Kai Li, Kien Hua Department of Computer Science University of Central Florida

28
Mobility-assisted Distributed Sensor Clustering for Energy- efficient Wireless Sensor Networks Kai Li, Kien Hua Department of Computer Science University of Central Florida

description

Mobility-assisted Distributed Sensor Clustering for Energy-efficient Wireless Sensor Networks. Kai Li, Kien Hua Department of Computer Science University of Central Florida. Traditional WSN: Energy issues. Internet. Sensors are energy constrained : Typically powered by AA batteries - PowerPoint PPT Presentation

Transcript of Kai Li, Kien Hua Department of Computer Science University of Central Florida

Page 1: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Mobility-assisted Distributed Sensor Clustering for Energy-efficient

Wireless Sensor Networks

Kai Li, Kien Hua

Department of Computer Science

University of Central Florida

Page 2: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Traditional WSN: Energy issues

Sensors are energy constrained: Typically powered by AA batteries

Communication consumes too much energy Data packet (self generated and

other sensor’s) Control packet (e.g. routing,

topology maintenance)

InternetSensors not only sense but also

relay data

Wireless sensor

Sink

Page 3: Kai Li, Kien Hua Department of Computer Science University of Central Florida

The aMANET Approach

Internet

Sensors transmit only

their own data

Autonomous MANET nodes collects and

forwards data to sink

Sink

The idea: Save sensor energy by separating sensing from communication

Page 4: Kai Li, Kien Hua Department of Computer Science University of Central Florida

aMANET is motivated by mobile connected robots researchMultiple mobile nodes cooperate to achieve some

common tasks (e.g. for energy-efficient data collection)Mobile nodes form a middle-layer network for data

collection and electronic transmission Our aMANET approach is different from existing

mobile elements approaches such as:Mobile sinks. An aMANET node doesn’t have to be as

advanced as mobile sinks (i.e. cost-effective). They don’t have to be connected to the internet.

Data mules. Data mules travel physically to deliver data to the sink, resulting in unpredictable latency. aMANET, However, exploits electronic data transmission.

The aMANET Approach

Page 5: Kai Li, Kien Hua Department of Computer Science University of Central Florida

The aMANET Challenge

Sink

Each aMANET node is responsible for a

sensor group

Need a clustering technique that can be performed in a

distributed mannercan save sensors energy

to extended their lifetime

Autonomous mobile node

Page 6: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Clustering in aMANET

Clustering in aMANET is different from traditional sensor clustering algorithms (e.g. LEACH, HEED, etc.)

In traditional sensor clustering, the cluster head (CH) is chosen from normal sensors. CH roles are rotated to distribute energy consumption.

It’s straightforward to let the aMANET nodes to assume the CH role, which is a energy consuming task.

Which one is more energy

efficient ?

Page 7: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Some Numerical AnalysisDenote the average energy consumption of a non-CH sensor in each round as , and that of a CH as . Let denote the ratio of CH. Compare the following two paradigms

Static Cluster Head (SCH). In SCH, sensors are chosen from normal sensors and CH roles are periodically rotated among sensors Sensor’s average energy consumption in each

round

Mobile Cluster Head (MCH). In MCH, a number of mobile aMANET nodes assume CH role Sensor’s average energy consumption in each

round

and can be estimated as

And

using the Two-ray Ground radio model. is the number of bits transmitted per round

According to the Two-ray Ground radio model, the energy consumption of transmitting and receiving one bit data over distance are given as

,

and

,

Where is the energy dissipation of the transceiver circuit, and is that of amplifier. Both and are dependent on distance: when is smaller than a threshold and otherwise

Page 8: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Numerical Results

Experiment Parameters:

5nJ/bit

100pJ/bit/

0.0013pJ/bit/

75m

Page 9: Kai Li, Kien Hua Department of Computer Science University of Central Florida

MADSEC: Problem Formulation

Sensors communicate with aMANET node in single hop, using adjustable transmission power

Each sensor has an initial energy of Joules and transits -bit data in each time-slot

Find a configuration to optimize some energy-oriented objective function.

Page 10: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Problem Formulation

The objective function

is the energy sensor n consumed by transmitting data to aMANET node k (located at ) in a round.

could be 1 or 0 depending on if sensor n is clustered to MCH k

is a function of the sensor’s state○ can be seen as a weight to each sensor based on sensor’s

state (e.g. residual energy).

When setting =1 this function measures the

energy consumption for the whole network in

each time-slot.

Page 11: Kai Li, Kien Hua Department of Computer Science University of Central Florida

The K-means Algorithm

The minimization of with respect to and , could be solved using the standard K-means algorithm

Initialization:

Randomly initialize for all

Iteration:

Repeat:

// Step 1: Assign sensor to the closest aMANET node (Cluster formation)

for for all

// Step 2: Update MCHs positions (Cluster Update)

for all

Until: converge

Output: and

Page 12: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Challenges The K-means algorithm is centralized

Can we make it distributed?

The algorithm requires location information , , which is generally not available in WSN, especially (i.e. the location of sensors) Can we make the least assumption about

location awareness?

Page 13: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Initialization:

Randomly initialize for all

Iteration:

Repeat:

// Step 1: Assign sensor to the closest aMANET node (Cluster formation)

for for all

// Step 2: Update MCHs positions (Cluster Update)

for all

Until: converge

Output: and

The K-means algorithm (revisit)

Step 1 can be approximated. Each mobile node sends out an invitation message. Sensor joins the one with the strongest received signal strength

How can a mobile node reposition himself to the right location in Step 2 without location information?

Page 14: Kai Li, Kien Hua Department of Computer Science University of Central Florida

MADSEC: The wAMRP metric The weighted Average Minimum Reachability

Power (wAMRP)

: number of sensors in a cluster

: the weight assigned to a sensor

: the minimum power for sensor to reach its Cluster Head○ Radio propagation model dictates that

𝑤𝐴𝑀𝑅𝑃=∑𝑖=1

𝑁

𝑤𝑖 ∙𝑀𝑖𝑛𝑅𝑝𝑖

𝑁

How does a mobile node compute the wAMRP at its current location ?

NEXT

Page 15: Kai Li, Kien Hua Department of Computer Science University of Central Florida

MADSEC: Computing wAMRP Each mobile node use the following protocol to compute

wAMRP within its cluster.1: ClusterInfo = {}

2: for power_level = 1 to MAX_POWER_LEVEL do

3: Set transmission power to Power(power_level)

4: Broadcast probe_msg(MyID)

5: for all received ask_msg(SensorID, ResEnergy, MyID) do

6: if SenorID ClusterInfo then

7: Add SensorID to ClusterInfo,

Compute weight according to ResEnergy, and

Record weight and Power(power_level)

8: endif

9: endfor

10: endfor

11: Compute wAMRP

Page 16: Kai Li, Kien Hua Department of Computer Science University of Central Florida

MADSEC: Relocation

How to locate the point where we get the minimum wARMP ?

We do not assume location awareness

Exhaustive search is not a feasible solution

Not interesting!!!

Page 17: Kai Li, Kien Hua Department of Computer Science University of Central Florida

MADSEC: Relocation

𝐮𝟎

𝐮𝟏

𝐮𝟐

𝐱∗

Random move

Random move

Directional move

𝒍1=𝑣𝑡0 [cos𝜃0sin𝜃0 ]𝒍2=𝑣𝑡1 [cos𝜃1sin𝜃1 ]

Initial location

Target location

We don’t know yet!

We actually could arrive at the optimal location with only three moves!

Page 18: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Formulation Let wAMRP = be the metric measured at location , then

We have the following equation array

=

{𝑔 (𝐮0 )−𝑔 (𝐮1 )=2𝐶 𝐥𝟏𝑇 [ 1𝑁 ∑

𝑛=1

𝑁

𝑤 𝑖𝐱𝐧−12(𝐮0+𝐮𝟏)]

𝑔 (𝐮1 )−𝑔 (𝐮2 )=2𝐶 𝐥𝟐𝑇 [ 1𝑁∑

𝑛=1

𝑁

𝑤𝑖𝐱𝐧−12(𝐮1+𝐮𝟐)]

Page 19: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Formulation We need to get to , but we only need to know .

Substituting this into the equation array, we get

Where

and

[c os𝜃0 sin𝜃0cos𝜃1 sin 𝜃1] ⌈ 𝑦1𝑦2 ⌉=⌈

𝑏1𝑏2⌉

𝑏1=𝑔 (𝐮0 )−𝑔 (𝐮1 )

2𝐶+ 12𝑣2(𝑡0

2+𝑡 12)

𝑏1=𝑔 (𝐮1)−𝑔 (𝐮2 )

2𝐶+ 12𝑣2(𝑡0

2+𝑡 12) [1+2cos (𝜃0−𝜃1) ]

Page 20: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Formulation The only requirement for a valid solution of the

equation array is simply

Which gives us

The two random moves should not be collinear!

𝑑𝑒𝑡 ([c os𝜃0 sin𝜃0cos𝜃1 sin𝜃1])≠0

𝜃1−𝜃0≠𝑘𝜋 ,𝑘∈𝑍

Page 21: Kai Li, Kien Hua Department of Computer Science University of Central Florida

MADSEC: Data Collection aMANET nodes schedule data aggregation after

clustering is finished

Each round of data collection is divided into a number of TDMA frames, in a similar way to LEACH

Each sensor will be allocated one time frame for data transmission

aMANET nodes fuses data collected from sensor, sends them over the aMANET and the sink.

Page 22: Kai Li, Kien Hua Department of Computer Science University of Central Florida

An overall review of MADSEC

…Iteration 1 Iteration 2 Iteration 3

Cluster formation

CH reposition

Clustering Phase …Data Collection Phase

𝑇 1 …

Round 1 Round 2

𝑇 2 𝑇 3

Page 23: Kai Li, Kien Hua Department of Computer Science University of Central Florida

NS2 Simulation Parameters

Type Parameter Value

Network Network size 100m 100m

Number of Sensors (N) 100

Sensor distribution Random

Sink Location (50, 0)

Application Sensor initial energy 1 Joule

aMANET node speed 2 m/s

Data rate 100 kb/s

Radio transceiver Maximum transmission power -1.58dBm

Receiver sensitivity -24dBm

5 nJ/bit

10 pJ/bit/

0.0013 pJ/bit/

MADSEC Round length 50s

Clustering frequency 5 rounds

Number of power levels 20

Page 24: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Simulation Results

Random Mobility: each MCH makes a random move, sensors join an

MCH with the minimum RSS

C-LEACH: a centralized version of LEACH, assuming a

centralized server holing information of the whole network

Unequal weights MADSEC, sets

compute the weight of a sensor as a

function of its residual energy:

Even random mobility can almost double

sensor network lifetime. And MADSEC does

even better!

Comparison of different clustering techniques:

Page 25: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Simulation Results (Contd.)

With smaller size clusters (more MCHs),

the computation of wAMRP is less accurate

More MCHs incurs more network

overhead

Comparison of variable number of MCHs

Clusters becomes smaller with more MCHs, therefore

sensors consumes less energy and live

longer

Page 26: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Simulation Results (Contd.)

Comparison of varying number of power levels

With more discrete power levels, the relocation

accuracy becomes higher, leading to closer results compared with optimal

Page 27: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Conclusions

MADSEC is a clustering technique designed for an aMANET for energy-efficient data collection. Its desirable features are: Energy-efficiency: sensor network lifetime are

remarkably improved over conventional clustering techniques.

Distributed Computing: each aMANET node runs the clustering algorithm in a distributed manner.

Few assumptions: we only need adjustable power levels. aMANET nodes don’t need GPSs for clustering updates.

Page 28: Kai Li, Kien Hua Department of Computer Science University of Central Florida

Thank [email protected]