Docstoc

An Artificial Immune System Based Sensor Network for Frost Warning and Prevention

Document Sample
An Artificial Immune System Based Sensor Network for Frost Warning and Prevention Powered By Docstoc
					                     An Artificial Immune System Based Sensor Network for
                                 Frost Warning and Prevention
                  Nigel P. A. Browne, Helia Mohammadi, Abdolreza Abhari and Marcus V. dos Santos
                                                Ryerson University
                                                 350 Victoria St.,
                                                 Toronto, Canada
                nbrowne@acm.org, helia@scs.ryerson.ca, aabhari@scs.ryerson.ca, m3santos@ryerson.ca


Keywords: Sensor Networks, Artificial Immune Systems,                  The solution presented here utilizes ZigBee wireless sen-
Pheromone Warning, ZigBee, Mesh Topology                           sors and an artificial immune system, inspired from several
                                                                   theories to develop an effective system.
Abstract
This paper presents an artificial immune system, using a Zig-       2.     BACKGROUND
Bee wireless mesh, to detect the onset of frost in an agricul-
                                                                      The system developed during this research draws on sev-
tural setting. The system was first tested using a small sim-
                                                                   eral areas, including artificial immune systems, pheromone
ulation and further validated by a physical implementation.
                                                                   warnings and sensor networks.
The implementation validated the hypothesis by testing the
application in a real-world setting, but also revealed several
limiting factors that were not detected during the simulation      2.1.    Artificial Immune System
phase. The implementation encompassed the development of              Artificial immune systems (AIS) are inspired by the func-
both the middleware and a monitoring and data-logging ap-          tions and structures of the human immune system (HIS). Fur-
plication.                                                         ther reading on the HIS can be found in [2, 3].
                                                                      The concept of an AIS was first introduced as Immune Net-
                                                                   works in [4, 5]. AIS was then introduced as a separate subject
1.   INTRODUCTION                                                  in [6, 7]. The core concept in AIS algorithms is the classi-
   This paper presents a potential solution to frost warning       fication of “self” and “non-self” or “dangerous” and “not-
and prevention in agriculture, using a wireless sensor network     dangerous”. A more detailed description of the different types
and an artificial immune system.                                    of AIS algorithms can be found in [8, 9]. For this work, we
   This application was targeted because there is currently a      draw upon several AIS theories, specifically, Danger Theory,
gap between the available technologies and the systems uti-        Pheromone Warnings and Network Theory.
lized by the agricultural industry. This is particularly signifi-
cant, because frost damage causes millions of dollars of dam-
                                                                   2.2.    Danger Theory
age each year [1]. By having a system that can automatically
                                                                      Danger Theory is based on a innate immune system re-
provide localized frost prediction and warnings, producers
                                                                   sponse found in all plants and animals. While AIS normally
would be able to better protect their investments. There are
                                                                   focuses on the differentiation of self/non-self cells [10], Dan-
various methods to prevent frost damage, once the threat is
                                                                   ger Theory states that only dangerous cells should be de-
established, but they are generally expensive to operate. Thus,
                                                                   stroyed. Otherwise, the self/non-self response mechanism
a producer would prefer to utilize them only when absolutely
                                                                   may lead to self destructive immune responses [11]. An ex-
required, as opposed to the current method of operating them
                                                                   ample of this is the human immune system’s ability to recog-
“just in case”. Examples of frost mitigation methods include
                                                                   nize food bacteria, which are non-self but non-dangerous and
torches, micro-sprayer systems and inversion fans. Unfortu-
                                                                   thus are not destroyed.
nately, the most effective mitigation system for sensitive and
high value crops, such as wine grapes, are costly inversion
fans. These fans resemble windmills and are used to mix and        2.3.    Network Theory
move the cooler air.                                                  Network Theory is a branch of AIS that focuses on the
   By developing an effective and intelligent sensor network       interactions of a network of immune cells for warning and
that was able to automatically operate the frost mitigation        classification. Each cell influences the reactions of the sur-
equipment, producers could maximize the survivability of           rounding cells to learn the patterns of the non-self cells. This
their crops, while simultaneously minimizing their operating       allows them to react faster when facing a similar pattern in
costs.                                                             the future.




     Copyright held by SCS.
2.4.   Pheromone Warning                                          simulated transmission delay in an attempt to make the sig-
    Another class of immune response is Pheromone Warning.        nal propagation through the simulated mesh more realistic.
It is an airborne communication mechanism, and is a key fea-      Based on the results of the simulation, the next step was the
ture in our implementation. Pheromone Warnings are often          physical implementation and testing, as described below.
observed in nature as a mechanism for stationary organisms
to communicate with other members of their species. An ex-        3.1.   Common Node Hardware
ample of this is when a tree is attacked by insects. The victim      All nodes within the network utilized a Texas Instruments
will release a pheromone warning that causes the surrounding      MCP430 CPU and a CC2480 ZigBee processor. The CPU
members of its species to produce toxins. This makes them         utilized two outputs to drive indicator LEDs, a photocell was
less palatable to the invaders. Another example of pheromone      connected to one of the analog-to-digital inputs and a push-
interactions and responses are demonstrated in [12].              button was connected to an interrupt-based digital input. Ad-
                                                                  ditional inputs and outputs exist on both chips, but were un-
2.5.   Sensor Networks                                            used for these experiments. The temperature sensor is built
  Sensor Networks consist of multiple independent sensors         directly into the ZigBee processor. An on-chip antenna from
which are connected using a wireless connection. These sen-       Johanson Technology was connected to the CPU.
sors are generally small and have low power requirements.
This is advantageous because they can operate for long peri-      3.2.   AIS Mesh Master Node
ods of time from either batteries or solar power sources.           The master node of the AIS Mesh is the coordinator of the
                                                                  ZigBee network and also serves as a gateway to the logging
2.6.   Mesh Topologies                                            application on the host computer.
   Mesh networks are a type of ad-hoc network. Normally, the        The coordinator is responsible for building the initial Zig-
nodes in this type of topology are not mobile. In this topol-     Bee network, setting the parameters and configuring the mesh
ogy, each node can communicate with every other node using        security. Once the network is initialized, the node functions as
multiple hops. Mesh Networks can still operate even when a        a message forwarding device, sending all received mesh mes-
node breaks down or a connection fails. This is done by “hop-     sages to the host application.
ping” from node to node, using multi-hop connections, until
the destination is reached. This makes Mesh Networks robust,      3.3.   Mesh Sensor Nodes
self-healing and suitable for continuous connections.
                                                                     The mesh nodes perform the majority of the work in this
                                                                  implementation. The system is composed of sensors, which
2.7.   ZigBee                                                     function similarly to an artificial immune system cell. Each
   ZigBee is an open mesh networking standard, based on the       cell is a ZigBee sensor node, containing a CPU, full ZigBee
IEEE 802.15.4 standard. ZigBee is low-cost, low-power and         stack, radio, on-chip antenna, temperature sensor, photo cell
wireless. ZigBee technology is used in industrial, scientific      and the custom node software. The node software is described
and medical environments. This technology is cheaper than         in the “Artificial Immune System Implementation” section.
other WPANs, such as Bluetooth. The outdoor range of some
ZigBee modules (Texas Instruments CC2591) is up to 2.5Km.         3.4.   Network Topology
An important feature of the ZigBee nodes is that they are self
                                                                     The ZigBee network was implemented as a mesh. Each
configuring and cooperative. If a nodes fails, the surrounding
                                                                  node was configured as a ZigBee router, allowing other nodes
nodes attempt to self cure and find a different route to connect
                                                                  to always connect to it. When a node was initialized it would
to the network.
                                                                  attempt to join an existing ZigBee network. If a network was
                                                                  unavailable it would wait and attempt a connection at a later
3.     METHODOLOGY                                                time. During the period that a node was not connected to the
   The system designed and implemented for this experiment        mesh, it was still able to function as a standalone, non-logging
is a synthesis of mesh networks, danger theory and network        sensor.
theory. The following section describes the components that          The physical test network utilized two sensor nodes and a
comprise the artificial immune system mesh and how they            master node.
work together.                                                       While designing the system, other topologies were consid-
   Prior to attempting to implement the artificial immune          ered and discarded. This was because only a mesh topology
mesh, the hypothesis was tested using a small scale simu-         would allow a hardware-based artificial immune system. Ad-
lation of the algorithm on a Windows computer. The simu-          ditionally, considering that a full scale network in an agri-
lation was a windows application written in C#. It utilized a     cultural setting could cover hundreds of hectares, using any
topology other than a mesh quickly becomes infeasible and           a node is triggered into a danger state, the other nodes imme-
would prevent AIS based solution from being effective.              diately surrounding it are also more likely to be triggered.
                                                                       The nodes determine their total danger level using the al-
                                                                    gorithm outlined in Program 1, utilizing influences from their
3.5.   Data Logging and Monitoring Application                      local sensors and the inputs from the surrounding mesh. As a
   The network coordinator node was connected directly to           final step, the nodes send a unicast message to the coordinator
the host computer. This permitted the development and use           with their local and total stimulus values.
of a monitoring, network debugging and data logging solu-              When the unicast message is received by the coordinator,
tion. The monitoring application also displayed the topology        the monitor application records the information to a local
of our small test network, the sensor outputs and the artificial     SQL database.
immune system’s status.                                                The communication and ZigBee configuration was built
                                                                    upon the ZigBee API provided by the chip manufacturer.
3.6.   AIS Implementation
   The AIS implementation is based on the basic algorithm           4.   RESULTS AND DISCUSSION
described by [9]. In particular, this implementation differs           The first results obtained from the series of experiments
from standard artificial immune systems, because each node           were those from the simulation of the potential AIS mesh
of the sensor mesh is a cell in the system. This creates a mesh     network. The simulated mesh assumed that each node was
that is intelligent and able to respond locally to distant stimu-   connected to the nodes surrounding it in a grid. The simu-
lus.                                                                lated signal delay was accomplished by calculating the ap-
   Each node generates a local danger value based on its sen-       proximate time a signal would take to cross the open space
sor inputs. This value is then broadcast to the mesh network,       between the nodes. This delay was then incorporated into the
and used to communicate the level of danger throughout the          node’s signal receiving code. While this method was a sim-
mesh, much like naturally occurring pheromone signaling.            plistic simulation, it provided an effective way to test the al-
                                                                    gorithm before attempting a physical implementation.
Program 1 Algorithm Pseudocode                                         The simulation was evaluated using a simulated 400m
  Initialize()                                                      by 300m (12 hectares) test plot. By introducing localized
  while stopping condition not true                                 changes in the simulated environments we were able to mon-
       SL = GetLocalStimulus()                                      itor the danger signal at each node and the propagation and
       SM = 0                                                       influence it causes on the surrounding nodes.
       for each SN in ReceivedBuffer                                   Experimental curiosity caused us to test a line topology and
            SM = SM + hops(SN) * CN * SN                            star topologies to compare them with the mesh topology. As
       end for                                                      expected, the line and star topologies did not propagate the
       ST = CL * SL + CN * SM                                       danger signals in a radial manner, which was required to send
       if ST >= D then                                              the danger signal to the surrounding nodes without knowing
            ActivateFrostCounterMeasures()                          explicitly where they were within the topology.
       end if                                                          Inspired by the results of the algorithm during the simula-
       Broadcast(ST)                                                tion phase, we proceeded to implement the algorithm in the
       UnicastToCoordinator(SL , ST)                                sensors nodes described in section three.
  end while                                                            The testing of the physical sensor network began by verify-
                                                                    ing the basic functionality and communication of each node.
   SL = Stimulus from local sensors                                 We quickly discovered that the on-chip antennas, while inex-
   SN = Stimulus from specific node                                 pensive and easy to work with, severely limited the transmis-
   SM = Stimulus from the mesh                                      sion range of the nodes. However, even with the limited range
   ST = Total stimulus                                              the nodes functioned as expected.
   ReceivedBuffer = Buffer of stimulus                                 To test the effectiveness of the sensors at detecting the on-
          from other nodes in the mesh                              set of frost, we acclimatized all of the nodes to 20 degrees
                                                                    Celsius before each experiment.
                                                                       The first experiment entailed gradually cooling a single
   When a node receives a danger level broadcast, it incorpo-       node to 1 degree Celsius. It was observed that as the node
rates the transmitting cell’s danger level with its own. After      cooled, its danger level increased and when it was broadcast
each hop, the original danger level is attenuated, reducing the     to the other nodes they successfully received the warning.
impact that it has to nodes further away. This means that when      Since the other nodes remained at 20 degrees Celsius, the
stimulus from the mesh was not enough to force remaining           speed monitor and a leaf moisture sensor would provide more
nodes to raise their danger levels.                                comprehensive information about the local environment.
   The second experiment involved cooling multiple sensors            The nodes used in the experimental implementation uti-
at the same time, but with slightly different rates. This exper-   lized on-chip antennas, which has a limited range. A de-
iment also successfully showed that the nodes detected the         ployed version of the node would require an external omni-
approaching freezing conditions as dangerous and then initi-       directional antenna. Additionally, the nodes could be outfitted
ated their warnings. The warning signals were successfully         with solar panels and a charging circuit to maintain power in-
broadcast throughout the network to the other nodes. This          definitely throughout the growing season.
time, since multiple nodes were cooling they always affected
the danger levels of the nodes surrounding them in the mesh.       REFERENCES
This was observed to cause the nodes to trigger their warning       [1] International herald tribune, Jan 18 2007.
signals sooner, as was anticipated. The danger levels can be
visualized as multiple waves meeting and causing construc-          [2] Gene Mayer. Microbiology and Immunology On-Line
tive interference.                                                      Textbook, chapter Immunology - Chapter One: Innate
   The third experiment involved acclimatizing one node to              (non-specific) Immunity. USC School of Medicine,
-4 degrees Celsius and then gradually cooling a neighboring             2006.
node to 5 degrees Celsius. This experiment had similar results
to the previous one, but the second node indicated danger sig-      [3] Jr. et al Janeway CA. Immunobiology. Garland Science,
nificantly sooner than in previous experiments. This was ex-             6th edition, 2005. ISBN 0-443-07310-4.
pected because of the sub-zero temperature of the first node.        [4] Packard N. Farmer, J.D. and Perelson A. The immune
   Based on the experiments, the Artificial Immune Sensor                system, adaptation and machine learning. Physica D,
Mesh appears to be effective. However, a significant factor              2:187–204, 1986.
that we were not able to successfully simulate in the labora-
tory was the consequence of other environmental conditions          [5] Varela F.J. Bersini H. Hints for adaptive problem solv-
on the effectiveness of the node communication in the mesh.             ing gleaned from immune networks. parallel problem
In particular, high humidity or heavy rain has previously been          solving from nature. First Workshop PPSW 1, Dort-
a problem with the reliability of signal transmission in the            mund, FRG, 1990.
field.
                                                                    [6] Perelson A.S. Allen L. Cherukuri R. Forrest, S. Self-
   An unexpected result of the experiments was that the mesh
                                                                        nonself discrimination in a computer. In Proceedings of
topology that formed was not as uniform as hoped. This af-
                                                                        the IEEE Symposium on Research in Security and Pri-
fected the danger signal propagation through the network.
                                                                        vacy, pages 202–212, 1994.
This may be caused by the small size of our physical test net-
work, since the nodes were able to reach and route through          [7] A Hofmeyr S, Forrest S. Architecture for an artificial
other nodes that they would not be able to transmit to in a real        immune system. Evolutionary Computation 8, 4:43–
deployment.                                                             473, 2000.

                                                                    [8] Dasgupta D., editor. Artificial Immune Systems and
5.   CONCLUSIONS AND FUTURE WORK                                        Their Applications. Springer-Verlag, Inc. Berlin, Jan-
   This paper presented a working implementation of a dis-              uary 1999. ISBN 3-540-64390-7.
tributed artificial immune system using a ZigBee mesh. The
primary hypothesis, that a ZigBee mesh could be used as a           [9] A.P. Engelbrecht. Computational Intelligence -An In-
Network Theory AIS was confirmed. Our secondary hypoth-                  troduction. Wiley, second edition, 2007.
esis, that the system could be used for the detection of frost
conditions, was also confirmed.                                     [10] Osborne B Goldsby R, Kindt T. Kuby Immunology, vol-
                                                                        ume Fourth. W H Freeman, 2000.
   The use of ZigBee for the mesh network was functional and
effective. However, after testing the implementation, it be-       [11] Uwe Aickelin. The danger theory and its application
came apparent that scaling the application may be constrained           to artificial immune systems. In University of Kent at
by the ZigBee network topology restrictions. Whether this               Canterbury, pages 141–148, 2002.
would impact a large scale deployment of an application
based on this system is still an open question.                    [12] G.V.P Reddy and A Guerrero. Interaction of insect
   An additional change that would improve its effectiveness            pheromones and plant semiochemicals. Trends in Plant
for detecting frost would be to incorporate additional types of         Science, 9, 2004.
sensors. In particular, the addition of a humidity sensor, wind

				
DOCUMENT INFO
Shared By:
Tags:
Stats:
views:26
posted:5/15/2012
language:English
pages:4