self identification was not an in-built feature insensors. However, recently, a new standard[20,21,22] promoted by the IEEE instrumentation and measurementworking group, enables sensors to have self identification capability. These standards are reviewed and enhancements are proposed in the next section. Figure   : Proposed OIF Model 3 IEEE 1451 Standards   Transducer Electronic Data Sheets (TEDS) The OIF model provides a framework for the developmentof datafusion systems, however, fundamentally,information fusion is carried out mathematically at alf levels of information fusion via some kind of uncertain reasoningframework. One of the most successful datafusion techniques is Bayesian data fusion that operates at almost all levels of data fusion. This technique is briefly introducedhere and highlights a key requirement of the continued success of this approach in the networked sensing context. Let 't' be the random variable that represents a parameter (e.g., temperature) to be estimated by asensor network. Let fY1, Y2 ..... Y} be the set of measurements from sensors 1,...k of any sensor network. The problem 721 C 0 N T E X T V  is to obtainthebestestimate of the unknown parameter 't', using all the measurements received from the sensors in a sensor network. Within theprobabilistic framework the parameter can be estimated from its conditional density using a simple rule: t = Jt.p(t Y, Y2, Yk )dt where p(t Y1, Y2,...., YJ) is the conditional probability density function and the estimate t is the bestestimate in a minimum variance sense [23]. The Bayesian Fusion rule provides avery elegant way toconstructthe conditional density from sensor measurements asillustrated below: t   Y,, Y2, ., Yk) =P(YJ, Y2, Yk   t) p(t) Assuming Y1, Y2 Yk are conditionally independent. AtjIj,Y2.,. k=- t)(Y2t).; (XIt)1 At) (1) where p(t) is the prior distribution of  t assumed by the application. The above formula of multiplying likelihoods pi (Yi   t) of individual sensors with the prior p(t) (and normalised using 6 ) is theso called Bayesian fusion rule[10]. The likelihood encapsulates the errorcharacteristic of the i-th sensor. Thus thelikelihood function (or some other uncertaintyrepresentationfunction like fuzzy membership function) that captures the sensors' uncertainty characteristic is vital for successful information fusion. Traditional IF systems use dedicated sensors with known a priorilikelihood functions. However in OIF systems, as sensors are part of a shared infrastructure and are not known a priori, this information on likelihood functions(uncertainty measures) needs to be gathered directly from sensors (upon discovery). In a networked sensing context, when sensors come in andgo out of service in an ad hoc manner, sensor discovery is not trivial if the sensors do not have self-identification and sensor service registration capability. Transducer Electronic Data Sheets (TEDS) is the heart and soul of a new IEEE 1451 family of smart transducer interface standards which are introduced by IEEE instrumentation andmeasurement working group for standardization of SmartTransducer Interfaces and a set of common interfaces [20,21]. TEDS contains the critical information needed by an instrument or measurementsystem to identify, characterize, interface and properlyuse the signal from the smart sensor. These key featuresare of great importance in the OIF context as theyprovide the sensors with self-identification capability on a sensornetwork. This self-identificationcapability is critical to opportunistic methods of sensorservice discovery. TEDS utilizes the concept of templates as shown in the Table 2 that definesthe specific properties for different sensor types. The Basic TEDS field has all the vital information like Manufacturers ID,sensor type etc that can enable context based sensor discovery. This research proposes to develop new methods of sensor service discovery by using the basic TEDS and user area information in conjunction with the contextual inputs to solve the relevant sensor discovery problem. From the BayesianFusion rule it is evident that thelikelihood information needs to be gathered from thesensors directly upon discovery. It is sometimes possible to approximate likelihood functions (uncertainty measures) from theavailable data from TEDS template using the Standard TEDS information (available in second field) along with the data generated by the sensor to servethe information fusion application. However, such derived functions are, at best, approximations and can lead to erroneous fusion results. As this is true sensor error characteristics and is theproprietary information of sensor manufacturer, it is better if it is a part of the standard and has it supplied directly by sensor suppliers/manufacturers as apart of standard TEDS template. As current approved standards do notincorporate such uncertainty descriptor functions, we propose the following modification to the standard TEDS field (showed in Bold in second field) to enable opportunistic information fusion. Basic TEDS Manufacturer ID 43 Model ID 7115 Version Letter B Serial Number 00731F Standard Calibration Date Jan. 29, 2000 Sensitivity @ 1.094E+03 mV/g TEDS Reference Freq, 100.0 Hz (IEEE Temp, Meas. Range, 23°C Electrical Output ±50 g Accelerometer Quality Factor ±5 V Sub-template) Temp Coefficient 300 E-3 Direction(X, Y, Z) -0.48 /°C Reasoning Frame Probability Likelihood Normal (0,1) Function. User Area Sensor Location Building 2-lvl:2 Calibration Due April 15,2002 Table2 - Proposed TEDS Template 722  4 Opportunistic Information Fusion with Heterogeneous Uncertainty Measures The freedom to supply uncertainty measures within the respective TEDS fields opens it up to heterogeneous choice of uncertaintymeasures. The uncertainty measures can be one of probability, possibility (fuzzy) orbelief measures that capture and represent the uncertainty in sensor measurements in different approximate reasoningframeworks. In such a scenario it is possible to encounter situations where sensors in the networked data/information fusion application canhave heterogeneous uncertainty measures. Seminal works byOxenham, Challa andMoreland [8] propose distributed information fusion methods that deal withheterogeneous uncertainty measures with built-in uncertaintypreservingtransformations to move fromone uncertainty reasoning framework to the other without changing the level of uncertainty. Forexample the pignistic and inversepignistic transformations proposed by Phillipe Smets [12,13] havebeen used in [8] to move between probability and belief measures before fusion as elucidated below. The Pignistic probability BetProb is calculated from a belief fiuction (F', m) by setting BetProb(xj)= EFEF'lx EF m(F ) / F'l (1 - m( )) for each xj in the frameof discernment Q [36, p. 202]. The Inverse Pignistic belief mass can be calculated by first letting the elements xj of Q2 be re-labelled such that: p(xjl) 2 p(xj2) > ... 2 p(xjk), where k= I for each r = 1,...,k defme the subset: Fr = {Xj1 X12 ,..., XI, }, and assign it the mass: Mt (Fr) =F r (p(xjr) - P(Xjr,l )), where p(xjk+l) =0 by convention. Then the focal elements of (F', m) are the (nested) sets Fr which have a non-zero mass m6(Fr) assigned to them.This work is, however, limited to Bayesian and Dempster-Shafer reasoning frameworks. To realize the full potential of the OIF paradigm, theseideas need to be extended to fuzzy/possibility theory frameworks as well. Sets framework. In the late 1990s, Ron Mahler proposed random sets as a meansof probabilistically modeling versions of crisp and fuzzy Dampster-Shafer theory in a way that is consistent with the Bayes' fusion rule [17, 18]. There are three steps in suchan approach to deal with data that has inherent uncertainties [15,16,17]. First, such evidence/data is modeled as a random closedsubset O of the underlying measurement space. Second, some modelling technique-fuzzy logic, Dempster-Shafer theory, rules-is used to construct (E)   Third, a generalized likelihood function p (E) Ix)   based on an ambiguous signature model base and a data-to-model matching technique, is used to hedge against uncertainties both in data-modeling and in the modeling of datageneration. The generalizedlikelihood function is then used in the Bayesian Fusion rule - like the one introduced in equation (1). Here we demonstrate modeling fuzzy membership function using random set. To construct useful models of 0 one could use a fuzzy-set modeling process, in which ambiguity in the dataz is modeled as a fuzzy membership function g(z) onmeasurement space. Let A be a uniformly distributed random number in [0,1]. Then the random subset 0 = Y-A(g)={z A < g (z)} containsthe same information as the fuzzy model g(z). A similar approach has been shown to convert a basic belief mass assignment to a random closed set in [15,16,17]. However, such modeling has not yet been used to fuse information from multiple sensors where each sensor represents information in a different uncertaintymeasure. We propose to use a basic methodology thatwill first convert the non-probabilisticuncertainty measure, like fuzzy membership function or basic belief mass assignment, into aclosed random subset and instantiate the Bayesian Fusion approach to combine the data. One of the key questions that has notyet been answered in literature andwhich has direct implication on the use of these techniques of information fusion is the relationship between level of uncertainty in the closed random set representation and the level of uncertainty in the fuzzy membership orbasicbelief mass assignment. 5 Middleware for Information Fusion Opportunistic One of the approaches we suggest in handling such heterogeneous uncertainty measures, including fuzzymeasures, is to consider Random The term Middleware relatesto software system infrastructure that constitutes a set of services that aim at facilitating the development of distributed applications 723  in heterogeneous environments. Middleware is a software layer that is placed above the operating system - including the basic communication protocols - and below the distribution applications thatinteract viathe network. Its primary objectives areto foster portability and interoperability of distributed application components. The middleware layer allowssoftware components to exchange data and interact with one another regardless of the underlying communication protocols, operating systems and hardware platforms on which the components reside. This is made possible through the use of standard application programming interfaces (APIs) and services. Mainstream middleware solutions traditionally offer a useful,distributed programming modelsand solutions that mask the heterogeneity of networks, end-systems,operating systems, programming languages and hardware. Middleware solutions such as CORBA, RMI, J2EE, .Net, etc. have been very successful in business applications; wrapping of legacy systems and many other tasks [24, 26, 28]. However,noneof these can adequately address the requirements posed by the new opportunistic information fusion paradigm such as: serendipitoussensorservice discovery, situation and context awareness, opportunistic information access and delivery, contextual interpretation and presentation,support for heterogeneous capabilities of individual network devices, and reliable support forprivacy, rights, security and trust. Hence a new Opportunistic Middleware model (OMM) is proposed. The Figure-2 depictsthe essence of the proposed opportunistic services in OMM. To obtain an opportunistic SN service the user issues a query which is analyzed in the Context Engine. If a query request is contextually feasible then therequest to discover sensor(s) is issued to the Opportunistic Data Fusion Framework. Next theprocessinvolvesestablishing relevance, feasibility and quality of service delivery. Here pivotal role in establishing service relevance is played by informationcontained in TEDS. The Relevance Correlation process aims at finding compatibility between information Fusion Rules, discovered (accessible) sensors and the TEDS fields which contain sensors' physical and logicalparameters. Upon findingopportunistic sensorservice and establishingrelevance, the service instance is registered as aSensor Service Registry. Figure - 2: The process of OpportunisticService Provisioiing 5.1 The Opportunistic Middleware Model The OMM is to support Sensor Network (SN) solutions thatare autonomic in nature, scalable, and flexible. The model should support autonomic computing [25, 27] that is characterized by a set of  self-x properties such as: self-organizing, self-configuring,self-healing,self-protecting and self-optimizing along with aspects of context and situation awareness. The proposed OMM aims to deliver software system framework that, apart from traditional middleware goals,attempts to achieve thefollowing: * Provide Support for Opportunistic Sensor Discovery Opportunistic sensorservice discovery aims to gather information wherever and whenever it is useful and relevant. Some sensors can provide duplicate information while others can befound by a chance. The proposed OMM needs to deal with aspects of sensors data redundancy and serendipity. * Provide Support for Opportunistic Data Access. Users should be able to interact with information/other users with ease and in various modes (i.e. fast data paths, slow data paths). High performancehardware should be supported to provide opportunistic services at Sensor Network Access Points. * Provide Support for Opportunistic Interpretation. Computer resources and computational tools should be used to make information and interactions meaningful to end users and to provide mechanisms forcontext interpretation and awareness. In contrastto the classical, adaptive, reflective and ad-hoc middleware models [29, 30, 31] adopted forarchitectures ofmost NS-based systems [28] theopportunistic middleware model is to promote notions of 724
Advertisement
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks