PaperlayerView exact revision on arXiv

Source-linked AI summary

MoRe-Fi: Motion-robust and Fine-grained Respiration Monitoring via Deep-Learning UWB Radar

Tianyue Zheng, Zhe Chen, Shujie Zhang, Chao Cai, Jun Luo

arXiv:2111.08195v1cs.LGeess.SP

TL;DR

MoRe-Fi addresses contact-free respiration monitoring when body movements undermine systems designed for static subjects. It combines IR-UWB radar, complex I/Q-domain augmentation, and an IQ Variational Encoder-Decoder to recover fine-grained respiratory waveforms, with experiments demonstrating accurate recovery under movement.

  • Problem

    Existing contact-free respiration systems often assume static subjects, limiting continuous monitoring despite the importance of respiratory waveform patterns for health assessment.

  • Method

    MoRe-Fi uses an IR-UWB radar, complex I/Q-domain data augmentation, and IQ-VED to recover respiratory waveforms under body-motion interference.

  • Results

    Experiments on a 66-hour dataset strongly confirm MoRe-Fi's ability to recover fine-grained respiratory waveforms under body movements.

  • Takeaways & Limitations

    MoRe-Fi expands contact-free respiration monitoring toward continuous use across body movements, including movements involving chest motion.

  • Takeaways & Limitations

    Disease-diagnosis performance was not fully evaluated because the study lacked subjects with related diseases and medical-professional support.

Abstract

from arXiv · show

Crucial for healthcare and biomedical applications, respiration monitoring often employs wearable sensors in practice, causing inconvenience due to their direct contact with human bodies. Therefore, researchers have been constantly searching for contact-free alternatives. Nonetheless, existing contact-free designs mostly require human subjects to remain static, largely confining their adoptions in everyday environments where body movements are inevitable. Fortunately, radio-frequency (RF) enabled contact-free sensing, though suffering motion interference inseparable by conventional filtering, may offer a potential to distill respiratory waveform with the help of deep learning. To realize this potential, we introduce MoRe-Fi to conduct fine-grained respiration monitoring under body movements. MoRe-Fi leverages an IR-UWB radar to achieve contact-free sensing, and it fully exploits the complex radar signal for data augmentation. The core of MoRe-Fi is a novel variational encoder-decoder network; it aims to single out the respiratory waveforms that are modulated by body movements in a non-linear manner. Our experiments with 12 subjects and 66-hour data demonstrate that MoRe-Fi accurately recovers respiratory waveform despite the interference caused by body movements. We also discuss potential applications of MoRe-Fi for pulmonary disease diagnoses.

1 INTRODUCTION

MoRe-Fi addresses the difficulty of recovering fine-grained respiratory waveforms during body movement, where existing RF systems often fail or produce noise-like readings. It combines IR-UWB sensing, complex-signal augmentation, and a variational encoder-decoder to enable motion-robust monitoring.

  • Motivation: Respiratory waveform patterns can support early disease detection, but wearable sensors are inconvenient because they require direct body contact.Respiratory rate and fine-grained waveform patterns reflect symptoms relevant to respiratory disease monitoring.
  • Motivation: MoRe-Fi targets fine-grained respiration monitoring during body movements, addressing the static-subject assumption that limits existing RF-sensing systems in real-life settings.Existing systems may suspend monitoring under strong movement or produce noise-like readings.
  • Challenges: Recovering waveforms under movement is challenging because conventional linear filtering may not handle nonlinear interference, while prior methods often use only phase or amplitude information.The paper identifies limitations in both signal processing assumptions and incomplete use of complex RF signals.
  • Approach: MoRe-Fi uses a commercial-grade IR-UWB radar, augments complex in-phase and quadrature signals, and feeds them to IQ-VED for robust respiratory waveform recovery.IQ-VED jointly captures complementary I/Q information in an interpretable latent representation.
  • Evaluation: Evaluations on a 66-hour dataset strongly confirm MoRe-Fi's waveform recovery ability under body movements.The contribution summary characterizes the system as fine-grained, low-complexity, and full-scale motion-robust.

2 BACKGROUND AND PRELIMINARIES

IR-UWB radar represents respiration in complex I/Q signals collected across repeated frames, but body movement scrambles these signals and defeats conventional one-dimensional recovery methods.

  • 2.1 Capturing Breath with IR-UWB Radar: Repeated radar frames form a signal matrix whose fast-time and slow-time dimensions expose periodic chest movements for respiration monitoring.The received complex signal is stacked across regularly transmitted frames, while the radar’s bandwidth separates objects at different distances.
  • 2.1 Capturing Breath with IR-UWB Radar: Neither amplitude nor phase alone faithfully recovers respiratory waveforms, producing distortions such as strength variation and missing cycles.These real-valued projections discard information contained in the complex I/Q representation.
  • 2.1 Capturing Breath with IR-UWB Radar: Respiration appears in the I/Q plane as a respiration-induced vector combined with a body-background-reflection offset, forming an elliptical trace during breathing.The reflected-signal strength, chest distance, carrier wavelength, and chest movement determine the I/Q components, while background reflection shifts the ellipse.
  • 2.2 Interference Caused by Body Movements: The studied movement scope keeps subjects within the radar FoV and near a constant mean distance while allowing limb, torso, exercise, and chest movements.It excludes drastic posture changes and major position changes that would require tracking.
  • 2.2 Interference Caused by Body Movements: Body movements scatter the I/Q trace by varying background reflections, radar distance, reflected-signal strength, and potentially multiple fast-time indices.These effects prevent a single stable ellipse from representing respiration.
  • 2.3 Conventional Respiration Recovery: Under strong movement, one-dimensional methods produce coarse waveforms with lost event details and incorrect rates because they cannot handle I/Q scrambling and fast-time crossing.Bandpass filtering, EEMD, and VMD therefore cannot support the defined large-movement scope, motivating a data-driven approach.

3 SYSTEM DESIGN

MoRe-Fi combines IR-UWB sensing, I/Q-domain augmentation, and IQ-VED to recover fine-grained respiratory waveforms during severe body-movement interference.

  • 3 SYSTEM DESIGN: MoRe-Fi extracts a subject-specific signal sub-matrix, augments it using I/Q rotation invariance, and applies IQ-VED to distill respiratory waveforms.IQ-VED is trained against wearable-sensor ground truth and is designed for recovery under severe movement interference.

3.1 Locating Respiration

MoRe-Fi locates respiration by removing static background clutter, detecting signal peaks, and selecting neighboring fast-time indices into a transposed sub-matrix.

  • 3.1 Locating Respiration: MoRe-Fi first applies a loopback filter to remove static background, then uses CFAR with an adaptive neighboring-index threshold to detect peaks.The filter weight β is empirically set to 0.9.
  • 3.1 Locating Respiration: Detected peaks are expanded to adjacent fast-time indices, forming a respiration sub-matrix that is transposed so slow time becomes the argument.This preserves multiple fast-time indices for subsequent waveform recovery.

3.2 Data Augmentation

MoRe-Fi augments complex I/Q radar data by rotating each complex element, enriching limited datasets while preserving respiration-induced motion patterns.

  • 3.2 Data Augmentation: Data augmentation addresses the difficulty of coordinating subjects, radar recordings, and wearable ground-truth collection while increasing training diversity.
  • 3.2 Data Augmentation: Rotating each complex element of the I/Q signal augments the dataset while preserving respiration traces and their overall distribution.The rotation changes distance d but not the chest’s respiration-induced periodic motion Δd.

3.3 Fine-Grained Waveform Recovery

IQ-VED adapts variational encoder-decoder modeling to complex I/Q radar signals, aligning separate latent representations to recover respiratory waveforms under motion interference.

  • 3.3 Fine-Grained Waveform Recovery: Variational encoding represents inputs as Gaussian latent distributions, providing a continuous latent space intended to reduce overfitting and improve handling of out-of-range inputs.The latent prior is standard Gaussian, and KL divergence regularizes the encoded distribution toward that prior.
  • 3.3 Fine-Grained Waveform Recovery: The encoder uses convolutional feature extraction to demix respiration from nonlinear motion interference before sampling latent representations for waveform recovery.The decoder reverses this process with transposed convolutions and produces a single-channel respiratory waveform.
  • 3.3 Fine-Grained Waveform Recovery: IQ-VED separately encodes I/Q components, aligns their latent distributions, and decodes the combined representation into a respiratory waveform.The two streams share structures because they are distinct perspectives of the same complex radar signal.
  • 3.3 Fine-Grained Waveform Recovery: Wasserstein alignment brings the I/Q latent distributions together while retaining discrepancies inherent to their distinct amplitude and phase information.Unlike KL divergence, Wasserstein distance provides useful gradients when distributions do not overlap.
  • 3.3 Fine-Grained Waveform Recovery: IQ-VED training combines reconstruction, I/Q regularization, and distribution-alignment losses, with separate weights for regularization and alignment.The reconstruction loss compares the decoded waveform with wearable-sensor ground truth using L2 error.

4 IMPLEMENTATION

The MoRe-Fi prototype uses a compact IR-UWB radar and synchronized wearable respiration sensor, with software and training configured for radar-based waveform recovery.

  • 4 IMPLEMENTATION: The prototype uses a Novelda X4M05 IR-UWB radar operating at 7.29 or 8.7GHz with 1.5GHz bandwidth and 50fps frame rate.A Raspberry Pi controls the radar, while a desktop with an Intel Xeon CPU, 16GB RAM, and RTX 2080 Ti supports processing.
  • 4 IMPLEMENTATION: A NeuLog respiration belt sensor records ground-truth waveforms at the same 50fps rate as the radar.Precision Time Protocol synchronizes the hardware clocks for aligning radar and wearable-sensor signals.
  • 4 IMPLEMENTATION: Training uses PyTorch-based software, 64-sample batches, IQ-VED loss, and stochastic-gradient-descent settings of 0.01 learning rate and 0.9 momentum.

5 EVALUATION

MoRe-Fi is evaluated on real-life activities with body movements using 12 healthy volunteers and 66 hours of radar and ground-truth recordings. It recovers respiratory waveforms and derived biomarkers accurately, outperforming a motion-aware baseline.

  • 5.1 Evaluation Setup: The evaluation uses 12 healthy volunteers performing seven activities and produces a balanced 66-hour dataset of RF and ground-truth recordings.The radar is placed 0.5–2 m from the subject at chest height; IQ-VED is trained with 20-second samples while its CNN encoder accepts arbitrary sample lengths in practice.
  • 5.3 Performance Results: MoRe-Fi recovers respiratory waveforms during sporadic and intensive movements, correcting phase shifts while preserving breathing patterns across activities.During turning over, it recovers respiration in both quasi-static and movement periods, with only minor transitional waveform deformation.
  • 5.3.1 Overall Performance: Mean cosine similarities exceed 0.95 for every movement type, while PP and SL have the least impact and WT and SS the greatest impact.Large movements involving body regions closer to the chest produce stronger interference and worse recovery.
  • 5.3.2 Comparison with Baseline Method: MoRe-Fi substantially outperforms BreathListener in waveform quality, cosine similarity, respiratory-rate estimation, and peak/valley-time estimation under intensive movement.The comparison attributes the advantage to IQ-VED’s integrated I/Q decomposition and reconstruction, whereas BreathListener’s EEMD-GAN pipeline can lose information and produce erroneous biomarkers.
  • 5.3.3 Estimation Errors of Indicators: Most respiratory-rate errors are below 0.1 bpm, and most mean peak and valley timing errors are below 0.1 s.Valley-time errors are noticeably larger than peak-time errors.
  • 5.3.3 Estimation Errors of Indicators: Mean tidal-volume errors remain below 3% for all seven movements, with more than 75% of errors below 5%.Sit-stand movements have the most adverse effect because chest displacement varies strongly along the radar propagation direction.

5.4 Impact of Practical Factors

MoRe-Fi remains effective across subjects and clothing, while performance depends on training size, latent dimension, loss weights, and sensing distance. The experiments identify practical operating choices and performance boundaries.

  • 5.4.1 Impact of Human Subjects: Mean cosine similarities exceed 0.95 across all 12 subjects, with more than 75% of similarities above 0.85.Waveform recovery is largely insensitive to physical differences among the subjects.
  • 5.4.2 Impact of Training Set Size: A training set of 6,000 samples yields cosine similarity above 0.9, after which increasing training size produces saturation.The study begins from 8,000 samples collected across different activities.
  • 5.4.3 Latent Space Dimension: Cosine similarity improves with latent dimension until 64, then degrades as training becomes harder; IQ-VED therefore uses 64 dimensions.The selected dimension balances latent expressiveness against compactness and training convergence.
  • 5.4.4 Weights of the Loss Function: The best waveform-recovery performance occurs at loss weights γ = 3 and η = 2e−4.The weights trade off latent continuity and disentanglement against I/Q-representation alignment and expressiveness.
  • 5.4.4 Impact of Subject Clothing: MoRe-Fi reaches an average cosine similarity above 0.9 across lightweight and heavyweight clothing combinations, performing slightly better with less clothing.The tested conditions include lightweight T-shirts, heavyweight T-shirts, and coat-plus-T-shirt combinations.
  • 5.4.5 Impact of Sensing Distance: Average cosine similarity remains above 0.8 under the most intensive movements at the maximum tested sensing distance of 2 m.Increasing sensing distance has a negative effect on performance.

6 POTENTIAL MEDICAL ADOPTIONS

MoRe-Fi’s recovered waveform can provide respiratory biomarkers and support discussion of pulmonary-disease monitoring. However, disease-diagnosis performance is not evaluated because relevant subjects and medical-professional support were unavailable.

  • 6.3 Limitations: The paper does not fully evaluate pulmonary-disease diagnosis because it lacks subjects with related diseases and support from medical professionals.Instead, it discusses potential medical adoption without measuring diagnostic performance.
  • 6.1 Biomarkers: MoRe-Fi’s biomarkers can indicate conditions including apnea, tachypnea, hyperpnea, dyspnea, and Cheyne–Stokes respiration.These biomarkers are derived from respiratory-waveform timestamps and amplitude.
  • 6.2 Flow-Volume Loops: Differentiating respiration volume and flow produces flow-volume loops that visualize inhalation and exhalation processes.The loop shape can provide diagnostic information for pulmonary diseases involving abnormal airflow.
  • 6.2 Flow-Volume Loops: Flow-volume loop patterns are illustrated for restrictive, obstructive, fixed upper-airway, and variable extrathoracic obstructions.The examples connect loop shapes with decreased lung volume, expiratory airflow obstruction, and airway obstruction scenarios.
  • 6.3 Potential Deployments: MoRe-Fi could support long-term respiratory monitoring in hospitals, nursing facilities, and homes for triage, care management, and awareness of respiratory status.These are proposed deployment scenarios rather than validated clinical outcomes.

7 RELATED WORK

Prior respiration-monitoring systems include wearable and contact-free approaches, but contact-free methods commonly assume static subjects or tolerate only limited motion. MoRe-Fi addresses this gap with motion-robust RF sensing and complex-signal processing.

  • 7.1 Contact-Free Monitoring: Wearable sensors inconvenience users through body contact, while contact-free systems seek to avoid adoption burdens and changes to users’ respiration habits.Contact-free approaches include RF and acoustic sensing.
  • 7.2 RF-Based Monitoring: Earlier RF systems estimate respiratory rate or coarse waveforms under static conditions and suspend monitoring when sudden motion interferes.Motion-robust proposals often require multiple synchronized radars or impose unrealistic assumptions about one-dimensional or quasi-static motion.
  • 7.4 Other Contact-Free Methods: Camera-based light-sensing methods are omitted because motion tracking is complicated and ineffective for the large-scale movements targeted by MoRe-Fi.The paper therefore focuses its comparison on RF and acoustic alternatives.
  • 7.3 Acoustic Monitoring: BreathListener targets waveform recovery during driving but addresses only small-scale body movements, whereas acoustic systems are vulnerable to ambient and motion-induced interference.Other acoustic systems primarily target respiratory rate while assuming static subjects.

8 CONCLUSION

MoRe-Fi enables motion-robust, contact-free respiration monitoring that recovers fine-grained waveforms rather than respiratory rate alone. Experiments on healthy subjects show promising waveform-recovery and long-term-monitoring performance, while clinical evaluation remains future work.

  • 8 CONCLUSION: MoRe-Fi extends contact-free vital-sign monitoring to full-scale body movements while targeting fine-grained respiratory waveforms.The system is designed for motion-robust, continuous monitoring rather than respiratory rate estimation alone.
  • 8 CONCLUSION: IQ-VED succeeds where conventional signal processing fails by using nonlinear decomposition to separate respiration from body-motion interference.The method addresses the composition of body-movement and respiration-induced chest-motion signals.
  • 8 CONCLUSION: Extensive experiments on healthy subjects demonstrate promising fine-grained waveform recovery and long-term respiration monitoring.
  • 8 CONCLUSION: Clinical validation and broader evaluation using spatial diversity from large antenna arrays are identified as future directions.The authors specifically plan to assess real-life clinical scenarios and pulmonary disease diagnosis applications.
Loading 2111.08195v1…