Z. Ghassemlooy & S Rajbhandari Performance of Diffuse Indoor Optical Wireless Links Employing Neural and Adaptive Linear Equalizers Z. Ghassemlooy & S Rajbhandari Optical Communications Research Group, School of Computing, Engineering & Information Sciences, University of Northumbria, Newcastle upon Tyne, UK ICICS 2007 Singapore
Outline Optical wireless – introduction Mutipath induces ISI ANN based equalizer Wavelet-ANN receiver Final comments
Optical Wireless Communication – What Does It Offer? Abundance bandwidth No multipath fading High data rates Protocol transparent Secure data transmission License free Free from electromagnetic interference Compatible with optical fibre (last mile bottle neck?) Low cost of deployment Easy to deploy Etc.
Power Spectra of Ambient Light Sources Wavelength (m) Normalised power/unit wavelength 0.2 0.4 0.6 0.8 1 1.2 0.3 0.5 0.7 0.9 1.0 1.1 1.3 1.4 1.5 Sun Incandescent x 10 1st window IR Fluorescent Pave)amb-light >> Pave)signal (Typically 30 dB with no optical filtering) 2nd window IR
Classification of Indoor OW Links RX TX (Diffuse) Directed Hybrid Non-directed Line-of-sight Non-line-of-sight
Indoor OWC - Challenges Causes (Possible ) Solutions Power limitation Eye and skin safety. Power efficient modulation techniques/holographic diffuser/ Transreceiver at 1500 nm band Noise Intense ambient light (artificial/ natural) Optical and electrical band pass filters, Error control codes Intersymbol interference (ISI) Multipath propagation (non-LOS links) Equalization, Multi-beam transmitter No/limited mobility Beam confined to small area. Wide angle optical transmitter , MIMO transceiver. Shadowing blocking LOS links Diffuse links/ cellular system/ wide angle optical transmitter Limited data rate Large area photo-detectors Bandwidth-efficient modulation techniques/Multiple small area photo-detector Strict link set-up LOS links Diffuse links/ wide angle transmitter
Modulation Techniques
Normalized Power and Bandwidth Requirement 2 3 4 5 6 7 8 10 12 14 16 18 20 Bit resolution, M Normalized bandwidth requirement PPM DH-PIM1 DPIM DH-PIM2 OOK PPM the most power efficient while requires the largest bandwidth DH-PIM2 is the most bandwidth efficient DH-PIM and DPIM shows almost identical bandwidth requirement and power requirement There is always a trade-off between power and bandwidth 2 3 4 5 6 7 8 -16 -14 -12 -10 -8 -6 -4 -2 Bit Resolution, M Normalized Power Requirement (dB) DH-PIM2 PPM DH-PIM1 DPIM
Power Spectral Density Notice the DC component:- when filtered will result in base line wander effect
Optical Wireless - Channel Model Basic system models – F. R. Gfeller et al 1979, J. M. Kahn et al 1995, Measurement studies - H. Hashemi et al 1994, J. M. Kahn et al 1995, - Diffuse + shadowing Statistical models - J.B. Carruthers et al 1997 Ray tracing techniques (to obtain simulated channel responses) - J.R. Barry, J.R., et al. 1995, F.J. Lopez-Hernandez, et al, 2000 Segmentation of reflecting surfaces + ray tracing techniques to calculate the intensity and temporal distributions - S. H. Khoo et al 2001 Fast multi-receiver channel estimation - J.B. Carruthers et al 2002 Fast Multireceiver Channel Estimation iterative site-based method for estimating the impulse response of optical wireless channels. simultaneous evaluation of channels for a number of receiver or transmitter locations, Offers significantly improved calculation times
Channel Model - Ceiling Bounce Model Developed by Carruthers and Kahn. Impulse response is: LOS Diffuse Diffuse shadowed LOS shadowed where u(t) is the unit step function and a is related to the RMS delay spread D
OWC - LOS Links Least path loss No multipath propagation High data rates Problems Noise is limiting factor Possibility of blocking/shadowing Tracking necessary No/limited mobility Rx Tx
Received signal for non-LOS Links OWC - Diffuse Links Different paths ─>Different path lengths ─> different delay ─>ISI. ISI ─> Delay Spread Drms ─> Room design and size Impulse response of channel Problems: High path loss Limited data rate due to ISI Power penalty due to ISI Rx Tx 2 4 6 8 10 -0.4 -0.2 0.2 0.4 0.6 0.8 1 1.2 Normalized Time Amplitude Received signal for non-LOS Links
How to Combat Noise and Dispersion? Noise Filtering: Optical or Electrical Match Filtering: Maximises signal-to-noise ratio, Modulation: Z. Ghassemlooy et al Coding: Block codes, Convolutional and Turbo codes. Spread Spectrum Tracking Transmitters: D. Wisely et al Imaging Receivers: J.M. Kahn et al Integrated Optical Wireless Transceivers: D.C. O’Brien Equalisation Diversity: S. H. Khoo et al 2001 Wavelet and AI based equalisers: Z. Ghassemlooy et al Some investigative work and preliminary simulations on MLSD undertaken but not pursued. MLSD can be implemented after any compensation technique.
Techniques to Mitigate the ISI Optimal solution - Maximum likelihood sequence detection. - Issues: complexity and delay Sub-optimal solution - Linear or decision feedback equalizer based on the finite impulse response (FIR) digital filter - The impulse response of filter c(f) = 1/h(f), where h(f) is the frequency response of channel
FIR Filter Equalizer (Classical Signal Processing Tool) Assumptions The statistics of noise is known (normally assume to be Gaussian) The channel is stationary or quasi-stationary The channel characteristics are known (at least partially) Signals are linear Problems: Non-linearity, time-varying and non-Gaussianity of real signals and channel Solution: Artificial neural network (ANN) based signal processing which takes into account non-linearity, time-varying and non-Gaussianity of signal and channel
ANN Output Neurons Hidden layer Input One or more hidden layer(s) Output is function of sum and product of many functions Useful tool because of learning and adaptability capabilities Extensively used as a classifier Application in many areas like engineering, medicine, financial, physics and so on Training is necessary to adjust the free parameters ( weight) before can be used as classifier Supervised and unsupervised learning (training) Hidden layer Input Neurons Output
ANN Activation Function f(.) Sigmoid function - x1 Inputs w1 x1w1 Weights Z xn xnwn wn ∑ f(.) Activation function Output Bias bi Activation Function f(.) Sigmoid function - Linear function - if , if if Any function that is differentiable
ANN Both the multilayer perceptrons (MLP) and the radial basic function (RBF) have been used for equalization RBF requires a larger number of hidden nodes at lower values of SNR The cascaded MLP and RBF outperform both the MLP and RBF in terms of the BER performance Learning rules for MLP The error-correction: {wij} are renewed after each iteration - the most simplest The Boltzmann Hebbian ………… Whichever training rule is used, the basic principle is to modify {wij} so that the error function is decreased after each iteration.
ANN Supervised Learning (Training) Target: to minimize the error en between target vector set tn and neural network output on for all input vector set in. Error signal en in Neural network on Comparator tn Algorithms: Compare tn and on to determine en (= tn-on) Adjust {wn} and bi to reduce the error en Continue the process until en is small
OWC System Block Diagram Input data X(t) Output data Tx h(t) Rx Equalizer Threshold detector ∑ n(t) ANN Equalizer Adaptive Linear Equalizer For a non-stationary environment
OWC Link A feedforward back propagation ANN PPM Encoder h(t) ∑ Neural Network Decision Device Optical Transmitter Receiver n(t) Decoder X(t) Matched Filter Zj Zj-1 . Zj-n Yj Z(t) M 0 0 1 0 Ts = M/LRb Xj 0 1 0 0 A feedforward back propagation ANN ANN is trained using a training sequence at the operating SNR Trained AAN is used for equalization
ANN Training Process The channel is time-varying To estimate channel parameters, a training sequence is transmitted at regular interval for tracking changes in the channel The information on channel is stored in the form of weights that are updated on receiving the training sequence The signal flows from input to the output (feedforward) while the error signal propagates backward, hence the name feedforward backpropagation NN The learning duration and the number of iteration required to adjust the NN parameters depends on the complexity of learning task Here the aim is not to optimize the learning task but to send a learning sequence of certain length to allow the NN to estimate new channel parameters
Simulation Flow Chart
Simulation Parameters Values Number of layers 2 Number of neurons in each layer 36,1 Activation function tan-sigmoid, log-sigmoid Training algorithm scaled conjugate gradient algorithm Minimum error 1-30 Minimum gradient
Simulation Parameters – Contd. Values OOK PPM DPIM Data rate, Rb (Mbps) 150 Bit resolution, M 3 Slot duration, Ts 1/ Rb M/( Rb .2M) 2M/(2M+1) Rb Training sequence 2000 bits 300 symbols 600 symbols RMS delay spread, Drms(ns) 10 5 2 Normalized time delay (Drms/Ts) 1.5 0.75 0.3 4 2.3 1.13 0.45 Delayed samples 8 22 11 6 13 7
Results and Discussion Error performance for LOS links (150 Mbps) 2 4 6 8 10 12 14 -6 -5 -4 -3 -2 -1 SNR( dB) SER 8-PPM 8-DPIM OOK PPM requires the least SNR to achieve a desirable slot error rate (SER) OOK shows the highest power requirement to achieve a desirable SER
Results and Discussion Unequalized (Rb = 150Mbps, Drms = 5ns) Unequalized OOK requires ~27dB more SNR compared to LOS link at SER of 10-5 For high values of normalized delay spread increasing the optical power will not improve error performance PPM suffers the most severely in a diffuse link because of the short pulse duration 5 10 15 20 25 30 35 40 -6 -5 -4 -3 -2 -1 SNR( dB) SER LOS PPM DPIM OOK Unequalized PPM Unequalized DPIM Unequalized OOK
Results and Discussion OOK performance (Rb = 150Mbps, Drms = 5ns) 5 10 15 20 25 30 35 40 -6 -5 -4 -3 -2 -1 SNR( dB) SER LOS Unequalized ANN equalizer Linear equalizer ANN equalizer and linear equalizer shows identical performance Power penalty is ~6.6 dB compared to LOS links at SER of 10-5 SNR gain is ~ 20 dB compared to unequalized performance at SER of 10-5
Results and Discussion ANN Equalizer (Rb = 150Mbps, Drms = 5ns) Performance of equalized DPIM and PPM is better than OOK even in highly dispersive channel DPIM show the best SER performance. Power penalty is ~14.3dB, 9.2dB, 6.7dB for equalized PPM, DPIM and OOK compared to corresponding LOS performance for a SER of 10-5 . 5 10 15 20 -6 -5 -4 -3 -2 -1 SNR( dB) SER DPIM PPM OOK ANN Equalized LOS
Results and Discussion ANN Equalizer (Rb = 150Mbps, Drms = 1, 2, &10 ns) 5 10 15 20 25 -6 -5 -4 -3 -2 -1 SNR(dB) SER DPIM PPM 10ns 2ns 1ns OOK Equalized PPM shows the best performance in less dispersive channel (Drms<2) Equalized DPIM shows the best SER performance in highly dispersive channel (Drms >2)
Wavelet-AI Receiver Signal decimated into 3 bit sliding windows. Transmitter filter g(t) Diffuse IR channel h(t) ADC Slicer Input bits Output XPavg X R Artificial Intelligence Anti- alias (LPF) noise n(t) + Wavelet Analysis Signal decimated into 3 bit sliding windows. Each window is transformed into wavelet coefficients by the CWT process. The coefficients are passed to the neural network for classification. Describe the general Wavelet-Ai receiver process and the need for training or adaptive learning.
Signal Sample ‘The Window’ 3 bit sliding window 1 2 8 3 For OOK signal decimated into 3 bit windows. Each window is processed into wavelet coefficients by the continuous wavelet transform (CWT). Introduce data window. Discuss its modification for the PPM binary symbol detector
Simulation Results - Multipath Propagation 3 Compare performances with equalised cases. Again briefly discuss Wavelet AI Binary Symbol Detector. Close by pointing out similar or better performance of Wlt-AI and its flexibility when implemented as the basis of software receiver. Equalised traditional receiver architecture & Wlt-AI reference (OOK RZ) Equalised traditional receiver architecture & Wlt-AI reference (PPM) Normalised to: 2.5Mb/s for BER 10-6 OOK RZ
Conclusions Artificial neural network as an equalizer shows similar error performance to the linear equalizer Equalized PPM shows the best performance in less dispersive channel while DPIM shows the best error performance in highly dispersive channel Power penalty for equalized OOK is ~11.5 dB in highly dispersive channel (Drms = 10 ns) at high data rate of 150Mbps making it feasible for practical implementation.
Issues and Future Works Higher sampling rate (at least 8 samples per bit) Hardware complexity The need for parallel processing, at the moment Adaptive error control decoding using neural network. Combine equalization and decoding as a single classification problem Wavelet network for equalization and decoding Development of high performance pointing, acquisition, and tracking.
Thank you!