Professor Paulraj and Bringing MIMO to Practice

Transcription

1 Professor Paulraj and Bringing MIMO to Practice Michael P. Fitz UnWiReD Laboratory-UCLA April 21, 24 UnWiReD Lab

2 A Little Reminiscence PhD in 1989 First research area after PhD was equalization Had to prove that you were better than Paulraj/Stanford Second research area was space-time modems Had to prove you were better than Paulraj/Stanford For me the standard against which to measure yourself has always been the Paulraj/Stanford group I am afraid I will always come out on the short end! Professor Paulraj has keep his group focussed on practical problems and has championed testbeds Slide 2

3 UnWiReD Laboratory Philosophy Slide 3

4 What Am I Working on These Days? Space-time coding Channel estimation and synchronization MIMO OFDM channel estimation Applicable techniques for WLAN Feedback communications Wireless channel modeling Slide 4

5 Methods to Estimate the Channel Transmitted reference Use transmitted reference to estimate channel Simple Bandwidth inefficient Joint demodulation and estimation Model and jointly estimate data and channel Bandwidth efficient but complex Need to resolve signal set ambiguities Noncoherent Model and average out the channel in forming the data demodulation architecture Performance and complexity depends on channel model Slide 5

6 Goals in Channel Estimation The channel estimates should have an SNR that is bigger than the observation SNR Traditional differential decoding has estimate SNR=Eb/No (3dB degradation) An upper bound on acceptable performance for small constellations Rules of thumb (less than 1 db degradation) Estimate SNR>5 Eb/No Error floor can occur if channel estimate SNR does not grow with observation SNR Slide 6

7 Transmitted References and Channel Estimation Preambles Pilot tones Mid-amble (GSM) Pilot subcarriers (OFDM) Pilot codes (IS-95) Pilot symbols Reference only needs to be orthogonal to data symbols Slide 7

8 A Transmitted Reference Leads to Simple Estimates If modulation is known then the only unknowns are Channel distortion Noise Both are well modeled as Gaussian processes Optimally estimating Gaussian processes (the channel) from Gaussian observations is a well studied problem Linear filter is optimal (Wiener filter, Kalman filter) Slide 8

9 Channel Estimation Filters Typically not all parameters of the processes (correlations) are known hence suboptimal or adaptive filters are implemented in practice Observed Reference Signals Linear Filter Channel Estimates Slide 9

10 Channel Models Y z Propagation produces channels selective in Space Frequency Time ( ) m p ( t, A) = H n exp j 2πΔ cos β α n + j2π cos( χ α λ n )f D t x( t τ n ) + W z t n= 1 ( ) Slide 1

11 Time Domain Multipath will produce a channel well characterized by a sum of sinusoids model m p n=1 H exp[ j2π cos( χ α )f t]x( t τ ) n n D n Fading in time is a function of Carrier frequency Speed of nodes or environment Angle spread at TX and RX Slide 11

12 Pilot Symbols and Nyquist Np P D D D D D D P D D P Finite number of degrees of freedom in time Slide 12

13 Frequency Domain Characteristics Multipath produces frequency selectivity Fading in frequency is a function of Delay spread m p n=1 H x( t τ ) n n Slide 13

14 Time Frequency Duality Time Response Frequency Response Delay spread has equivalence to Doppler spread Finite number of degrees of freedom in frequency domain Slide 14

15 Optimal Nyquist Sampling Rate of insertion is a function of Doppler spread (time) Delay spread (frequency) Should be periodic so that interpolation filters can be reused Edge effects change the optimal pattern Slide 15

16 Space Domain Multipath will produce a channel well characterized by a sum of sinusoids model R A 2 m p n=1 ( ) = H exp j 2πΔ n Fading in space is a function of Carrier frequency Separation of antennas Angle spread at TX and RX Note time and space is a function of the same parameters Time variations are due to motion through a spatial standing wave λ cos ( β α ) n Slide 16

17 Spatial Standing Wave Distance in wavelengths Distance in wavelengths Slide 17

18 Space-Time Correlation There can be unexpected space-time correlations due to movement through the spatial standing wave Slide 18

19 What is different about MIMO channel estimation? Each received signal is a mixture of what was sent on each of the transmit antennas Y j k t ( ) = H ji i=1 X( k) + W( k) For instance in a 2x2 system you get two observation sequences to estimate 4 channel characteristics There is an observable issue that has to be dealt with by appropriate preamble design Under normal operational characteristics the optimal sequences for channel estimation are orthogonal sequences (Guey:96) Slide 19

20 Channel Characteristics Conclusions A sum of sinusoids is an accurate model for variations in Time Space Frequency Time and space can be coupled Frequency seems to be independent of the others This finite parameterization should be exploited in channel estimation in MIMO radio Model based estimation Slide 2

21 Wireless LAN Traffic Data Messages Typically are long 15 bytes is nominal Protection capacity approaching codes (e.g., LDPC) Service Messages ACK, NACK, RTS, CTS 14-2 bytes in length Protected by short block length low rate codes (Space-time codes) Slide 21

22 Channel Estimation Paradigms Target system for the results presented in this talk is 4x4 Preamble based estimation Long packet based estimation Short packet based estimation Slide 22

23 Preamble Based Estimation The participants tend to favor preamble based channel estimation No delay needed in decoding Simple structures Preamble processing is often done in simple ways Each subcarrier estimated in an independent fashion Significant performance improvements can be achieved by understanding the channel characteristics Smoothing can be achieved in time and frequency Slide 23

24 Preamble Design Constraints Some contraints Channel F (large delay spread) produces rapidly varying channel Channel F requires a frequency domain spacing of at least 1 in 2 in 82.11a subcarriers Hard to get antenna orthogonality in frequency domain (ICI) Looked at preambles of length 1, 2, and 4 symbols for 4 transmit antennas Two issues - SNR and ICI Slide 24

25 Four Symbol Preamble x4 Super Orthogonal, 64 QAM, 16 Cosets, Rate 5.5 BPCU Preamble based PSAM (Length = 4) designed for channel model F PCSI, Channel Model D PSAM, Channel model D PCSI, Channel Model E PSAM, Channel model E PCSI, Channel Model F PSAM, Channel model F R E B Eb/N, db Slide 25

26 Two Symbol Preamble x4 Super Orthogonal, 64 QAM, 16 Cosets, Rate 5.5 BPCU Preamble based PSAM (Length = 2) with Modified PS Power designed for channel model F PCSI, Channel Model D PSAM, Channel model D PCSI, Channel Model E PSAM, Channel model E PCSI, Channel Model F PSAM, Channel model F R E B Eb/N, db Slide 26

27 Powerful Codes (with R. Wesel) Slide 27

28 Example Data Frame Design PLCP Preamble Signal Payload 82.11a 82.11a/82.11n Time, OFDM symbols Slide 28

29 Characteristics of Design Subsamples channel Same efficiency as 82.11a but estimates 16 times more channels Can track and interpolate Frequency selective fading Time selective fading Frequency offset and phase noise Block orthogonal pilot codes Can allow simpler implementations Reuse across space and reduce number of taps Slide 29

30 Channel Estimation Performance Slide 3

31 Some Insights into Pilot Based Channel Estimation Channel estimation techniques require robustness Straight Wiener filter solutions do not work directly Error floors can occur due to variations in the channel impulse response Error floors can occur due to residual frequency offset or phase noise Robustness can be achieved with appropriate selection of interpolation coefficients At this point we do not have a good theory but have found solutions that work Slide 31

32 Short Packets Are Important in 82.11n Short packets are needed for RTS and CTS TCP/IP ACKs 14-2 bytes in length Short packets should have a different structure than data packets LDPC optimal only for long blocks Channel estimation will take a larger percentage of the bandwidth with short packets Transmission should be reliable at a wide variety of channel conditions (1 bpcu -- 2 bpcu) Slide 32

33 Short Packet Design (4x4) 4 OFDM symbols for both training and data Slide 33

34 Short Packet Opt I integrated with STBC-MTCM Slide 34

35 Preamble Insights So Far Do not need a preamble of length equal to the number of transmit antenna Urban legend The performance of shorter preambles will be worse due to loss of orthogonality Additional Orthogonality is achieved in the frequency domain and there is significant variations in the frequency domain to cause ICI Frequency domain interleaving of data is possible Loss in SNR compared to preamble based scheme only due to frequency domain interpolation Slide 35

36 Pilot Placement and Pilot Processing Pilot insertion characteristics determined by worst case channels Worst case Doppler and delay spread determine insertion rates in OFDM Adaptive insertion rates would be possible with feedback Pilot processing can be adaptive to the current channel conditions Estimates of the current delay or Doppler spread could allow more accurate channel estimation Goal: Use the preamble to simultaneously estimate the channel characteristics and estimate the channel more accurately Slide 36

37 Adaptive Channel Estimation Performance can be improved by adaptation Performance can be significantly degraded by channel estimation mismatch 1 2x2 Alamouti, 64 QAM, Channel B 1-1 PCSI PSAM (B) PSAM (C) PSAM (D) PSAM (E) PSAM (F) PSAM (ML) 1 2x2 Alamouti, 64 QAM, Channel E 1-1 PCSI PSAM (E) PSAM (F) R E B 1-2 R E B Eb/N, db Eb/N, db Slide 37

38 Approach to Adaptive Channel Estimation Postulate that one of N possible channel conditions exists Adaptive estimation becomes a hypothesis test Processing can become very simple In MIMO OFDM the major parameter of importance is delay spread Need only to estimate what interpolator bandwidth is appropriate Slide 38

39 Example Results Used 82.11n channels for the possible hypothesis Design SNR=Actual SNR = 5dB Design SNR=Actual SNR = 1dB Detected Channel Detected Channel A B C D E F A B C D E F A A 1698 Actual Channel B C D E Actual Channel B C D E F F Slide 39

40 Do The Things Paul Has Pioneered Work in Practice Slide 4