PaperlayerView exact revision on arXiv

Source-linked AI summary

Framework for a Perceptive Mobile Network using Joint Communication and Radar Sensing

Md. Lushanur Rahman, J. Andrew Zhang, Xiaojing Huang, Y. Jay Guo, Robert W. Heath

arXiv:1901.05558v1cs.NI

TL;DR

The paper addresses sensing-parameter estimation in mobile networks that combine communication and radar sensing amid sophisticated signals and rich multipath. It develops a unified platform, compressive-sensing formulations, and recursive background subtraction, demonstrating feasibility and efficient operation in simulations while identifying full-duplex hardware challenges for downlink sensing.

  • Problem

    Sensing-parameter estimation in perceptive mobile networks must handle sophisticated mobile signals, rich multipath, and unwanted clutter while sharing communication transmissions.

  • Method

    The paper proposes a unified platform for uplink and downlink sensing, 1D compressive-sensing estimation schemes, and recursive background subtraction for clutter suppression.

  • Results

    The proposed scheme works efficiently and demonstrates feasibility in simulations, with robust and accurate estimates when received SNR is sufficiently high.

  • Takeaways & Limitations

    Perceptive mobile networks can integrate sensing with communication and potentially support applications in smart cities, homes, cars, and transportation.

  • Takeaways & Limitations

    Downlink sensing requires full-duplex transceivers, while full-duplex MIMO remains challenging to realize in practice because of antenna cross-talk and coupling.

Abstract

from arXiv · show

In this paper, we develop a framework for a novel perceptive mobile/cellular network that integrates radar sensing function into the mobile communication network. We propose a unified system platform that enables downlink and uplink sensing, sharing the same transmitted signals with communications. We aim to tackle the fundamental sensing parameter estimation problem in perceptive mobile networks, by addressing two key challenges associated with sophisticated mobile signals and rich multipath in mobile networks. To extract sensing parameters from orthogonal frequency division multiple access (OFDMA) and spatial division multiple access (SDMA) communication signals, we propose two approaches to formulate it to problems that can be solved by compressive sensing techniques. Most sensing algorithms have limits on the number of multipath signals for their inputs. To reduce the multipath signals, as well as removing unwanted clutter signals, we propose a background subtraction method based on simple recursive computation, and provide a closed-form expression for performance characterization. The effectiveness of these methods is validated in simulations.

I. INTRODUCTION

The paper proposes a perceptive mobile network that integrates communication and radar sensing through a unified platform, while addressing sensing-parameter estimation challenges from sophisticated signals and rich multipath. It develops compressive-sensing-based estimation schemes and recursive clutter suppression, with simulations validating the framework and algorithms.

  • Motivation: The framework integrates radar sensing into current communication-only mobile networks while retaining simultaneous communication service.It uses shared transmitted signals and many existing communication hardware and signal-processing modules.
  • Challenges: Modern mobile signals create estimation challenges because multiuser-MIMO and OFDMA signals may be randomly modulated and discontinuous across time, frequency, or space.These properties make many existing sensing techniques not directly applicable.
  • Challenges: Rich mobile-network multipath limits sensing algorithms, motivating preprocessing that reduces multipath and separates unwanted clutter before estimation.The paper defines clutter as unwanted multipath containing little new information, mainly from permanent or long-period static objects.
  • Framework: The unified platform supports three sensing types and represents their signal formulations with a common expression, enabling common sensing algorithms.The paper also describes required hardware and system changes for existing mobile networks.
  • Methods: Two estimation schemes address OFDMA and multiuser-MIMO signals: direct estimation uses received signals, while indirect estimation strips demodulated data symbols and decorrelates signals.The direct scheme assumes transmitted information symbols are known; the indirect scheme simplifies sensing inputs through signal stripping.
  • Methods: A low-complexity recursive background-subtraction method reconstructs clutter, reduces sensing inputs, and provides closed-form performance expressions relating reconstruction and noise to recursion parameters.The method can also separate signals with largely separated Doppler frequencies.
  • Evaluation: The paper presents simulations to validate the effectiveness of the proposed framework and sensing algorithms.The paper is positioned as an initial feasibility and methodology study whose algorithms are not yet optimized for complexity and performance.

A. System Model

The proposed perceptive mobile network integrates communication and radar sensing in a CRAN architecture, supporting uplink and downlink sensing with shared communication signals. Downlink sensing offers potential accuracy advantages but requires hardware changes, especially for full-duplex operation.

  • System Architecture: The CRAN platform centrally processes communication and sensing signals collected by synchronized, cooperative RRUs.The architecture includes a BBU pool for communication and a sensing processing unit.
  • Supported Sensing Operations: Communication signals from BSs and MSs are reused for sensing, enabling uplink, downlink active, and downlink passive sensing modes.Downlink active sensing uses a RRU’s own echoes, while passive sensing uses signals from other RRUs.
  • Supported Sensing Operations: Uplink sensing estimates relative rather than absolute delays because MS transmitters and RRU receivers are typically unsynchronized.Triangulation-based localization techniques may remove this timing ambiguity.
  • Supported Sensing Operations: Downlink sensing can potentially be more accurate and less privacy-sensitive because BS transmitters are more capable and downlink data symbols are centrally known.The sensed results are not directly linked to MSs.
  • Signal Sources: The framework focuses on always-available whole data payloads because random DMRS signals may be insufficient for high-resolution sensing.Signals may also be jointly optimized for communication and radar sensing.
  • Required System Modification: Uplink sensing can operate on current architectures with timing ambiguity, whereas downlink sensing requires hardware changes and same-time transmission and reception.Full-duplex leakage can overwhelm reflected echoes, and practical full-duplex MIMO remains challenging because of antenna cross-talk and coupling.
  • Required System Modification: Two near-term downlink options are separated transmit/receive antennas or receiving-only RRUs, but both require hardware changes and added deployment considerations.The paper states that downlink sensing is more cost-effective in TDD than FDD systems.
  • Open Challenges: Further joint optimization of waveforms, antenna placement, and sparsity remains a research opportunity for integrated systems.These design challenges are additional to the system modifications required for sensing integration.

III. FORMULATION OF SENSING PARAMETER ESTIMATION

This section develops common communication-and-sensing channel models for uplink and downlink operation, then introduces an on-grid delay representation for parameter estimation. The models expose detailed multipath parameters that communications typically do not need to resolve.

  • III. FORMULATION OF SENSING PARAMETER ESTIMATION: The paper extends general channel models to uplink and downlink sensing and constructs a common model for shared sensing algorithms.The formulation quantizes delay before presenting the direct and indirect estimation schemes.
  • A. General System and Channel Models: The assumed CRAN system uses cooperative RRUs, multiuser-MIMO, OFDMA, and user-specific shared subcarriers.Each user may occupy only part of the total subcarriers.
  • A. General System and Channel Models: Under a planar-wave assumption, each ULA response is parameterized by an AoD or AoA.The array response uses the antenna count and angular parameter.
  • A. General System and Channel Models: Each multipath component is characterized by complex amplitude, propagation delay, Doppler frequency, AoA, and AoD.The channel impulse response uses these parameters with a Kronecker-product spatial representation.
  • A. General System and Channel Models: The amplitude is assumed frequency independent because its variation is small across the bandwidth when cellular fractional bandwidth is small.This assumption simplifies the channel representation.
  • A. General System and Channel Models: Communications generally estimate composite channel values, whereas sensing must resolve detailed channel structure and individual sensing parameters.The sensing parameters include delay, Doppler, angles, and multipath amplitude.
  • A. General System and Channel Models: Extended sensing based primarily on machine learning, where these parameters may not be explicitly needed, is outside the paper’s scope.
  • A. General System and Channel Models: The received signal is converted to the frequency domain for processing on each OFDM block and subcarrier.The frequency-domain channel matrix corresponds to the time-domain impulse-response model.

B. Formulation for Downlink Sensing

Downlink and uplink sensing are formulated through related received-signal models and unified in a delay-quantized sparse representation. The resulting model supports common compressive-sensing processing while retaining node-specific multipath structure.

  • B. Formulation for Downlink Sensing: Downlink sensing is formulated separately at each RRU because each node observes echoes from its own and other RRUs.Joint processing may help when propagation channels are highly correlated, but its benefits are not generally obvious.
  • B. Formulation for Downlink Sensing: A downlink RRU receives reflected signals from all cooperative RRUs, including its own transmission.The received signal is indexed by subcarrier and OFDM block.
  • B. Formulation for Downlink Sensing: The downlink model uses structured array matrices to represent transmit and receive spatial responses across RRUs.The block-diagonal matrix U maps transmit-side spatial responses for the RRUs.
  • B. Formulation for Downlink Sensing: Packing observations from multiple RRUs increases measurement length but also increases the number of unknown parameters.Channel reciprocity can nevertheless support joint processing across RRUs.
  • C. Formulation for Uplink Sensing: The uplink received-signal model has the same structural form as the downlink model, with different symbols and parameter values.This correspondence enables a common on-grid formulation.
  • D. Generalized Delay-Quantized On-grid Formulation: Delay is quantized onto a grid with resolution 1/(gB), assuming sufficiently many subcarriers and small quantization error.The grid uses Np multipath positions, with only L nonzero signals among them.
  • D. Generalized Delay-Quantized On-grid Formulation: Both sensing directions are converted into a sparse delay-on-grid model with Np ≫ L candidate multipath signals and only L active components.A permutation matrix maps user- or RRU-specific multipath signals into the generalized representation.

E. Selection of Compressive Sensing Algorithms

The paper selects one-dimensional compressive sensing to estimate parameters from cellular signals with limited measurements in several domains. It uses direct and indirect signal formulations, then applies MMV block-sparse recovery to estimate delays and recover remaining parameters.

  • E. Selection of Compressive Sensing Algorithms: 1D compressive sensing is used to estimate sensing parameters from a generalized delay-on-grid model.The five parameters may otherwise be estimated individually or jointly through 1D-to-4D CS models.
  • E. Selection of Compressive Sensing Algorithms: High-dimensional on-grid CS can suffer large Doppler and angular quantization errors when channel coherence time and antenna counts limit measurements.Cellular signals provide many subcarriers but fewer measurements in other domains.
  • E. Selection of Compressive Sensing Algorithms: Off-grid CS methods are not yet extended to the required high-dimensional, MMV, and block-CS settings without complexity or estimation-range constraints.The cited approaches also impose constraints on parameter separation.
  • E. Selection of Compressive Sensing Algorithms: The 1D method is presented as a basis for future off-grid and higher-dimensional extensions, trading potentially better estimation performance against higher complexity.
  • Direct Estimation: The direct scheme assumes known symbols, obtained centrally in downlink or by demodulation in uplink.Sensing can tolerate more delay than communication, allowing uplink symbol demodulation before sensing.
  • Direct Estimation: Received signals are organized so 1D CS first estimates delays, after which associated amplitude estimates provide the other sensing parameters.
  • Direct Estimation: The organized observations form an MMV block-sparse problem whose nonzero rows identify quantized delays.BSBL-FM is used to recover the block-sparse signals, and recovered amplitudes support further parameter estimation.
  • Direct Estimation: When multiple multipath signals share a delay bin, estimation is straightforward across different RRUs but requires more complex techniques otherwise.The paper states that a method for separating these cases has yet to be developed.

A. Single Multipath for Each Delay

For one multipath signal per quantized delay, block compressive sensing identifies delay-support blocks, after which their amplitudes and matrix structure yield other sensing parameters.

  • BSBL-FM estimates the nonzero blocks whose indexes correspond to quantized delay values.In noisy cases, blocks can be sorted by estimated magnitude and thresholded to filter multipath signals.
  • Only one column of each nonzero block contains the multipath coefficient, Doppler phase, and steering-vector structure.The block-diagonal signal model makes the remaining columns sparse across antenna-related vectors.
  • AoA or AoD is obtained from cross-correlations of the relevant block submatrix, depending on the calculation order.The same block structure supports either angle-of-arrival or angle-of-departure estimation.
  • Doppler frequency is estimated across multiple OFDM blocks using cross-correlation of the corresponding submatrices.The method forms block estimates from successive OFDM block signals before computing the Doppler estimate.

B. Multiple Multipath Signals with the Same Delay

When multipath signals share a delay, the MMV formulation can represent their differing angles, while signal stripping simplifies estimation by separating node channels and removing modulation.

  • Multiple Multipath Signals with the Same Delay: Multiple multipath signals with the same delay and different angles appear in the MMV estimates using a single dictionary entry for that delay.Signals from different users or remote radio units produce multiple nonzero node-specific submatrices.
  • Multiple Multipath Signals with the Same Delay: If same-delay multipaths originate from the same user, their estimation requires the corresponding same-user formulation.The passage introduces this case separately from multipaths originating at different users.
  • Multiple Multipath Signals with the Same Delay: For a small number of multipaths at a given delay, 2D-ESPRIT or 2D-MUSIC can estimate angles, while 3D spectrum analysis can additionally estimate Doppler.These methods operate on each estimated block submatrix and can extend across multiple OFDM blocks.
  • Signal Stripping: Signal stripping addresses high-complexity block CS by removing modulation and separating channels for individual uplink users or downlink remote radio units.It retains as few as a single node's composite channel matrix as the sensing input.
  • Signal Stripping: Accurate composite channel estimates are required for signal stripping to reduce sensing parameters, complexity, and potentially improve performance.The paper does not provide detailed refinement algorithms and evaluates channel-estimation error separately.

B. Estimation of Sensing Parameters

The channel-matrix approach converts a delay-on-grid multipath model into a multiple-measurement-vector compressive-sensing problem, then derives angles, magnitude, and Doppler from its estimates.

  • B. Estimation of Sensing Parameters: The reconstructed channel matrix is modeled with quantized delays and steering matrices for the receive and transmit arrays.Diagonal matrices encode path coefficients, Doppler phase, and delay-dependent frequency phase.
  • B. Estimation of Sensing Parameters: Stacking channel-derived row vectors across usable subcarriers produces a matrix whose columns encode delay-dependent Fourier structure.At least two columns are needed in the simple example to estimate angle of departure.
  • B. Estimation of Sensing Parameters: The model becomes an MMV compressive-sensing problem with a partial DFT dictionary, solvable using OMP or Bayesian CS.The sparse representation uses Np ≫ L candidate multipath signals, with only L nonzero components.
  • B. Estimation of Sensing Parameters: Once delays and the coefficient matrix are estimated, cross-correlation provides AoA and AoD estimates.Path magnitudes can be obtained during angle estimation and used to threshold effective delay bins in noisy channels.
  • B. Estimation of Sensing Parameters: Doppler shift is estimated by cross-correlating refined coefficient-matrix estimates obtained at sufficiently separated times during channel stability.The time separation preserves channel stability while exposing Doppler phase changes.

VI. CLUTTER REDUCTION

The clutter-reduction method recursively averages refined channel estimates and subtracts the resulting clutter estimate, with the averaging interval critical because Doppler and sensing-parameter changes both alter radio channels.

  • VI. CLUTTER REDUCTION: Near-zero-Doppler echoes are treated as clutter, whereas nonzero-Doppler echoes are called dynamic multipath.The distinction defines which components the background estimate is intended to capture or suppress.
  • VI. CLUTTER REDUCTION: Background subtraction averages channel estimates over a long period, then subtracts the clutter estimate from current and future refined estimates.The method requires static clutter parameters and is suited to indirect sensing, or direct sensing using training signals.
  • VI. CLUTTER REDUCTION: The recursion uses a learning rate α, with the initial clutter estimate set to zero or an average of initial channel estimates.The estimate is updated every Th seconds.
  • VI. CLUTTER REDUCTION: The averaging interval Th is critical because radio-channel differences reflect both Doppler shifts and changes in sensing parameters.Small Th can fail to reduce dynamic multipath, so larger Th is needed for a clear clutter estimate.
  • VI. CLUTTER REDUCTION: p = 500 is approximately needed for ρ(p) to approach 1 when fD = 0, while larger α improves performance.Figure 2 depicts ρ(p) across Doppler frequencies from 0 to 400 Hz; Figure 3 compares Th values of 240Ts, 120Ts, and 60Ts.
  • VI. CLUTTER REDUCTION: When α = 0.99, the recursive output noise becomes 0.005σ2 and simulations show convergence at approximately p = 150.The method therefore suppresses noise, and clutter subtraction almost does not increase noise.
  • VI. CLUTTER REDUCTION: Adjusting recursion parameters can separate multipath signals with different Doppler frequencies into different groups.This extends background subtraction beyond clutter removal.

VII. SIMULATION RESULTS

The simulations evaluate the proposed sensing framework under a multiuser-MIMO mobile-network setup with clustered multipath and repeated parameter realizations. They use sparse Bayesian learning methods to solve the formulated direct and indirect MMV problems.

  • The evaluation uses Block Sparse Bayesian Learning for direct estimation and Sparse Bayesian Learning for indirect estimation.These methods are used to solve the respective MMV problems.
  • The simulated system has 4 RRUs serving 4 users, with 4 antennas per RRU, 1 antenna per mobile station, a 2.35 GHz carrier, and 100 MHz bandwidth.
  • Downlink simulations use all N = 512 subcarriers, whereas uplink simulations use 128 randomly indexed subcarriers shared by four users through multiuser-MIMO.
  • Multipath channels are generated in clusters with pathloss factors 4 for downlink and 2 for uplink, while transmission powers are 30 dBm for RRUs and 25 dBm for mobile stations.
  • The simulations vary clustered multipath over path count, direction, distance, and Doppler distributions, with additional inter-cluster offsets.
  • Each figure plots 10 implementations with fixed sensing parameters but changed data symbols and noise, using pluses for estimates and circles for actual multipath parameters.

A. Direct Estimation

The sensing schemes estimate delay, AoA, distance, and speed from direct or indirect mobile-network signals, with indirect processing generally improving separation and resolution under suitable conditions. Off-grid delays remain manageable for delay and AoA, while speed estimates are more sensitive.

  • Direct Estimation: Downlink and uplink AoA-distance estimates are reported as robust and accurate in the direct-sensing results.The depicted distance is total transmitter-to-receiver signal-travel distance and may require across-RRU synthesis to map to object distance.
  • Indirect Estimation based on Signal Stripping: Indirect uplink sensing produces accurate, channel-error-robust delay and AoA estimates, while moving-speed estimates have relatively large errors.Speed errors are attributed to small Doppler phase values and sensitivity to noise and the sampling interval T.
  • Indirect Estimation based on Signal Stripping: When channel estimation error is sufficiently small, indirect methods can outperform direct methods by efficiently separating different users’ channels and signals.
  • Indirect Estimation based on Signal Stripping: With continuous off-grid delays, downlink delay and AoA remain identifiable with degraded but acceptable accuracy, whereas speed estimates vary significantly across realizations.
  • Indirect Estimation based on Signal Stripping: The proposed 1D CS method achieves much better resolution than classical 2D DFT for the comparable downlink AoA-distance setup.
  • Indirect Estimation based on Signal Stripping: Under a practical 5G NR subcarrier allocation with limited delay-domain measurements, 1D CS achieves better distance, AoA, and speed resolution for most multipath channels.

C. Effect of Clutter Suppression

The background-subtraction method is evaluated for suppressing near-zero-Doppler clutter before sensing. Its reconstruction quality depends on the learning rate and the number of samples used for clutter estimation.

  • Effect of Clutter Suppression: The clutter simulations use near-zero Doppler frequencies to represent clutter signals, and only the background-subtraction method is evaluated.
  • Effect of Clutter Suppression: α > 0.99 provides a good balance between reconstruction difference and convergence time in the recursive clutter-reconstruction algorithm.
  • Effect of Clutter Suppression: Increasing the clutter-estimation parameter p improves subtraction: low p leaves missed dynamic-path estimates and residual multipath, while high p can remove clutter completely.
  • Effect of Clutter Suppression: The conclusion identifies background subtraction as the proposed clutter-suppression method within the perceptive mobile-network framework.
Loading 1901.05558v1…