An Experimental Receiver Design For Diffuse IR Channels Based

Download Report

Transcript An Experimental Receiver Design For Diffuse IR Channels Based

An Experimental Receiver Design
For Diffuse IR Channels Based on
Wavelet Analysis & Artificial
Intelligence
R J Dickenson and Z Ghassemlooy
Optical Communication Research Group
Sheffield Hallam University
www.shu.ac.uk/ocr
Contents
•
•
•
•
•
•
•
Diffuse IR indoor multipath channel
Compensating schemes
Traditional receivers
Wavelet and AI based receiver
Proposed receiver
Simulation results
Conclusions
Diffuse IR System - Major Performance
Limiting Factors
 Inter Symbol
Interference
 Noise
 Power Limitations
Tx
Rx
Compensating Methods
 Modulation Schemes
– DH-PIM
– DPIM
– PPM
 Diversity
– Angle
– Multi-beam
Rx
Tx
Rx
Rx
Rx
Rx
Rx
Traditional Receiver Concepts
12
 ZFE
 DFE
 Coding
- Block
- Convolutional
- Turbo
Normalised optical power requirements (dB)
10
8
6
4
2
OOK-NRZ
0
-2
32-DH-PIM2
-4
32-DPIM
-6
-8
-10
-3
10
32-DH-PIM1
32-PPM
-2
-1
10
10
0
10
DT
Normalised optical power requirements Vs.
normalised delay spread for various modulation schemes
Alternative Techniques - Wavelet
Analysis & Artificial Intelligence







De-noising
Image Compression
Earthquake
Electrical Fault Detection
Mechanical Plant Fault Prediction
Apple Ripeness
Communications
What Is A Wavelet?
Simple Description:
 A finite duration
waveform
 Has an average value
of zero
 Is a basis function, just
like a sine wave in
Fourier analysis
Fourier Analysis And The Wavelet
Transform
3 sine waves at different
frequencies and times.
Frequency spectrum
The peaks will remain statically
located regardless of where in time
the frequencies occur
Fourier Analysis And The Wavelet
Transform
Wavelet results
In the wavelet domain we have both a representation of
frequency (scale), and also an indication of where the
frequency occurs in time.
Neural Networks
x
1
w
1
 Loosely based on
biological neuron
 Neural networks come
in many flavours
 Used extensively as
classifiers
 Supervised and
unsupervised learning
x
x
w
Σ
2
2
F
Out
w
n
n
Input
Layer
Hidden Hidden
Layer 1 Layer 2
Output
Channel Model & Receiver Structure
Receiver
…1 0 1 0
Tx
CHANNEL

Rx Filter
Feature
Extraction
Pattern
Recognition
WAVELET
ANALYSIS
NEURAL
NETWORK
Thresholder
1 0 1 0...
NOISE
• Input data format: OOK NRZ
• Channel: Carruthers & Kahn Channel Model, with impulse
response of:
6a 6
h(t , a) 
t  a 
where u(t) is the unit step function
7
u (t )
Simulation Flow Chart
Incoming Data
n bits long.
Low Pass Filter
Decimate
Stream it to 5
Bit windows
CWT at 4
scales on every
window
• CWT:
- 5 bit sliding window
- coif1 mother wavelet
- Operating scales of 60,
80, 100 and 120 using
Bit To Detect
Decimate each
set of
coefficients to
100 sample
points
5 Bit Window
Pack samples
into a 100xn
matrix
Offer each
column to the
neuronal
classifier
Threshold the
output to 1 or 0
• ANN:
- 4 layers with 176 neurons
- 3 different activation functions, trained to detect the
value of the centre bit from a 5 bit length window
Simulation Results – BER V. SNR


•
•
Data rate: 40 and 50 Mb/s
Normalised delay spread: 0.44
and 0.55
for BER of 10-5 the wavelet-AI
scheme offers SNR improvement
of:
- ~ 8 dB at 40 Mbps
- ~ 15 dB at 50 Mbps
over the filtered threshold
scheme
For the wavelet-AI scheme the
penalty for increasing the data
rate by 10 Mbps is ~ 5dB whilst it
is around 15dB for the basic
scheme.
Conclusions
 A novel technique to combat multipath
dispersion
 Improvement of ~ 8 dB in SNR compared
with the threshold based detection scheme
 Promising results, however, significant further
work is required.
 Not intended to replace coding methods
Any Questions?
• Thank you for your kind attention.
• I will attempt to answer any questions you
have.