PaperlayerView exact revision on arXiv

Source-linked AI summary

SR-TTA: Spatial-Redundancy Test-Time Adaptation for Interference-Robust Respiration Sensing

Jingyuan Liu, Zheng Chang, Haoqiu Xiong, Zhuangzhuang Cui, Sofie Pollin

arXiv:2610.11755v1cs.LG

TL;DR

Strong in-band motion interference can defeat hand-crafted respiration fusion because its frequency overlaps breathing. The paper combines a learned complex-weight beamformer with label-free spatial-redundancy test-time adaptation. It reports lower respiration-rate error under simulated interference and fewer failures on real testbed recordings after calibration and adaptation.

  • Problem

    Strong motion whose frequency falls inside the respiration band makes hand-crafted fusion unable to distinguish the interferer from breathing by power or frequency.

    Source: §I · p. 1

  • Method

    A learned complex-weight beamformer is adapted per recording without labels by maximizing antenna-subset consistency under an out-of-band spectral veto.

    Source: §I · p. 1

  • Results

    Respiration-rate error falls from 5.8 to 0.8 bpm under simulated in-band interference, and real-data failures fall from 47% to 7% after calibration and SR-TTA.

    Source: §I · p. 1, D. Cross-Domain Service Quality · p. 6

  • Takeaways & Limitations

    SR-TTA brings the calibrated learned front-end to the hand-crafted baseline’s level on real testbed recordings using label-free adaptation.

    Source: D. Cross-Domain Service Quality · p. 6, §V · p. 6

    Real moving-interferer experiments and an implemented O-RAN control loop remain future work.

    Source: §V · p. 6

Abstract

from arXiv · show

Future 6G networks aim to expose sensing as a native service by reusing communication infrastructure. We study respiration sensing on a cell-free massive multiple-input multiple-output (MIMO) base station, where a 64-antenna channel must be fused into a breathing waveform. The state-of-the-art hand-crafted fusion is near-optimal in benign conditions. It collapses, however, under strong in-band motion interference, whose frequency falls inside the respiration band. We show that a learned complex-weight beamformer recovers respiration by spatial nulling, and that the remaining gap to a per-recording oracle can be closed at deployment by label-free test-time adaptation. Crucially, we identify which label-free signal makes this work. Frequency- and variance-based criteria cannot separate an in-band interferer from breathing. Our spatial-redundancy test-time adaptation (SR-TTA), which maximizes consistency across random antenna subsets under an out-of-band spectral veto, preserves benign performance in our tests. The respiration-rate error drops from 5.8 to 0.8 breaths per minute (bpm) under simulated in-band interference, and the pipeline maps onto the Open Radio Access Network (O-RAN) architecture as O-RAN distributed-unit (O-DU) range-gating, an adaptation xApp, and a calibration rApp. On real testbed recordings, a one-time cross-subject calibration plus SR-TTA reduces failures from 47% to 7%, drawing level with the hand-crafted combiner using label-free test-time adaptation.

I. INTRODUCTION

Hand-crafted fusion handles static subjects but struggles when strong motion falls within the respiration band, where power and frequency cannot distinguish interference from breathing. The paper combines learned complex-weight beamforming with label-free spatial-redundancy adaptation and reports improved respiration recovery under interference.

I. INTRODUCTION

  • Strong, slow motion can fall inside the 0.1–0.5 Hz respiration band, defeating power-based weighting and frequency separation.The hand-crafted front-ends are near-optimal for a single static subject, but co-occurring motion is a recognized failure mode.Source: §I · p. 1
  • Prior deep wireless-sensing models are vulnerable to domain shift, while standard label-free test-time adaptation objectives target classification rather than waveform recovery.The cited examples include unseen subjects, geometries, and environments.Source: §I · p. 1
  • The proposed pipeline maps range-gating to the O-DU, BeamNet and SR-TTA to a near-RT RIC xApp, and one-time calibration to a non-RT rApp.The mapping feeds an exposed respiration service.Source: §I · p. 1, Figure 1 · p. 2
  • BeamNet learns complex weights for spatial nulling, and SR-TTA adapts them per recording using antenna-subset consistency under an out-of-band spectral veto.The adaptation is label-free and uses neither ground truth nor an interferer detector.Source: §I · p. 1
  • Respiration-rate error falls from 5.8 to 0.8 bpm, while absolute Pearson correlation rises from 0.67 to 0.95 under in-band interference.The contribution passage reports these comparisons for the proposed front-end against SSIR-WAC.Source: §I · p. 1

II. SYSTEM MODEL AND PROBLEM

The system reuses a cell-free massive-MIMO uplink for passive respiration sensing. Distributed access points observe the target’s breathing-modulated chest reflection alongside direct and moving-body reflections.

A. Cell-Free MIMO Uplink Sensing

  • A cell-free massive-MIMO base station uses uplink pilots from a single-antenna user equipment to estimate full-band CSI for passive sensing.The scene includes distributed access points and a seated subject, with a second moving body acting as an interferer.Source: A. Cell-Free MIMO Uplink Sensing · p. 2, §II · p. 2
  • The channel comprises M distributed antennas grouped into uniform linear arrays at the access points, K subcarriers across bandwidth B at carrier frequency fc, and I frames.The seated subject modulates the channel through the chest-reflection path.Source: A. Cell-Free MIMO Uplink Sensing · p. 2

B. Range-Channel Extraction

Range-channel extraction reduces the CSI to one complex channel per antenna and frame, but does not separate the sources. A moving body contributes an interferer when its motion spectrum overlaps respiration.

B. Range-Channel Extraction

  • The inverse DFT retains one range bin per antenna for data reduction, not source separation, because path-length resolution is too coarse for this scene.Static-component removal suppresses the background before producing one complex range-channel per antenna and downsampled frame.Source: B. Range-Channel Extraction · p. 2
  • Breathing traces a small complex-plane arc, but the informative real or imaginary projection varies by antenna and many antennas are individually unreliable.The extracted per-antenna channels are sampled at a decimated waveform rate.Source: B. Range-Channel Extraction · p. 2

C. In-Band Interference

  • Each range-channel contains a mixture of breathing, moving-body interference, and noise when a second scatterer is present.The interferer’s motion is in-band when its spectrum overlaps the respiration band B=[0.1, 0.5] Hz.Source: C. In-Band Interference · p. 2

D. Problem Statement

The task is to recover a breathing waveform from antenna range-channels despite in-band interference and without deployment-specific ground truth. The paper uses a learned per-recording spatial combiner whose complex weights can represent interference-cancelling phase relationships.

D. Problem Statement

  • The objective is to recover the breathing waveform from the range-channels up to sign and scale, robustly and without per-deployment ground truth.Source: D. Problem Statement · p. 2

III. INTERFERENCE-ROBUST SENSING FRONT-END

A. Range-Channel and the Combiner Function Class

  • The complex-weight function class matches the per-recording supervised oracle, which recovers the waveform almost perfectly under interference when fit to that recording’s ground truth.Complex cross-antenna phase cancellation can place a spatial null, unlike nonnegative scalar weighting.Source: A. Range-Channel and the Combiner… · p. 3

B. BeamNet: Learned Spatial Filtering

BeamNet predicts an antenna combiner from the input, then a temporal convolutional network denoises the combined trace into a breathing estimate. It is pretrained on synthetic recordings and calibrated once on clean recordings at a new site, but a single feed-forward beam does not reach the per-recording oracle.

B. BeamNet: Learned Spatial Filtering

  • BeamNet predicts antenna weights from per-antenna range-channel descriptors using shared convolutional encoding, positional embeddings, self-attention, and a linear head.The combined trace is denoised by a temporal convolutional network to produce the breathing estimate.Source: B. BeamNet: Learned Spatial Filtering · p. 3
  • BeamNet is pretrained on synthetic recordings with a physics-based in-band interferer and calibrated once on clean recordings from other subjects at the deployment site.The monitored subject and interference are never seen in training.Source: B. BeamNet: Learned Spatial Filtering · p. 3
  • A single feed-forward beam does not reach the per-recording oracle because it must generalize to unseen subjects and interferer geometries.The training objective maximizes temporal correlation with the ground-truth waveform.Source: B. BeamNet: Learned Spatial Filtering · p. 3

C. Adaptation Criteria and Proposed SR-TTA

At deployment, SR-TTA addresses the per-recording oracle gap without ground truth by adapting the combiner while keeping the denoising network fixed. The adaptation criterion must distinguish breathing from an in-band interferer.

C. Adaptation Criteria and Proposed SR-TTA

1) Frequency Criteria Fail in Band:

Frequency attraction and dominant-variance subspace nulling fail when breathing and interference overlap. SR-TTA instead uses agreement across antenna subsets to favor spatially redundant breathing.

1) Frequency Criteria Fail in Band:

  • In-band frequency surrogates can lock onto the interferer, and empirically degrade the beam.Frequency is safer as an out-of-band veto than as an in-band attractor.Source: 1) Frequency Criteria Fail in Band · p. 4

2) Subspace Nulling Fails under Overlap:

  • When breathing and interference subspaces overlap, nulling dominant-variance directions removes breathing energy too, and performance degrades monotonically as k grows.The approach assumes the dominant-variance directions belong to the interferer alone.Source: 2) Subspace Nulling Fails under Overlap · p. 4

3) Proposed Spatial-Redundancy Consistency:

  • SR-TTA draws random antenna subsets and maximizes pairwise agreement, favoring breathing that is spatially redundant over an interferer reliant on few antennas.The resulting beam is robust to antenna dropout, unlike one that rides the localized interferer.Source: 3) Proposed Spatial-Redundancy… · p. 4

4) Out-of-Band Spectral Veto:

SR-TTA combines spatial consistency with a penalty on out-of-band energy. The veto suppresses broadband leakage but cannot reject interference that is both spatially redundant and in-band.

4) Out-of-Band Spectral Veto:

  • The adaptation objective combines spatial consistency with a penalty on output energy outside the service band Bsv.The service band is 5–50 bpm.Source: 4) Out-of-Band Spectral Veto · p. 4
  • The veto penalizes out-of-band energy without rewarding an in-band peak, but neither term guarantees rejection of an interferer that is spatially redundant and in-band.Spatial consistency distinguishes sources within the band, while the veto suppresses broadband leakage.Source: 4) Out-of-Band Spectral Veto · p. 4
  • SR-TTA initializes at the network beam, redraws subset masks at each of Nstep Adam updates, and returns the full-array output under adapted weights.Test-time adaptation is label-free and uses no interferer detector; pretraining and one-time calibration use labels.Source: 4) Out-of-Band Spectral Veto · p. 4

D. Complexity and O-RAN Placement

The proposed O-RAN placement assigns range-channel extraction to the O-DU, BeamNet and SR-TTA to a near-RT RIC xApp, and offline calibration to a non-RT rApp. Evaluation uses simulated moving-scatterer interference, clean synthetic and real recordings, and specified respiration metrics.

D. Complexity and O-RAN Placement

  • The proposed O-RAN mapping places range-channel extraction at the O-DU, BeamNet and SR-TTA in a near-RT RIC xApp, and offline calibration in a non-RT rApp.The mapping is architectural; O-RAN integration, interface traffic, and end-to-end latency were not benchmarked. Adaptation updates 2M weights, with computation scaling as NstepS TCN forward/backward passes per window.Source: D. Complexity and O-RAN Placement · p. 4

1) Data:

  • The simulated study uses 64 antennas across eight APs and injects a moving scatterer with frequency drawn inside the respiration band into independently generated recordings.It also evaluates 240 clean synthetic windows and 44 real single-person recordings; real+interference regimes inject simulated motion into the real recordings.Source: 1) Data · p. 4
  • All interfering motion is simulated, and the real recordings contain no second moving body.Source: 1) Data · p. 5
  • Performance is reported using absolute Pearson correlation, failure rate for windows with |ρ|<0.7, and respiration-rate MAE in bpm.Rates are spectral peaks compared with ground-truth-derived rates; MAEs are rounded to 0.1 bpm, and figure bands show ±one SEM.Source: 2) Metrics · p. 5

2) Metrics:

Across the interferer-frequency sweep, only SR-TTA tracks the per-recording oracle as interference enters the respiration band. The section also defines the supervised oracle reference and gives the fixed BeamNet and SR-TTA configurations used across regimes.

3) Baselines and Ceiling:

  • The oracle is a ridge-regularized per-recording linear combiner fit to that recording’s ground truth, providing a supervised reference for linear fusion.SSIR-WAC is the training-free physics baseline, and BeamNet without adaptation is the feed-forward ablation.Source: 3) Baselines and Ceiling · p. 5
  • BeamNet uses a per-antenna encoder, self-attention over 64 antennas, and a dilated TCN; SR-TTA uses 40 Adam steps and the same configuration in every regime without an interferer detector.Real-testbed evaluation uses leave-one-group-out calibration on other subjects’ clean recordings; synthetic regimes use the pretrained model without calibration.Source: 4) Implementation · p. 5

B. Mechanism: Which Adaptation Criterion Survives In-Band?

  • At 0.2 Hz, SR-TTA reaches 0.97 while SSIR-WAC falls to 0.49, and SR-TTA stays at 0.97–1.00 across the sweep, tracking the per-recording oracle.Frequency TTA reaches 0.91 at 0.2 Hz, while dominant-variance subspace nulling remains at 0.75–0.86; Fig. 3 shows the frequency sweep and criterion comparison.Source: Figure 3 · p. 5, B. Mechanism: Which Adaptation… · p. 5
  • In the illustrated 20 s in-band recording, SSIR-WAC follows interferer bursts, the feed-forward beam drifts, and SR-TTA re-locks onto the true rhythm without labels.Source: Figure 4 · p. 5, B. Mechanism: Which Adaptation… · p. 5

C. Spatial-Redundancy Scaling Law

SR-TTA improves respiration sensing under in-band interference on real and simulated recordings, while scaling with the number of distributed APs. The conclusion notes that real moving-interferer experiments and an implemented O-RAN control loop remain future work.

C. Spatial-Redundancy Scaling Law

  • As distributed APs increase from 1 to 8, SR-TTA rises from 0.78 to 0.96, tracking the oracle ceiling from 0.92 to 1.00.SSIR-WAC improves only from 0.54 to 0.67, while achieved consistency remains ≥0.98 throughout the sweep.Source: C. Spatial-Redundancy Scaling Law · p. 5, Figure 5 · p. 6

D. Cross-Domain Service Quality

  • Under strong in-band interference, SR-TTA holds waveform fidelity at 0.95 as the physics baseline breaks down, while respiration-rate error falls from SSIR-WAC’s 5.8 bpm to 0.8 bpm; clean synthetic failures stay ≤2%.Source: D. Cross-Domain Service Quality · pp. 5–6
  • After calibration, SR-TTA reaches |ρ|=0.90 with 7% failures on real testbed recordings, drawing level with SSIR-WAC at 0.89 and 8%.Without calibration, SR-TTA reaches 0.89 at 5–9% failures.Source: D. Cross-Domain Service Quality · p. 6

V. CONCLUSION

  • Real moving-interferer experiments and an implemented O-RAN control loop remain future work.Source: §V · p. 6
Loading 2610.11755v1…