Sunteți pe pagina 1din 6

Adaptive Pseudo-OFDM based WLAN Systems using Neural Network

Sauptik Dhar1, Madhumita Paul2


1 Wipro Technologies, Department of Electronics and Telecommunication Engineering, National Institute of Technology, Silchar.

Email: sauptik.dhar@wipro.com1 , mita_paul@rediffmail.com2

Abstract In this paper we have presented an adaptive Orthogonal Frequency Division Multiplexing (OFDM) system by relating its equivalence with a proposed neural network architecture. We have analyzed the Euclidean Distance (ED) between the transmitted and the received signal and shifted the orthogonal points of the carriers to provide a better channel response. Thus the orthogonal property of the system is changed to a Pseudo-orthogonal property to provide a better system performance. Hence we move towards a new system of Pseudo Orthogonal Division Multiplexing (POFDM).

I. INTRODUCTION Orthogonal Frequency Division Multiplexing (OFDM) is one of the most popular technologies used in Broadband Fixed Wireless Access (BFWA). In the conventional OFDM system the Orthogonality in Time Domain (TD) is preserved by using Cyclic Prefix (CP). Moreover, for an overlapped bandlimited orthogonal signal multicarrier modulation (MCM) a Raised Cosine Filter (Window) is used to limit the signal to a certain band. This window function results in convolution of the signal in (TD) and further spreads the signal, making it even more robust to the time dispersive channel effect. If the effective signal duration is T and TCP and Twin represents the CP duration and the duration induced by the window, then the transmission efficiency is reduced by a factor given by T/ (T+ TCP + Twin). In discrete form this may be given by N/ (N+ NCP + N win ) [2], where N, NCP, N win represents the samples equivalent to the time durations T ,T CP , Twin respectively. Thus the increase in Guard Interval (GI) NGI=Ncp + N win results in decrease in the power efficiency of the system. Among some of the major constraints that an OFDM system suffers are sensitivity to carrier frequency offset, phase noise caused by tuning oscillator, Doppler shifts and selective fading of the OFDM signal in Frequency Domain (FD) induced by the channel. In this

paper we will concentrate more on the selective fading part of the OFDM signal transmitted. Off late much emphasis has been laid upon some other aspects, which includes sub band blocking [4] for the Peak to Average Power ratio (PAPR) problem or the preequalised OFDM [5]. In our paper we have reconsidered the preequalisation of the signals but have changed the orthogonality in Frequency Domain (FD) to provide a Pseudo-OFDM (POFDM) system for a better system performance. An abridged version of our work has been highlighted in the following sections. In section II we present a general overview of the conventional OFDM and a new angle to view the system. In section III we provide a complete analysis of our proposed model. In section IV we give the simulation results and conclude in favor of our model. II.CONVENTIONAL OFDM SYSTEM ANALYSIS A brief analysis of the conventional OFDM system has been presented in this section. A. Signal Generation The generated OFDM signal is given by
NSC

s(t)= x n exp(j2(n f)t) n=1


NSC

,where f=T sc

= xnexp(j2(n /Tsc)t) Tsc=subcarrier duration n=1 On sampling the signal at the k th sample where Tsamp is the sampling period, we get
NSC

sk= xn,k exp(j2(n /Tsc)k Tsamp)


n=1
NSC

sk = xn,k exp(j2(n k/Nsc)


n=1

(1)

From (1) we see that passing the complex signals through an IFFT block can generate the OFDM symbol. This in the vector form can be given by S=XT W1(NSC). (2)

The band over which the channel variations are highly correlated ( i.e RH(f,0)/ RH(0,0)>1) is called the coherence bandwidth(fch) If fch of the channel is small compared to the bandwidth of the transmitted signal, the channel is called to be frequency selective [ 3] Thus from this we obtain a measure for the frequency selective nature of the channel. Besides this the channel correlation function bears quiet a great deal of information that can be derived on changing our insights. Without loss of generality we can associate a channel impulse response to each path (equivalent to the sample function of the channel random process). At any particular instant a delay () will be representative of any particular path. An analysis of the following cases provides a new view If RpcP (f,t)/ RpcP(0,0)>1 then it is a nonselective channel; If RpcP(f,t)/ RpcP(0,t)>1 (8) then the channel is predictive over a duration t along the frequency domain. This itself is the very basis that can support the concept of a preequalised system. If we set d( 1,*) as the difference function of any delay path from the Line of sight signal(LOS,1) in TD, then for a proper reception we shall have to set a limit (1*) on the receiver side to which the signal can stray from the LOS i.e d(1 ,2 ) 12 (9) Applying the threshold on all the delayed signals d(1 ,j ) ; =max{ 1j } (10)
J=over all the paths

S=Generated Signal X=Modulated Signal W=Weight matrix of the IFFT However a complete detailed representation of the OFDM signal is given from [1]
NSC/2 1

s(tkTsamp)=Re{w(tkTsamp)xn,kexp(j2(fc+n/TFFT)(t-k Tsamp)) n=-NSC /2 for kTsamp - Twin TCP t kTsamp + Twin + Twin =0; else T: TFFT: TCP: Twin: fc : NSC : w (t): (3)

Symbol length; time between two consecutive OFDM symbols; FFT time; effective part of the OFDM symbol; Duration of the Cyclic Prefix; Window interval for spectral shaping; Center frequency of the occupied frequency spectrum FFT length; number of FFT points; Pulse shaping function;

(7) p= the paths concerned; P= the set of Paths

B. Channel Response The overall channel response for the Multipath effect and Additive White Gaussian Noise (AWGN) can be given by
max

r(t)= h(,t)*s(t)+n(t) = h(,t)s(t) d +n(t) ;


0

(4)

h(,t) : The channel impulse response due to the multipath effect; n(t) : The sample function of the AWGN Random Process; max : Maximum delay spread allowed. The channel correlation functions may be given by the relation

(10 ) sets the condition for a orthogonal signal in time domain. C. Received Signal Analysis. A brief study of the received signal analysis has been presented in [1]. For convenience we are reproducing some of the major derivations under different situations 1. Signal is completely synchronized yi,k= xi,k hi,k + ni,k 2.FFT time synchronisation error yi,k= xi,k hi,k exp(j2it/TFFT) + ni,k 3. Carrier synchronisation error yi,k= xi,k hi,k sinc(f TFFT) exp(j[+2if(kT+TFFT/2)) + ni,k (13) (11) (12)

RH(f,t) = Rh(,t) exp( j2f) d


-

(5)

RH(f,t): The correlation function of the channel Transfer Function(T.F); Rh(,t) : The correlation function of the channel Impulse Response(CIR) Thus the spaced frequency correlation function of the channel is

RH(f,0) = Rh(,0) exp( j2f) d


-

(6)

The remedy to the synchronization error is dealt to some extent by the Schimdl and Cox Algorithm (SCA)[6]. III. ADAPTIVE FEATURE OF THE OFDM SYSTEM

SNsc (p)=Transmitted Signal

W-1(NSC) IFFT

H(f,t) h(,t) AWGN Learning Algorithm

W(NSC) FFT

The basis of our model depends mostly on our interpretation of the autocorrelation function in frequency domain (as presented in section II). From (8) we can assume a stationary signal over a time period. Thus pre-equalization over an instant explores the constant channel response over an interval. The basic block diagram of our system is provided in Fig.1.In our model we have manipulated the orthogonality of the sub carriers (SC) and have proposed towards a POFDM system. This in turn will introduce the Inter-Carrier Interference (ICI) among the SCs in the FD but will provide a robust system against multipath affect (selective fading). It can be shown that the overall system performance is improved. The selection of a proper performance metric is necessary to display the situation. We chose Euclidean Distance (ED) to be a proper performance metric for the analysis of the system performance. The neural network architecture for the system is presented in FIG.2. In this section we have provided a brief analysis of the proposed block diagram. Passing the data through the IFFT weights generates the OFDM signal. This signal is then allowed to pass through the channel function. A FFT block then decodes the corrupted signal. This decoded output is then checked with the transmitted signal, which provides our ED. The ED is used as the performance metric and is applied over a proper learning rule, which in turn is used to provide proper shift in the IFFT weights. Thus we shifted the orthogonal points of the signal. This shift in the orthogonal points results in ICI of the signals; but the overall system performance improves. Thus if we apply a proper learning algorithm and learning rate, depending upon the no of iterations, the system performance can be improved. This total number of iterations also determines the degree of Non-orthogonality of the symbols. The extra term resulting due to the destruction of the orthogonal property of the symbol under synchronization may be given by (12). However synchronization errors that may come up are equivalently given as in Section I along with the extra phase deviation of each SCs due to the shift in the orthogonal points. The deviation of the orthogonal points if limited to a certain threshold can provide an optimum performance in terms of the overall system performance measured in terms of the ED performance metric. The detailed mathematical analysis of the total deviation has been provided below. A. Analysis of the proposed system and the total shift in weights for a generic learning algorithm. From FIG.1 we observe that at any iteration point (p) the transmitted forward signal is given by SNsc (p) = XNsc T(p) * WNsc (p) ; (14)

FIG. 1 The generalized block diagram


A1(p)exp(j1(p)) w11j(p) xk(p)

x1(p) w1j(p) x2(p) j(p) . wij(p) xi(p) vj(p) sj wnj(p) xn(p)

k(p) j(p) vk(p) dk(p) ek(p)

Aj(p)exp(jj(p))

FIG. 2 The Neural Network architecture TABLE 1 Symbol meanings used in FIG 2 Symbols xi(p) wij(p) vj(p) j(p) sj(p) j(p) dk(p) ek(p) A1(p) 1(p) Meanings ith component of the input signal at pth iteration Weight from ith to the jth neuron at pth iteration Local induced field of the jth neuron Activation field of the jth neuron Output symbol from the jth neuron Corrupted symbol of the jth neuron Decoded symbol of the jth neuron Deviation of the decoded symbol from the desired Attenuation factor of the jth path Phase deviation of the jth path

XNsc T(p)=Input data signal presented in the IFFT block input. WNsc (p) =Weight presented by the IFFT block. Now, the corrupted signal due to the multiplicative noise can be given by Total paths Nsc (p) = {Aj (p) * exp (jj (p)) * XNsc (p) * WNsc (p)} (15) j=1 Thus the decoded signal can be given as DNsc(p)= Nsc (p)* WNsc 1(p). where, WNsc1 (p) =Weight presented by the FFT block. (16)

If however we apply a learning function f (.) along the signal backward propagation, we get the decoded output of the signal after the feedback has been applied ; DNsc(p)= {[XNsc T(p)*f(WNsc (p)) ]*H(p)}*WNsc 1(p). (17) Thus by a proper selection of the learning function we can achieve the ED of the transmitted signal X Nsc T(p) and the received signal after the learning has been applied for k iterations DNsc(p+k) within a desirable threshold. On applying the learning function on the weights fp(w) the weights deviate from orthogonality and becomes non-orthogonal in nature. This will result in ICI. If we allow this deviation to such so that the overall effect of both ICI and channel response is optimum we will get a better system performance. For a better representation we construct the weight space, which will be of NFFT (Nsc) dimension FIG 3a. In the weight space the ED between the initial weight (w in) and the final desired weight (wfin) gives the radius of the contour over which the learning function is applied. The wfin is taken as the center of the circle over which f(w) is taken to be analytic. From this we can break the learning function using Taylors theorem as f(w)=f(wfin)+ (w-wfin) f1(wfin)+ (w-wfin)2 f2(wfin)+.. +(w-wfin)n fn(wfin)++Rn f (wfin)+ (w-wfin) f1(wfin) , Neglecting the higher terms w= f(w)f(wfin)/ f1(wfin) For w=win wtot= f (win)f (wfin)/ f1(wfin) (18)

do obtain the ED between the transmitted signal and the decoded signal.

Desired Weight wfin

w wwww

Initial weight win

Distance between the weights wtot

FIG 3a. Weight space of NFFT dimension

Weight at zeroth iteration Weight after first Iteration ww Weight after kth iteration FIG 3b. Change in weights at different steps Distance between final weights.

(19)

This gives the total deviation of the weight from orthogonality. If however the desired weight is obtained after spanning through k iterations (as shown in FIG 3b) then the total deviation is given by the cumulative deviation at each iteration. wtot= f (win)f(w1)/ f1(w1) + f(w1)f(w2)/ f1(w2)+.. + f(wk-1)f(wk)/ f1(wk) (20)

This provides if the learning function is analytic over the contour. Otherwise the Laurent series may be applied which will give a weight much nearer to the desired weight. From (20) we see that the total deviation from orthogonality and the number of iterations required to obtained wfin depends upon the learning function. Thus the search for a proper learning function and step size is much desired for. A generalized algorithm for a better system performance is Step1. Initialize the weights as W=IFFT, W-1=FFT Step2. Pass the signal through Weight (W) block. Step3. Pass the signal through the channel. Step4. Pass the corrupted signal through (W-1 ) block. Step5. While stopping criteria not obtained

using the ED as a parameter, apply the learning function on the weight block W. Repeat steps 2 to 4 Step 6.Finally transmit the data signals through the final weight block. IV.SIMULATION RESULTS The simulation parameters were taken similar to that of IEEE 802.11a. NFFT=64,NGI=16. The neural architecture was designed keeping no bias level. The learning algorithm taken into consideration was the delta-learning algorithm. The details of the POFDM system parameters and the neural network architecture parameters have been provided in TABLE 2 and TABLE 3 respectively. All the simulations have been done in

MATLAB. We have utilized the simulation block diagram provided by [1] and modified the blocks to our system architecture. We deviated from our generalized algorithm by using a variable learning rate so that the learning does not oscillate in between the local minimas. Initial simulation of the model using a static learning rate over iterations proved that the ED between the transmitted and decoded signal due to ICI TABLE 2 Pseudo-OFDM symbol parameters Parameter Description Modulation technique Number of SC Number of IFFT/FFT points Symbol period Carrier period to elementary period ratio Parameter Value QPSK 52 64 224e-6 1/10

undesirable. However, by keeping a variable learning rate the step size of the weight change could be limited and hence provided an optimum response and decreasing ED. The waveform of the transmitted signal and that of the PseudoOFDM after first iteration has been presented in FIG 4a and FIG 4b respectively. The simulated output of the ED at various iterations has been presented in FIG 5a. It can be clearly seen that the ED is maximum for the zeroth iteration and limited for

Euclidean Distance 2.00E+05 Distance 1.50E+05 1.00E+05 5.00E+04 0.00E+00 1 8 15 22 29 36 43 50 57 64 Sam ples 0th_iteration 1st_iteration 2nd_iteration 3rd_iteration

TABLE 3 Neural network architecture parameters Parameter Description Parameter Value Neural Architecture FIG 2 Performance Metric ED Learning Rule Similar to delta rule Learning Rate (initial=0.5) Variable over iteration

FIG 5a.Comparision of the ED at different iterations.


Euclidean Distance
2.50E+04 2.00E+04 Distance 1.50E+04 1.00E+04 5.00E+03 0.00E+00 1 8 15 22 29 36 43 50 57 64 Samples 1st_iteration 2nd_iteration 3rd_iteration

FIG 4a.The orthogonal signal transmitted

FIG 5b.Comparision of the ED at first, second and third iterations. the first, second and third iterations. To provide a better view of the ED at the first, second and third iterations the FIG 5b has been provided. It can clearly be seen that the ED at the first iteration is almost comparable to that of the second and the third iterations. This result may be attributed to the fact that the overall nullification of the channel affect by the change in the weights is annulled by the ICI affect due to the change in orthogonality. Thus from the simulated results we find the system performance to be optimum after the first iteration .The ED at the first iteration is shown by the blue solid line.

FIG 4b.The Pseudo-Orthogonal signal after first iteration. and the channel response was mostly dominated by the ICI so a further change in orthogonality after the first iteration was

IV. CONCLUSION In this paper we compared the conventional OFDM system against the proposed Pseudo OFDM system. The POFDM system proved to be better than the conventional OFDM system as long as the number of iterations and the learning algorithm were properly chosen. The POFDM system proved almost 10 times better than the conventional OFDM system on the ED performance metric. Thus we propose that a slight shift from the orthogonal point may increase the ICI but may prove better in the overall system performance. The need for a better learning algorithm and a proper selection of the learning rate is always desired, as we have seen in (20).

REFERENCES [1] Ramjee Prasad, OFDM for Wireless Communication Systems. Artech House Publishers, 2004. [2] J.G.Proakis ,Digital Communications, 3rd Edition.McGraw Hill ,1995 [3] Shinsuke Hara, Ramjee Prasad, MULTICARRIER TECHNIQUES FOR 4G MOBILE COMMUNICATIONS. Artech House Publishers, 2003. [4] L.Hanso, W. Webb & T.Keller, Single and Multi-carrier Quadrature Amplitude ModulationPrinciples and Applications for Personal Communications, WLANs and Broadcasting, England :John Wiley & sons,2000 [5] L .Hanzo, M.Mnster, B.J.Choi, T.Keller, OFDM and MC CDMA for Broadband MultiUser Communications, WLANs and Broadcasting,England: John Wiley & sons,2003 [6] T.M Schmidl and D.Cox, Robust frequency and timing synchronization for OFDM, IEEE Transactions on Communications, vol 45,pp. 1613-1621, December 1997.

S-ar putea să vă placă și