Presentation on theme: "Z. Ghassemlooy, S Rajbhandari and M Angelova"— Presentation transcript:
1Z. Ghassemlooy, S Rajbhandari and M Angelova Signal Detection and Adaptive Equalization Using Discrete Wavelet Transform - Artificial Neural Network for OOK Indoor Optical Wireless LinksZ. Ghassemlooy, S Rajbhandari and M AngelovaSchool of Computing, Engineering & Information Sciences, University of Northumbria, Newcastle upon Tyne, UK
2Outline Optical Wireless – Key issues Digital Signal Detection EqualizationWavelet ANN Based ReceiverResults and Conclusion
3Indoor Optical Wireless Links The key issues:Eye safetyshift from 900 nm to 1550 nm - eye retina is less sensitive to optical radiationpower efficient modulation techniquesMobility and blockingdiffuse configuration instead of line of sight, but at cost of:reduced data rateincreased path lossmultipath induced inter-symbol-interference (ISI)
4Digital Signal Detection - The Classical Approach The discrete-time impulse response of the cascaded systemoptical channel (ceiling bounce)
5Digital Signal Detection - The Classical Approach OOK - the average probability of error:the probability of error for the penultimate bit in ai:.where opt is the optimum threshold level, set to the midway value of RPave (Tb)0.5
6Digital Signal Detection - The Classical Approach Matched filter is difficult to realized when channel is time varying.Maximising the SNR based on the assumption that noise statistics is known.SNR is sensitive to the sampling instants.In non-dispersive channel, the optimum sampling point is at the end of each bit period.In dispersive channel, the optimum sampling point changes as the severity of ISI changes.
7Digital Signal Detection - The Classical Approach For higher values of normalized delay spread (> 0.52)- bit error rate cannot be improved simply by increasing the transmitter powerTo mitigate the ISI, optimum solutions are:- Maximum likelihood sequence detector- Equalizers1-3 - A practical solution(i) Inverse filter problemThe frequency response of the equalizing filter is the inverse of the channel response.Adaptive equalization is preferred if the channel conditions are not known in advance.Two classes : linear and decision feedback equalizer.(ii) Classification problem1- J. M. Kahn and J. R. Barry, Proceedings of IEEE, 85 (2), pp , 19972- G. W. Marsh and J. M. Kahn, IEEE Photonics Technology letters, 6(10), pp , 19943- D. C. Lee and J. M. Kahn, IEEE Transaction on Communication, 47(2), pp , 1999
8Equalization - A Classification Problem Dispersion induced by channel is nonlinear in natureReceived signal at each sampling instant may be considered as a nonlinear function of the past values of the transmitted symbolsChannel is non-stationary- overall channel response becomes a nonlinear dynamic mapping
9Equalization: A Classification Problem Classification capability of FIR filter equalizer is limited to a linear decision boundary (a non-optimum classification1)FIR bases equalizers suffer from severe performance degradation in time varying and non-linear channels2The optimum strategy - to have a nonlinear decision boundary for classification- ANN- with capability to form complex nonlinear decision regions- In fact both the linear and DFE are a class of ANN3 .- Wavelet41- L.Hanzo, et al, Adaptive wireless transceivers: Wiley-IEEE Press, 2002, pp2- C. Ching-Haur, et al , Signal Processing,vol. 47, no. 2, pp3- S. Haykin, Communications Magazine, IEEE , vol.38, no.12, pp , Dec. 20004- D. Cariolaro et al, IEEE Intern. Conf. on Communications, New York, NY, USA, pp , 2000.
10Receiver - Classification Based OpticalSignalOptical ReceiverFeature ExtractionPattern ClassificationPost-ProcessingWavelet TransformNeural NetworkModular based receiver:Feature extraction (wavelet transform) - for efficient classificationPattern classification (ANN).WT-ANN based receiver outperforms the traditional equalizers1.1- R. J. Dickenson and Z. Ghassemlooy, International Journal of Communications Systems, Vol. 18, No. 3, pp , 2005.
12CWT vs. DWT CWT DWT - Infinite scale - but with redundant coefficients - no redundancy as in CWT- easier to implement using filter banks (high pass and low pass)- reduced computational time- possibility signal denoising by thresholding the wavelet coefficient
13Discrete Wavelet Transform Level 1DWTcoefficientsLevel 2DWTcoefficientsDown-samplingFilteringcD1h[n]2cD2Signalh[n]2x[n]cA1. . .g[n]2cA2g[n]2DWT coefficient - obtained by successive filtering and down samplingSignal is decomposed:- using high pass h[n] and a low pass g[n] filtersfilters are related to each other and are known as the quadrature mirror filter.- down sampling by 2
14WT- ANN Based Receiver Model 8-sample per bitSignal is decimated into W-bit discrete sliding window. (i.e. each window contains a total of 8W-bit discrete samples )Information content of the window is changed by one bit3-level DWT for each window is determinedDWT coefficients are denoised by:i) Thresholding : A threshold is set and ‘soft’ or ‘hard’ thresholding are used for detail coefficientsii) Discarding coefficients: detail coefficients are completely discardedDenoised coefficient are applied to ANNANN is trained to classify signal into two binary classed based on DWT coefficients
15Denoising Signal using DWT Hard thresholdingSoft thresholdingThe threshold level for universal threshold scheme:: variance of the wavelet coefficientDenoised signalwhere -1 is the inverse WT
16Simulation Parameters ValueData rate Rb155 MbpsChannel RMS delay spread Drms10 nsNo. of samples per bit8Mother waveletDiscrete MeyerANN typeFeedforward back propagationNo. of neural layers2No. of neurons in 1st layer4No. of neurons in 2nd layer1ANN activation functionlog-sigmoid, tan-sigmoidANN training algorithmScaled conjugate gradient algorithmANN training sequence400 bitsMinimum error1-30Minimum gradientDWT levels3
17Results – BER for OOK @ 150 Mb/s Maximum performance of ~6 dB compared to linear equalizer.Performance depends on the mother wavelets.Discrete Meyer gives the best performance and Haar the worst performance among studied mother wavelet.Figure: The Performance of OOK at 150Mbps for diffused channel with Drms of 10ns
18Results - BER for OOK @ 150 & 200 Mb/s 510152025-5-4-3-2-1SNR (dB)BERANN(155Mbps, W=3)Unequalized 155MbpsLinear Equalizer(200Mbps)ANN(155Mbps, W=1)Linear Equalizer(155Mbps)ANN(200Mbps, W=3)ANN(155Mbps, W=5)The DWT-ANN based receiver showed a significant improvement compared to linear equalizerSNR gain of ~6 dB at BER of 10-5 for W = 33-bit window is the optimumReduced complexity compared to CWT based receiver without any degradation in performanceFigure: The BER performance of OOK linear and DWT-ANN base receiver at 155 and 200 Mbps for diffused channel with Drms of 10ns
19ConclusionsThe traditional tool for signal detection and equalization is inadequate in time-varying non-linear channel.Digital signal detection can be reformulated as feature extraction and pattern classification.Both discrete and continuous wavelet transform is used for feature extraction.Artificial Neural Network is trained for classify received signal into binary classes.3-bit window size is adequate for feature extraction.Enhance performance compared to the traditional FIR equalizer ( a gain of ~ 6dB at BER of 10-5.Reduced complexity using DWT compared to CWT based receiver with identical perfromance.