mmVib: Micrometer-Level Vibration Measurement with β¦
Transcript of mmVib: Micrometer-Level Vibration Measurement with β¦
mmVib: Micrometer-Level Vibration Measurement withmmWave Radar
Chengkun JiangSchool of Software and BNRist,
Tsinghua [email protected]
Junchen GuoSchool of Software and BNRist,
Tsinghua [email protected]
Yuan HeSchool of Software and BNRist,
Tsinghua [email protected]
Meng JinSchool of Software and BNRist,
Tsinghua [email protected]
Shuai LiSchool of Software and BNRist,
Tsinghua [email protected]
Yunhao LiuTsinghua University & [email protected]
ABSTRACTVibrationmeasurement is a crucial task in industrial systems, wherevibration characteristics reflect the health and indicate anomalies ofthe objects. Previous approaches either work in an intrusive manneror fail to capture the micrometer-level vibrations. In this work, wepropose mmVib, a practical approach to measure micrometer-levelvibrations with mmWave radar. By introducing a Multi-Signal Con-solidation (MSC) model to describe the properties of the reflectedsignals, we exploit the inherent consistency among those signalsto accurately recover the vibration characteristics. We implementa prototype of mmVib, and the experiments show that this designachieves 8.2% relative amplitude error and 0.5% relative frequencyerror in median. Typically, the median amplitude error is 3.4π’πfor the 100π’π-amplitude vibration. Compared to two existing ap-proaches, mmVib reduces the 80π‘β-percentile amplitude error by62.9% and 68.9% respectively.
CCS CONCEPTSβ’NetworksβCyber-physical networks; β’Computer systemsorganizationβ Embedded and cyber-physical systems.
KEYWORDSWireless Sensing, Millimeter Wave, Vibration Measurement
ACM Reference Format:Chengkun Jiang, Junchen Guo, Yuan He, Meng Jin, Shuai Li, and Yun-hao Liu. 2020. mmVib: Micrometer-Level Vibration Measurement withmmWave Radar. In The 26th Annual International Conference on MobileComputing and Networking (MobiCom β20), September 21β25, 2020, Lon-don, United Kingdom. ACM, New York, NY, USA, 13 pages. https://doi.org/10.1145/3372224.3419202
Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than ACMmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,to post on servers or to redistribute to lists, requires prior specific permission and/or afee. Request permissions from [email protected] β20, September 21β25, 2020, London, United KingdomΒ© 2020 Association for Computing Machinery.ACM ISBN 978-1-4503-7085-1/20/09. . . $15.00https://doi.org/10.1145/3372224.3419202
ππ-level amplitudeVibrating
objects
Background
reflections
mmWave
Device Cost Deployment Cost
Sensor-based Vibrometer Low (~$πππ+) High
Laser-based Vibrometer High (~$ππππ+) High
mmVib Moderate ($πππ) Low
Figure 1: Vibration measurement with mmVib
1 INTRODUCTIONVibration is the most common phenomenon in industry. Vibra-tion of the industrial objects generally reflects their internal states.Damage or malfunction of the objects usually leads to abnormalchanges in the vibration characteristics [8]. Vibration measurement,namely to measure the vibration amplitude and frequency, is a cru-cial task in various industrial scenarios for checking machineryhealth, identifying anomalies, and diagnosing faults [1, 8, 13].
Conventional approaches for vibration measurement rely onspecialized sensors like piezoelectric sensors [9, 20] or optical de-vices like laser vibrometers [6, 27]. Specialized sensors require to bedirectly installed on the vibrating object, which means high com-plexity in deployment and maintenance. Optical devices often havehigh precision and accuracy, but their prohibitive cost preventsthem from being widely used in real application scenarios. Thetable in Fig. 1 shows a brief comparison of vibration measurementapproaches [6, 9].
With the rapid progress in wireless sensing, recent works pro-pose to exploit wireless signals, e.g. acoustic signal [21, 31] andradio frequency (RF) signal [18, 32, 36], for vibration measurement.The vibrating object is a physical reflector of the wireless signal, sothat the vibration affects the propagation of the reflected signal. Bymeasuring and analyzing the reflected signal, one can obtain thevibration characteristics. Compared to conventional approaches,
MobiCom β20, September 21β25, 2020, London, United Kingdom Chengkun Jiang, Junchen Guo, Yuan He, Meng Jin, Shuai Li, and Yunhao Liu
the measurement based on wireless sensing is low-cost and easy todeploy in practice. However, its precision is often limited, due to therelatively long wavelengths of the employed wireless technologies,while the vibration amplitudes of a large number of machines inthe industry do not exceed 100π’π [1, 8].
mmWave is a promising technology for measuring tiny displace-ments, owing to its short wavelength. Recent works propose to usemmWave in different sensing applications [7, 11, 16, 19, 23, 29, 33,35, 37, 42]. But those approaches cannot provide highly precise andaccurate vibration measurement in the industrial scenarios. Fig. 1shows a typical industrial environment, from which we can per-ceive the following challenges. First, the industrial environmentsare multipath-rich environments. Although mmWave has betterdirectionality than the conventional wireless signals, the receivedsignal at the antenna is still a mix of the signals reflected from boththe vibrating object and other reflectors in the environment. Second,the vibration of the industrial objects is often at the micrometerlevel [1, 8]. The signal changes caused by such vibration are easilyaffected by the noise in the signal. Due to the above reasons, thevibration-induced changes in the reflected signal are obscured anddistorted, making it extremely difficult to extract accurate vibrationcharacteristics.
In this work, we address the above challenges and proposemmVib, a practical approach to measure the micrometer-level vibra-tion with mmWave radar. We propose a multi-signal consolidationmodel (MSC) that describes the properties of the reflected signals inthe In-phase and Quadrature (IQ) domain and exploit the inherentconsistency among those signals to accurately recover the vibrationcharacteristics. Our contributions are summarized as follows:
β’ We propose MSC, a signal model that comprehensively describesthe composition of the reflected mmWave signals. MSC capturesthe multi-frequency and multi-antenna properties of the reflectedsignal from the vibrating object in the multipath-rich environ-ments.
β’ Based on MSC, the design of mmVib addresses critical challengesin achieving the micrometer-level accuracy: (i) pinpointing thevibrating object in the mixed reflected signal; (ii) recovering themicrometer-level vibration under the influence of noise and otherreflected signals.
β’ We implement mmVib on the commercial off-the-shelf (COTS)mmWave radar and evaluate its performance in both lab andreal-world environments. The results show that mmVib achieves8.2% relative amplitude error and 0.5% relative frequency errorin median. Typically, the median amplitude error is 3.4π’π forthe 100π’π-amplitude vibration. Compared to two existing ap-proaches, mmVib reduces the 80π‘β-percentile amplitude error by62.9% and 68.9% respectively.
The rest of the paper is organized as follows: Β§2 introduces thepreliminaries of vibration measurement with mmWave radar. Wepresent the MSC model in Β§3 and the design of mmVib in Β§4. Β§5discusses the limitations of mmVib. Β§6 presents implementationdetails and evaluation results. Β§7 discusses the related works. Weconclude mmVib and discuss future works in Β§8.
30 50 100 200Amplitude (um)
0
25
50
75
100
125
Rel
ativ
e Er
ror (
%)
84.96% 84.95% 85.05% 85.86%
0.30% 0.31% 0.31% 0.34%
Amp. = 30~200um, Freq. = 100HzAmplitude Frequency
(a) High amplitude estimation errors
Clean Other Bins Same BinLocation of Reflectors
0
25
50
75
100
125
Rel
ativ
e Er
ror (
%)
82.90% 82.01%
96.83%
0.45% 0.39% 0.47%
Amp. = 100um, Freq. = 100HzAmplitude Frequency
(b) Impact of metal reflectors
Figure 2: Direct estimation based on phase change
2 PRELIMINARIESIn this section, we first introduce the principles of usingmmWave tomeasure the displacement and then analyze why it canβt be directlyused for the micrometer-level vibration measurement in industry.
2.1 Estimate Displacement with mmWaveThe core idea to track the displacement of the target with wirelesstechnologies is to extract the phase changes of the received signalreflected from the target:
π₯π = _π₯π
2π (1)
where _ is the wavelength. The phase change π₯π acts as a stableand accurate indicator of the propagation distance change π₯π , com-pared to other signal features like RSS. Therefore, the millimeter-level wavelength of mmWave endows it with over 25Γ and 65Γhigher sensitivity to tiny displacements, compared to WiFi andRFID respectively. Thatβs why mmWave is deemed to be a promis-ing technology for measuring mechanical vibrations in industrialscenarios, where the vibration amplitude typically does not exceed100π’π. However, our experiments provide a different observation.We configure a vibration calibrator (the vibrating object we use andintroduce in Β§6.1) to vibrate with different amplitudes (30 βΌ 200π’π)and 100π»π§ frequency. Fig. 2(a) shows the errors of the vibrationcharacteristics that are directly estimated based on phase change.The results show that the frequency estimation is consistently accu-rate while the relative errors of the amplitude estimation are over80%. The accuracy of the frequency estimation indicates that wehave indeed extracted the vibration signal. But the amplitude errorsare too large to correctly characterize the vibration. In order to findout the reason behind, we take a deep look at the signal processingof mmWave radar in the next subsection.
2.2 Phase Extracted from mmWave RadarmmWave radar usually adopts frequency-modulated continuouswave (FMCW) chirp signals for distance measurement, as shown inFig. 3(a). The frequency difference between the transmitted signal(Tx) and the received signal (Rx) indicates the signal propagationtime, which can be used to determine the object distance. Denotingthe time-variant distance between the antenna and the vibratingobject by π (π‘), the transmitted and received signal can be expressedby:
πππ₯ (π‘) = exp[ π (2π πππ‘ + ππΎπ‘2)]ππ π₯ (π‘) = πΌπππ₯ [π‘ β 2π (π‘)/π]
(2)
mmVib: Micrometer-Level Vibration Measurement with mmWave Radar MobiCom β20, September 21β25, 2020, London, United Kingdom
Fre
quency
Fast Time
Slow Time
Range B
ins
Range FFT
Slow-Time
Samples
Fast-Time Samples
Fast Time
π = ππ + πΎπ‘
πππ₯(π‘)ππ π₯(π‘)π (π‘)
(a) FMCW chirp signals
Symmetrical velocities of reciprocating motion
Static object
(b) Range-Doppler spectrum
Figure 3: FMCW and range-Doppler processing
where πΌ is the path loss. ππ and πΎ are the starting frequency andthe chirp slope of FMCW signal respectively. Then a mixer is usedto eliminate the carrier wave in the received signal and obtain theso-called beat frequency signal π (π‘) as:
π (π‘) = πππ₯ (π‘)πβπ π₯ (π‘) β πΌ exp[ π4π (ππ + πΎπ‘)π (π‘)/π] (3)
whose phase values contain the distance information π (π‘).In practice, there exist reflected signals from different distances,
which cause different frequency components in π (π‘). In order toextract the desired signal from π (π‘), a Range-FFT operation [7, 19]is conducted on the samples of π (π‘) within a chirp (denoted as thefast-time samples) for the signal separation. As shown in Fig. 3(a),this operation maps the frequency spectrum of π (π‘) to the rangespectrum indicating the distances of different reflectors. Since theelapsed time of a chirp is around 0.1ππ , the displacement duringthis period can be neglected and we only focus on the displacementsacross consecutive chirps. Therefore, in the Range-FFT results ofeach chirp, we pick up one sample in each range bin. Then, in acertain range bin, combining those samples (denoted as slow-timesamples) forms a sample sequence of 10πΎπ»π§ sampling rate. If werewrite the object range π (π‘) as π (π‘) = π 0 + π₯ (π‘), where π₯ (π‘) is thevibration displacement within the range-bin resolution and π 0 isthe object-radar distance, the reflected signal π (π‘) from the objectrange bin can be represented as:
π (π‘)Range-FFT
ββββββββββββββββat object range bin
π (π‘) = πΌ exp[ π4π ππ (π 0 + π₯ (π‘))/π] (4)
Then, we can estimate the velocity of π₯ (π‘) by performing an-other FFT operation called Doppler-FFT on π (π‘). Fig. 3(b) showsthe Range-Doppler spectrum that can detect the existence of thereciprocating motion of a vibrating object. So why canβt the phasechange correctly recover the vibration displacement? Actually, π (π‘)is composed of all the reflected signals from the range bin of the vi-brating object. Thus, π₯ (π‘) canβt be directly derived with the phasevalues of π (π‘). Fig. 2(b) shows that the estimation error increaseswhen we deliberately place multiple metal reflectors in the samerange bin, which further verifies the above statement.
Another often neglected fact is that the measurement accuracyis also seriously affected by the signal noise, especially for tinyvibrations. For instance, a 100π’π vibration only results in about0.12πππ phase change, which is very sensitive to the signal noise.
Therefore, to achieve accurate vibration measurement, two criti-cal factors should be properly and carefully considered: (i) otherreflected signals entangled with the vibration reflection; (ii) thenoise in the phase values.
3 MODELING THE VIBRATIONIn this section, we introduce the MSC model, which acts as thetheoretical foundation of mmVib.
3.1 Identifying the Signal in the IQ domainAccording to Eq. 4, the reciprocating motion can change the signalphase within a certain range, which means the signal samples plot-ted in the IQ domain can form an arc-shaped trajectory centered atthe origin of coordinates. To verify this, we measure the vibrationsof 100π’π and 200π’π amplitudes with 1π and 2π distances awayfrom the radar. The corresponding IQ signal samples are shown inFig. 5(a).
There indeed exist arc-shaped trajectories: the central angles ofthe arcs are proportional to their amplitudes, but their centers areobviously not at the origin. The relationship between the centralangle and the vibration amplitude is reasonable because the vibra-tion displacement linearly changes the signal phase. For example,the central angle of the 100π’π vibration is about the half of thecentral angle of the 200π’π vibration. As for the center coordinates,their offsets from the origin are due to the impact of other reflectedsignals from the same range bin of the vibrating object. Actually,the theoretical derivation of Eq. 4 only considers the vibration re-flection from the object. So in MSC, we modify it to also considerthe background reflections:
π β²(π‘) = πΌππ₯π [ π4π ππ (π 0 + π₯ (π‘))/π] +βπ
πΌ[π ]π΅ππ₯π [ π4π πππ [π ]
π΅/π]
= πΌππ₯π [ π4π ππ (π 0 + π₯ (π‘))/π] + πΌπ΅ππ₯π [ π4π πππ π΅/π](5)
where MSC regards all the background reflections as one compositereflection ππ΅ from one single virtual reflector. π π΅ and πΌπ΅ representthe distance from the virtual reflector to the radar and the signalstrength of ππ΅ , respectively.
As shown in Fig. 4(a), we use the vector form of the signals inthe IQ domain to provide an intuitive representation of the signalsuperimposition:
ββπ β² =
ββπ + ββ
ππ΅ .ββπ β² rotates synchronously with ββ
π ,which explains the good accuracy of the frequency estimation inFig. 2. However, since \ β² differs from the real phase variation \ , suchdirection extraction suffers from errors in the amplitude estimation.
An intuitive way to extract the correct vibration reflection ββπ
fromββπ β² is to fit a circle based on the signal samples. Then the static
component ββππ΅ can be eliminated by translating the circle center tothe origin of coordinates, and the phase change is proportional tothe vibration displacement [23]. However, itβs very challenging forthis simple circle fitting to deal with tiny vibrations, as shown inFig. 6. The low signal-to-noise ratio (SNR) could lead to unstablefitting results and inaccurate vibration measurements.
MobiCom β20, September 21β25, 2020, London, United Kingdom Chengkun Jiang, Junchen Guo, Yuan He, Meng Jin, Shuai Li, and Yunhao Liu
VibratingObject
I
Q
π
π!π"
π
π"
(a) Basic signal model
I
Q Chirp #1
Chirp #2Chirp #3
βπ
(b) Ideal multi-chirp signals
I
Q
Chirp #1
Chirp #2Chirp #3
βπ!
(c) Practical multi-chirp signals
I
QAnt #1 Ant #2
Ant #3
(d) Multi-antenna signals
Figure 4: The MSC model
0 20000 40000 60000I
60000
50000
40000
30000
20000
10000
0
Q
Dist. = 1m, Amp. = 200umDist. = 1m, Amp. = 100umDist. = 2m, Amp. = 200um
(a) Samples of different vibrations
40000 20000 0 20000 40000I
40000
20000
0
20000
40000
Q
Chirp-1Chirp-2Chirp-3Chirp-4Chirp-5Chirp-6Chirp-7Chirp-8
(b) Samples of 8 chirps
5000 2500 0 2500 5000 7500I
6000
4000
2000
0
2000
4000
6000
Q
Chirp-1Chirp-2Chirp-3Chirp-4Chirp-5Chirp-6Chirp-7Chirp-8
(c) Translated samples of 8 chirps
5000 0 5000 10000 15000 20000 25000I
10000
5000
0
5000
10000
15000
Q
Ant #1Ant #2Ant #3Ant #4
(d) Samples from different antennas
Figure 5: The real IQ samples under different conditions
3.2 Properties of MSCTo cope with low-SNR signals, our insight is that increasing thenumber of observations of the same vibration helps to control mea-surement errors. We exploit the multi-frequency and multi-antennaproperties to provide multiple observations.
3.2.1 Multi-frequency Property of MSC. Eq. 4 indicates the signalphase changes when either ππ or π₯ (π‘) changes. The vibration signalπ₯ (π‘) is out of our control, but the starting frequency of a chirp ππcan be set in the mmWave radar. When we change ππ and keepπ₯ (π‘) the same, the phase of the reflected signal changes. In IQdomain, this results in the rotation of the reflected signal aroundthe origin of coordinates, as shown in Fig. 4(b). Therefore, if wecan enable a chirp group with different starting frequencies tosimultaneously measure the same vibration, ideally the reflectedsignal corresponding to each chirp will rotate around the originof coordinates and form a large arc. The rotation angle betweenreflected signals of two chirps can be represented by:
π₯π = 4ππ₯ππ (π 0 + π₯ (π‘))/π (6)
π₯ππ is the starting frequency gap between the chirps. Therefore,we can fit a better circle using the combined signals.
However, due to the existence of the background reflection, π₯ππleads to the rotation of the both signal components, as shown inFig. 4(c). In order to show that, we set a group of 8 chirps with6ππ»π§ starting frequency gap and the same bandwidth. The firstchirp has 77πΊπ»π§ starting frequency and we measure the vibrationwith 50π»π§ frequency and 100π’π amplitude. The vibrating objectis placed 1π from the radar and the IQ signals are shown in Fig.5(b). The vibration reflections (colored lines) rotate around their
circle centers (grey dotted circles) and the background reflections(grey lines) also rotate around the origin of coordinates. Since thedistance π π΅ β π 0 + π₯ (π‘), the rotation angle of the backgroundreflections π₯ππ΅ = 4ππ₯ππ (π π΅)/π is different from π₯π .
Suppose we find the correct circle center of each vibration re-flection, we can eliminate the background reflection by movingthe circle centers to the origin of the coordinates. As shown inFig. 5(c), we observe that the combined vibration reflection is robustagainst the noise by forming a large arc-shaped trajectory. However,to utilize this property, we still face two critical challenges: (i) howto create such a chirp group with a COTS FMCW radar; (ii) how tofind the correct circle centers with the knowledge of this property.We address these challenges in Β§4.
3.2.2 Multi-antenna Property of MSC. Generally, a mmWave radarhas multiple antennas. The signals received from multiple Rx an-tennas can be utilized to pinpoint the vibrating object and refinethe measurement.
Due to the half-wavelength spacing of Rx antennas, the propaga-tion distances of the reflected signals from the vibrating object todifferent antennas are different from each other, as described in Fig.4(d). Eq. 5 tells us that this will cause the rotations of the vibrationreflections as well as background reflections. Fig. 5(d) shows the IQsamples from 4 Rx antennas when measuring a 100π’π vibration at1π distance and 10β¦ angle of arrival (AoA) relative to the first Rxantenna. We can see that the vibration reflections (colored lines)and the background reflections (grey lines) are different.
Note that the measured signal is along the AoA of the vibrationreflection, which may differ from the vibrating direction. The mea-sured signal is actually a projection of the real vibration to the AoA
mmVib: Micrometer-Level Vibration Measurement with mmWave Radar MobiCom β20, September 21β25, 2020, London, United Kingdom
ππ΅
I
Q
πΤ¦π
Large arc controls
the fitting
Correct
phase
change
πβ²
Incorrect phase change
πβ²
(a) Large amplitude case
Q
π
Small arc causes
fitting ambiguity
Inaccurate
phase change
Τ¦π
ππ΅
I
(b) Small amplitude case
Figure 6: Signal fitting with different amplitudes
of reflection. Therefore, this multi-antenna property not only pro-vides multiple observations but also offers an opportunity to refinethe measurement. To estimate AoAs of the vibration reflectionsfrom different Rx antennas, we only consider their phase changes,i.e. the geometric rotations of the colored lines in Fig. 5(d). Thebackground reflections also rotate when the propagation distanceπ π΅ changes, but we are not interested in their AoAs. How to iden-tify the correct AoAs of the vibration reflections is also a criticalchallenge to be addressed by mmVib.
3.3 MSC Summarywe conclude the following properties of MSC that can improve thevibration measurement:β’ The chirp group provides us the opportunity for better signalfitting by creating multiple observations on the same vibrationthat can consolidate the vibration extraction.
β’ The starting frequency gap in the chirp group determines therotation angle π₯π between two vibration reflections from twosuccessive chirps. π₯π should be consistent in the whole chirpgroup, which can be utilized to further improve the signal fitting.
β’ The spacing of the Rx antenna array differentiates the receivedsignals due to the different propagation distances. The rotationof the vibration reflections of different antennas can infer theirdifferent AoAs.In the next section, we elaborate in detail how we exploit these
properties in our design of the robust and accurate measurementsystem.
4 MMVIB DESIGNThis section introduces the design of mmVib. Fig. 7 shows theworkflow of mmVib, which consists of three main modules.β’ VibrationDetection (VD):VD takes the raw samples ofmmWavesignals as inputs. By analyzing the Range-Doppler spectrum ofthe signals, it detects the candidate range bins of vibrating ob-jects.
β’ Robust Vibration Extraction (RVE): RVE takes the extractedsignals reflected from each candidate range bin as inputs. Thenit exploits MSC to combine the vibration signals from multiplechirps for accurate signal extraction.
β’ Vibration Refinement (VR): VR takes into account the multi-antenna signals to calculate the AoAs and refine the vibration
Rx
Tx
Vibration
Detection
Vibration
Robust
Vibration
Extraction
Vibration
Refinement
vibration
signal
mmVib
Multi-chirp Multi-antenna
Figure 7: Overview of mmVib
deconv.
conv.
Possible
bins
Positive
velocitiesNegative
velocities
Velocity (m/s)
Ra
ng
e (
m)
-1 0 0.9
1.7
0-0
Accurate bin
Half of the
spectrum
Velocity (m/s)
Ra
ng
e (
m)
0 0.9
1.7
0
1
bin
bin
Te
mp
late
Vibration transmission
Vibration detection
Figure 8: Vibration detection
measurement. It aggregates all the measurements to determinethe vibration characteristics.
4.1 Vibration DetectionThe Range-FFT can separate signals based on their propagationdistance. Only specific range bin contains the target vibration signaland we refer it as the vibration bin. Therefore, VDmodule examinesthe Range-Doppler spectrum to search for candidate vibration bins.
In the spectrum, a higher magnitude at a certain range bin and acertain Doppler bin indicates the higher probability of the existenceof a vibrating object. A candidate vibration bin should contain bothpositive and negative velocities due to its reciprocating motion.However, as observed in Fig. 8 where only one object is vibrating,the number of candidate bins is much larger than the number ofvibrating objects.
To understand this observation, we re-examine our experimentsetup (Fig. 12) and find that, besides the vibrator, the body of thecalibrator as well as the table are also vibrating. The above ob-servation is the result of the vibration transmission effect [12]: thevibration signal in a certain bin can be transmitted to adjacent binssymmetrically. According to the vibration transmission model [12],the vibration velocity decays by π1.5 during its transmission, whereπ is the distance between the vibration bin and its adjacent bins.
We model the impact of the vibration transmission on the Range-Doppler spectrum as a convolution operation: the vibration at acertain bin transmits to its adjacent bins through a 1D convolu-tion template. Therefore, the detection process can be modeled asa deconvolution operation, which is the inverse operation of theconvolution [17].
We illustrate the vibration transmission and detection in Fig.8. The convolution template can be obtained with the vibrationtransmission model: we set the coefficient of the middle bin to 1,and calculate the coefficients of the adjacent bins by multiplyingan attenuation factor 1/π1.5. After the deconvolution, VD selects
MobiCom β20, September 21β25, 2020, London, United Kingdom Chengkun Jiang, Junchen Guo, Yuan He, Meng Jin, Shuai Li, and Yunhao Liu
Fre
quency
Slow Time
Sam
e R
ange B
in
Range FFT
Fast-Time Samples
Simultaneous
Slow-Time
Samples
Fast Time
Figure 9: Chirp group generation
the top-E range bins with the largest magnitude as the candidatevibration bins. According the evaluation results in Β§6.4.1, we mayempirically set E to the number of vibrating objects plus one toachieve a relatively high accuracy in practice.
4.2 Robust Vibration ExtractionThe RVE module extracts the vibration signal from each detectedvibration bin. To guarantee the accurate extraction under low SNR,it first generates a chirp group with different starting frequencies toprovide multiple observations, and then eliminates the backgroundreflections with a consolidated vibration extraction algorithm basedon MSC.
4.2.1 Chirp Group Generation. A typical FMCW radar transmitsonly one chirp signal at a time. Then, how to generate a chirp groupthat can simultaneously measure the same vibration as we expected?
Our key insight is to rearrange the fast-time samples of the beatfrequency signal π (π‘), defined in Eq. 3. Recall that the traditionalRange-FFT operation takes all fast-time samples as inputs and gen-erates one slow-time sample. If we separate fast-time samples intodifferent groups and perform Range-FFT in each group, we canobtain multiple simultaneous slow-time samples for each range bin.Fig. 9 illustrates this process: we use a sliding window of size 4 on6 fast-time samples, and obtain two fast-time sample groups with asliding step of 2 samples. This is equivalent to generate two shorterchirps (Chirp #1, Chirp #2) with different starting frequencies fromthe original long chirp. The chirp group has two appealing charac-teristics: (i) Since slow-time samples are much longer than fast-timesamples, the chirps in a group can be regarded simultaneous toeach other. (ii) Different chirps starts at different frequencies, whichresult in diverse but consistent observations of the same vibration.
There are two key parameters to generate the chirp group. In-creasing the number of chirps leads to more observations, at the costof increased computation complexity. Increasing the shift frequencyenlarges the difference among chirps, but sacrifices the bandwidthof one chirp as well as its range resolution. In practice, we may firstset the number of chirps, according to the constraints in compu-tation complexity and the required measurement latency. In ourimplementation, the number of chirps is 8. Then, shift frequencyis set by considering the complexity of the measurement environ-ment. For example, in our laboratory experiments where there isonly one vibrating object, the shift frequency is set at 200ππ»π§. Inthe field experiments where multiple vibrating objects co-exist, we
I
Q
Inherent
Circle
Ξπ ΞπΞπ
Minimize
average
distance
Scale to the
correct positionTranslate to the
correct position
Figure 10: Consolidated circle fitting
set a relatively small shift frequency of 13ππ»π§ to preserve a highrange resolution to differentiate reflections from different vibratingobjects.
4.2.2 Consolidated Vibration Extraction. To extract the vibrationsignal, the primary task is to estimate the background reflections.With the multiple IQ signals provided by the generated chirp group,we can improve the vibration extraction process under low SNR.
The multi-frequency property of MSC tells us that, due to theirinherent consistency, those signals form a large arc around the ori-gin of coordinates when the background reflections are eliminated.Hence, the consolidated vibration extraction algorithm runs in thefollowing steps: (i) It first estimates the background reflection ofeach chirp signal through basic circle fitting step; (ii) Then con-solidated circle fitting step generates a fitting constraint for eachchirp signal that in turn improves the first step; (iii) The first twosteps iteratively run until a perfect large arc is obtained. Finally weextract and aggregate the vibration signals through the vibrationsignal extraction step. Below are the details of the algorithm.
Step 1 - Basic Circle Fitting: Let πΏ = {ππ,π}πΏΓπ , ππ,π β R2
denote the IQ samples from πΏ chirps with π samples per chirp. Forthe π-th chirp, the fitting is turned into an optimization problemto obtain a circle with radius ππ and center ππ that minimizes thesummed geometric distance from every sample to the circle:
πβπ, πβπ= argmin
ππ ,ππ
πβπ=1
(β₯ππ,π β ππ β₯ β ππ
)2, π β [1, πΏ] (7)
It is a nonlinear least squares optimization problem and can besolved with the Levenberg-Marquardt (LM) algorithm [10]. Whenthe SNR is low, however, the basic circle fitting is error-prone with-out a proper constraint on the radius.
Step 2 - Consolidated Circle Fitting: The first step gives abasic but not always accurate estimation of the background reflec-tion of each chirp signal. Therefore, combining multiple translatedchirps signals after the background elimination probably wonβtform a perfect large arc as expected. Suppose the large arc fallson an intrinsic circle, Fig. 10 shows the two cases that each chirpsignal might not necessarily fall on it: (i) translation-needed case:an improperly fitted radius will make the IQ samples of a chirpfall inside (blue samples) or outside the circle; (ii) scaling-neededcase: a stronger or weaker signal strength of a chirp will make itsIQ samples fall on other concentric circles of the intrinsic circle(yellow samples). Therefore, by properly translating and scaling
mmVib: Micrometer-Level Vibration Measurement with mmWave Radar MobiCom β20, September 21β25, 2020, London, United Kingdom
Radar Plane
(ππ , ππ)
π·π π·ππ· β π·π
Projection
Vibrating
direction
(π, π)
π·π
π«πΉπ = πΉπ β πΉπ
Rx4 Rx3 Rx2 Rx1
π/π π/ππ/π
π·π
π· = π
Figure 11: AoA estimation and signal projection
the IQ samples of each chirp, we can finally get a perfect large arc.The following is consolidated circle fitting process:
β’ First, we eliminate the background reflection of each chirp signal.For the IQ samples of the π-th chirp {ππ,π}ππ=1, the elimination isdone by subtracting the center coordinate ππ of the chirp signal:{π β²
π,π}ππ=1 = {ππ,π β πβ
π}ππ=1.
β’ Second, for each chirp signal, we derive its translation directionΞπ as the unit vector of the vector from the origin of coordinatesto the average sample point 1
π
βππ=1 π
β²π,π
.β’ Third, we simultaneously solve the radius of the intrinsic circleπ as well as the translation factors π = {ππ }πΏπ=1 and the scalingfactorsπΈ = {πΎ}πΏ
π=1 by minimizing the average geometric distanceof every sample to the intrinsic circle:
πβ,πβ,πΈβ = argminπ,π ,πΈ
1πΏπ
πΏβπ=1
πβπ=1
(β₯πΎππ β²π,π + πππ₯ππ β₯ β π
)2(8)
We also add a penalty term Ξ Β· (βπΏπ=1 πΎπ β πΏ)
2 to the above lossfunction for regularization, where Ξ is the penalty factor. Withestimated πβ,πβ
πandπΎβ
πfor π-th chirp, its radius can be constrained
around its expected radius πβ/πΎβπ+ πβ
ππ₯ππ in the next iteration.
The iteration stops after the relative change in π is less than asmall threshold (e.g. 1%). Denote the time cost of the fitting processfor each chirp signal by πΌ1 and that of the fitting process for theintrinsic circle by πΌ2, the time cost of π iteration is (πΏπΌ1 + πΌ2) Β· π .According to our experience, π is usually less than 3. So, the totalprocessing time is mainly determined by the number of chirps πΏ.
Step 3 - Vibration Signal Extraction For π-th chirp, with thefinal translated sequence {ππ,π β πβ
π}ππ=1 whose phase sequence
is {ππ,π}ππ=1, we obtain one observation of the vibration signal{ππ,π}ππ=1:
ππ,π =π
4π πππunwrap(ππ,π) β π 0, π β [1, π ] (9)
where πππ is the starting frequency of π-th chirp. By aggregatingall the observations from the chirp group, we can obtain the finalmeasurement. We utilize the inter-quartile mean (IQM) algorithm[34] for the signal aggregation, which calculates the truncated meanof the data within its inter-quartile range: ππ = IQM({ππ,π}), π β[1, π ].
4.3 Vibration RefinementIn this module, we estimate the AoA of each Rx antenna and refinethe vibration measurement, so as to obtain the correct measurementalong its actual vibrating direction.
4.3.1 AoA Estimation. The conventional AoA estimation method,which exploits the phase differences among Rxs, might not workhere due to the unawareness of background reflections. Our ideais to directly estimate the AoAs of vibration reflections with theirrotation angles in the IQ domain. Note that the rotation angles of IQsamples from different Rx antennas reveal their phase differences,which are induced by their different propagation distances. Supposewe derive the difference between the propagation distance of theπ-th antenna compared to that of the first Rx antenna as π₯π π = π π βπ 1,π β [2, π], the basic model for AoA estimation is illustrated inFig. 11: (i) we assume the non-parallelism of the arriving waves thatdescribe different AoAs at different Rx antennas; (ii) suppose ππ =(β_2 (π β 1), 0
)β€is the location ofπ-th Rx and π = (π1, π2)β€ is the
location of the vibrating object, we can compute π π as β₯π β ππ β₯.With π₯π π = π π β π 1 and π 1 = β₯πβ₯, we formalize the followingoptimization problem to solve πβ:
πβ = argminπ
πβπ=2
(β₯π β ππ β₯ β β₯πβ₯ β π₯π π
)2 (10)
Then, the AoAs of π Rx antennas {π½π}ππ=1 can be calculated ac-cording to their geometric relationship with πβ.
4.3.2 Direction-aware Vibration Refinement. Since mmWave radarcan only sense the displacement along the LOS direction towardsthe vibrating objects, the measurement from the RVE module isjust a projection of the vibration signal to this direction. mmVibexploits the multi-antenna property to recover the vibration signalalong its real vibrating direction.
Suppose the angle between the real vibrating direction and thenorm direction of the antenna array is π½ and the AoA ofπ-th an-tenna is π½π , the measurement fromπ-th antenna is a projectionof the vibration signal with an angle π½ β π½π . Denote the measuredvibration amplitudes by {Xπ}π
π=1 and the correct vibration ampli-tude by X. We can estimate π½ and X together with the followingoptimization problem:
Xβ, π½β = argminX,π½
πβπ=1
β₯X cos (π½ β π½π) β Xπ β₯2 (11)
A larger antenna array with more antennas can lead to a betterestimation of the final vibration. It is also feasible to obtain morevibration measurements from different AoAs by combining theresults of multiple synchronized radars.
5 DISCUSSION5.1 Multi-object MeasurementItβs easy for mmVib to handle the multi-object measurement ifthese vibrating objects fall into different range bins: their reflectionsignals can be separated through the Range-FFT. In Β§6.5, our casestudy shows that we can configure the deployment position ofthe radar so that multiple vibrating objects are located in different
MobiCom β20, September 21β25, 2020, London, United Kingdom Chengkun Jiang, Junchen Guo, Yuan He, Meng Jin, Shuai Li, and Yunhao Liu
β’ Distance
(b) Measurement Distance = 700cm
Low-SNR Output
Vibration
Calibrator
Metal Vibrator Control Panel
β£ AnglemmWave Radar
4Rx 2Tx
β Amplitude
β‘ Frequency High-SNR Output
(a) Measurement Distance = 100cm
Figure 12: Experiment setup of mmVib
range bins. However, it is possible that we canβt achieve this idealdeployment condition and several vibrating objects might fall intothe same range bin. In such cases, spatial spectrum analyses, e.g.receiver beamforming technologies [28] or blind signal separationalgorithms [2] can be adopted to separate these reflections.
5.2 NLOS MeasurementmmWave signals have limited penetration capability. The evalua-tion in Β§6.3.3 shows that the performance of mmVib doesnβt degradeunder thin and non-metallic blockages since mmWave signal canpenetrate them. Therefore, we may enclose the mmWave board andits on-board antenna to improve the devicesβ durability. However,in the practical deployment of mmVib, we should avoid thick ormetal blockages, e.g. walls and pillars.
5.3 Phase NoiseThe phase noise in the vibration signal actually has two sources:multiplicative noise and additive noise [26]. The multiplicativenoise is induced by the device circuits such as the oscillator andthe mixer. The additive noise is introduced by the wireless channel,which is typically treated as the Additive White Gaussian Noise(AWGN) [30]. The consolidated observations provided by the chirpgroup in mmVib cope with the additive noise, which is particularlyeffective in the scenarios where the SNR is low. However, the mul-tiple observations canβt eliminate the inherent multiplicative noise.For even higher accuracy of vibration measurement, one may resortto device calibration prior to deployment.
6 EVALUATIONIn this section, we implement mmVib and evaluate it in both thelab environment and a real steel plant.
6.1 Implementation and MethodologyImplementation:We implementmmVib on a commercial mmWaveradar board, Texas Instruments (TI) IWR1642 BoosterPack [14].IWR1642 chip works on the 77πΊπ»π§ millimeter-wave frequencyband (77 βΌ 81πΊπ»π§). It integrates 6 on-board antennas (2 Tx anten-nas and 4 Rx antennas). We let Tx1 send the FMCW signal with2.5πΊπ»π§ bandwidth, and Rx1βΌRx4 receive the reflected signal. Theraw sampling rate (fast-time samples) is around 6ππ»π§ and thechirp sampling rate (slow-time samples) is 10ππ»π§. The raw fast-time samples are captured through a TI DCA1000EVM [15] data
0 20 40 60 80 100Relative Error (%)
0.0
0.2
0.4
0.6
0.8
1.0
CD
F
mmVibCircFitRadar
(a) CDF of amplitude estimation
0.0 0.5 1.0 1.5 2.0Relative Error (%)
0.0
0.2
0.4
0.6
0.8
1.0
CD
F
mmVibCircFitRadar
(b) CDF of frequency estimation
Figure 13: Overall Performance
acquisition board in the high-speed and real-time manner. The dataprocessing coded in Python runs on a computer with an Intel i7-8550U processor and 16πΊπ΅ memory. The mmWave board costs $299while its core chip only costs $40.
Ground Truth: The experiments are conducted in both our laband a steel plant. In the lab, we use a vibration calibrator to generatetunable vibrations with 20π»π§ to 500π»π§ (Β±1%) frequency and 5π’πto 500π’π (Β±1%) amplitude. These parameters describe typical vibra-tions of industrial objects. The ground truth of those measurementsin the steel plant is provided by a piezoelectric vibration sensor.
Experiment setting: Fig. 12 shows the experiment setup in ahallway of 2.4π Γ 10π. The vibration calibrator is placed on a tablewhile the mmWave radar is placed on a tripod. We evaluate mmVibin terms of vibration amplitude and frequency, measurement dis-tance and angle, etc. For each setting, we collect at least 40 tracesof raw mmWave data1.
Comparisons:We compare mmVib with two mmWave-basedvibration measurement approaches introduced before: the theoreti-cal phase-based method proposed in [7] (denoted by Radar) and thebasic fitting-based method proposed in [23] (denoted by CircFit).To ensure fairness, the three approaches use the same data andpre-processing methods.
Metrics: In the experiments, we evaluate the performance interms of the errors in amplitude and frequency estimation: the latterone indicates the correctness of the measured vibration signalswhile the former one stands for the accuracy.
1The dataset is available at http://tns.thss.tsinghua.edu.cn/sun/
mmVib: Micrometer-Level Vibration Measurement with mmWave Radar MobiCom β20, September 21β25, 2020, London, United Kingdom
10 20 30 50 100 200Vibration Amplitude (um)
101
100
101
102
103
Erro
r (um
)
Dist. = 300cm, Freq. = 50Hz
mmVib CircFit Radar
10 20 30 50 100 200Vibration Amplitude (um)
101
100
101
102
103
Erro
r (um
)
Dist. = 600cm, Freq. = 50Hz
mmVib CircFit Radar
(a) Impact of vibration amplitudeson amplitude errors
20 50 100 200 300 500Vibration Frequency (Hz)
101
100
101
102
Erro
r (um
)
Amp. = 10um, Dist. = 100cm
N/A
mmVib CircFit Radar
20 50 100 200 300 500Vibration Frequency (Hz)
101
100
101
102
Erro
r (um
)Amp. = 100um, Dist. = 100cm
N/A N/A N/A
mmVib CircFit Radar
(b) Impact of vibration frequencieson amplitude errors
50 100 200 300 400 500 600 700Measurement Distance (um)
101
100
101
102
Erro
r (um
)
Amp. = 30um, Freq. = 50Hz
mmVib CircFit Radar
50 100 200 300 400 500 600 700Measurement Distance (um)
101
100
101
102
Erro
r (um
)
Amp. = 100um, Freq. = 50Hz
mmVib CircFit Radar
(c) Impact of measurement distanceson amplitude errors
10 20 30 50 100 200Vibration Amplitude (um)
101
100
101
102
Erro
r (H
z)
Dist. = 300cm, Freq. = 50Hz
mmVib CircFit Radar
10 20 30 50 100 200Vibration Amplitude (um)
101
100
101
102
Erro
r (H
z)
Dist. = 600cm, Freq. = 50Hz
mmVib CircFit Radar
(d) frequency errors
Figure 14: Accuracy of amplitude and frequency estimation in the lab environment
6.2 Evaluation on Vibration MeasurementIn this experiment, we evaluate the performance by changing theamplitude (from 10π’π to 200π’π), frequency (from 20π»π§ to 500π»π§)and distance (from 50ππ to 700ππ). The calibrator is placed directlyin front of the radar with a vibrating direction along the radarβsnorm direction. In these cases, smaller amplitude and farther dis-tance mean lower SNR.
6.2.1 Overall performance: Fig. 13 shows the overall performanceof all the settings: mmVib achieves 8.2% relative amplitude error and0.5% relative frequency error in median. Typically, mmVib achievesa median amplitude error of 3.4π’π for the 100π’π-amplitude vibra-tion. The comparisons indicate that mmVib outperforms state-of-the-art approaches by (i) significantly reducing the error of ampli-tude estimation; (ii) improving the stability of frequency estimation.For all settings, mmVib reduces the 80π‘β-percentile amplitude errorby 62.9% and 68.9%, compared to CircFit and Radar respectively.
Next, we examine the impact of different factors on the estima-tion accuracy and stability: in Fig. 14, the first row presents thelower-SNR cases while the second row presents the higher-SNRcases. Note that the Y-axis uses the logarithmic scale.
6.2.2 Amplitude accuracy under different amplitudes. In this exper-iment, we keep the frequency to 50π»π§ and change the amplitudesfrom 10π’π to 200π’π at two distances 300ππ and 600ππ respec-tively. From the results in Fig. 14(a), we can see that: (i) mmVib canaccurately measure the tiny vibrations at a relative far distance: fora typical 300ππ-30π’π case, it achieves an average amplitude errorof 2.7π’π. (ii) For two fitting-based approaches, their performancesare basically in proportional to the SNR. (iii) Since CircFit can beeasily affected by noises, the improvement of mmVib is more signif-icant when the SNR is lower, i.e. the lower amplitude or the fartherdistance. (iv) For extremely low-SNR cases, e.g. the 600ππ-10π’πcase, mmVib seems to have great improvement in the amplitudeestimation. In fact, in these cases mmVib actually doesnβt extractthe correct vibration signals, which will be explained in Β§6.2.5.
6.2.3 Amplitude accuracy under different frequencies. In this exper-iment, we keep the distance to 100ππ and change the frequenciesfrom 20π»π§ to 500π»π§ at two amplitudes 10π’π and 100π’π respec-tively. Due to power limitations, our calibrator cannot generatevibration signals of a large amplitude at a high frequency or a smallamplitude at a low frequency. We can see from Fig. 14(b) that: (i)For 10π’π cases, mmVib achieves a low amplitude error at a largefrequency range, i.e. 2.1um in average, which significantly outper-forms other approaches. (ii) For higher-SNR cases in lower Fig.14(b), mmVib also outperforms the other two approaches, but theperformance gap is relatively small.
6.2.4 Amplitude accuracy under different distances. In this exper-iment, we keep the frequency to 50π»π§ and respectively measure30π’π and 100π’π vibrations at a distance from 50ππ to 700ππ.We can see from Fig. 14(c) that: (i) mmVib can work with a rel-atively long measurement distance: the average error is 4.7π’π forthe 500ππ-30π’π case while 6.9π’π for the 700ππ-100π’π case. (ii)mmVib outperforms the other two approaches at all the distances.
6.2.5 Frequency accuracy under different conditions. We evaluatethe performance of frequency estimation under different vibrationamplitudes and measurement distances with the same data in Β§6.2.2and Β§6.2.4. The results in 14(d) show that mmVib achieves theabsolute frequency error less than 0.3π»π§. It is also worth noticingthat when the distance is 600ππ and the amplitudes is not largerthan 20π’π, mmVib has a relatively large frequency estimation error,which means mmVib doesnβt extract the correct vibration signalsin those cases. That implies the limitation of mmVib in handlingvibration signals with extremely low SNR.
6.3 Impact of practical factorsWe evaluate the impact of several practical factors that are relatedto the applicability of mmVib in practice.
6.3.1 Multipath conditions. The multipath condition can signif-icantly affect the mmWave signal. The experiment is conductedin our office where tables, chairs and computers act as multipath
MobiCom β20, September 21β25, 2020, London, United Kingdom Chengkun Jiang, Junchen Guo, Yuan He, Meng Jin, Shuai Li, and Yunhao Liu
clean light heavyMultipath Condition (same bin)
0
20
40
60
80
100
Erro
r (um
)
mmVib CircFit Radar
(a) Reflectors in the same bin
clean light heavyMultipath Condition (other bin)
0
20
40
60
80
100
Erro
r (um
)
mmVib CircFit Radar
(b) Reflectors in the other bins
0 10 20 30 40Measurement Angle (degree)
0
10
20
30
40
50
60
70
80
Erro
r (um
)
mmVib w/ refinementmmVib w/o refinement
(c) Impact of measurement angle
0 20 40 60 80 100Error (um)
0.0
0.2
0.4
0.6
0.8
1.0
CD
F
Amp. = 100um, Freq. = 50Hz, Dist. = 100cm
airclothplastic
foamphonelaptop
(d) Impact of LOS blockages
Figure 15: Impact of practical factors
reflectors. Besides, we place extra metal plates at different locationsto make the multipath signals fall into the same bin of vibrationreflections or the other bins. Denote the cases with 0, 2 or 6 metalplates by clean, light and heavy. The amplitude, frequency anddistance are fixed to 100π’π, 50π»π§ and 100ππ. The results in Fig.15(a-b) show that (i) mmVib outperforms other approaches for allthe cases; (ii) in the case that strong multipath signal interfereswith the vibration reflection, mmVibβs performance degrades. Inpractice, we can improve the spatial resolution by reducing thefrequency shift in the chirp group.
6.3.2 Measurement angles. We evaluate the impact of measure-ment angles. The measurement angle is defined as the AoA of thefirst RX antenna. In this experiment, we keep the vibrating direc-tion along the norm direction of the antenna array, and translatethe calibrator to control the measurement angles from 0β¦ to 40β¦,as shown in Fig. 12(a). The amplitude and frequency are set to100π’π and 50π»π§. Fig. 15(c) shows the amplitude errors of mmVibwith and without the VR module under different angles. We cansee that: (i) The amplitude error increases with the angle, sincethe measurement distance increases and leads to lower SNR. (ii)For different angles, mmVib achieves the average amplitude errorless than 13.7π’π. (iii) Our VR module brings performance gain,especially when the angle is relatively large.
6.3.3 LOS-path blockages. We evaluate the ability ofmmVib to dealwith LOS-path blockages. We place various objects with differentmaterials but similar thickness (βΌ 1.5ππ) in front of the mmWaveradar (βΌ 20ππ), and evaluate the amplitude errors. Fig. 15(d) showsthe results that (i) the metal materials, e.g. laptops and phones,will block the mmWave signal and make the measurement resultsinapplicable; (ii) the radiation-absorbent material, e.g. the foam,will distort the mmWave signal and greatly degrade the systemperformance; (iii) due to the penetration property of mmWave,other materials wonβt obviously degrade the system performanceand only introduce 10π’π βΌ 20π’π errors in most cases. Based onthis result, we may enclose the mmWave board and its on-boardantenna to improve the devicesβ durability.
6.4 System Micro-benchmarks6.4.1 Vibration detection. This experiment is conducted to illus-trate how to select the parameter E in the VD module. We place4 speakers as vibrating objects in 4 adjacent range bins with themeasurement distances of 89ππ, 95ππ, 101ππ and 108ππ respec-tively. We randomly select different numbers of speakers to play
1 2 3 4 5 6# Candidates
1
2
3
4
# Vibratio
n Sou
rces
(um)
0.90 0.97 0.99 1.00 1.00 1.00
0.00 0.83 0.96 0.99 1.00 1.00
0.00 0.00 0.84 0.93 1.00 1.00
0.00 0.00 0.00 0.95 1.00 1.00
0.0 0.2 0.4 0.6 0.8 1.0Accuracy
(a) Vibration detection
0 1 2 3 4 5Angle Error ( )
0.0
0.2
0.4
0.6
0.8
1.0
CD
F
010203040
(b) AoA estimation
Figure 16: System micro-benchmarks
the same single-tone sound and evaluate how many range bins weshould inspect, i.e E, to correctly find N vibration bins. From Fig.16(a), we observe that the VD module can efficiently and accuratelylocalize the vibration bins by setting E to N + 1.
6.4.2 AoA Estimation. We evaluate the AoA estimation with thesame setup and data in Β§6.3.2. Fig. 16(b) plots the CDFs of AoAestimation errors of Rx1 corresponding to different deploymentangles. We observe that: (i) mmVib achieves a relatively low 80π‘β-percentile AoA estimation error of about 1.4β¦. Since the size of themetal vibrator is relatively small, the results prove that our AoAestimation algorithm is effective by only considering the vibrationreflections. (ii) Because the radarβs field of view (FoV) is about Β±40β¦,the SNR degrades significantly in the 40β¦ cases.
6.4.3 Processing Efficiency. We evaluate the processing efficiencyof mmVib. We mainly consider 3 major components of mmVib:pre-processing (chirp group generation and Range-FFT), vibrationsignal extraction, and post-processing (refinement and aggrega-tion). The median processing time of these components is 103.4ππ ,428.1ππ and 2.5ππ for each mmWave data frame. This is an ac-ceptable time cost and we believe it can be further improved byprocessing all the chirps in parallel.
6.5 Field StudyWe conduct a field study to deploy and evaluate mmVib in a steelplant, the real-world industrial environment. Fig. 17 shows thedeployment, where the vibrating objects are the bearings of a trans-mission system containing the descaling pump, speed reducer andmain motor. The system works in two modes: low-speed operationand high-speed operation, with different rotating frequencies. Theamplitudes of the vibrations are `π-level. The plant installs the
mmVib: Micrometer-Level Vibration Measurement with mmWave Radar MobiCom β20, September 21β25, 2020, London, United Kingdom
mmWave RadarVibrating
Bearing
Descaling
Pump
(a) Non-contact measurement
Bearing of
Speed Reducer
Bearing of
Main Motor
mmWave
Radar
(b) Multi-object measurement
Figure 17: Field study of mmVib
piezoelectric vibration sensors on different parts of the target de-vices and the sensor readings are sent back to the console of themonitoring room via wires. We use those readings as the groundtruth.
Non-contactmeasurement:mmVib outperforms conventionalapproaches due to its non-contact measurement mechanism with-out any disturbance on the running machines or extra deploymentoverhead. Thus, what we are most curious about is whether it worksin practice and how far the measurement distance can be. Fig. 18shows the estimation stability (median and quartiles) and accuracyof the vibration amplitude and frequency of a descaling pump intwo operation modes. Taking the high-speed mode for example,when the distance varies from 100ππ to 500ππ, the average ampli-tude errors are 3.6π’π, 6.7π’π, 5.3π’π, 17.4π’π respectively while theaverage frequency errors are less than 0.4π»π§. This indicates thatmmVib is able to sense the `π-level vibration in practice and itsmeasurement is accurate and consistent when the distance β€ 3π.
Multi-object measurement: The second appealing character-istic of mmVib is its capability of measuring multiple vibratingobjects simultaneously. In this experiment, we place the radar infront of the speed reducer and main motor, and ensure that thebearings of these two machines fall into different range bins. Fig.19 shows that: (i) mmVib captures the fact that, although the ampli-tudes of the speed reducer and main motor differ from each other,their frequencies are nearly identical due to their direct connection(ii) The relatively small inter-quartile ranges and acceptable esti-mation errors demonstrate the stability and accuracy of mmVib formulti-object measurement.
0
50
100
Ampl
itude
(um
)
Low Speed High Speed
100 200 300 500Distance (cm)
0
10
20
Erro
r (um
)
(a) Amplitude Estimation
0
20
40
60
80
Freq
uenc
y (H
z)
Low Speed High Speed
100 200 300 500Distance (cm)
0
2
4
Erro
r (H
z)
(b) Frequency Estimation
Figure 18: Non-contact measurement results
0
20
40
Ampl
itude
(um
)
Low Speed High Speed
Speed Reducer Main MotorMechanical Devices
0
5
10
Erro
r (um
)
(a) Amplitude Estimation
0
20
40
Freq
uenc
y (H
z)
Low Speed High Speed
Speed Reducer Main MotorMechanical Devices
0
5
Erro
r (H
z)
(b) Frequency Estimation
Figure 19: Multi-object measurement results
7 RELATEDWORKSIn this section, we review the related literature of mmVib.
Vibrationmeasurement approaches.Conventional approachesfor vibration measurement are based on specialized sensors or opti-cal devices. Piezoelectric accelerometer is designed for vibrationmeasurement based on the piezoelectric effect [5, 9, 20]. It requiresto be installed on the surface of the vibration source, which couldintroduce non-trivial deployment and maintenance cost. Laser vi-brometer is also a promising solution for high-accuracy vibrationmeasurement [3, 6, 27]. However, its has the high device cost andstrict requirement for the LOS path. A recent work Vibrosight pro-poses to employ a low-cost laser sensor as a long-range vibrometer[38]. It mainly focuses on leveraging vibration spectrums for objectrecognition rather than restoring vibration signals.
RF-based vibration measurement. RF-based approaches arepromising in measuring a targetβs displacement by measuring thechange of the RF signals in a non-intrusive manner [4, 18, 32, 36, 39].ART exploits the 2.4πΊπ»π§ signals and models the relationship be-tween signal features and vibration parameters [32]. However, themm-level accuracy canβt satisfy the micrometer-level requirementin industry. Tagbeat and TagSound exploit the 915ππ»π§ UHF RFIDfor vibration measurement [18, 36]. Similarly, they face the samelimitation on the vibration amplitude. It is still impossible to esti-mate micrometer-level vibrations.
mmWave-based sensing. Compared with common wirelesssignals such as RFID, WiFi and acoustic signals, mmWave is highlysensitive to tiny displacements due to the mm-level wavelength.Lots of works exploit mmWave for high-precision tracking [33],
MobiCom β20, September 21β25, 2020, London, United Kingdom Chengkun Jiang, Junchen Guo, Yuan He, Meng Jin, Shuai Li, and Yunhao Liu
hand gesture and human activity recognition [16, 19], object imag-ing and recognition [25, 41, 42], localization and map construction[11, 22, 24, 40], vital signal monitoring [7, 23, 37], noise-resistantspeech sensing [35] and water-to-air wireless communication [29].
mmWave-based vibration measurement. mmWave is there-fore a promising solution for micrometer-level vibration measure-ment [7, 23, 29, 37]. Ding et al. have proposed a theoretical signalmodel which translates the signal characteristics of FMCW to thevibration parameters [7]. However, without considering the mul-tipath effect, the model fails to extract the correct tiny vibrationamplitude. Mikhelson et al. overcome the above problem by analyz-ing the mmWave signal in IQ domain and introduce circle fittingmethod [23]. However, this work canβt deal with the tiny vibrationbecause of the ambiguity in the small arc fitting.
Compared with the existing mmWave-based approaches, mmVibparticularly addresses the challenges in multipath-rich and noisyenvironment for highly accurate vibrationmeasurement. Built uponthe COTS mmWave radar, mmVib exploits the multi-frequency andmulti-antenna properties of the reflected signals and shows superiorperformance, especially in low-SNR environments.
8 CONCLUSION AND FUTUREWORKIn this paper, we present mmVib for micrometer-level vibrationmeasurement. A multi-signal consolidation model is propose toguide the robust and accurate extraction of tiny vibrations underlow SNR conditions. We evaluate mmVib in the laboratory as wellas field environment. mmVib achieves 8.2% relative amplitude errorand 0.5% relative frequency error in median. Typically, the medianamplitude error is 3.4π’π for the 100π’π-amplitude vibration.
As of the time of publication of this paper, the real-world de-ployment and application of mmVib are on the way. The feature ofnon-invasive measurement and the consistently high measurementaccuracy of mmVib have attracted the attention from industry. Inthe future, we will collaborate with industrial partners to deploytens of measurement devices in the plants and further extend theresearch on mmVib in the following aspects:
Solving the practical issues: We plan to make mmVib a morepractical solution to support the continuous and multi-object vi-bration measurement. How to deal with the dynamic interferencefrom surrounding objects and walking people is also an importantissue to study.
Extending the sensing capabilities: Through the discussionwith industrial partners, we find that some machines have morecomplex vibration characteristics. For instances, different parts ofa machine may vibrate differently. The vibration of a machine mayform a 2D trajectory rather than 1D movement. We will addressthese problems in the future.
ACKNOWLEDGMENTSWe are grateful to the anonymous reviewers for their valuable andconstructive comments. The field study conducted during this workwas supported by Nanjing Nangang Iron & Steel United Co., Ltdand Jiangsu Jinheng Information Technology Co., Ltd.
This work was jointly supported by National Key R&D Programof China No. 2017YFB1003000, National Natural Science Foundationof China No. 61772306 and No. 61902213. Chengkun Jiang andJunchen Guo are co-primary authors of this paper.
REFERENCES[1] E. Peter Carden and Paul Fanning. 2004. Vibration based condition monitoring:
a review. Structural health monitoring 3, 4 (2004), 355β377.[2] J. F. Cardoso. 1998. Blind signal separation: statistical principles. Proc. IEEE 86,
10 (1998), 2009β2025.[3] P. Castellini, M. Martarelli, and EP. Tomasini. 2006. Laser Doppler Vibrometry:
Development of advanced solutions answering to technologyβs needs. Mechanicalsystems and signal processing 20, 6 (2006), 1265β1285.
[4] Bo Chen, Vivek Yenamandra, and Kannan Srinivasan. 2015. Tracking keystrokesusing wireless signals. In Proceedings of ACM MobiSys. 31β44.
[5] Wenqiang Chen, Maoning Guan, Yandao Huang, Lu Wang, Rukhsana Ruby, WenHu, and Kaishun Wu. 2018. ViType: A Cost Efficient On-Body Typing Systemthrough Vibration. In Proceedings of IEEE SECON. 1β9.
[6] Keyence Corporation. 2020. Keyenceβs Laser Displacement Sensors. https://www.keyence.com/products/measure/laser-1d/.
[7] Lei Ding, Murtaza Ali, Sujeet Patole, and Anand Dabak. 2016. Vibration parameterestimation using FMCW radar. In Proceedings of IEEE ICASSP. 2224β2228.
[8] Scott W. Doebling, Charles R. Farrar, Michael B. Prime, and Daniel W. Shevitz.1996. Damage identification and health monitoring of structural and mechan-ical systems from changes in their vibration characteristics: a literature review.Technical Report. Los Alamos National Lab., NM (United States).
[9] Banner Engineering. 2020. NIβs Vibration Sensor White Paper.http://www.ni.com/en-us/innovations/white-papers/06/measuring-vibration-with-accelerometers.html.
[10] Walter Gander, Gene H. Golub, and Rolf Strebel. 1994. Least-squares fitting ofcircles and ellipses. BIT Numerical Mathematics 34, 4 (1994), 558β578.
[11] Francesco Guidi, Anna Guerra, and Davide Dardari. 2015. Personal mobile radarswith millimeter-wave massive arrays for indoor mapping. IEEE Transactions onMobile Computing 15, 6 (2015), 1471β1484.
[12] Carl E. Hanson, David A. Towers, and Lance D. Meister. 2006. Transit noise andvibration impact assessment. Technical Report.
[13] Yuan He, Junchen Guo, and Xiaolong Zheng. 2018. From surveillance to digitaltwin: Challenges and recent advances of signal processing for industrial internetof things. IEEE Signal Processing Magazine 35, 5 (2018), 120β129.
[14] Texas Instruments Incorporated. 2020. IWR1642: Single-chip 76-GHz to 81-GHz mmWave sensor integrating DSP and MCU. http://www.ti.com/product/IWR1642.
[15] Texas Instruments Incorporated. 2020. Real-time data-capture adapter for radarsensing evaluation module. http://www.ti.com/tool/DCA1000EVM.
[16] Wenjun Jiang, Chenglin Miao, Fenglong Ma, Shuochao Yao, Yaqing Wang, YeYuan, Hongfei Xue, Chen Song, Xin Ma, Dimitrios Koutsonikolas, et al. 2018.Towards Environment Independent Device Free Human Activity Recognition. InProceedings of ACM MobiCom. 289β304.
[17] M. K. Khan, S. Morigi, L. Reichel, and F. Sgallari. 2013. Iterative methods ofRichardson-Lucy-type for image deblurring. Numerical Mathematics: Theory,Methods and Applications 6, 1 (2013), 262β275.
[18] Ping Li, Zhenlin An, Lei Yang, and Panlong Yang. 2019. Towards Physical-LayerVibration Sensing with RFIDs. In Proceedings of IEEE INFOCOM. 892β900.
[19] Jaime Lien, Nicholas Gillian, M. Emre Karagozler, Patrick Amihood, CarstenSchwesig, Erik Olson, Hakim Raja, and Ivan Poupyrev. 2016. Soli: Ubiquitousgesture sensing with millimeter wave radar. ACM Transactions on Graphics 35, 4(2016), 142.
[20] Jian Liu, ChenWang, Yingying Chen, and Nitesh Saxena. 2017. VibWrite: Towardsfinger-input authentication on ubiquitous surfaces via physical vibration. InProceedings of ACM CCS. 73β87.
[21] Jian Liu, Yan Wang, Gorkem Kar, Yingying Chen, Jie Yang, and Marco Gruteser.2015. Snooping keystrokes with mm-level audio ranging on a single phone. InProceedings of ACM MobiCom. 142β154.
[22] Chris Xiaoxuan Lu, Stefano Rosa, Peijun Zhao, Bing Wang, Changhao Chen,Niki Trigoni, and Andrew Markham. 2020. See Through Smoke: Robust IndoorMapping with Low-cost mmWave Radar. In Proceedings of ACM MobiSys.
[23] Ilya V. Mikhelson, Sasan Bakhtiari, Thomas W. Elmer, Alan V. Sahakian, et al.2011. Remote sensing of heart rate and patterns of respiration on a stationarysubject using 94-GHz millimeter-wave interferometry. IEEE Transactions onBiomedical Engineering 58, 6 (2011), 1671β1677.
[24] Ioannis Pefkianakis and Kyu-Han Kim. 2018. Accurate 3D localization for 60GHz networks. In Proceedings of ACM SenSys. 120β131.
[25] Akarsh Prabhakara, Vaibhav Singh, Swarun Kumar, and Anthony Rowe. 2020.Osprey: a mmWave approach to tire wear sensing. In Proceedings of ACMMobiSys.28β41.
[26] K. V. Puglia. 2002. Phase noise analysis of component cascades. IEEE MicrowaveMagazine 3, 4 (2002), 71β75.
[27] Lorenzo Scalise, Yanguang Yu, Guido Giuliani, Guy Plantier, and Thierry Bosch.2004. Self-mixing laser diode velocimetry: application to vibration and velocity
mmVib: Micrometer-Level Vibration Measurement with mmWave Radar MobiCom β20, September 21β25, 2020, London, United Kingdom
measurement. IEEE Transactions on Instrumentation and Measurement 53, 1 (2004),223β232.
[28] Petre Stoica, Randolph L. Moses, et al. 2005. Spectral analysis of signals. (2005).[29] Francesco Tonolini and Fadel Adib. 2018. Networking across boundaries: enabling
wireless communication through the water-air interface. In Proceedings of ACMSIGCOMM. 117β131.
[30] David Tse and Pramod Viswanath. 2005. Fundamentals of wireless communication.Cambridge university press.
[31] Junjue Wang, Kaichen Zhao, Xinyu Zhang, and Chunyi Peng. 2014. Ubiquitouskeyboard for small mobile devices: harnessing multipath fading for fine-grainedkeystroke localization. In Proceedings of ACM MobiSys. 14β27.
[32] TengWei, ShuWang, Anfu Zhou, and Xinyu Zhang. 2015. Acoustic eavesdroppingthrough wireless vibrometry. In Proceedings of ACM MobiCom. 130β141.
[33] TengWei and Xinyu Zhang. 2015. mTrack: High-precision passive tracking usingmillimeter wave radios. In Proceedings of ACM MobiCom. 117β129.
[34] Wikipedia. 2020. Interquartile Mean. https://en.wikipedia.org/wiki/Interquartile_mean.
[35] Chenhan Xu, Zhengxiong Li, Hanbin Zhang, Aditya Singh Rathore, Huining Li,Chen Song, Kun Wang, and Wenyao Xu. 2019. WaveEar: Exploring a mmWave-based Noise-resistant Speech Sensing for Voice-User Interface. In Proceedings of
ACM MobiSys. 14β26.[36] Lei Yang, Yao Li, Qiongzheng Lin, Huanyu Jia, Xiang-Yang Li, and Yunhao Liu.
2017. Tagbeat: Sensing mechanical vibration period with COTS RFID systems.IEEE/ACM Transactions on Networking 25, 6 (2017), 3823β3835.
[37] Zhicheng Yang, Parth H. Pathak, Yunze Zeng, Xixi Liran, and Prasant Mohapatra.2017. Vital sign and sleep monitoring using millimeter wave. ACM Transactionson Sensor Networks 13, 2 (2017), 1β32.
[38] Yang Zhang, Gierad Laput, and Chris Harrison. 2018. Vibrosight: Long-RangeVibrometry for Smart Environment Sensing. In Proceedings of ACM UIST. 225β236.
[39] Mingmin Zhao, Fadel Adib, and Dina Katabi. 2016. Emotion recognition usingwireless signals. In Proceedings of ACM MobiCom. 95β108.
[40] Anfu Zhou, Shaoyuan Yang, Yi Yang, Yuhang Fan, and Huadong Ma. 2019. Au-tonomous Environment Mapping Using Commodity Millimeter-wave NetworkDevice. In Proceedings of IEEE INFOCOM. 1126β1134.
[41] Yanzi Zhu, Yuanshun Yao, Ben Y. Zhao, and Haitao Zheng. 2017. Object recog-nition and navigation using a single networking device. In Proceedings of ACMMobiSys. 265β277.
[42] Yanzi Zhu, Yibo Zhu, Ben Y. Zhao, and Haitao Zheng. 2015. Reusing 60 GHzradios for mobile radar imaging. In Proceedings of ACM MobiCom. 103β116.