A Practical Channel Estimation Scheme for Indoor 60GHz Massive MIMO System Arumugam Nallanathan King s College London
Performance and Efficiency of 5G Performance Requirements 0.1~1Gbps user rates Tens of Tbps/km 2 traffic Tens of Gbps peak rates <10 ms delay >500 km/h mobility >10 6 /km 2 connection Efficiency Requirements 5~15x spectrum efficiency 100x energy efficiency 100x cost efficiency
Massive MIMO for 5G High spatial resolution High spectral efficiency High energy efficiency Pilot contamination (uplink) High overhead of training (downlink) High computational complexity Channel nonreciprocity (FDD)
Millimeter wave for 5G
Millimeter wave for 5G 60 GHz propagation limitations High path loss Poor penetration Absorption of the radio signal by water vapor or oxygen Therefore.. The 60GHz spectrum: proposed for short- range wireless systems. The primary propagation: limited to line-of-sight (LOS) and the first-order reflection paths.
Channel Dimension in Massive MIMO Facts (few local scattering around BS) Antenna spacing of massive antenna array as small as halfwave length BS with large-scale antenna array at the top of high buildings In millimetre wave band, the severe path loss makes sure only a few reflecting paths could arrive at BS (angular spread seen by BS is small) Therefore. Channel contains strong spatial sparsity CCM: rank deficient and possess low-rank property Methods: CCM based, CS based
This image cannot currently be displayed. This image cannot currently be displayed. Methods of Channel Estimation (1) Low-Rank Approximation Channel Model Add Title Array structure based channel model Low-rank Covariance Narrow angular spread Drawbacks High overhead for covariance High calculation complexity Low-rank Approximation H. Yin, D. Gesbert, M. Filippou, and Y. Liu, A coordinated approach to channel estimation in large-scale multiple-antenna systems, IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 264 273, Feb. 2013. A. Adhikary, J. Nam, J. Y. Ahn, and G. Caire, Joint spatial division and multiplexing the large-scale array regime, IEEE Trans. Inf. Theory, vol. 59, no. 10, pp. 6441 6463, Oct. 2013.
Methods of Channel Estimation (2) Compressive Sensing Still low-rank model Channel Model Add Title Sparse channel assumption [Rao, 2014]: : Non-zero coefficients : Zero coefficients Channel Estimation Drawbacks High complexity Non-linear optimization X. Rao and V. K. Lau, Distributed compressive CSIT estimation and feedback for FDD multi-user massive MIMO systems, IEEE Trans. Signal Process., vol. 62, no. 12, pp. 3261 3271, June. 2014.
Summary : Low-Rank Model Low-Rank Covariance & Compressive Sensing Where channel sparsity holds? BS is elevated at Low-rank high buildings. Covariance Few local scatters surrounding BS. mmwave massive MIMO Unmanned Aerial Vehicle (UAV) Communication Only way to reduce the channel dimensionality!!!
Some Thoughts about Massive MIMO Massive MIMO + Broad AS MIMO + Narrow AS Massive MIMO High spatial resolution Narrow Angular Spread
Some Thoughts about Massive MIMO Massive MIMO Narrow Angular Spread High spatial resolution Why not directly look into angle domain and explore it?
System model System and Channel Model Sparse Channel Model Estimate Estimate Estimate MUSIC, ESPRIT, too complex for massive FFT (DFT)
Properties of ULA and Channels Property 1: For the simplest 1-path case, normalized DFT of is approximately sparse. Take DFT of (a) Single incident ray with (b) Single incident ray with
Properties of UPA and Channels Take 2D DFT of An example of single path 2D-DFT sparse characteristics, where BS array has 100 80 antennas.
Properties of ULA and Channels Property 2: Considering the multi-paths case, i.e., then normalized 2D DFT of is still sparse. An example of multi-paths 2D-DFT sparse characteristics, where BS array has 100 80 antennas.
Phase Rotation (a) Single incident ray with (b) Single incident ray with ULA UPA
Proposed Method Simplified Channel Estimation Estimate Estimate Estimate Estimate: DFT point: Max Power Phase shifter parameter
Channel Estimation Strategy The Communication Protocol Uplink Preamble Uplink Training with User Grouping Angle Reciprocity Downlink Training with User Grouping
The Communication Protocol Each cell is allocated orthogonal training sequences of length. To obtain for each user 19
Uplink Preamble LTE frame length: User velocity: 1 ms 80 km/h User moves only 0.22m in 10 frames The relative position to BS is not changed. can be viewed as unchanged Spatial signatures help with subsequent training 20
Uplink Preamble Obtain Spatial Signatures (DOA) for All Users Schedule users for orthogonal training with conventional methods: Take DFT of, and slide a window of length over to find the optimal DFT point and phase shifter Obtain to prepare for subsequent channel estimation.
Uplink Training with User Grouping Uplink User Grouping Divide users into groups according to their spatial signatures, : guard interval since is much smaller than. Need orthogonal pilots Users with different colors can reuse the same pilot
Uplink Training with User Grouping Uplink Training After user scheduling, all of conventional methods can be adopted: Pilot contamination is reduced
Angle Reciprocity Clustered Response Model One-ring Model The propagation path of electromagnetic wave is reciprocal. Uplink DOA and downlink DOD are reciprocal.
Angle Reciprocity Denote and as uplink angular set, at and, respectively As long as Electromagnetic characteristics do not change in several dozens of GHz.
Reciprocity The propagation path of electromagnetic wave is reciprocal. Electromagnetic characteristics do not change in several dozens of GHz. Uplink DOA and downlink DOD are reciprocal even for FDD systems. Channel Gain Reciprocity Angle Reciprocity TDD FDD Frequency Discrepancy
Angle Reciprocity Why is angle reciprocity not important for conventional MIMO? Channels can be estimated directly in small-scale antenna systems, namely, no need for decomposition. Small-scale antenna arrays have much lower spatial resolution and the beams are too broad for user separation
Downlink Channel Model Downlink Channel Representation Denote the downlink channel from BS to user-k as Similar to uplink Remaining unknowns: 1. Spatial signature (DOD): 2. Coefficients: Similar to channel estimation algorithm in uplink
Downlink Training with User Downlink User Grouping Step 1: Grouping users with same Step 2: same as uplink grouping Grouping All these users can be scheduled simultaneously. Downlink Channel With user grouping and reduced-dimensional channels, conventional channel estimation methods can be adopted directly. Users only have to feed back 6 components to BS Significantly reduced feedback
Comparison with Existing Works Comparison of different methods All Problems Solved JSDM [Caire] [Yin, 2013] Compressive sensing Proposed FDD Reciprocity only downlink only uplink Without Covariance Complexity High-dimension EVD High-dimension EVD Non-linear optimization FFT & partial FFT Uplink Pilot Decontamination Reduced Downlink Training Overhead
RMT Roadmap of Massive MIMO Channel Estimation Eigen-space Beam space Angle space 2010 2012 2013 2015 2016 Massive MIMO Rich Scattering & Larger Spacing Asymptotically orthogonality mmwave Massive MIMO Non-reciprocal Covariance & High Complexity Low-rank Channel Covariance Compressive Sensing Non-reciprocal Support & Non-linear Optimization & Power Leakage & RIP problem Array Signal Processing Location Aware Incorrect angle information Angle Estimation Angle Reciprocity Low Complexity
Simulation Results The Simulation Parameters Performances of Uplink/Downlink Training Effects of Systems Parameters
Simulation Parameters Simulation Scenario UPA users are gathered in 5 clusters Default SNR Performance Metric respectively
Performance of UL/DL Training Comparison of uplink/downlink MSE
Performance of UL Training Uplink MSE performance comparison Proposed method LS method Power constraint for each user:
Performance of UL Training Downlink MSE performance comparison Proposed method LS method Power constraint for each user:
Performance of UL/DL Training Comparison of downlink MSE with different BS antennas
Performance of DL Sum-rate Average achievable sum rate (AASR) Proposed method LS method Total power constraint:
Comparison with Covariance Based Method To keep fair, assume that the total power and time for training are the same for all three methods.
Conclusions A new channel estimation from array processing point of view The same angular assumption as exiting methods Could handle the pilot contamination and downlink large training overhead problem Low overhead Low computational complexity Suitable for both TDD and FDD
Acknowledgement Thanks to my research collaborators D. Fan (King s College London/ Beijing Jiaotong University) F. Gao (Tsinghua University) G. Wang and Z. Zhong (Beijing Jiaotong University) This work was supported by the U.K. Engineering and Physical Sciences Research Council (EPSRC) under Grant EP/M016145/1