Chapter 2 Overview of WiMAX System
2.8 Frequency and Timing Requirements
Timing Requirements
For any duplexing, all SSs shall acquire and adjust their timing such that all uplink OFDMA symbols arrive time coincident at the BS to an accuracy of ±25% of the minimum guard interval or better. For example, this translates into ±8 samples in the case of 1024-FFT OFDMA.
Frequency Requirements
At the BS, the transmitted center frequency, receive center frequency, and the symbol clock frequency shall be derived from the same reference oscillator. At the BS, the reference frequency accuracy shall be better than ± ×2 10−6.
At the SS, both the transmitted center frequency and the sampling frequency shall be derived from the same reference oscillator. Thereby, the SS uplink transmission shall be locked to the
BS, so that its center frequency shall deviate no more than 2% of the subcarrier spacing, compared to the BS center frequency.
During the synchronization period, the SS shall acquire frequency synchronization within the specified tolerance before attempting any uplink transmission. During normal operation, the SS shall track the frequency changes by estimating the downlink frequency offset and shall defer any transmission if synchronization is lost. To determine the transmit frequency, the SS shall accumulate the frequency offset corrections transmitted by the BS (for example in the RNG-RSP message), and may add to the accumulated offset an estimated UL frequency offset based on the downlink signal.
Chapter 3 Decision Directed Channel Estimation
This section describes some channel estimation schemes to deal with different environments. After finding the packet starting point, channel estimation is performed to recover the channel frequency response. In this Chapter, the channel estimation will be introduced in detail. Then, we give a brief overview of the relevant technique.
3.1 Channel Estimation Structure
The simple structure of the channel estimation is show as Fig. 3-1. The received signal is modeled as Y m
( )
=H m X m( ) ( )
+N m( )
, Y is the received signal, X is QPSK i.i.d.transmitted signal, H is with zero mean complex Gaussian distribution, N is white Gaussian noise, and m is subcarrier index.
Rx_signal
Initial Channel
Iteration-Decision directed
Final result ˆH H Y
Fig. 3-1 Channel estimation structure
There are two operations in the structure. One is the initial channel response estimation, the other is the iteration steps to refine the channel response until convergence.
3.2 Obtain The Initial Channel
We have some different choices to obtain the initial channel responses, namely, Pilot Symbol Aided, Linear Prediction, and Linear Interpolation.
Pilot Symbol Aided [17][18]
We add pilots XNp to a symbol block. Denote N as the pilot numbers, and p reasonably assume pilots XNp are known. We can obtain the initial channel HNp, then put the initial channel HNp into a low pass filter. The output of the low pass filter is the final result ˆH .
Linear Prediction [15][16]
Suppose the last p s symbol channels are known perfectly, we predict the ' p+ 1 channel from the previous 1 ~ p channels.
Linear Interpolation
We need a buffer to save a block of data. We use Pilot Symbol Aided in the first and last symbol to obtain channel, and the initial channels of the other symbols are obtained by the Linear Interpolation.
3.2.1 Pilot Symbol Aided
Channel estimators usually need some kind of pilot information as a point of reference. A
pilot
Fig. 3-2 Pilot Symbol Aided
fading channel requires constant tracking, so pilot information has to be transmitted more or less continuously. Decision-directed channel estimation can also be used. But even in these types of schemes, pilot information has to be transmitted regularly to mitigate error propagation. The structure of the Pilot Symbol Aided is show as the Fig. 3-2. The channel of every symbol was estimated by pilot. The initial channel estimation of every symbol was shown in the follow description,
( ) ( ) ( )
The illustration is shown in Fig. 3-3
( )
0Fig. 3-3 Initial channel HNp
W is the DFT matrix. N is the pilot numbers. The sub-carrier index p m is in the pilot
The complete channel estimation ˆH of all the sub-carriers is computed from hL+1 by
( ) 1
The operations of the equations are the same as a low pass filter [23]. The equation is shown as Fig. 3-4 Low pass filter
IFFT
The parameter N has large correlation with the length of the channel impulse response p
L and cyclic prefix length N . Now, we discuss the relationship of the pilot number cp N , p the channel length L , and the cyclic prefix length N .If the real channel impulse response cp and channel frequency response are shown as Fig. 3-5 and Fig. 3-6
0 1 2 L M −1
Fig. 3-5 Real channel impulse response
0 1 2 m M −1
Fig. 3-6 Real channel frequency response
0 1 2 L M −1
Fig. 3-7 Initial channel impulse responsehNp ( )0
Fig. 3-8 Initial channel frequency responseHNp
In OFDM system, we assume the channel impulse responses are time invariant during a symbol time, and the cyclic prefix length N is always larger than channel impulse cp
response length L . Because we want to reduce the inter-symbol interference (ISI) perfectly, so, the cyclic prefix length N is larger than the channel length L always. Than we just cp discuss the relationship between N and cp N during a symbol time. p
Ncp ≤Np
According to the sampling theory, the relationship between channel impulse response and channel frequency response can be shown in Fig. 3-7 and Fig. 3-8. According to the figures, we know that the initial channel HNp through to the low pass filter still close to the
real channel impulse response.
Np <Ncp
Similarly, the relationship between channel impulse response and channel frequency response are shown as Fig. 3-9 and Fig. 3-10.
We know that, if sampling rate is smaller than Nyquist frequency in frequency domain, the aliasing will happen, in the time domain.
According to Fig. 3-9 and Fig. 3-10, we know that the initial channel HNp through to the low pass filter not close to the real channel. Because there will be aliasing on impulse
0 1 2 L M−1
0 n
( )
0h
( )
h L
0 0 0
Fig. 3-9 Initial channel impulse responsehNp
( )
0H H m
( )
m
0 1 2 m M −1
0 0 0 0 0 0 0 0 0
Fig. 3-10 Initial channel frequency responseHNp
0 1 2 L M−1
Fig. 3-11 Final resource of the channel impulse response.
response. According to the discussion, the parameter pilot number Np must be large than
N or equal to cp N . cp
3.2.2 Linear Prediction
In this section, we introduce autoregressive (AR) model. When the p th-order AR (AR
( )
p ) hypermodel for the frequency-selective fading channel is employed, we obtain a linear dynamic system which can be described by the following state-space model, the structure of the Pilot Symbol Aided is show as the Fig. 3-12, , , ,
pilot
Fig. 3-12 Linear Prediction
is the Gaussian process noise, and it is assumed to be circular and independent of the receiver noise n. Because multi-path are independent and uncorrelated, so the weight numbers a l i, are calculated separately. We want to predict the t th symbol channel the from p th-order channel hl t, 1−,hl t, 2− ,hl t, 3− ,...,hl t p,− as
( ) ( ) ( ) ( )
The above optimality equations are called the Wiener-Hopf equations for linear prediction. It can be rewritten in matrix form as
[ ] [ ] [ ]
Optimal solution is given by[ ] [ ] [ ]
Toeplitz (square) matrix Any square matrix of the form
0 1 1
Is said to be Toeplitz.
3.2.3 Linear Interpolation
If we have B symbols in an OFDM block, we use Pilot Symbol Aided in first symbol of the block, and obtain this symbol channel. Than, we need a buffer to save a block data, and the first symbol of the next block. The initial channels of the other symbols are obtained by the Linear Interpolation. The structure is shown as Fig. 3-13
The first channel of the block1 is ˆH t , the first channel of the block2 is
( )
0 ˆH t . We save( )
1Fig. 3-13 Linear Interpolation
(
0) (
0) (
0)
After the last action, the initial channel is obtained. If the OFDM system active without the iteration step, just take the initial channel as a final result, the result is not close to the real channel. If the initial channel is obtained by Pilot Symbol Aided, the final result is not very terrible. But if the linear prediction is used, the error between initial channel and real channel will be accumulating to next initial channel. So, we do some iteration to close the real channel.
We have two difference choice to use. One is Decision directed, the other is Expectation Maximization (EM) Algorithm.
3.3 Iteration- Decision Directed
The iteration processing was introduced as follows. Before doing the iteration, the input of the lowpass filter is called the initial channel estimation H , but after the iterative is used, the ( )0 output of the lowpass filter is the pth iterativel result ˆH( )p . The iteration process is shown in the Fig. 3-14. The initial channel estimation H is the input of the iteration processing, pass ( )0 by lowpass filter, some of the AWGN is reduced. The receiver signal Y and the output of the lowpass filter ˆH are used to calculate the detection signal ( )0 ˆX . The detection signal ( )0
( )0
ˆX are used to be new pilots. After p times iteration, the output of the LPF ˆH(p+1) is the final channel. The next iterative result of the detection signal is ˆX(p+1).
LPF Quantize
Fig. 3-14 Decision directed iteration process
The iteration stop until the channel estimation ˆH(p+1) is stable.
3.4 Decision Directed Channel Estimator
Base on the last discussion, we have three methods to obtain the initial channel, and two iteration methods.
Initial channel
1. Pilot Symbol Aided 2. Linear Prediction 3. Linear Interpolation
Iteration- Decision directed
We combine the methods to be six channel estimators 1. Decision directed Pilot Symbol Aided (DD Pilot) 2. Decision directed Linear Prediction (DD AR) 3. Decision directed Linear Interpolation (DD LI)
3.5 Channel System Environment
In this section, we first show the simulation environment for our study, including the system parameters and channel environments.
3.5.1 System Parameters
We have to specify the system parameters from all usable values provided by 802.16e so that the simulation environment could be constructed. The standard is very flexible in choice of bandwidth and cyclic prefix length. However, it would be difficult to conduct the simulation and implementation study without some particular sets of parameters. Hence we pick some sets of parameters shown in this section. The system profile we select is Wireless random generated binary data. The frame duration could be 5, or 10ms. Other parameter values are
Table 3-1 System Parameters Used in Our Study
Parameters Values
System Channel Bandwidth (MHz) 10 10 20
Sampling Frequency (MHz) 11.2 11.2 22.4
FFT Size 1024 2048 2048
Sub-carrier Spacing (kHz) 10.94 5.47 10.94
Useful Symbol Time (μsec) 91.4 182.8 91.4
Guard Time (μsec) 11.4 22.8 11.4
OFDM Symbol Time (μsec) 102.9 205.7 102.9
in Table 3-1
3.5.2 Channel Environments
Typical models of the wireless communication channel include additive noise and multi-path fading. For channel simulation, noise and multipath fading are described as random processes, so they can be algorithmically generated as well as mathematically analyzed.
3.5.2.1 Gaussian Noise
The simplest kind of channel is the additive white Gaussian noise (AWGN) channel, where the received signal is only subject to added noise. A major source of this noise is the thermal noise in the amplifiers which may be modeled as Gaussian with zero mean and constant variance. Incomputer simulations, random number generators may be used to generate Gaussian noise of given power to obtain a particular signal-to-noise ratio (SNR).
Table 3-2 Terrain Type vs. SUI Channels
Terrain Type SUI Channels
A: hilly terrain, heavy tree SUI-5, SUI-6 B: between A and C SUI-3, SUI-4 C: flat terrain, light tree SUI-1, SUI-2
Table 3-3 General characteristics of SUI channels
Doppler Low delay spread Moderate delay spread High delay spread
Low SUI-1, SUI-2, SUI3 SUI-5
High SUI-4 SUI-6
3.5.2.2 Slow Fading Channel
In slow fading, multipath propagation may exist, but the channel coefficients do not change significantly over a relatively long transmission period. The channel impulse response over a short time period can be modeled as
( )
1( )
parameters are time-invariant in a short enough time period.
3.5.2.3 Fast Fading Channel
With sufficiently fast motion of either the transmitter or the receiver, the coefficient of each propagation path becomes time varying. The equivalent baseband channel impulse response can then be better modeled as
( )
1( )
( )( )
Note that αl and θl are now functions of time. But τl is still time-invariant, because the path delays usually change at a much slower pace than the path coefficients. The channel coefficients are often modeled as complex independent stochastic processes. If there is no line-of-sight (LOS) path between the transmitter and the receiver, each path may be made of the superposition of many reflected paths, yielding a Rayleigh fading characteristic. A commonly used method to simulate Rayleigh fading is Jakes’ fading model, which is a deterministic method for simulating time-correlated Rayleigh fading waveforms.
Table 3-4 SUI-1 Channel Model
Relative delay ( sμ or sample number) Average power
Tap ( sμ ) (normal) (dB) (normalized)
1 0 0 0 0.9610
2 0.4 4 -15 0.0303
3 0.9 10 -20 0.0096
Table 3-5 SUI-2 Channel Model
Relative delay (μs or sample number) Average power
Tap ( sμ ) (normal) (dB) (normalized)
1 0 0 0 0.9135
2 0.4 4 -12 0.0576
3 1.1 12 -15 0.0289
Table 3-6 SUI-3 Channel Model
Relative delay ( sμ or sample number) Average power
Tap ( sμ ) (normal) (dB) (normalized)
1 0 0 0 0.7061
2 0.4 4 -5 0.2233
3 0.9 10 -10 0.0706
Table 3-7 SUI-4 Channel Model
Relative delay (μs or sample number) Average power
Tap ( sμ ) (normal) (dB) (normalized)
1 0 0 0 0.6424
2 1.5 4 -4 0.2557
3 4 16 -8 0.1018
Table 3-8 SUI-5 Channel Model
Relative delay ( sμ or sample number) Average power
Tap ( sμ ) (normal) (dB) (normalized)
1 0 0 0 0.7061
2 4 45 -5 0.2233
3 10 112 -10 0.0706
Table 3-9 SUI-6 Channel Model
Relative delay ( sμ or sample number) Average power
Tap ( sμ ) (normal) (dB) (normalized)
1 0 0 0 0.8773
2 14 156 -10 0.0877
3 20 224 -14 0.0349
Table 3-10 ETSI “Vehicular A” Channel Model in Different Units
Relative delay (μs or sample number) Average power
Tap ( sμ ) (normal) (dB) (normalized)
1 0 0 0 0.4850
2 0.31 3 -1 0.3852
3 0.71 8 -9 0.061
4 1.09 12 -10 0.0485
5 1.73 19 -15 0.0153
6 2.51 28 -20 0.0049
3.5.2.4 Power-Delay Profile Model
For simplicity in analysis and simulation, the delay τl in the above two models can be discretized to have a certain easily manageable granularity. This results in a tapped-delay-line model for the channel impulse response, where the spacing between any two taps is an integer multiple of the chosen granularity. For convenience, one may excise the initial delay and
make τ0 = . Often, it is convenient to normalize the path powers relative to the strongest 0 path. And, often, the first path has the highest average power.
The channel model used here is a modification of the Stanford University Interim (SUI) channel models propsed in [19]. These models, for a collect of scenarios, provide the parameters to model the various random phenomena involved with a simulation; of course there are many possible combinations of these parameters to obtain such channel descriptions.
A set of 6 typical channels was selected for the three most common terrain categories that are typical of the continental United States. The parametric view of the SUI channels is summarized in Table 3-2 and Table 3-3
Scenario for SUI channels
z Cell size: 7km
z Base station antenna height: 30m z Receiver antenna height: 6m
z Base station antenna beamwidth: 120
z Receiver antenna beamwidth: omnidirectional
(
360 and)
30z Vertical polarization only
We list all of the SUI channel models in Table 3-4 to Table 3-9 that will be used in our simulations below.
Another channel model chosen from one of the channel environments defined by ETSI is used in this thesis. The model is as shown in Table 3-10. This is a channel model for the vehicular test environment, which the tested speed is from 120km/h to500km/h. This
environment is characterized by larger cells and higher transmit power, and is valid for NLOS case only and describes worse case propagation. Channel A is the low delay spread case that occurs frequently. See [20] for more details.
3.6 Simulation
We show the simulation result as in Fig. 3-15~Fig. 3-20 , there are four cases to be discussing,
SUI3 for v=100km/hr, v=240km/hr
SUI5 for v=100km/hr, v=240km/hr
Vehicular A for v=100km/hr, v=240km/hr
Ts is the sampling time, and F is the Doppler frequency, where d
d c
F =vF c
F Doppler d frequency
F carrier c frequency
v velocity of mobile
c velocity of light
This simulation are base on SUI-3 channel, SUI-5 channel and Vehicular A. The subcarrier number M is 1024, cyclic prefix length Ncp =M 8 . Carrier frequency
c 3.5
F = GHz. Bandwidth BW =11.2 MHz, block length B= symbols in Linear 4 Interpolation. When Linear Interpolation is used, block length B should not too large, the symbol channels of this block are still linear, so the Linear Interpolation can work. If B is too large, the channels are not linear, so the Linear Interpolation can not work effective.
3.6.1 SUI3
Decision directed in 240 km/hr
Ideal Channel DD Pilot DD AR DD LI
Fig. 3-15 SER performance for DD in 240 km/hr
10 15 20 25 30 35 40
Decision directed in 100 km/hr
Ideal Channel DD Pilot DD AR DD LI
Fig. 3-16 SER performance for DD in 100 km/hr
3.6.2 SUI5
Decision directed in 240 km/hr
SNR
Fig. 3-17 SER performance for DD in 240 km/hr
10 15 20 25 30 35 40
Decision directed in 100 km/hr
SNR
Fig. 3-18 SER performance for DD in 100 km/hr
3.6.3 Vehicular A
Decision directed in 240 km/hr
SNR
Fig. 3-19 SER performance for DD in 240 km/hr
10 15 20 25 30 35 40
Decision directed in 100 km/hr
SNR
Fig. 3-20 SER performance for DD in 100 km/hr
Accordimg to the simulations, we can see that the Decision directed-AR estiamtor can not work in high speed for SUI-3 and SUI-5. The Decision directed-Pilot Symbol Aided has the best performance of the symbol error rate (SER), but it has the highest pilot number. The Decision directed-Linear Interpolation can work in each environments, but it need to use one block data to estimate channel.
Chapter 4
Expectation Maximization Channel Estimation
4.1 Introduction of Expectation Maximization (EM) [16]
The EM algorithm [28, [29] is a technique for finding maximum likelihood (ML) estimates of system parameters in a broad range of problems where observed data are incomplete. The EM algorithm consists of two iterative steps: the expectation (E) step and the maximization (M) step. The E-step is performed with respect to the unknown underlying parameters, using current estimates of the parameters, conditioned upon the incomplete observations.
TheM-step then provides new estimates of the parameters that maximize the expectation of the log-likelihood function defined over complete data, conditioned on the most recent observation and the last estimate. These two steps are iterated until the estimated values converge.
The EM algorithm [28], [29] is an iterative method to find the ML estimates of parameters in the presence of unobserved data. The idea behind the algorithm is to augment the observed data with latent data, which can be eithermissing data or parameter values, so that the likelihood function conditioned on the data and the latent data has a form that is easy to manipulate. The algorithm can be broken down into two steps: the E-step and the M-step. We assume that the data Z (“complete” data) can be separated into two components,
(
,)
Z = X Y , where Y are the observed data (“incomplete” data) and X are the missing data. We denote θ as the unknown parameter we try to estimate from X .
The E-step finds Q
(
θ θ| ( )p)
, the expected value of the log-likelihood of θ , logf Z(
|θ ,)
where the expectation is taken with respect to X conditioned on Y and the latest estimate of θ , θ( )p
(
( )) { ( )
( )}
Q θ θ| p =E log f Z|θ | ,Y θ p .
The M-step then finds θ(p+1), the value of θ that maximizes Q
(
θ θ| ( )p)
over allpossible values of θ :
(p 1) arg max Qθ
(
| ( )p)
θ + = θ θ
This procedure is repeated until the sequence θ , ( )0 θ , ( )1 θ ,…converges. The EM ( )2 algorithm is constructed in such a way that the sequence of θ ’s converges to the ML ( )p estimate of θ .
Under our system model:
( ) ( ) ( ) ( )
Y m =H m X m +N m
Z is “unobserved data” or “complete data”.
Y is “observed data” or “incomplete data”.
H is unknown parameter we try to estimates.
X is “latent data” or “missing data”.
4.2 Iteration- Expectation Maximization (EM)
Fig. 4-1 EM Algorithm iteration process
Due to the Gaussian noise assumption, the probability density function of Y given X and H is given by
By assuming that all C symbols are equally likely and averaging the condition pdf over the variable X , we obtain the pdf of Y given H as follows
We assume that the frequency-domain signal X of a given sub-carrier represents a QPSK signal with constellation size C (=4).
(2 1)
consists of the following two steps E-step
Before doing Expectation Maximization (EM) Algorithm, the initial channel estimation H ( )0 is the input of the EM step, the output of the EM Algorithm H is shown as ( )1
( )1
( ) ( )
4 ( )0( )
The final result of the detection signal is ˆX(p+1). The iteration stop until the channel estimation ˆH(p+1) is stable. The final result ˆH(p+1) is shown as
4.3 Estimate Noise Power
Base on the Expectation Maximization (EM) Algorithm, we know that the noise power
2
σn is necessary. We can obtain the parameter easily. When we received one symbol signal, the receiver signal are shown as
( ) ( ) ( ) ( )
Y m =H m X m +N m (4.9)
We can obtain the noise power as
4.4 Expectation Maximization Channel Estimator
Base on the last discussion, we have three methods to obtain the initial channel, and iteration method.
Initial channel
4. Pilot Symbol Aided 5. Linear Prediction 6. Linear Interpolation
4. Pilot Symbol Aided 5. Linear Prediction 6. Linear Interpolation