Network Repair. (Under the direction of Professor Edward Grant).
Wireless Sensor Networks (WSNs) have the potential to provide a wealth of high resolution sensory data, both temporally and spatially, over large areas and for long periods of time, but can be limited in effectiveness when a sensor node loses power or becomes damaged. The quality of the sensor network data is also reliant on the underlying network connectivity and can be degraded by imprecise deployments, and unforeseen changes in the network structure over time such as changes in weather conditions. The ability to use autonomous mobile robotic platforms to repair or replace bad sensor nodes, or to map out WSNs to identify weak nodes, has potential to enhance the performance of WSNs and improve their robustness. This dissertation investigates: 1) WSN connectivity issues over the lifetime of a network, and 2) identifying and repairing disconnects within a WSN using an autonomous robot.
The effects of asymmetric links between WSN nodes and the best methods to model networks composed of asymmetric nodes were studied in depth. It was found that for networks requiring bidirectional links that the use of a disk model was optimal; however, for networks with asymmetric links, elliptical or irregular models were preferred. Thus in situations where asymmetries are permitted, more efficient network connectivity is obtained using elliptical or irregular models.
by
Kyle Anthony Luthy
A dissertation submitted to the Graduate Faculty of North Carolina State University
in partial fullfillment of the requirements for the Degree of
Doctor of Philosophy
Computer Engineering
Raleigh, North Carolina
2009
APPROVED BY:
Dr. Thomas C. Henderson Dr. John F. Muth
Dr. H. Troy Nagle Dr. Mihail Sichitiu
Dr. Edward Grant
BIOGRAPHY
TABLE OF CONTENTS
LIST OF TABLES . . . vi
LIST OF FIGURES . . . viii
LIST OF ABBREVIATIONS . . . xiv
1 Introduction . . . 1
2 A Review of Resource Constrained Wireless Sensor Networks for Envi-ronmental Monitoring . . . 4
2.1 Introduction . . . 5
2.2 Sensor Node Characterization . . . 6
2.2.1 Method of Deployment . . . 6
2.2.2 Battery Life . . . 8
2.2.3 RF Communication . . . 8
2.2.4 Environmental Effects . . . 11
2.3 Experience of Large Area Deployments . . . 14
2.3.1 Great Duck Island [1] . . . 16
2.3.2 Redwood Climate Monitoring [2] . . . 16
2.3.3 Wildfires . . . 17
2.3.4 Volcano Monitoring [3, 4, 5] . . . 19
2.3.5 Glacsweb [6] . . . 20
2.3.6 Soil Monitoring . . . 21
2.3.7 Vineyard Monitoring [7] . . . 24
2.4 Discussion . . . 25
2.5 Conclusion . . . 29
3 Connectivity Assessment of the Disc RF Radiation Model . . . 31
3.1 Introduction . . . 31
3.2 Related Work . . . 32
3.3 Connectivity Analysis of Isotropic RF Models . . . 33
3.4 Connectivity Analysis of Anisotropic RF Models . . . 44
3.5 Conclusions and Future Work . . . 47
4 Using Received Signal Strength from Wireless Sensor Networks for Test-ing Collision Detection and Obstacle Avoidance with Autonomous Mobile Robots . . . 49
4.1 Introduction . . . 50
4.2 Related Work . . . 52
4.3 Robotic Test Platform . . . 53
4.4.1 RSS Variation Over Time and the Collision Threshold . . . 57
4.4.2 Robot/Object Collision Detection Scenarios . . . 62
4.5 Received Signal Strength (RSS) for Obstacle Avoidance . . . 64
4.6 Conclusions and Future Work . . . 69
5 Experiments in the Autonomous Robotic Repair of Wireless Sensor Net-works . . . 72
5.1 Introduction . . . 73
5.2 Related Work . . . 74
5.2.1 Robotic Deployment and Repair . . . 74
5.2.2 WSN Node Based Robots . . . 75
5.3 Hardware Platform . . . 75
5.4 Proposed System Philosophy and Methodology . . . 78
5.5 Robotic Repair . . . 81
5.6 Implementation . . . 85
5.7 Improvements . . . 88
5.8 Conclusions and Future Work . . . 91
6 Perimeter Detection in Wireless Sensor Networks . . . 94
6.1 Introduction . . . 95
6.2 Related Work . . . 96
6.2.1 Centralized Techniques . . . 97
6.2.2 Localized Techniques . . . 98
6.3 Problem Definition or Definition of Boundary Nodes . . . 99
6.4 Analysis of Boundary Detection Techniques . . . 100
6.4.1 Local Image Modeling . . . 101
6.4.2 The Iso-Contour Technique of [8] . . . 105
6.4.3 Degree Analysis for Boundary Detection . . . 107
6.4.4 Convex Hull Boundary Detection . . . 108
6.5 Future Work . . . 115
6.5.1 The Localization Problem . . . 115
6.5.2 Topological Analysis . . . 116
6.6 Conclusions . . . 117
7 Conclusions and Future Work . . . 118
Bibliography . . . 120
Appendices . . . 132
A Derivation of an Empirical RF Radiation Model . . . 133
C Collision Detection and Obstacle Avoidance Case Studies . . . 142
C.1 Near-Field Sensors . . . 142
C.1.1 Near-Field Extrinsic . . . 143
C.1.2 Near-Field Intrinsic . . . 146
C.2 Far Field Sensors . . . 148
C.2.1 LIDAR Case Study . . . 149
C.3 Beacon-Based Sensors . . . 152
LIST OF TABLES
Table 2.1 Summary of the deployment infrastructure. . . 15
Table 2.2 Summary of problems and solutions for actual deployments.. . . 26
Table 3.1 The density and average number of neighbors for various annuli when 95% of nodes are members of the same network . . . 40
Table 3.2 ENC and number of networks of elliptical nodes with increasing aspect ratio. 45
Table 4.1 Comparison of sensors used for collision detection and obstacle avoidance. . . . 56
Table 4.2 The variation over time of all pairs at all -10 dBm. . . 58
Table 4.3 The variation over time at different transmission power levels for node pair
A→B. . . 58
Table 4.4 The variation over time for different transmission frequencies. 1000 packets are sent in 10, 5, and 1 minute intervals. . . 58
Table 4.5 The variation over time for different averaging windows. . . 59
Table 4.6 Success rate of collision detection trials with different transmit powers and measurement distances for in-line and orthogonal stationary node placements. . . 61
Table 4.7 Average received signal strength comparison of omnidirectional and directional antennas in 10 m increments for 3 transmit powers. The left antenna is closest to transmitter. . . 66
Table 4.8 Obstacle avoidance performance for both steel and wooden obstacles at all start/destination pairings. S-Success (modified with R-passed on right, L-passed on left, O-roll, H-hit obstacle, C-clipped obstacle, M-moderate confusion, X-excessive confusion), F-Failure.) . . . 69
Table 5.1 Distances traveled for the 3 cases averaged over 15 trials demonstrating RSSI variation due to environmental changes. . . 81
Table 5.3 Comparison of repair methods to the Moore’s Path repair traversal. The metrics compiled over 50 trials are listed and show the percentage of trials resulting in shorter repair distances as well as the average distance improvement percentage averaged over all 50 trials. . . 92
Table 6.1 List of definitions. . . 97
Table 6.2 Table comparing the performance metrics of the evaluated boundary detection algorithms. . . 101
LIST OF FIGURES
Figure 1.1 Components of the WSN repair system. . . 2
Figure 2.1 100 nodes randomly distributed across an area of 1 km2. Connected nodes share the same letter. . . 7
Figure 2.2 Example of asymmetry using the Radio Irregularity Model (RIM) [9]. Node A falls within communication range of node B, but node B cannot hear node A. . . 9
Figure 2.3 An array of Tmote-sky nodes attempting to resolve node classifications in a leadership protocol. . . 10
Figure 2.4 Reception variation among three Tmote-sky modules with power ranging from -25 dBm to 0 dBm (3.16µW to 1 mW).. . . 12
Figure 2.5 First 3 Fresnel zones for a 2.4 GHz link with a separation of 100 m. To avoid significant interference from ground reflections objects must be kept outside of the inner 60% (dashed red oval) of the first Fresnel zone. Therefore to reduce ground interference, the node must be raised 1.06 m off the ground. . . 13
Figure 2.6 Signal strength mappings (red→purple is high→low) with different eleva-tions (the static node marked by the red pin): (a) the static node is on the ground, (b) the static node is 1 m high, and (c) the static node is 1 m high and the measure-ment node on the robot is at 0.5 m. Images created with Google Maps and GPS Visualizer. . . 28
Figure 3.1 Result of a sample branching process simulation. Only the viable nodes are displayed, each represented by a colored circle and its corresponding id number. . . 34
Figure 3.2 Sample uniform random distributions of different nodal densities. . . 36
Figure 3.3 Analysis of network connectivity while varying the node density: (a) The average number of disjoint networks, (b) The ratio of the average number of disjoint networks to the the number of nodes in the deployment, and (c) the size of the largest network. . . 37
Figure 3.5 The number of disjoint networks vs nodal density for annuli of several
dif-ferent inner radii but the same coverage area. . . 39
Figure 3.6 Sample networks of different aspect ratios; circle nodes with (a) ratio of 1:2, (b) ratio of 1:4 and annular nodes with (c) ratio of 1:2, (d) ratio of 1:4. . . 41
Figure 3.7 Connectivity analysis comparing deployment areas or side ratios 1:1, 1:2, and 1:4 for (a) circular nodes and (b) annular nodes. . . 42
Figure 3.8 Comparison of the connectivity of circular nodes to that of annular nodes for aspect ratios of (a) 1:2 and (b) 1:4, showing that regardless of shape, the annulus still outperforms the circle. . . 43
Figure 3.9 An empirically derived RF model of the Tmote-sky. . . 45
Figure 3.10 Sample of corresponding simulations for (a) empirical shape, only counting bi-directional paths yielding 26 networks, (b) empirical shape with asymmetric links counted yielding 18 networks, and (c) circles with comparable radius yielding 22 networks . . . 46
Figure 3.11 . . . 47
Figure 3.12 The number of networks vs the number of nodes deployed in a given area for elliptical nodes of increasing aspect ratio . . . 48
Figure 3.13 Disjoint networks formed through a percolation branching process. . . 48
Figure 4.1 The piecewise transfer function of a typical logarithmic amplifier, (a), and the structure of one type of logarithmic amplifier that depends on parallel non-linear and unity gain amplification stages, (b).. . . 51
Figure 4.2 Top (left) and bottom (right) views of the modified Roomba. . . 53
Figure 4.3 Modified Roomba equipped with a GPS and directional antennas. . . 55
Figure 4.4 Variation in signal strength over time for 4 power levels. . . 59
Figure 4.5 Distance vs. signal strength. . . 60
Figure 4.6 A successful collision detection (a) and a premature turn (b). . . 61
Figure 4.8 Slippage test rigs where (a) wheels completely lose contact and (b) the wheels spin on rollers. . . 65
Figure 4.9 The signal strength shadow cast by a vehicle. The beacon node is signified by the red pin, two meters from the car, denoted by black triangles. The signal strength ranges from -83 dBm to -41 dBm. Images created with Google Maps and GPS Visualizer. . . 66
Figure 4.10 Signal strength tracking to avoid a concrete barrier in a parking structure. Image created with Google Maps and GPS Visualizer. . . 67
Figure 4.11 Layout of obstacle avoidance test with robot start positions S1-S4 and des-tinations D1-D6. . . 68
Figure 4.12 Successful time lapse photo of obstacle avoidance of (a) filing cabinets and an unsuccessful avoidance of (b) a table. . . 70
Figure 5.1 The original sensors and control alterations of the Roomba platform. . . 76
Figure 5.2 The signal reception mote and GPS tower addition to the Mote-Bot. . . 77
Figure 5.3 Sample uniform random distributions of different nodal densities. Each node is represented by a letter with the encompassing circle specifying its communication range. Nodes identified with the same letter are members of the same network. . . . 79
Figure 5.4 Comparison of distances traveled for the 3 cases I) lab A II) lab A, enclosed III) lab B. . . 81
Figure 5.5 Control algorithm for basic robot RSSI navigation . . . 82
Figure 5.6 Time lapse photo of RSSI navigation taken at time steps of approximately 2 s. From the stationary mote the robot moves out and then clockwise around it, staying within the specified RSSI range. . . 82
Figure 5.7 Plot of a 5thorder polynomial fit to the disjoint networks curve, 5.7(a), and 1st and 2nd derivatives of the fitting function showing maxima, minima, and the inflection point, 5.7(b) . . . 83
Figure 5.8 An example of the networks defined in a 110 node uniform random distri-bution. . . 84
Figure 5.9 Pseudocode of the single node disconnect repair process . . . 86
Figure 5.11 An example of network repair using 6 nodes and the Mote-Bot. The layout is presented in (a) where the red node, 6, hears the fire but is not connected to nodes 1-5 until the Mote-Bot repairs the link. This corresponds with the photograph of the experiment, (b). . . 88
Figure 5.12 Moore’s Curve superimposed on a sample network with repair points marked with white stars. . . 89
Figure 5.13 An example of simple perimeter detection (a) with poor search time and (b) with fast detection. . . 90
Figure 5.14 An example of network edge detection for network repair. . . 91
Figure 6.1 Example of autonomous network repair by following communication perime-ters. . . 95
Figure 6.2 Demonstration of the proposed imaging technique for boundary detection. 6.2(a) The boundary nodes (green circles with large green halos) as differentiated from the remaining nodes (small blue circles) by analyzing the network communica-tion range image of 6.2(b) where individual nodes are labeled with different colors on a common background color. . . 100
Figure 6.3 The average recall and precision over 50 uniform random networks of 1024 nodes using localized image analysis for different resolutions. . . 102
Figure 6.4 An example square grid of 1024 nodes with unit placement and commu-nication range of radius 1.5 units. The node color signifies nodes with the same degree. . . 103
Figure 6.5 The effect of normal placement error on recall and precision for (a) one-hop and (b) two-hop local images. . . 104
Figure 6.6 The effect of normal localization error on recall and precision for (a) one-hop and (b) two-hop local images. . . 104
Figure 6.7 An example of hop iso-contours using Funke’s technique. Strips of consistent color denote nodes that are the same hop count from the beacon source. The ends of these contour segments are identified as boundary nodes. . . 106
Figure 6.8 Performance of hop-contours with 1-6 source nodes. . . 107
Figure 6.9 The effect of placement error on the performance of the iso-contour technique with 4 sources applying variances from 0 to 0.5 in steps of 0.1. . . 108
with a variance of 0.3. The colors represent the degree of the node, blue being low-est, and dark circles surround nodes that are identified as boundary nodes. . . 109
Figure 6.11 The recall, (a), and the precision, (b), of the degree threshold boundary detection technique considering placement error and the degree threshold. . . 110
Figure 6.12 The Graham Scan process, edge nodes are labeled red, non-edges, black, and undetermined, blue. (a) Node angles are determine with respect to the node with the lowest y-coordinate. (b) A left hand turn through a node results in that node being labeled an edge. (c) A right hand turn through a node results in a non-edge node. (d) Another left hand turn labeling an edge node. . . 111
Figure 6.13 The Graham Scan Algorithm [10] . . . 112
Figure 6.14 The recall, (a), and the precision, (b), of the convex hull boundary detection technique considering placement error and increasing perturbation. This is repeated considering two-hop neighborhoods resulting in the recall profile of (c), and precision profile of (d). . . 113
Figure 6.15 The recall, (a), and the precision, (b), of the convex hull boundary detection technique considering localization error and increasing perturbation values. . . 114
Figure 6.16 The real-world implementation (a) layout and (b) connectivity graph. . . 115
Figure A.1 Variation in signal strength at the receiving node as a function of orientation for 4 different transmission strengths averaged over 3 different transmitter/receiver pairs with a 99 % confidence interval. . . 134
Figure A.2 2-D empirical distance-valued radiation model. . . 136
Figure B.1 26 networks formed by 50 circular nodes of areaπdistributed uniformly over a 10x10 unit area. . . 138
Figure B.2 50 node distribution of randomly oriented ellipses with aspect ratio 1:2 resulting in 32 subnetworks when only considering bi-directional communication paths between networks. . . 139
Figure B.3 50 node distribution of randomly oriented ellipses with aspect ratio 1:2 resulting in 18 subnetworks when only considering uni-directional communication paths between networks. Note that the labels of the networks in this situation are not consecutive.. . . 139
Figure B.4 Simulation results with labeling based on bi-directional network links. . . 140
Figure C.1 The importance of near-field sensor coverage where the upper images depict (a) successful detection of a wall by bump sensors during a forward collision and (b) unsuccessful detection of a wall with bump sensors when in reverse. The lower images show (c) successful detection of a wall using an IR wall detector and (d) unsuccessful detection using the IR wall detector. . . 144
Figure C.2 The detection of a cliff obstacle using both (a) IR sensors and (b) wheeldrop sensors. . . 145
Figure C.3 Extrinsic near-field sensors can successfully detect slippage scenarios in ex-treme cases such as (a), but are not reliable for true slippage such as the treadmill of figure (b). . . 146
Figure C.4 Time-lapse photo of a robot using encoders to avoid an obstacle whose position is known a-priori. . . 147
Figure C.5 Modified Roomba equipped with a SICK LMS-200 LIDAR. . . 150
Figure C.6 Detectable object sizes with distances for a LIDAR with an 0.5 degree step size. . . 151
Figure C.7 . . . 154
Figure C.8 Slippage cannot be detected with the scenario of (a) and (b), but can be if obstacles are introduced as in (c) and (d). . . 155
LIST OF ABBREVIATIONS
CSMA - Carrier Sense Multiple Access
dBm - Power in decibels referenced to 1 mW
ENC - Expected Number of Connections
GPS - Global Positioning System
LIDAR - Light Detection and Ranging
MAC - Media Access Control
PODR - Point of Diminishing Return
RSS - Received Signal Strength
RSSI - Received Signal Strength Indicator
SCI - Serial Control Interface
Chapter 1
Introduction
This dissertation focuses on the symbiotic relationship between autonomous robots using minimalist sensing and control technology and wireless sensor networks (WSNs). Information gathered from the network is used to help the robot navigate or perform a given task. Conversely, the robot can provide services to the WSN, such as localization. One application that takes advantage this synergistic relationship is wireless sensor network repair. In this research, techniques are developed that take advantage of the state of the network to allow the mobile robot to navigate the WSN and to complete a repair task.
The enabling feature of WSNs for use in ubiquitous monitoring is that they are low cost. In keeping with the low cost theme, a minimalist approach to mobile robotic repair was adopted. It is demonstrated herein that the WSN repair task can be performed using received signal strength (RSS) measurements alone, by swapping messages with the nodes in the WSN. This meets the minimalist requirement by taking advantage of the communication infrastructure of the network, a task which requires minimal processing.
Figure 1.1: Components of the WSN repair system.
robotic platform and the WSN, ultimately resulted in the development of control strategies for repairing a WSN.
The remainder of this dissertation is outlined as follows:
Chapter 2 is a survey of environmental wireless sensor network (WSN) deploy-ments. These deployments cover large areas and are exposed to the eledeploy-ments. This makes these WSNs ideal for studying WSN connectivity issues. The experiences of several research groups are presented and tabulated, identifying the challenges that were faced and how they were overcome. This information should be useful to researchers planning to use WSNs for future deployments.
here that the disc is only the worst-case model if bidirectional communication between all nodes in the WSN is not a requirement. Bidirectional communication is important when mobile robotic navigation is being undertaken within the WSNs. Without bidirectional communication, a mobile robot can be misdirected by the network to a point of no return. Since the RF shape of a WSN node is unknown, and changes depending on the environment, it is important to develop navigation techniques that are insensitive to models based on a distance metric. Therefore, Chapter 4 researches mobile robotic navigation techniques that rely only on the received signal strength (RSS) measurements of the RF transceiver. This chapter demonstrates that the local gradient in RSS can be used for node tracking. Obstacle avoidance is also introduced as a derivative behavior of this tracking ability. Collision detection using RSS is also demonstrated.
Having established a tracking technique, Chapter 5 employs this to develop al-gorithms for WSN repair. By tracking at a specified RSS around the network, the mobile robot can extend the network’s communication range. This simple technique can detect and repair single hop disconnects between disjoint networks. This technique was implemented on a 6 node network where the mobile robot platform used two directional antennas for de-tecting signal gradients. A separate technique was then developed that improved the repair time by having the mobile robot drive to the perimeter of the network before commencing its search.
While perimeter detection is shown to be advantageous in Chapter 5, it is not trivial to implement within a WSN. Chapter 6 defines edge nodes by applying image analysis to the network. Several edge detection techniques are then applied and a comparative study presented in terms of their ability to successfully identify perimeter nodes in the presence of both localization and placement error. Of the techniques presented, one that locally computes a perturbed convex hull is shown to perform best. The perturbed convex hull technique is therefore implemented on a twenty-five node WSN.
Chapter 2
A Review of Resource Constrained
Wireless Sensor Networks for
Environmental Monitoring
Abstract
2.1
Introduction
Wireless sensor networks (WSNs) have proved themselves an effective means of monitoring the environment. They have been deployed in nursing homes in an attempt to improve the care for the elderly [11], in buildings to aid rescue workers in case of fires [12], to track domestic and wild animals [13, 14], and on redwood trees to provide high resolution insight into spatial and temporal climate changes [2]. This revolution in ubiquitous sensing and computation has largely been fueled by the Berkeley Mote (mote) and its associated operating system, TinyOS, introduced by Kris Pister, Jason Hill, and David Culler [15, 16]. The mote has been joined by similar sensor node platforms including Medusa MK2 [17], Fleck [18], and Nymph [19] to name a few. As sensor network technology evolves, these devices are becoming cheaper, more power efficient, smaller, and generally more capable. This will enable researchers to expand the range of applications of sensor networks, e.g., to include large area deployments that blanket an area with hundreds, or thousands, of nodes. However, such deployments must overcome the practical problems related to packet collisions on the wireless channel, route discovery and efficiency, and power management. In many ways these issues are related and can scale with increased network size. With larger networks, more nodes attempt to access the wireless channel simultaneously so more com-plex routing is required to accommodate a large number of nodes. Collisions and increased data flow mean nodes remain active, resulting in increased power consumption. This is often seen with vital nodes, those near base stations or gateways [20]. With all traffic being routed through these vital nodes, they are often the first to lose power. Physically placing nodes within the network can reduce such problems, and thereby provide the infrastructure for more efficient operation. In large area deployments however, the ability to make precise educated placements is no longer practical since nodes cannot be placed by hand.
discusses the measures taken to overcome them. This information is distilled and displayed in Table 2.2. A discussion of the prominent issues is then presented in Section 2.4. Lastly, Section 2.5 presents the conclusions and closing remarks.
2.2
Sensor Node Characterization
When deploying any WSN, it is useful to know metrics related to the physical layout of the nodes and how the environment can impact their performance. The physical layout directly impacts the choice of routing protocol, or clustering algorithm, needed to achieve efficient communication in the network. Many of the issues that impact the perfor-mance of large scale networks apply to all wireless networks. Some, like weather conditions, are particularly important in environmental sensing applications. This section serves to highlight issues encountered in the deployment of environmental sensor networks through a rigorous survey of sensor node characterization studies.
The treatment of this subject is broken down into 4 categories:
Method of Deployment- how the method of deployment can affect network connectivity. Battery Life- how power requirements vary with sensor payload as well as communication and routing protocols.
RF Communication - how frequency of operation, polarization, asymmetry, and free space loss combine to inhibit communication.
Environmental Effects - the ability of mother nature to impact network effectiveness.
The elements of this classification are not mutually exclusive. For example, adverse weather conditions can increase the signal loss of a radio link. This in turn requires the nodes to transmit at a higher power to maintain communication, reducing the battery life.
2.2.1 Method of Deployment
autonomous helicopter [26]. To date, no large area sensor networks have been deployed by air.
With an aerial deployment it is difficult to precisely place nodes. This will lead to network disconnects when nodes fail to land within the communication range of one another. For example, consider the scenario of Figure 2.1 where 100 nodes with 100 m range are randomly dropped in an area of 1 km2. Nodes in a related network are assigned the same letter, and their range is denoted by a circle. The sample network results in 15 distinct networks, ‘a’ through ‘o’, the largest of which consists of 31 nodes (network ‘b’) while the smallest is a single node with no neighbors (network ‘o’). Although in this example there is good radio coverage, the deployment has limited potential in terms of providing a comprehensive sensory coverage of the area.
It should also be noted that random deployments can result in redundancy, as exemplified by network ‘k’, (see Figure 2.1) which is composed of 6 nodes that are closely clustered. Although not beneficial in terms of coverage or connectivity, redundancy is an important issue when considering the impact that node death has on the deployed system, as additional routing pathways will exist. As will be demonstrated in the review of real-world deployments, routing redundancy mechanisms are beneficial even when the deployment is not random.
2.2.2 Battery Life
Battery life is the major cause of node death. Power efficiency commonly elicits the most attention as it is affected by network behaviors, i.e., the amount of processing done, the activity of the media access control (MAC) layer [27, 28, 29, 30], and routing protocols (see [31] for a survey of routing protocols). As previously mentioned, nodes that are in high traffic areas, such as near the base station, will receive and transmit more and sleep less, causing them to consume more power [20].
The choice of a battery, its chemical composition, or other potential power sources can also be considered. A good overview of battery technologies and power scavenging systems, including solar cells, temperature gradients, human power, air flow, pressure vari-ations, and a thorough treatment of vibrvari-ations, is presented in [32]. According to [32] solar, vibration, and wind offer attractive power scavenging alternatives for sensor networks.
As mentioned previously, redundancy may reduce the impact of a dead node by providing an alternate route. The deployment of more capable gateway nodes has been suggested to lessen the impact of power drain on the supporting node infrastructure [33].
2.2.3 RF Communication
There are several issues relating to inconsistencies in radiation from a transmitter that make the modeling of wireless networks difficult. Not only do radiation variations exist among nodes, but the behavior of an individual node can change with changes in the environment. It has even been proposed that the use of overly simplistic models that do not account for variations (such as in Figure 2.1) frequently result in erroneous performance measurements in simulated protocols [34].
Polarization
Asymmetry
Since isotropic radiation is currently unachievable with a low cost system, it is conceivable that nodes can demonstrate asymmetric linkages. In instances such as the one depicted in Figure 2.2 which employs the Radio Irregularity Model[9], node A may hear node B but node B cannot hear node A. Such occurrences have been reported in detail in [38] which showed no significant correlation between distance and the asymmetric links. Rather the findings demonstrate that reception rate is a better measure of symmetry. Specifically, links with high or low reception rates are considerably more likely to be symmetric than those links with medium reception rates. The study discussed in [40] reports occurrences of link asymmetry to be as high as 5-15% in large network deployments. These links can impact a variety of routing and clustering protocols as discussed in [9].
Figure 2.2: Example of asymmetry using the Radio Irregularity Model (RIM) [9]. Node A falls within communication range of node B, but node B cannot hear node A.
The 90 node test setup is shown in Figure 2.3. In this test the radio transmission strength was software limited and on the ground had a range of<1 m.
Figure 2.3: An array of Tmote-sky nodes attempting to resolve node classifications in a leadership protocol.
Free Space Loss
Assuming no external influences on the system, radio propagation will follow the Friis free space equation, Equation 2.1 [35]. The power received, Pr, is related to the power transmitted, Pt, taking into account the frequency, λ, and the antenna gains Ga and Gb. Assuming these parameters remain constant, the power received is then inversely proportional to the square of the distance, r, between the two nodes. Again, this applies only for the ideal case while in practice one must take into account a number of losses due to environmental factors. There are several models of such losses which use a higher exponent of proportionality, thereby resulting in decreased received power [35].
Pr =GaGb
λ
4πr
2
Pt (2.1)
Hardware Variation
ASIC during production [35]. Empirical studies and statistical evaluations of variations in both receivers and transmitters are presented in [9], [37], and [38].
Hardware variation in reception of Tmote-sky nodes is demonstrated with signal strength measurements between node pairs. For this evaluation, four motes were configured to use the SMA antenna connection rather than the microstrip F-antenna to remove any environmental variables that can interfere with the wireless link. This test therefore did not include variation due to the printed antenna, just the radio and the matching network. To measure receive mode variation, three nodes, A, B, and C were connected in turn via a 50 ohm shielded cable to a single transmitting node. The transmission node was configured to send 100 28 byte messages at each of the eight programmable power levels of the CC2420: 0 dBm, -1 dBm, -3 dBm, -5 dBm, -7 dBm, -10 dBm, -15 dBm, and -25 dBm. The received signal strength of each message was recorded on each receiving node for analysis. The results are shown in Figure 2.4 which plots the programmed output power with what was measured on the receiving end with a 99% confidence interval. There is little variation in the measurements for each individual mote at the higher levels, but it is substantial at the lower powers with standard deviations reaching 0.45 dBm. When comparing node to node, it can be seen that mote C is the weakest of the test set, demonstrating lower than average signal detection for all transmit powers. Mote B on the other hand was generally the highest power node. The far right bar of each set depicts the average across all 3 nodes. Given any node out of this set with a given power setting it can have a 99% confidence interval as large as 6 dBm, or quarter power. This implies that hardware variation should not be ignored when modeling sensor networks.
2.2.4 Environmental Effects
Figure 2.4: Reception variation among three Tmote-sky modules with power ranging from -25 dBm to 0 dBm (3.16µW to 1 mW).
Obstructions
One of the most important considerations is the presence of obstructions. Line of sight is particularly important at the high frequencies often employed by sensor nodes. Even objects not in the line of sight can considerably attenuate the signal if they lie within the Fresnel zone [35]. Reflections caused by objects lying within the Fresnel zone introduce a signal that is out of phase with the original, causing destructive interference. The Fresnel zone is defined by Equation 2.2 which relates the wavelength,λ, to the distancesd1 andd2 of the respective endpoints from a point along the line of sight between the two nodes [35]. There are different Fresnel levels, n, which are regions where the radio waves are out of phase with neighboring regions.
fr≈ r
nλd1d2
d1+d2
(2.2)
be achieved if the antennas are positioned approximately 1 m above ground level. Raising nodes or antenna above the ground is a common practice and this will be demonstrated in the next section. Both [38] and [36] provide evidence that indicates an increase in packet reception as nodes are raised above the ground. In the real-world examples presented in the next section, nodes are routinely elevated to increase transmission range.
Figure 2.5: First 3 Fresnel zones for a 2.4 GHz link with a separation of 100 m. To avoid significant interference from ground reflections objects must be kept outside of the inner 60% (dashed red oval) of the first Fresnel zone. Therefore to reduce ground interference, the node must be raised 1.06 m off the ground.
propagation models is found in [35].
Weather
Even if one could accurately model a given deployment’s connectivity in a specific environment there is no assurance that this will be stable over time. Weather conditions such as rain, snow, and fog can play a dramatic role in attenuating the strength of a link, increasing the aforementioned Friis equation exponent of proportionality. This is demonstrated empirically in [36] where a marked decrease of packet reception in moisture rich conditions is shown. This is caused by absorption and scattering due to water droplets and other particles in the air. Absorption is the largest contributer to attenuation under 10 GHz. An in depth mathematical treatment of the subject is provided in [35]. Some sensor nodes also use microstrip or printed antennas which [44] claims have a greater sensitivity to factors such as humidity and temperature.
Incidental Impairment
Accessibility and environmental impact are only two reasons to use WSNs. Often the subject area is inhospitable and dangerous. Naturally, scientists can expect some node casualties over the course of the deployment. Fire, flood, rockfall, and overly curious wildlife can all cause node failures.
2.3
Experience of Large Area Deployments
Table 2.1: Summary of the deployment infrastructure.
Project Nodes Used Number of Nodes Sensors Used Distance Covered Length of
Deployment
Great Duck Island [1] Mica 43 temperature gateway 4 months
Jul - Nov 2002 humidity 350 ft.
pressure light
infrared (in burrows)
Sonoma Redwoods [2] MicaDot 33 temperature
e2 m spacing 44 days
Apr - June 2004 humidity over 55 m
incident photosynthetic light
reflected photosynthetic light
FireWxNet [45] Mica2 site1 - 6 temperature longest horizontal 5 days
Sept 2005 site2 - 5 relative humidity
e430 m
site3 - 2 wind speed/direction longest vertical
e
180 m
FireBug [46] Mica2 test 1 - 10 temperature not reported 1 day
Sept 2004 test 2 - 12 relative humidity 1 day
barometric pressure GPS
Volc ´an Tungurahua [3] Mica2 3 - mic only infrasonic mircophone longest 10.7 m 3 days
July 2004 1 - GPS only GPS
1 - no sensor
Volc ´an Reventador [4] TMote Sky 16 seismometer
e3 km 3 weeks
Aug 2005 MicaZ 2 microphone
GPS (1 MicaZ)
Glacsweb [6] in house node pressure 100 m through ice >1 year
2003-2005 PIC based temperature gateway - 2.5 km
868 MHz radio 9 3D orientation
433 MHz radio not reported conductivity
strain
Life Under Your Feet [47] MicaZ 10 light 9 in a 2 m x 2 m grid 147 days
Sept 2005-Feb 2006 temperature which single hop to a
soil moisture basestation 35 m away
soil thermistor
Soil Moisture [48] Mica2 9 rain gauge (1) over 1 hectare 30 days
Banksia MDA300 4 sensing soil moisture (3)
1 base 4 intermediary
Vineyard Monitoring [7] Mica2 65 internal temperature
e
30 m spacing 6 months
2.3.1 Great Duck Island [1]
The first environmental wireless sensor deployment using a mote platform, is the University of California at Berkeley’s experiment on Maine’s Great Duck Island in the summer of 2002 [1]. This system consisted of 43 Mica nodes and was tasked with gathering information on the nesting habits of Leach’s Storm Petrels. The nodes were placed both inside and outside burrows and measured humidity, pressure, temperature, and light. The burrow nodes were also equipped with an infrared sensor to detect the presence of a bird. The network was arranged to be single hop with all nodes reporting to a gateway node which was fitted with a long-range antenna, allowing it to communicate over 350 ft back to a base station. The base station provided an internet connection allowing researchers to monitor the network remotely.
This deployment was active for 123 days, during which time a number of issues were addressed. Due to the length and location of the deployment, packaging was a big concern since many of the sensors had to be exposed to the elements in order to take measurements. To protect against environmental conditions, everything except the sensors was covered with a protective coating of parylene. The above ground motes were also placed in a ventilated acrylic casing to provide added protection. Despite these precautions, water penetration still led to failures. Szewczyk et al. also evaluated the network performance in detail, highlighting the problems caused by collisions and unexpected backoff in the CSMA MAC protocol used [1].
2.3.2 Redwood Climate Monitoring [2]
data was routed to a base station which stored the data in a TinyDB [51] database and provided a General Packet Radio Service (GPRS) link for long-distance remote monitoring of the network. Each node also logged measurements into local memory for downloading at a later time as a backup to ensure that all the data was gathered.
The network ran successfully for 44 days, showing discernible trends in both tem-perature and light through the course of the day. However, the data yield was not ideal. It was later discovered that some nodes had filled up their local memory and lost subsequent measurements. Some of this data was recovered at the base node. Erroneous temperature measurements were also noticed throughout the deployment. This occurred when the bat-tery voltage dropped below a threshold. The node still operated, but the sensor reported abnormally high values. Tolle et al. [2] point to the dual roles of the deployment, as a network and as a data-logger, as a check system that allows the performance of both to be evaluated. For other deployments they recommend using software to monitor the status of the nodes in the network, not just the sensor measurements that the nodes provide [2].
2.3.3 Wildfires
A popular application for large scale environmental WSN deployments has been for wildfire monitoring. These deployments can be in areas with varying natural topographies, which affects the significance of measurements. Such variation is demonstrated by the following two deployments.
FireWxNet [45]
The FireWxNet [45] project from the University of Colorado monitored actual wildfire activity in Idaho’s Bitteroot National Forest during September of 2005. They used Mica2 motes to sense temperature, relative humidity, wind speed, and wind direction to help predict fire behavior. Specifically, these nodes were able to monitor for temperature inversions which greatly affect a fire’s behavior [45]. This task is not possible with current remote weather stations, because they do not have the spatial resolution required to sense such events.
nodes were mounted in a protective enclosure that was approximately 1.5 m off the ground. This served to protect the nodes as well as to extend their communication range. A visual cue system utilizing the mote’s LEDs was used for deployment to avoid asymmetric links and ensure network connectivity. Hartung et al. [45] reported improved communication range due to the drastic elevation changes inherent in their mountain deployment. This effectively reduced interference within the Fresnel zone and allowed the nodes to achieve links of approximately 400 m although the quoted range of the Mica2 mote is only 300 m. During the experiment they monitored the battery status and reported a reduced capacity at night due to lower temperatures. Naturally, this translated into a reduced performance for the network, particularly with the base station; which consumes more power through the programming board that connected it to a computer. The authors also reported on interference which they blamed on their overly cautious routing protocol sending each packet 60 times in a minute. Since they used a CSMA based MAC, backoff times escalated with each failed transmission until the frame is over [45].
The FireWxNet deployment has inspired two projects to further aid in the under-standing of sensor network deployments and management. SWARMS [52] (Sensornet Wide Area Remote Management System) is a middleware system that allows multiple disparate and heterogeneous sensor networks to be monitored through a single interface. NodeMD [53] provides the ability to recover from software errors on deployed nodes by preventing failures such as stack overflow, before a node enters an un-recoverable state.
FireBug [46]
2.3.4 Volcano Monitoring [3, 4, 5]
A collaboration of researchers from Harvard University, the University of North Carolina, the University of New Hampshire, and the Instituto Geof´ısico in Ecuador have used sensor networks to aid in the study of two volcanoes in Ecuador [3, 4]. The first deployment was on Volc´an Tungurahua and consisted of 5 Mica2 nodes [3]. Three of the nodes were fitted with low frequency microphones, one was equipped with a GPS unit to provide accurate system timing, and the last was connected to a relay transmitter to communicate 9 km back to base. The three acoustic nodes were housed in waterproof cases to protect them from the weather. These acoustic nodes were then placed in trees to improve radio communication and to keep them away from livestock. One acoustic node was placed at the relay node, while the others were placed 6.3 m, and 10.7 m away from the relay node. Flora prevented longer distances between the acoustic nodes.
This deployment lasted for 3 days and yielded acoustic data that was commen-surate with measurements from a traditional seismic monitoring station nearby. In their network analysis the researchers point to weather influence and antenna orientation as the suspected cause of communication failure and packet loss. The researchers also point out that radio transmission interferes with ADC readings. To combat this they instituted a simple filtering algorithm. However, they also reported that this was not an issue when the Telos mote platform was used.
an event occurs,Fetch iterates through each node in the network, requesting the data each node has sampled. When a request is received by a sampling node, the node stops sampling in order to transmit [5]. Time in the network is synchronized from the GPS equipped node using the Flooding Time Synchronization Protocol [54] which compares and resolves each node’s local time to the global system time of the GPS clock. Any necessary in field updates to the system are performed using Deluge [55].
The Reventador network sampled 229 seismic occurrences over a 19 day period, providing researchers with 107 Mb of data. However, more events were measured using the traditional sensor system. Many events occurred in rapid succession, so while data was being retrieved via Fetch, some events were lost. The authors also discovered a bug in Deluge that caused nodes to reboot. This bug has since been fixed in Deluge updates. A TinyOS clock driver error caused time synchronization errors in FTSP which was then corrected through a lengthy timestamp filtering and rectification process. The research team also experienced power outages at their base camp, which caused more data to be lost. In regard to network performance they point to asymmetrical links and varying clock rates across nodes as their primary issues. They also report the loss of one node, whose antenna was destroyed by volcanic debris [5].
2.3.5 Glacsweb [6]
The initial system was deployed in Briksdalsbreen, Norway in 2003 and consisted of 9 nodes 50 m to 80 m deep [56]. Most nodes in this deployment were unable to achieve communication. One node however got stuck approximately 20 m below the surface and was able to provide insightful pressure data until it eventually worked free and fell out of range. Updated nodes were deployed in 2004 and were successfully able to provide data for over a year from within and underneath the glacier [6]. During this deployment a trend of poor connectivity during rain or cold was reported with respect to the base station and the internet access point. Their internet access point was also once taken offline by an avalanche. Future work on this project looks to address the connectivity issues by further reducing the node frequency to 173 MHz and by switching to an ad hoc form where nodes can use multi-hop communication to relay information back to the base station. This will provide greater adaptability to changes in radio variation in this environment.
2.3.6 Soil Monitoring
To date, there have been two deployments to monitor soil moisture, one in the Eastern United States and the other in Western Australia. Although similar, they offer different philosophies and approaches to data collection and management. The knowledge gained in both has spawned the development of new communication techniques and deploy-ment planning tools.
Life Under your Feet [47] [57]
Researchers at Johns Hopkins University and Microsoft have collaborated on the “Life Under your Feet” project which seeks to provide insight into soil ecology. The water content of the soil affects the organisms that live there, so in studying one, information is gained about the other.
track of any packets it missed during the lump transfer and repeatedly queried the node for the missing information until it successfully had all of the data. Maintaining this high level of raw data integrity allows for alternative scientific analysis than that for which it was originally intended, making it potentially more useful to the scientific community.
The deployment successfully collected data for 147 days and the data was verified against traditional, manual field measurements. The 6 million collected data points were in agreement with local temperature and rainfall trends.
Although it achieved positive results, the project did not run without difficulties. A large number of packet drops were observed and it was determined through packet error and link quality measurements that the link between the nodes and the base station was inadequate. Fortunately, researchers were able to bring a mobile base station closer to the deployment site to gather the stored data. As changes in the network software were desired, each node was manually reprogrammed. This proved to be not only time consuming, but also detrimental to experimentation by human disruption of the deployment environment. Future plans are to reprogram the network wirelessly via tools such as Deluge [55]. Re-programming also necessitated opening the enclosure which adversely affected its integrity, allowing water to seep into some of the containers. Prior to deployment, moisture tests were performed with the enclosure and it was found that silica beads were helpful in keep-ing the instrumentation dry. Data was also lost due to poor connections with the MicaZ which is a common complaint with these motes [5]. This connector does however provide comprehensive access to all of the mote’s peripherals.
Banksia Woodland [48]
In the same vein as the “Life Under your Feet” project, researchers at the Univer-sity of Western Australia and the Australian National UniverUniver-sity have deployed a network in the Banksia woodland in Western Australia to monitor groundwater variations during rainfall events. This information allows for groundwater recharge assessment for ground-water management. What is unique about this deployment is that the focus is on creating a reactive sampling network, sampling more frequently during events of interest and less often otherwise. This more selective sampling regime reduces the amount of data that needs to be transmitted and subsequently increases network lifetime. This is in direct contrast to “Life Under your Feet” which requires all of the raw data to be gathered at a steady temporal frequency.
The network consisted of 9 Mica2 motes fitted with MDA300 sensor boards that were classified into 4 types. The first was a single node serving as a base station and connected to a GSM gateway. Co-located with the base station was a node outfitted with a rain gauge. The rain gauge node signaled rain events to three soil moisture nodes through a series of 4 intermediary routing nodes. SMAC [30] was chosen for this application as it promises lower energy usage through coordinated sleeping cycles. The antennas of all nodes were extended over a meter above the ground to improve the range of radio communication. The nodes were also placed in custom water-tight enclosures.
The network successfully reported temporal and spatial variations in ground water during rain events over a 16 day period. The rain gauge however was operational for the entire 30 days of the field test. Upon deployment a communication calibration test was con-ducted to ensure that symmetric communication was possible between neighboring nodes. Despite conducting this communication test, not all packets were successfully transmitted during the study. Also, repeated messages hinted at dropped acknowledgments and by ex-tension, link asymmetry. These behaviors fluctuated significantly over the duration of the deployment. Some packet loss was also attributed to the GSM gateway and known poor GSM coverage in the deployment area. Lastly, as with many other deployments, mother nature found a way through the mote enclosure causing some sensors to return erroneous measurements.
of information gain depending on how many additional nodes would be required. These researchers also believe that using a better antenna on the GSM gateway would improve their system’s performance. Cardell-Oliver has also examined the network lifetime issue in more detail and has developed ROPE (Reactive, Opportunistic Protocol for Environmental monitoring) [60]. ROPE uses both data compression and network strength analysis to transmit data only when it is best to do so. More reliable transmission of smaller packets (with the same data content) results in less power consumption by the radio.
2.3.7 Vineyard Monitoring [7]
Although most deployed networks are used to provide information about catas-trophic events or for discipline specific insight, researchers at Intel Research, Cassia Tech-nology, and Pacific Agri-Business Research Center have applied sensor networks to help farmers with crop management. More specifically, they have concentrated their efforts on grape fields used in wine production. The temperature and moisture content in vineyards is important in determining what kinds of grapes grow best there and subsequently are an indicator of the profit potential of the land. Traditionally there are only a few tempera-ture logging sensors placed throughout vineyards and the results are extended to an entire region. However, in this project, Beckwith et al. [7] have shown that there is significant variation within such regions.
the placement is in communication range of neighboring nodes. To further ensure message delivery, all messages were sent to the base station 5 times via different routes.
Of the 65 nodes, 16 were used for sampling and routing while the remaining nodes were only concerned with sampling. The nodes sampled the internal temperature, the external temperature, and the battery voltage and reported to the base station every 5 minutes. In contrast, a traditional monitoring station samples every 15 minutes and the data must be downloaded via a wired serial link. 16 such stations were also deployed for comparison. It is interesting to note that the deployment of the traditional sensors took several days longer than the entire 65 node mote network. During operation, the mote base station sends out a time synchronization beacon every 5 minutes to ensure network timing consistency.
The network successfully provided high resolution measurements, spatially and temporally, over 6 months. The temperature variations throughout showed that more cost effective grapes could be grown in different areas than previously thought. The network also proved to be effective in reporting cold events that could destroy the grape crop if not handled correctly. The authors foresee that, in the future, the system would automatically control the irrigation of the fields to handle frost abatement. Analysis of the transmitted packets shows that the message and routing repetition were warranted as only 77% of the data was received. Periods of instability in the network requiring a reset were also reported.
2.4
Discussion
The research deployments presented in Section 2.3 bring attention to several prob-lems associated with WSNs used for environmental monitoring. However, three of these problems are commonly seen throughout the WSN deployments presented: 1) Communi-cation Range, 2) Link Asymmetry, and 3) Weather. These are significant challenges for the WSN community and are not mutually exclusive. For example, inclement weather can cause reduced communication range, resulting in the formation of asymmetric links.
Table 2.2: Summary of problems and solutions for actual deployments.
Project Problems Faced Corrective Measures/Suggestions
Great Duck Island [1] electronics sensitive to moisture coat boards in parylene poor transmission from within nests place another node at burrow entrance
MAC layer backoff from collisions
Sonoma Redwoods [2] bad temperature readings maintain sufficient voltage
logging memory filled use network monitoring software
moisture and dirt custom HDPE case
FireWxNet [45] range mounted on 1.5m pole
temp affects battery use external power for base
asymmetric links visual LED deployment system interference use different routing algorithm
FireBug [46] range mounted on 0.5 m pole
battery connection use heat resistant battery casing
Volc ´an Tungurahua [3] fauna and range put nodes in trees
reduced range due to flora deploy in clear areas
packet loss due to weather none
packet loss due to orientation none
RF interferes with ADC switch to Telos family
Volc ´an Reventador [4] radio range mounted on 1.5 m pole
asymmetrical links none
Glacsweb [6] transmission through ice/snow/water reduce RF frequency from 863 MHz to 433 MHz
loss of link to 173 MHz
movement and orientation use multi-hop communication to adapt
poor connectivity during rain or cold none
electronics sensitive to moisture placed in a protective capsule
Life Under Your Feet [47] poor communication with base station used a mobile base station
develop better network models for deployment planning [59] water seepage into enclosure wireless reprogramming to prevent wear
silica
Soil Moisture [48] poor communication link use multiple routing paths for messages
Banksia including asymmetry
poor GSM coverage use a better antenna
network lifetime use power efficient communication such as ROPE [60]
FireWxNet, the Banksia soil monitoring deployment, and the volcano monitoring projects. The dependence of radio range on node height is visually demonstrated in Fig-ure 2.6. A single Tmote-sky mote was placed in the center of a parking area while a robot equipped with another mote, mounted at a height of 25 cm. The robot was also equipped with a GPS which it used to traverse a set of predefined waypoints covering a portion of the parking structure. At each waypoint, the robot sent 10 ping messages to the static node and recorded the average. This was done for 3 cases: 1) the static node is placed on the ground, Figure 2.6(a), 2) the static node is raised 1 m above the ground, Figure 2.6(b), and 3) the static node is placed at a height of 1 m and the robot’s node is placed at 0.5 m, Figure 2.6(c). Each trial of Figure 2.6 has the same color scale that varies from red, signifying high signal strength, to violet, denoting low signal strength. The signal strength increases significantly when the static node is raised from 0 m to 1 m. Raising the receiving node shows another small improvement, noticeable in the area near the static node.
Weather also proved to be a hindrance, both with the electronics and the RF char-acteristics of the networks. Rain was shown to be a problem in both the Great Duck Island and Glacsweb projects. Glacsweb also had to deal with ice which not surprisingly caused them communication difficulties. To counter this, the researchers lowered the communica-tion frequency with each new node they deployed. The low night temperatures experienced by the FireWxNet team reduced the output of their base station battery causing inter-ruptions in service. Battery life caused several nodes to go offline or in some cases report erroneous data.
A few of the applications reviewed sampled large amounts of data and had difficulty maintaining the data integrity. It was suggested in [2] that data be logged locally as well as transmitted to the base station. Doing so serves as a backup in case of network disruptions. In our own wireless sensing work we have found that micro secure digital (µSD) cards are a cost-effective alternative to the mote’s standard on-board flash memory. SD cards typically provide gigabytes of storage compared to the 1 Mb available on the Tmote-sky and do not adversely affect the mote’s footprint.
(a) (b)
(c)
To further study large scale deployments, UC Berkeley, Arched Rock Corporation, and Moteiv Corporation have introduced TRIO [61]. TRIO lies in an area of approximately 50,000 m2and consists of 557 custom, environmentally protected nodes. These nodes consist of a Telos mote, an XSM module, and a solar power unit. The first test of this system was for multi-target tracking and used a microphone, an infrared sensor, and a magnetometer. Although the primary purpose of this network is to examine service in high density deploy-ments it will likely also provide insight into large, random, environmental deploydeploy-ments as well. In fact, the TRIO project has already had issues with environmental factors, including birds, who perch on the nodes and cover the solar cells with excrement. This warranted a change in design, making the nodes a less attractive perch. The solar panels were also oriented to face the sun to improve their effectiveness.
2.5
Conclusion
This paper has provided a survey of the problems affecting the performance of environmental WSNs. The problems are reinforced by reviewing research in nodal charac-terization in a variety of environments. Recent large area deployments were also reviewed to demonstrate that these issues are also common in the field. Furthermore, lessons learned from these deployments and techniques applied to correct them are given to aid others considering future deployments. It should be noted that this is only a sampling of deployed mote-based environmental sensor networks. Work is ongoing in many of these projects and new projects, such as the prediction of landslides in India [62] and the detection of avalanche conditions in the mountains of Utah [63, 64], are in development.
As deployments continue to branch into large scale environmental monitoring, a point will be reached when manual positioning is no longer feasible. At this point the WSN community will need to prepare for and predict how the presented issues will affect such deployments. This will require better modeling techniques, deployment strategies, and methods for repairing networks when the communication fails.
Chapter 3
Connectivity Assessment of the
Disc RF Radiation Model
Abstract
Deployment is one of the major challenges facing wireless sensor networks. The goal is to be able to rapidly deploy nodes to form networks over large areas. However, techniques for large area deployments, such as air drop, are imprecise, and connectivity issues will emerge. This study focuses on the analysis of connectivity trends in large area random network deployments. Specifically, this research addresses the role of the disc RF model in wireless sensor network simulations. This is done through a comparative evaluation of connectivity in uniform random deployments against the results of the percolation studies presented in [65]. Franceschetti et al. [65] demonstrate that the circle is the worst-case shape. Our research demonstrates similar trends and conclusions when isotropic models are considered, but does not agree when the RF model is anisotropic and bidirectional communication is required throughout the network. Ultimately, the circle is neither the worst-case nor the best-case model.
3.1
Introduction
such as air drops, will be used. Air drop deployments cannot be adjusted on-site to ensure ideal deployment and so they will be imprecise and spatially random.
The argument has been made in [65] that the disc is the worst-case model to use when a two-dimensional network is being considered. Rather, in [65], diffuse probabilistic, annuli, and even asymmetric radiation patterns have been demonstrated to be more effective in terms of network connectivity than a disc having the same effective radiation area. The comparative studies presented in Section 3.3 validate that Franceschetti’s conclusion also holds for uniform random network deployments using isotropic RF models. Section 3.4, however, shows that this theory breaks down when asymmetric links are introduced by anisotropic RF models and bidirectional communication is required throughout the network. An overview of related work is provided in Section 3.2. Section 3.3 takes a close look at the connectivity of randomly distributed networks. Specifically, the research com-pares the connectivity of uniform, randomly distributed networks, in a variety of isotropic shapes, to the standard disc communication model used by many researchers. Uniform random networks are assumed, meaning that nodes are randomly placed independently of the placement of other nodes. Section 3.4 compares an empirically derived RF model of the Tmote-sky to the connectivity profiles developed in Section 3.3. Section 3.5 draws conclusion on this study and suggests avenues for future work.
3.2
Related Work
As previously mentioned, the most likely form of deployment of nodes across large areas is by air drop. Researchers at UC Berkeley have implemented and tested an air drop scheme using an unmanned aerial vehicle [25]. They dropped several Rene motes for use in vehicle tracking exercises. Corke et al. focus both on WSN deployment and repair using an autonomous helicopter [26]. The helicopter is guided by differential GPS to repair points where it drops motes into network holes to affect a repair.
Other research has focused on deployment geometries for achieving connectivity and coverage. A grid layout method that accounts for different sensor and communication ranges is presented in [68]. Likewise, [69] examines decentralized protocols that guarantee a degree of coverage while maintaining connectivity. Similar analysis of coverage is shown in [70] to be increasingly difficult in three dimensions.
The Tmote-sky mote [71] is currently the newest mote marketed by the Moteiv Corporation. Different aspects of this mote have been characterized. In [72] the authors provide a hardware characterization of the Tmote-sky where they focus on the design de-cisions made in relation to its production. The authors of [72] also conducted tests that characterized the received signal strength indicator (RSSI). Work on RSSI and link quality indication (LQI) is carried out in [73] using the MicaZ mote which has the same radio as the Tmote-sky. A more comprehensive study of the Tmote-sky is performed in [39]. The results reported show that LQI is a good predictor of packet yield and demonstrates a correlation between RSSI and relative distance. Node height above the ground provides increased packet reception, but poor performance is reported on the ground. Experimenta-tion performed in [74] uses external high gain antennas to achieve better link stability and range.
3.3
Connectivity Analysis of Isotropic RF Models
Franceschetti et al. [65] evaluated several different nodes using a branching per-colation process and determined the expected number of connections per node (ENC) at which the network being formed percolates, or grows without bound. Branching percola-tion processes begin with a single node, and that node has a defined area over which it can communicate. A Poisson process, with a given mean and standard deviation of λ, is then used to determine how many children will be created. These children are then placed in a uniform random manner within the area covered by the parent. This technique is then repeated on the children. If a child is placed within the area of an ancestor other than its parent (or in some implementations, its own siblings) then that node is eliminated from future consideration. An example of a network resulting from a branching process is shown in Figure 3.1 using discs and a λof 7.
Figure 3.1: Result of a sample branching process simulation. Only the viable nodes are displayed, each represented by a colored circle and its corresponding id number.
implies that a more diffuse shape, or even an asymmetric shape, would be more desirable than a disc pattern in that each node in the network requires fewer neighbors to maintain connectivity. In other words, for the process outlined here, the circle is the worst-case network model. However, a large area air-drop deployment will not behave according to a branching process. The next step therefore, is to see if the observed trend holds for uniform random populations.
Coverage and redundancy can also be visually evaluated in Figures 3.2(a)-3.2(c).
For each population of N nodes the number of independent networks formed in the region is counted. This metric provides a measure of how connected the region is. Repeating this simulation 100 times and averaging the number of networks formed yields the trend of Figure 3.3(a), plotted with 95% confidence intervals. As expected, there are fewer disjoint networks with increasing nodal density. However, according to this plot, a distribution of 5 nodes has about the same number of disjoint networks as 200 nodes. In terms of connectivity, there are far more connected nodes in the 200 node deployment, which results in improved coverage and general usefulness since a larger network can provide more information about the environment. Another way to look at this figure is to compare the ratio of nodes to the number of disjoint networks as shown in Figure 3.3(b). In this figure, the difference between a 5 node and a 200 node network is clear as it takes into account the number of nodes that comprise the networks. With only 5 nodes, all nodes are in different networks, whereas for 200 nodes there are approximately 10 different networks and for 50 about 25 different networks. Figure 3.3(b) also tells a little about network sizes. For the 200 node case, the inference is that there is one network that is considerably larger than the others. This is best seen in Figure 3.3(c) which shows the largest networks for the various nodal densities averaged over all trials. This corroborates the above inference showing that with 200 nodes there is a network of around 185 nodes that is significantly larger than the others.
(a) 30 nodes (b) 100 nodes
(c) 300 nodes
(a)
(b)
(c)
Figure 3.1, only the viable nodes are presented. If all nodes that were considered are shown, the jumble of Figure 3.4 results. Of the 2500 nodes considered in this sample branching process, an average of 2093 were discarded and 407 were considered viable. By cutting these nodes out, the addition of neighbors via a redundant branch are eliminated, keeping the ENC pure. This selectivity is not a luxury afforded to a uniform random distribution, and redundancy abounds. This can be beneficial however, as redundancy can help ease traffic bottlenecks and helps compensate for node failures.
Figure 3.4: All nodes placed in a sample branching process simulation. The colored circles and numbers represent viable nodes while the black numbers are node placements that were discarded.