MidFusion: An adaptive middleware for information fusion in sensor network applications

Abstract

Applications and services are increasingly dependent on networks of smart sensors embedded in the environment to constantly sense and react to events. In a typical sensor network application, information is collected from a large number of distributed and heterogeneous sensor nodes. Information fusion in such applications is a challenging research issue due to the dynamicity, heterogeneity, and resource limitations of sensor networks. We present MidFusion, an adaptive middleware architecture to facilitate information fusion in sensor network applications. MidFusion discovers and selects the best set of sensors or sensor agents on behalf of applications (transparently), depending on the quality of service (QoS) guarantees and the cost of information acquisition. We also provide the theoretical foundation for MidFusion to select the best set of sensors using the principles of Bayesian and Decision theories. A sensor selection algorithm (SSA) for selecting the best set of sensors is presented in this paper. Our theoretical findings are validated through simulation of the SSA algorithm on an example scenario.

Introduction

A sensor is a device that has the capability to sense, process and communicate [1]. Network of smart sensors are being embedded in the environment to provide information to applications that are required to react to events taking place in the environment. These sensor devices include very small special-purpose sensors such as security tags, generic sensor nodes such as motion sensors and high-bandwidth sensors such as cameras and microphones [1]. Sensor network applications rely on a combination of available sensors of all categories as the sources of information. The aggregation of dissimilar data is often referred to as the problem of information fusion. Therefore, all sensor network applications share similar requirements of information fusion. A simple example of a sensor network application scenario is presented.

Scenario: Consider an intruder detection system for a building. The building is divided into three areas based on the security level of the region—Area1, Area2 and Area3. Various sensor devices are strategically placed in the three different areas. For the purpose of this example, let us consider two RFID tag readers, RFID1 and RFID2, three Video surveillance devices, Video1, Video2 and Video3 and a Biometric Reader. The presence of an intruder in the building is detected by a combination of these devices. For example, RFID1 can read the tags only in Area1, but RFID2 can signal the presence of an intruder in all the areas. However, the possibility of false positives and cost of device usage vary for each device, even though they may cover the same area.

As given in the scenario, today’s sensor network applications are mostly customized to use a predetermined set of sensors in order to acquire sensory information. However in a dynamic pervasive computing environment, a large number of sensors may be available in the network to ensure fault-tolerance and to provide adequate coverage in terms of space as well as type of sensory information needed for the applications. In reality, at any given time, information from only a small subset of sensors may be meaningful, because many sensors can potentially make simultaneous and similar observations. In effect, there exists redundant information with varying quality and cost of acquisition. Therefore, our premise is that sensory information can be more efficiently exploited if application services are able to dynamically discover and intelligently select the best set of sensors at any instant.

Sensor networks are different from conventional networks, because of many inherent characteristics and limitations. Typically, sensor networks have huge number of heterogeneous nodes, and they are subjected to mobility, failures, environmental obstructions and power constraints [2]. In addition to the conventional overheads associated with traditional networks, applications services need to spend extra effort and time to deal with additional challenges due to the resource constrained nodes of sensor networks.

Current research has led to middleware services that enable sensor networks and applications to interact with each other and adapt transparently. Depending upon the application and the context, in addition to task scheduling, the middleware must have the capability to query and monitor sensor nodes, clusters of sensor nodes, or sensor agents managing a group sensor nodes. Typically, a sensor network middleware has a layered structure that is distributed among the network of sensors and interfaces with the application as depicted in Fig. 1.

Desirable characteristics of a middleware architecture for sensor network applications can be summarized as follows [2], [3]:

• support the development, maintenance, deployment and information fusion requirements of diverse sensing applications;

• create a runtime environment that can support and coordinate multiple applications;

• provide appropriate abstractions to the applications to deal with heterogeneity, energy constraints, and dynamicity of sensor networks;

• provide knowledge about the applications to the sensor networks in order to perform application specific data processing;

• provide data centric communications;

• adapt to dynamic and real-time changes and manage resources efficiently; and

• support automatic configuration and error handling in the sensor networks.

There has been considerable research to seek low level solutions for dealing with resource constraints of sensor networks. Solutions have been proposed for minimizing power requirements through clustering, topology control, routing and MAC-level protocols such as LEACH [4], Directed Diffusion [5], PAMAS [6], S-MAC9 [7], ASCENT [8], SPAN [9], STEM [10], Lint [11], and others [12], [13].

Although research in the area of sensor network middleware is still in its infancy, notable endeavours that focus on management and efficient use of sensor networks include Cougar [14], SINA [15], DSWare [16], Impala [17], Mires [18], and Mate [19]. In Cougar [14], a database approach is used and sensor readings are considered as virtual relational database tables. Cougar efficiently manages the power consumption by distributing the query among sensor nodes to minimize the energy consumed to collect the data. Sensor information networking architecture (SINA) [15] also provides a database abstraction that allows for resource efficient data aggregation through location aware and hierarchical clustering of sensors. Data services middleware (DSWare) [16] employs a group based approach to provide data abstraction for application to improve the performance of real-time execution and reduction in communication requirements, while hiding the faulty nature of sensor operations and wireless communications. Application services adapt to the underlying sensor network characteristics in Impala [17]. Applications in Impala are modular and exploit mobile code techniques for adaptation and repairability. In Mires [18], a message-oriented middleware design is employed in which a publish/subscribe communication paradigm allows the sensors to provide resource efficient and application specific data processing. Mate [19] is a tiny virtual machine that executes on top of TinyOS [23] and is designed for providing simple programming interface to sensor nodes.

While it is important for middleware services to deal with resource constraints of sensor networks, it is also critical to consider the requirements of information fusion from an applications perspective. DFuse [21] is a data fusion framework that facilitates transfer of different areas of application level information fusion into the network to save power. DFuse does this transfer dynamically by determining the cost of using the network using cost functions. Adaptive middleware [20], is a middleware for context-aware applications in smart-home setups. In this scheme, the middleware matches the quality of context (QoC) requirements with the QoC available with the sensors. The application’s QoC requirements are mapped to a utility function using the QoC attributes of the sensors available. MiLAN [22] is also similar to the adaptive middleware scheme. In MiLAN, applications QoS requirements are matched with the QoS provided by the sensor networks. However in this scheme, the QoS requirements of the applications are some predetermined numbers which the applications should know in advance in addition to the quality associated with the type of sensors it can use. In dynamic and pervasive computing environments, the number and type of sensors available to the applications may vary. Therefore it is impractical to include knowledge about all the different sensor nodes that an application can potentially make use of. In addition, all these sensors come at various levels of cost and benefit to the application.

In this paper, we propose MidFusion, a middleware for sensor network applications that performs fusion of sensory information. MidFusion discovers and selects the best set of sensors on behalf of applications, depending on the quality of service (QoS) requirements, the QoS that can be provided by the sensor networks and the information acquisition cost, in a transparent way. Adaptive middleware and MiLAN also match the requirements of the application with what can be provided by the network. Both the schemes however do not consider the information acquisition cost. In adaptive middleware scheme [20], the application’s QoC requirements are mapped to a utility function using the QoC attributes of the sensors available. In contrast, MidFusion maps the QoS availability of the sensors to a utility function based on the QoS parameters of the application and the cost of information acquisition. The characteristics of MidFusion architecture are compared with those of some known relevant sensor network middleware designs in Table 1.

The inherent characteristics of sensors bring a level of uncertainty to such information sources. Information fusion in the face of uncertainties is a challenging research problem. Bayesian networks [24], [25] provide an efficient tool for performing information fusion and decision making under conditions of uncertainty. In addition, Bayesian networks also provide an efficient way to specify the sensor requirements to the middleware.

MidFusion is a middleware that works on the principles of Bayesian and Decision theoretical paradigms to provide a portable abstraction to applications. In other words, MidFusion supports multiple sensor network applications. Furthermore, MidFusion can be used by applications that are already modelled or can be modelled using Bayesian networks. The design objectives of MidFusion include:

• select the best set of sensors by ensuring sufficiency and efficiency,

• allow applications to specify their sensory requirements,

• customize services for diverse applications,

• provide application transparency to the underlying network structure, and

• inject application information into the sensor network to manage their resources better.

This paper is organized as follows. The design features of MidFusion are presented in Section 2. We then apply the formulations in the context of the example scenario to analyze MidFusion in Section 3. Section 4 provides a discussion of ongoing and future improvements to MidFusion, and concludes the paper.