ICI Mitigation for Mobile OFDM with Application to DVB-H
Outline Background and Motivation Coherent Mobile OFDM Detection DVB-H System Description Hybrid Frequency/Time-Domain Channel Estimation Conclusions
OFDM Principle Orthogonal Frequency Division Multiplexing Method to multiplex many narrowband signals into an aggregate wideband data stream All sub-channels are orthogonal to each other with a spacing of 1/T sym
OFDM over Doubly-Selective Channels Frequency Selectivity Due to long multipath delay spread Causes inter-block interference (IBI) and inter-carrier interference (ICI) if cyclic prefix (CP) length is shorter than channel memory Time Selectivity Due to mobility Coherent time inversely proportional to Doppler frequency Causes ICI
Banded Frequency-Domain Channel Matrix banded structure For time-invariant channel, the frequency domain channel matrix is diagonal For time-varying channel, it is not diagonal Frequency-domain channel matrix approximately banded
Frequency-Domain Equalizer (FEQ) Conventional 1-tap equalizer is simple: inverse of G matrix diagonal D+1 4D+1 D D 1-tap FEQ is suboptimal in presence of Doppler G m 2D+1 Q-tap (Q=2D+1) MMSE FEQ H H 2 1 w = g ( G G + σ I ) m where m g m m m is the middle column of Q G m m Selecting Q entails a performancecomplexity tradeoff Frequency-domain channel matrix G
Frequency-Domain ICI Mitigation We consider a frequency-domain equalizer (FEQ) Based on minimum mean square error (MMSE) An FIR filter with few taps per subchannel (less than 5 taps) Detect sub-carriers one-by-one Reduced complexity
Frequency-Domain ICI Mitigation FIR MMSE-FEQ combined with Ordered Successive Interference Cancellation (OSIC) Detected sub-channel symbols are fed-back and subtracted from received vector detect the best sub-channel first Process repeated until all sub-channels are detected Reduces error floor of FIR MMSE-FEQ
FIR MMSE-FEQ-OSIC : 10% Doppler size-64 FFT with perfect receiver CSI
Time-domain Channel Estimation Mark a small number of rows (M) of the timedomain channel matrix to estimate Insert P pilot tones in equi-spaced groups P ML c (L c is # of channel taps) Least-squares estimate for M marker rows Linear (or Wiener) interpolation between marker rows to estimate the entire timedomain channel matrix
Iterative Joint Channel and Data Estimation Use the estimated channel matrix to detect the OFDM symbol as before The whole estimated OFDM symbol is used as pilot symbol to obtain a refined channel estimate Since the second iteration has more pilots, channel estimation is more reliable Improves performance significantly
Channel Estimation : N=64, 10% Doppler Even 2 iterations improve performance significantly
Application to DVB-H Case study : Digital Video Broadcastinghandheld (DVB-H) FFT size can be 2048,4096, or 8192 Large FFT size needed to handle long channel delay spread (micro-seconds) Designed to operate under mobile environment Eliminates need for TEQ (CP overhead is negligible) but ICI cancellation is critical
DVB-H System Diagram
DVB-H Simulation Parameters 8 MHz channel bandwidth Baseband simulation QPSK modulation Rate ½ convolutional coding with bit interleaving 1 pilot every 12 subcarriers Reed-Solomon coding with byte interleaving (optional)
DVB-H Simulation Environment Channel model COST 207, TU6 channel Approx. 50-tap channel after sampling Guard interval longer than the channel memory Doppler spectrum based on Jake s model Normalized (w.r.t. subcarrier spacing) Doppler Up to 3% Doppler for 2k mode (110 Km/h at 800 MHz carrier frequency) Up to 10% Doppler for 8k mode (100 Km/h at 800MHz carrier frequency)
DVB Channel Estimation Challenges Difficult to estimate full channel matrix reliably in the time domain Too long CIR and too many parameters One-tap frequency-domain channel estimation Suboptimal at high Doppler Difficult to estimate 3-tap channel reliably off-diagonal elements of G have relatively low energy Severe performance degradation at High Doppler (8k mode, 10% Doppler)
Piecewise Linear Model A good approximation for channel time-variations
Hybrid F/T Domain Estimation 1. Ignoring ICI, obtain an estimate of diagonal of G. 2. Transform to time-domain by IFFT, and get an estimate of middle row of time-domain channel matrix
Hybrid F/T Domain Estimation 3. Repeat step 1 and 2 for three consecutive blocks, obtaining three middle rows 4. Using 3 middle rows, calculate the slope for each channel tap based on linear model 5. Using the middle point and the slope for each channel tap, estimate the linearly time-varying channel 6. Convert the time-varying channel model to frequencydomain by FFT, and obtain a full matrix estimate of G
Hybrid Channel Estimation (8k mode, 10% Doppler) 3-tap FEQ from estimated channel eliminates the error floor
Sparse Channel Impulse Response An example of TU06 channel model
Channel Estimation Complexity Reduction Time-domain (sparse structure) Many channel taps can be zeroed out Choose a number of K channel taps with the largest absolute value Estimate K instead of L slopes (K<<L) Frequency-domain (banded structure) Calculate 3 diagonals of G instead of whole matrix Symmetry between super and sub-diagonals for large N
Reduced-complexity channel estimation results (8k mode, 10% Doppler) Number of taps entails a complexity/performance tradeoff
Hybrid Channel Estimation Summary Hybrid frequency/time domain channel estimation algorithm suppresses ICI effectively at practical complexity Further complexity reductions Sparse nature of TD CIR Banded structure of FD channel matrix FIR-MMSE FEQ coefficients are computed based on 3 or 5 estimated diagonals