Run Zhao, Dong Wang, Qian Zhang, - PowerPoint Presentation

Slide1

Run Zhao, Dong Wang, Qian Zhang, Haonan Chen, Anna Huang

CRH: A Contactless Respiration and Heartbeat Monitoring System with COTS RFID Tags

Shanghai Jiao Tong University

zhaorun@cs.sjtu.edu.cnSlide2

01

Motivation

Challenges

System Design

Conclusion

Implementation and Evaluation

Outline

02

03

04

05Slide3

Motivation

1

PART ONESlide4

Respiration rate (RR) & Heart rate (HR) monitoring

Disease prediction Emotion estimation

Exercise/Sleep quality

Fatigue warning

SIDSSlide5

RR and HR Monitoring

ObtrusiveSpecial and expensiveInconvenient and cumbersomeUncomfortableSpecial devicesBounded monitoring regionMultiple users

InconvenientUncomfortable

Easily removed

Wearable

Sensor

RF-based

Clinical Device

[6]

[Zephyr2016]

Chest

R: 1cm

H: 0.6mmSlide6

Monitoring: RF-based Slide7

Monitoring: RFID-based

Relatively free monitoring region: simply attaching tags in the space with great coverage of RFID antennas

RFID devices are

pervasive

and

non-intrusive

to users in homes and offices.

Tiny-size

Low-cost

Battery-free

Flexible

[16]

Phase Resolution

0.0015 radian (4096 bits) -> 0.038mm

Multi tags to distinguish the

individual

respiration and heartbeat signals with the nature of RFID

Clothes with tagReader in light bulbSlide8

Key Insight

The chest motion, resulting from respiration and heartbeat, can modulate the backscattered RFID signal.

Device-Free(tag near body)

On-Body

(tag on body)

RFID reading rate per tag (almost 50 Hz)

Typical respiratory rate (12-20 bpm)

Normal heart rate (50-90 bpm)

Line of Sight (LOS)

Static or Dynamic

Static

Multipath

Dynamic Multipath

Propagation pathsSlide9

Received signal

𝒔(𝒕)Device-FreeOn-Body

Static ones

M dynamic ones

Time-varying due to

the chest or other motions

Attenuation and initial phase offset

Dynamic components caused by respiration and heartbeat

quasi-periodic variations in LOS path

quasi-periodic variations in indirect path reflected by body

Rhythmic chest motion

 ripple-like pattern in RFID phase

Multiple tags

 boost the robustness

 Slide10

Preliminary experiment

Device-free scenario: noticeable quasi-periodicitySlide11

Preliminary experiment

a bandpass filter for RR estimationSlide12

Challenges

2

PART TWOSlide13

Challenges and Solutions

C1S1

Environmental noise

Hiding the respiration and heartbeat signals

More insensitive to the minute chest movement

The respiration signal is much stronger

Their frequency characteristics

Bandpass filter for each tag in each sliding window

Extract time-frequency domain features for RR and HR respectivelySlide14

Challenges and Solutions

C2S2

Frequency hopping spread spectrum

Discontinuous phase stream

Data Smoothing

A smoothing method to neglect the phase jumpSlide15

Challenges and Solutions

C3S3

Intense motions

Overwhelming the respiration and heartbeat signals

Motion Interference Detection

The strong power of intense motions

on-body

device-free

Intense motion Slide16

System Design

3

PART THREESlide17

System Overview

Raw phase Measurements

Smoothing

Filtering

Preprocessing

Motion Interference Detection

Intense Motion Detector

Tachypnea Bradypnea Apnea

Apnea Detector

Coarse-grained Signal Extractor

RR and HR Estimator

Respiration & Heartbeat Extractor

Respiration

HeartbeatSlide18

Preprocessing

Smoothing (unwrapping)Slide19

Preprocessing

hopping

first sample at frequency j without hopping

last two sample points at frequency

i

first two points at frequency j

Smoothing (for hopping)Slide20

Preprocessing

Smoothing (unwrapping)FilteringUnwrappingFor hopping

Hampel identifier

Linear interpolation

Exponentially weighted moving average removalSlide21

Intense Motion Detector

Motion interference  hide the quasi-periodic signal

CRH uses hypothesis testing to detect intense motions via a threshold of the variance. Slide22

Respiration and Heartbeat Extractor

Coarse-grained Signal ExtractorSources of phase fluctuation: respiration, heartbeat and noise

RR and HR Estimator

more dominant with different frequency ranges

spreads over the

whole frequency domain

two 10-order Butterworth filters with cut-off frequencies (0.1-0.5 Hz and 0.7-1.8 Hz) Slide23

Respiration and Heartbeat Extractor

RR and HR EstimatorEmpirical mode decomposition (EMD) separate respiration/heartbeat signal and preserve characteristics of varying frequency

EMD decomposes a signal into intrinsic mode functions (IMF) with a trend.

Num of extrema ≈ zero-crossings

Upper and lower envelopes: symmetric

Simple oscillatory mode

Corresponding to respiration/heartbeat or noise

Data-driven

No need a predefined mother wavelet

Num of source > num of measurementSlide24

Respiration and Heartbeat Extractor

RR and HR EstimatorEmpirical mode decomposition (EMD)

EMD decomposes a signal into intrinsic mode functions (IMF) with a trend.

Num of extrema ≈ zero-crossings

Upper and lower envelopes: symmetric

Simple oscillatory mode

Corresponding to respiration/heartbeat or noise

:

mean of two envelopes

:

phase stream

S

ifting

repeats k times

,

until becomes an IMF

:

S

’=

N

IMFs

Then sifting on

 Slide25

Respiration and Heartbeat Extractor

RR and HR EstimatorNum of IMF ↑, corresponding signal frequency ↓

Hilbert Transform: amplitude and frequency of each IMF

Strongest IMF



respiration or heartbeat of the nearest user

Multi-tag EMD method

Segmented by zero-crossing & extrema

Whether

Weighted mean with variances (

)

Sliding mean filter

Mode mixing caused by intermittent noiseSlide26

Apnea Detector

Outside normal scope of RR for 30 s Tachypnea alarm (>20 bpm)Bradypnea alarm (<12 bpm)

Apnea

Cessation of respiration for at least 10 s in duration

Oscillatory of IMF: frequency of each IMF ≠ 0

sliding moving window (10s with 50% overlap)

majority voteSlide27

Implementation and Evaluation

4

PART FOURSlide28

Implementation and Experimental Methodology

Slide29

Implementation and Experimental Methodology

3 Tags on wall

3 Tags on desk

Antenna

User interface

Tags array

Ground truth

Metronome

Garmin HR monitor

Reading rate about 55HzSlide30

Evaluation (On-Body Scenario)

Sampling ParametersDistance

0.05 bpm

0.75 bpm

Latency

0.2 sSlide31

Evaluation (Device-Free Scenario)

DistanceMultiple Users

Through Wood Wall

In: -48

dbm

Ouside

: -52 dBm Slide32

Evaluation (Daily Monitoring)

RespirationHeartbeathigher HR20 bpm on average

0.131.6Slide33

Conclusion

5

PART FIVESlide34

Conclusion

Propose the design, implementation and evaluation of CRH, a contactless respiration and heartbeat monitoring system with COTS RFID devicesSeparate the respiration and heartbeat signals from body motions and noiseEstimate the fine-grained RR and HR via multi-tag EMD Achieve a high accuracy of less than 0.13 bpm and 1.6 bpm mean errors for RR and HR respectivelySlide35

Future Work

Not a replacement for medical systemsIn futureExtract the fine-grained respiration and heartbeat signals at any time even with motion interferenceMake an exploration on exercise and sleep monitoring, as well as combining CRH with the indoor localization to provide more applicationsSlide36

CRH: A Contactless Respiration and Heartbeat Monitoring System with COTS RFID Tags

Q & A